Asian Journal of Psychiatry 87 (2023) 103705

Contents lists available at ScienceDirect

Asian Journal of Psychiatry

journal homepage: www.elsevier.com/locate/ajp

Artificial intelligence in psychiatry research, diagnosis, and therapy

Jie Sun a, b, 1, Qun-Xi Dong c, 1, San-Wang Wang b, d, 1, Yong-Bo Zheng b, e, 1, Xiao-Xing Liu b,
Tang-Sheng Lu f, Kai Yuan b, Jie Shi f, Bin Hu c, *, Lin Lu b, e, **, Ying Han f, ***
a
Pain Medicine Center, Peking University Third Hospital, Beijing 100191, China
b
Peking University Sixth Hospital, Peking University Institute of Mental Health, NHC Key Laboratory of Mental Health (Peking University), National Clinical Research
Center for Mental Disorders (Peking University Sixth Hospital), Beijing 100191, China
c
School of Medical Technology, Beijing Institute of Technology, Beijing 100081, China
d
Department of Psychiatry, Renmin Hospital of Wuhan University, Wuhan 430060, China
e
Peking-Tsinghua Center for Life Sciences and PKU-IDG/McGovern Institute for Brain Research, Peking University, Beijing 100871, China
f
National Institute on Drug Dependence and Beijing Key Laboratory of Drug Dependence Research, Peking University, Beijing 100191, China

A R T I C L E I N F O A B S T R A C T

Keywords: Psychiatric disorders are now responsible for the largest proportion of the global burden of disease, and even
Psychiatric disorders more challenges have been seen during the COVID-19 pandemic. Artificial intelligence (AI) is commonly used to
Artificial intelligence facilitate the early detection of disease, understand disease progression, and discover new treatments in the fields
Diagnosis
of both physical and mental health. The present review provides a broad overview of AI methodology and its
Prognosis
Treatment
applications in data acquisition and processing, feature extraction and characterization, psychiatric disorder
classification, potential biomarker detection, real-time monitoring, and interventions in psychiatric disorders.
We also comprehensively summarize AI applications with regard to the early warning, diagnosis, prognosis, and
treatment of specific psychiatric disorders, including depression, schizophrenia, autism spectrum disorder,
attention-deficit/hyperactivity disorder, addiction, sleep disorders, and Alzheimer’s disease. The advantages and
disadvantages of AI in psychiatry are clarified. We foresee a new wave of research opportunities to facilitate and
improve AI technology and its long-term implications in psychiatry during and after the COVID-19 era.

1. Introduction psychiatric disorders at an earlier prodromal stage and more personal­

ized treatment, which may be achieved by utilizing high-performance
Psychiatric disorders are among the greatest challenges for public computing capabilities of artificial intelligence (AI) to analyze multidi­
health worldwide. According to the global burden of disease, psychiatric mensional clinical and biological data (Richards et al., 2022).
disorders contribute to 4.9% of global disability adjusted life-years Traditional mental health makes decisions entirely based on sub­
(DALYs) and an age-standardized DALY rate of 1566.2 per 100,000 jective reports of patients’ cognitive and emotional states and their
people, which are increasingly recognized as leading causes of disease course of symptoms. Overreliance on incomplete impressions and su­
burden (GBD, 2022). Patients with psychiatric disorders present phys­ perficial appraisals of patients may impede precise diagnosis and
ical morbidity and premature mortality (Plana-Ripoll et al., 2019), optimal treatment. Dynamic and lifelong gene-environment interactions
which impose a great burden on their daily life and society as a whole. that are modulated by multilevel biological networks made psychiatric
There has been an increase in mental health challenges during the disorders a continuous spectrum of characteristics. Given the high
COVID-19 pandemic, with increases in the incidence of depression, clinical heterogeneity of mental illness, clinicians still need to do
anxiety, and posttraumatic stress disorder, among others (Dedoncker cumbersome trial-and-error searches for the best therapeutic strategies,
et al., 2021). This large burden calls for the accurate identification of which may exacerbate and prolong disease pathology and precipitate

* Correspondence to: School of Medical Technology, Beijing Institute of Technology, 5 Zhongguancun South Street, Haidian District, Beijing 100081, China.
** Correspondence to: Institute of Mental Health/Peking University Sixth Hospital, 51 Huayuanbei Road, Haidian District, Beijing 100191, China.
*** Correspondence to: National Institute on Drug Dependence and Beijing Key Laboratory on Drug Dependence Research, Peking University, 38 Xueyuan Road,
Haidian District, Beijing 100191, China.
E-mail addresses: [email protected] (B. Hu), [email protected] (L. Lu), [email protected] (Y. Han).
1
These authors contributed equally to this work.

https://doi.org/10.1016/j.ajp.2023.103705
Received 4 April 2023; Received in revised form 16 July 2023; Accepted 20 July 2023
Available online 22 July 2023
1876-2018/© 2023 Elsevier B.V. All rights reserved.
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

poor outcomes. There is an emerging consensus that new AI approaches neurons, and they can also be considered artificial neural networks that
are necessary to move the field forward. The human brain acts as are composed of artificial neurons or nodes (Hopfield, 1982). Therefore,
analogous, machines are digital. Humans learn from different experi­ neural networks are biological neural networks or artificial neural net­
ences that are backed by motivation and logic, and the human brain has works that comprise real biological neurons to solve AI problems.
much more powerful thinking capacity and problem-solving skills to Connections of biological neurons are modeled as weights, where posi­
adapt to the crux of a situation, which is not constricted to a particular tive values reflect excitatory connections and negative values represent
pattern (Richards et al., 2022). With the capacity of providing a style of inhibitory connections. All inputs are weighted and summed. These
work efficiency, the machine can solve problems readily. AI enables neural networks can be used for predictive modeling, adaptive control,
machines to learn and recognize patterns and relationships from large and applications that can be trained using datasets. Experience-based
representative examples, mimicking human cognitive activity. These self-learning can occur in the network, and it can draw conclusions
features can serve as revolutionary tools for psychiatric research and from complex, seemingly unrelated sets of information.
precision medicine. The comprehensive actualization of Deep learning is a branch of machine learning that developed from
endotype-specific psychiatric disorders requires AI algorithms that can artificial neural networks. It is constructed with three or more layers of
competently incorporate complex, heterogeneous, and multidimen­ neurons to gradually extract advanced features from the original input
sional data. AI-based applications have been rapidly developed in psy­ data and finally generate category or prediction results. Deep learning
chiatric research and diagnosis (Dwyer et al., 2018; Esteva et al., 2019). has been used in many AI applications, such as computer vision, self-
Research on psychiatric disorders that employs AI approaches has driving cars, natural language processing (NLP), and online adver­
increased in recent years through increasing interest and ubiquitous tising (LeCun et al., 2015). ChatGPT (“generative pretrained trans­
needs to address the world’s growing mental health crisis. Various former”) is a specific implementation of a language model that relies
definitions and analytical AI methods are introduced in the present re­ heavily on NLP techniques, which simulates human conversation in
view. We synthesize the most up-to-date publications to provide insights response to prompts or questions based on the transformer architecture.
into current AI characteristics, performance, applications, and chal­ Deep learning enables facial recognition to be more accurate and robust,
lenges to identify future directions in the field of advancing AI appli­ and it allows medical scans to be interpreted without human analysis.
cations in psychiatric disorders. Deep learning is particularly suited to diagnostic classification that re­
quires distinctions of subtle brain structural and functional abnormal­
2. Artificial intelligence methodology and its implications in ities between psychiatric patients and healthy subjects. It can effectively
psychiatry capture higher-order features and integrate complementary information
from multimodal data, such as magnetic resonance imaging (MRI) and
John McCarthy defined the term “artificial intelligence” as “the sci­ positron emission computed tomography (PET), to enable more accurate
ence and engineering of creating intelligent machines” (Ray et al., and reliable diagnostic predictions.
2022). Alan Turing, who wrote, “Computing Machinery and Intelli­ Predictive analytics encompass data mining, predictive modeling,
gence” in 1950, discussed the conditions under which a machine can be and machine learning statistical techniques that analyze current and
considered intelligent (Turing, 1950). Today, AI is commonly leveraged past facts to predict future or otherwise unknown events (Mijderwijk
to facilitate the early detection of diseases, understand disease pro­ and Steiger, 2022). The goal of predictive analytics is to go beyond
gression, optimize drug selection and dosing, and discover new treat­ knowing what has happened to providing a best assessment of what will
ments (Topol, 2019). The greatest strengths of AI are its ability to happen in the future.
perform rapid diagnostic analyses of large biomedical datasets. The
medical fields that have been the most successful in using pattern 2.2. Implications of AI-based technology in precision medicine in
recognition include ophthalmology, cancer detection, and imaging. In psychiatry
these areas, AI can perform as well as professional clinicians in assessing
image abnormalities and nuances that are otherwise difficult to inspect. AI algorithms are breaking into multiple aspects of psychiatric
These nuances can be better detected by AI methodologies than expe­ research, including data collection and structuring, feature extraction
rienced clinicians (Brinker et al., 2019). and characterization, psychiatric disorder classification, potential
biomarker detection, real-time monitoring, and optimal treatments for
2.1. Methodology of AI psychiatric disorders (Fig. 1). Its rapid development offers breakthrough
opportunities to achieve precision medicine in psychiatry.
Machine learning is the study of computer algorithms that are
automatically improved through experience. Machine learning algo­ 2.2.1. Data collection and structuring for psychiatry
rithms build models based on sample data, known as “training data”, In the medical field, data acquisition and processing are fundamental
and make predictions or decisions without explicit programming (Bzdok to research design. Medical imaging is an area where statistical learning
and Meyer-Lindenberg, 2018). Machine learning algorithms are used in methods are most widely used for big data, and structured methods for
a wide variety of fields, such as email filtering and computer vision, analyzing and integrating such large-scale datasets may prove increas­
where it is difficult or unfeasible to develop conventional algorithms to ingly valuable for disease diagnosis. Structured data analysis methods
perform the needed tasks. Machine learning tends to focus more on based on electroencephalography (EEG) could be used as part of the
hypothesis generation than hypothesis driving (Graham et al., 2019). diagnostic workflow in clinical practice (Mohammadi et al., 2015).
Machine learning approaches are traditionally divided into three broad Additionally, with the development of digital technologies, the appli­
categories, depending on the nature of the “signal” or “feedback” that is cation of machine learning techniques to big data may be a viable option
available to the learning system: supervised learning, unsupervised for dealing with vast amounts of semi-structured information in
learning, and reinforcement learning. Machine learning capabilities healthcare, which can provide objective and comparable indicators of
promise to inform disease models, discover and develop new treatments accuracy and facilitate the creation of large repositories of medical re­
for diseases and prevention strategies in psychiatry, build predictive cords (Fischbach and Lord, 2010). With the development of AI and
models of useful outcomes, and provide objective and comparable in­ precision medicine, it will become easier, less expensive, and more
dicators of accuracy. Additionally, building a deeper understanding of convenient to collect these physiological data. Furthermore, large vol­
system tasks with a focus on important individual differences is neces­ umes of medical data, updated in real time, will help improve the pre­
sary to facilitate the development of AI techniques in psychiatry. dictive accuracy of models, reduce misdiagnosis rates, and assist
Neural networks are networks or circuits that are composed of psychiatrists in diagnosing and treating psychiatric disorders.

2
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

Fig. 1. Implications of artificial intelligence-based technology in precision medicine in psychiatry. Artificial intelligence and machine learning techniques are
breaking into psychiatry research and mental healthcare, including data collection and structuring, feature extraction and characterization, diagnosis and subtype
classification, potential biomarker identification, real-time monitoring, and optimal treatments for psychiatric disorders.

2.2.2. Psychiatric feature extraction and characterization approaches potentially useful for continuous data influx and updates,
Applying machine learning tools to currently available clinical data such as hospital patient databases, allowing minor modifications of
in psychiatry has enabled data validity mining and characterization. For predictive analyses based on current conditions rather than potentially
example, Dipnall et al. leveraged data mining to identify over 250 outdated bodies of data while maintaining their prior knowledge of
different biomarkers for 5546 subjects who were affected by depressive relevance. Machine learning has begun to use these new diagnostic tools
syndromes, leading to a more accurate construction of predictive func­ to define parameters for specific psychiatric disorders. Carpenter et al.
tions (Dipnall et al., 2016). Abbas et al. applied machine learning to gold utilized a decision tree algorithm to test whether a child was likely to
standard clinical data that were obtained from thousands of children have generalized anxiety disorder or separation anxiety disorder. Their
who were at risk for autism spectrum disorder (ASD), applying new results showed that decision trees predicted both of these types of anx­
feature selection, engineering, and coding techniques with experimental iety with 96% accuracy (Carpenter et al., 2016). Moreover, deep
results to demonstrate significant accuracy improvements, including learning has achieved good development in the diagnosis and prognosis
sensitivity and specificity, compared with standard screening tools of other psychiatry disorders, such as ASD, sleep disorders, and
(Abbas et al., 2018). Additionally, feature mining for big data can be addiction.
used in psychiatry through machine learning techniques to compile and
analyze multidimensional and multilevel disease models to develop 2.2.4. Detection of potential biomarkers in psychiatry
more effective treatments for psychiatric disorders. Moreover, Research AI offers new approaches for diagnosis and prognosis in psychiatry
Domain Criteria is a platform to classify large amounts of data that are and the detection of potential biomarkers in psychiatric disorders.
used for machine learning implementations (Insel et al., 2010). Merging Substantial heterogeneity is generally seen in the pathophysiology of
validated correlation matrices, such as Research Domain Criteria, with psychiatric disorders. The identification of novel biomarkers will help
AI techniques may help integrate various large psychopathological do­ provide more objective and refined definitions of these illnesses.
mains into an effective diagnostic structure. Biomarker innovation is a promising way to improve the differential
diagnosis of mental illness and individualized patient care (Chen et al.,
2.2.3. Classification of psychiatric disorders 2014). For example, to pinpoint a new risk biomarker, called the Alz­
Machine learning currently offers the opportunity to analyze disease heimer’s disease (AD) pattern similarity score, a study proposed a
models for complex, multifactorial disease states and may inform high-dimensional architecture based on AI and machine learning that
treatment options for existing therapies. The flexibility and potential for can concurrently achieve AD prediction and the integration of neuro­
structural evolution based on deep learning also makes neural network imaging data and multi-omics data (Casanova et al., 2018). The study

3
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

indicated that the AD pattern similarity score can serve as an AD risk development, privacy concerns, and interpretability, highlighting the
biomarker because of its strong association with cognitive status, age, complexity of implementing AI in digital psychiatry. Ultimately, these
and cognitive function. advanced analytical methods enable more precise risk assessment, early
A previous study also proposed an ensemble clustering approach to detection of mental health conditions, and personalized treatment rec­
identify subtypes across the psychotic-affective disorder spectrum, ommendations, leading to better patient outcomes in the field of
including schizophrenia, bipolar disorder, and major depressive disor­ psychiatry.
der (MDD), in a trans-diagnostic sample of major psychiatric disorders
(MPDs) (Chang et al., 2021). Using the whole-brain amplitude of 3. Performance of AI-driven technologies in the diagnosis and
low-frequency fluctuations (ALFFs), the authors found that archetypal treatment of psychiatric disorders
MPDs are associated with an increase in frontal and decreased posterior
ALFF, and atypical MPDs are associated with a decrease in frontal ALFF The use of AI technology in the medical field is becoming more
and an increase in posterior ALFF. Their findings suggested that frontal common, but its application in psychiatric disciplines has been relatively
vs. posterior functional imbalance is a novel core trans-diagnostic slow to develop. Psychiatric clinical data are often provided in the form
biomarker that differentiates subtypes of MPDs, which could have im­ of subjective patient perceptions and qualitative statements or written
plications for precision medicine in psychiatry. records. However, there are still many areas of psychiatry that can
benefit from AI technology (Shatte et al., 2019). AI has great potential to
2.2.5. Real-time monitoring and interventions for psychiatric disorders redefine our diagnosis and understanding of psychiatric disorders.
In the field of psychiatry, in addition to improving disease models, Generally, there is substantial heterogeneity in the pathophysiology of
physical status monitoring and prevention are imperative for those who psychiatric disorders. The identification of biomarkers will help provide
are potentially at risk. Advances in AI have made scalable, rapid, and more objective and refined definitions of these illnesses. AI technologies
automated treatments and interventions possible, and AI-based tech­ can develop better pre-diagnostic screening tools and risk modeling to
nologies enable multiple smart devices (e.g., mobile phones and tablets) determine an individual’s susceptibility to, or risk of developing, a
to scale in a relatively seamless manner, laying the foundation for mo­ psychiatric disorder (Shatte et al., 2019). We briefly introduce appli­
bile AI detection and treatment solutions while collecting data for cations of AI in the early warning, diagnosis, prognosis, and treatment of
personalized protection and treatment. GerAmi, for example, integrates psychiatric disorders, including depression, schizophrenia, ASD,
AI into a multi-subject management system for geriatric wards, an attention-deficit/hyperactivity disorder (ADHD), addiction, sleep dis­
intelligent integration tool that helps provide physical and cognitive orders, and AD (Fig. 2).
support to healthcare workers and predict potentially dangerous situa­
tions in these geriatric facilities (Corchado et al., 2008). mindLAMP, a 3.1. Mood disorders
customizable application, can collect various streams of patient data,
such as surveys, cognitive tests, and movement information, enabling Mood disorders, including MDD and bipolar disorder (BD), are
the ability to use smartphones and their embedded sensors to help among the most prevalent and costly ailments, currently the second
predict and understand people’s experiences of living with mental largest contributor to the global disease burden (Disease et al., 2018).
illness and rehabilitation (Torous et al., 2019). Because of the lack of objective markers, the prediction and diagnosis of
Overall, by combining these techniques, AI-based tools can effec­ diseases are completely based on patients’ subjective statements and
tively analyze large volumes of multidimensional and multimodal data clinicians’ observations and interpretations (Bzdok and
in digital psychiatry. A variety of customized multisensory and multi­ Meyer-Lindenberg, 2018). As a result of the similarity in clinical
dimensional data used in AI-based digital psychiatry present a complex symptoms during the initial stages, approximately 60% of individuals
landscape, encompassing physiological sensors (e.g., heart rate), with BD are initially misdiagnosed as MDD. Unfortunately, this misdi­
movement and activity sensors, facial expression analysis, voice and agnosis can lead to a prolonged waiting period of 5–10 years before
speech analysis, eye-tracking devices, smartphone data, wearable bio­ receiving the correct diagnosis (Goodwin and Jamison, 2007; Hirsch­
sensors, virtual reality, and more (Cosic et al., 2016; Cosic et al., 2021). feld, 2014). Such delays in obtaining an accurate diagnosis have severe
By continuously tracking various parameters, wearable devices can consequences, including inadequate treatments and a poorer prognosis
provide a rich source of data that captures the holistic picture of an for the affected individuals. Encouragingly, however, with improve­
individual’s well-being. Edge computing offers the advantage of pro­ ments in computer processing speed and power, AI technology enables
cessing and analyzing this data locally. The integration of wearable statistical models that can incorporate more influential factors into an­
devices and edge computing facilitates applications in remote moni­ alyses to approximate the actual situation and extract more
toring, wearable devices, and mobile mental health interventions. It high-dimensional information from data for better diagnosis and
enables real-time analysis of data from sensors and devices, such as treatment.
monitoring physiological signals or behavioral patterns, facilitating AI has been widely used in research on the diagnosis of mood dis­
early warning systems for potential mental health crises or relapse orders. Liu et al. analyzed 1384 psychiatric patients (322 depressive
prevention, allowing for timely interventions without relying heavily on patients) and sequenced whole genomes. Use of the burden of genomic
cloud-based infrastructure. Furthermore, the continuous monitoring variants in coding/non-coding regions as the feature vector in the deep
provided by wearable devices in the aftermath of major mental health learning algorithm significantly distinguished patients from the control
stressors generates enormous amounts of longitudinal data. This meta group (Liu et al., 2022a). Additionally, as an objective biomarker, MRI
data can be utilized to develop personalized predictive model, offer a combined with AI technology is most likely to make breakthroughs in
promising avenue for early warning systems in mental health. The depression. Rubin-Falcone et al. used the gray matter volume of a sup­
continuous monitoring of individuals’ behavioral, social, and physio­ port vector machine to classify BD and MDD at the individual level
logical features, particularly during challenging periods, can provide (Rubin-Falcone et al., 2018). Zhao et al. proposed functional network
valuable data for the development of personalized predictive models connectivity-based generative adaptive networks to distinguish
and timely interventions (Cosic et al., 2020a), ultimately improving depressed patients from healthy controls (Zhao et al., 2020). The strat­
mental health outcomes and promoting overall well-being. This ification of depression has always been complexed for clinicians, but it
comprehensive approach allows for the discovery of complex associa­ can be done well by integrating disease severity, the history of mental
tions, identification of relevant features, and improvement of prediction disorders, disease complications, genetics, and medical birth data
accuracy. This diverse array of sensors and features poses development through deep learning (Allesoe et al., 2022). Researchers also use
challenges such as data integration, quality assurance, algorithm external devices to record patients’ physiological and behavioral data,

4
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

Fig. 2. Application prospects and potential of artificial intelligence in the detection, diagnosis, and treatment of psychiatric disorders. Multidimensional clinical and
biological data of patients with psychiatric disorders are collected in the laboratory via magnetic resonance imaging (MRI), electroencephalography (EEG), and gene
sequencing, among others, and monitored by intelligent devices, including daily life monitoring, face and speech recognition, etc. AI algorithms offer a breakthrough
opportunity to more accurately detect, diagnosis, and classify psychiatric disorders, including depression, schizophrenia, autism spectrum disorder (ASD), attention-
deficit/hyperactivity disorder (ADHD), addiction, sleep disorders, and Alzheimer’s disease (AD), and improve prognosis and treatment. Furthermore, AI robots, such
as Woebot, RoboTherapy, Nao, and Kaspar, are developing rapidly and could be a major aid to the clinical treatment of psychiatric disorders, especially depression
and ASD.

including audio, video, eye tracking, and motion posture, and extract learning algorithms and advanced data analysis techniques, AI can assist
corresponding features to diagnose depression. Stanford University re­ in identifying early warning signs and personalized intervention stra­
searchers combined voice recognition, computer vision, and NLP tech­ tegies to mitigate the risk of relapse. By monitoring and continuously
nologies to determine whether a single individual suffered from analyzing these data streams, AI algorithms can detect subtle changes or
depression through expression and language (Albert Haque et al., 2018). patterns, such as deviations from baseline symptomatology, changes in
Another study proposed the Deep Learning-assisted Integrated Predic­ sleep patterns, social withdrawal, or early signs of medication
tion Model (DLIPM) for the early prediction and diagnosis of mood non-adherence, that may indicate an increased risk of relapse. With this
disorders in children, which was ahead of other studies technologically information, AI can provide timely alerts and notifications to both pa­
in terms of sensitivity rate, specificity rate, recall ratio, and precision tients and healthcare providers, enabling proactive interventions.
ratio (Zhang et al., 2021). Candidate biological, clinical and neuropsy­ Furthermore, AI can also assist in personalizing relapse prevention
chological markers were applied to classify BD from MDD or healthy strategies based on individual patient characteristics. Wu et al. proposed
control with overall classification accuracy for BD was 0.77 (Colombo a latent-space machine learning algorithm that was tailored to
et al., 2022). One of the key advantages of AI in early diagnosis is its resting-state EEG signals and found that EEG biomarkers that were
ability to identify subtle patterns and risk factors that may not be readily related to prefrontal nerve responsiveness effectively predicted antide­
apparent to human clinicians. AI algorithms can process complex mul­ pressant efficacy (Wu et al., 2020). Many studies have also shown that
tiple dataset sources, extract meaningful information and employ so­ machine learning combined with genomics and plasma metabolomics
phisticated analytical techniques that can help distinguish between BD and structural MRI was able to predict the therapeutic response to an­
and MDD, even in cases where the clinical presentation overlaps. tidepressants and electroconvulsive therapy (Chekroud et al., 2016;
Relapse is a significant concern for patients with mood disorders, and Joyce et al., 2021). Furthermore, a research team identified a person­
its treatment has individual differences. There is currently a lack of alized symptom-specific biomarkers through the cross-validation of a
objective individualized strategies when formulating treatment plans for machine learning model and implemented biomarker-driven deep brain
depression. Worsening this situation, the efficacy rate of antidepressants stimulation closed-loop treatment (Scangos et al., 2021). With the wide
is estimated to be only 42–53% (Taliaz et al., 2021). Thus, MDD treat­ application of AI and popularity of wearable devices, the diagnosis and
ment mainly depends on repeated trials. AI technology has the potential treatment of depression will rely on a multimodal integration model
to play a crucial role in relapse prevention and build prediction treat­ based on objective biomarkers.
ment models, which will reduce trial-and-error strategies with regard to The treatment of depression continues to see problems, such as
drug prescriptions to improve efficiency. By leveraging machine inconvenience and high treatment costs for patients in areas with poor

5
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

medical conditions. However, psychotherapy that is provided through diagnosis and treatment efficacy prediction for schizophrenia is ex­
smartphone applications and chat robots using AI can effectively relieve pected. Based on neuroimaging data from schizophrenia patients in a
depressive symptoms, which helps offset some limitations of existing large sample among multiple centers, a Chinese research team found
traditional therapies (Linardon et al., 2019; Torous et al., 2021; Vai­ that impairments in striatal function could be used as a new biological
dyam et al., 2021). Increasingly, more virtual and robotic AI agents are quantitative indicator that can accurately screen schizophrenia patients
being developed to perform high-level therapeutic interventions (Fiske from healthy people, providing an important biological basis for
et al., 2019). Woebot is a chatbot that works over short messaging ser­ assisting doctors with selecting antipsychotic drugs with different
vices, WhatsApp, or internet platforms. It works like a virtual psycho­ mechanisms of action (Li et al., 2020). Altogether, AI plays an important
therapist to help patients recognize their emotions and reduce anxiety. role in individualized accurate diagnosis and treatment efficacy pre­
In a randomized controlled trial, 70 depression patients were randomly diction for schizophrenia, thus addressing the lack of effective, reliable,
divided into a Woebot group and an e-book reading group. Compared and instructive objective quantitative biomarkers in clinical practice.
with the e-book group, depression was significantly relieved in the
Woebot group (Fitzpatrick et al., 2017).
3.3. ASD and ADHD
3.2. Schizophrenia
AI has successfully been used to diagnose and treat children with
Schizophrenia is a severe and disabling mental disease. The early ASD (Wei et al., 2023). Early detection is known to improve long-term
diagnosis of schizophrenia and effective interventions and treatments outcomes in individuals with ASD. A study that used machine learning
can improve the cure rate and prevent the disease course. An effective and deep learning approaches to investigate the epigenetic basis of
computer-aided diagnosis system to help doctors would be very helpful autism and identify early biomarkers for disease prediction found the
(Feyaerts et al., 2021). In recent years, researchers have identified ab­ significant enrichment of molecular pathways that are involved in
normalities in brain structure and function in schizophrenia patients neuroinflammation, synaptic long-term potentiation, serotonin degra­
using brain imaging technology, including changes in gray matter vol­ dation, and mammalian/mechanistic target of rapamycin and
ume and density, white matter abnormalities, and slight differences in Rho-family guanosine triphosphatase signaling (Bahado-Singh et al.,
spatial distribution patterns of functional brain connections, thus 2019). Zhou et al. applied a deep learning-based framework to 1790
providing a basis for building an accurate and objective diagnosis model ASD simplex families and found that noncoding mutations-ASD pro­
of schizophrenia (Li et al., 2020). For example, researchers have used bands harbored both transcriptional and post-transcriptional regu­
machine learning classification methods to explore ways to identify lation-disrupting de novo mutations contributed to autism risk (Zhou
people with schizophrenia from the healthy subjects (Winterburn et al., et al., 2019). Jaganathan et al. predicted splicing from primary se­
2019), differentiate different subtypes of schizophrenia (Chand et al., quences with a deep neural network and found that new mutations that
2020), distinguish individuals who will become schizophrenics from predicted the consequences of splicing alterations were significantly
high-risk groups (Zarogianni et al., 2017), and distinguish clinical pa­ enriched in ASD and mentally retarded patients, and 21 of 28 of these
tients at different disease stages (Guo et al., 2016). Machine learning patients were validated by RNA sequencing (Jaganathan et al., 2019).
technology combined with a patient’s language (Chan et al., 2023; Fu Machine learning technology has also been successfully used to distin­
et al., 2021) and facial video (Huang et al., 2022) can help doctors guish adult ADHD and healthy controls and classify ADHD subtypes
differentiate schizophrenia from healthy controls. A three-dimensional based on the power spectrum, measured by EEG (Tenev et al., 2014).
convolutional neural network (CNN) framework was proposed to The findings showed that this AI approach improved discrimination
discriminate schizophrenia patients and normal healthy controls based between ADHD and control groups and between ADHD subtypes.
on MRI of the brain, achieving 99.82% accuracy (Qureshi et al., 2019). Apart from assisting with ASD diagnosis, AI has also been applied to
These findings have the potential to be applied as an adjunct to assist treat ASD. Artificial intelligence robots, intelligent mechanical devices,
clinicians in initial screening for schizophrenia. and intelligent wearable devices could help promote social skills and
Furthermore, AI plays a vital role in the early prediction as well as living abilities among ASD patients (Huijnen et al., 2019; Pennisi et al.,
relapse prevention of schizophrenia. Sensors on smartphones and 2016). AI robots, such as RoboTherapy, Nao, and Kaspar, help children
wearable devices have become tools for evaluating behavioral patterns with ASD develop social skills (Fiske et al., 2019). RoboTherapy is an
in a range of people and obtaining additional objective behavioral data. example of socially assistive robotics that was designed to help children
Barnett et al. used a passive monitoring application that was installed on with ASD develop social skills (Libin and Libin, 2003). Such in­
smartphones to follow 17 psychiatric patients for up to 3 months. They terventions mainly aim to improve social skills, including imitation,
found abnormal mobile patterns and social behaviors within 2 weeks taking turns, staying engaged, and empathy. These children can then
that could predict relapse in schizophrenia (Barnett et al., 2018). apply these learned skills to their relationships with human peers
Another study of 83 schizophrenia patients using digital markers found (Mengoni et al., 2017; Pennisi et al., 2016). ASD patients performed
similar results (Henson et al., 2021). Ben-Zeev et al. determined changes better with their robot partners than with human therapists and
in sensor data among individuals with schizophrenia-spectrum disorders exhibited improvements in spontaneous language during therapy ses­
in the days before relapse, including physical activity, geographical sions (Pennisi et al., 2016). A study that used social robots as embedded
location, the duration of cell phone unlocking, and voice frequency and reinforcers of social behavior in children with autism found that social
duration, and identified unique digital indicators of psychotic relapse robots may be developed into useful tools to improve social skills and
(Ben-Zeev et al., 2017). In addition, machine learning has been applied communication, specifically by embedding social interaction into
to determine the predictors of psychotic relapse and follow-up with­ intrinsic reinforcers and motivators (Kim et al., 2013). Huijnen et al.
drawal with a percentage of correct prediction of 52.4%, sensitivity used the Kaspar robot among children with ASD (Huijnen et al., 2019).
62%, specificity 38.7% (Walter et al., 2019). Strengths of Kaspar are related to personalization possibilities, its
The prediction of drug response has always been a focus of research playfulness, the action-reaction principle, its neutral expression, the
on schizophrenia. Recent studies showed that the degree center based on consistent and repetitive application of actions, the ability to vary
graph theory combined with a machine learning algorithm predicted the behavior in a controlled manner. Such robots are developing rapidly and
response of first-episode schizophrenia patients to antipsychotic drug could be a major aid to clinical treatment in the future. However, certain
therapy (Liu et al., 2022b). Wang et al. used MRI and polygenic risk challenges, such as limited reaction possibilities, the likelihood of chil­
scores to predict the treatment response in schizophrenia, with an ac­ dren being afraid of AI robots, difficulties with generalization or trans­
curacy rate of 86% (Wang et al., 2022). A quantitative indicator for fer, and potential dependence on AI robots, have also been identified.

6
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

3.4. Addiction people daily, with a high prevalence and detrimental physiological ef­
fects. Insomnia is a risk factor for the development and exacerbation of
In the past decade, some biomedical research fields, including psychiatric disorders, which occurs in up to one-third of adults world­
addiction research, have successfully applied machine learning wide (Perlis et al., 2022). The detection of insomnia and sleep apnea has
methods. Machine learning assembles a broad set of methods and attracted the most research attention within the field of sleep disorders.
techniques to solve a wide range of problems, such as identifying in­ Electroencephalographic signals and fMRI images were used for
dividuals with substance use disorders, finding patterns in neuro­ insomnia detection (Qu et al., 2021). Machine learning and deep
imaging, understanding prognostic factors and their association, and learning were incorporated for insomnia detection. Support vector ma­
determining the genetic basis of substance use disorders (Afshar et al., chine classifiers achieved the best performance. OSA is a phenotypic and
2022). prognostic heterogeneous disease for which there is no common uniform
Afshar et al. applied NLP to clinical notes that were collected from therapeutic approach (Zinchuk et al., 2017). Intricate polygenetic con­
electronic health records to accurately screen for substance misuse and tributions and complex environmental factors generate a combination of
developed a novel and accurate approach to leverage the first 24 h of various anatomical and functional factors that distinguish individuals.
electronic health record notes to screen multiple types of substance Different anatomical and functional impairments, such as high loop
misuse (Afshar et al., 2022). Researchers also applied state-of-the-art gain, low arousal threshold, and upper airway muscle dysfunction,
deep and recurrent neural network models to a dataset of more than require precise diagnostic tools. By analyzing metadata from snoring,
100,000 opioid users to classify patients as opioid users (Che et al., detecting patterns of craniofacial abnormalities by AI could help predict
2018). Moreover, a machine learning framework was used to predict the presence of OSA and propose precision therapy. Snoring frequency
substance use disorder treatment success (Acion et al., 2017). band ratio, spectral flatness, and periodicity were shown by machine
Drug repurposing is a strategy that can help identify potential new learning algorithms to have high sensitivity with regard to screening for
treatments for complex diseases. An integrated drug repurposing strat­ OSA, but their diagnostic value was low because of low specificity
egy that combined computational-based drug prediction, patient elec­ (Alshaer et al., 2019). A combined deep CNN and bi-directional long-­
tronic health record-based clinical corroboration, and mechanism of short term memory network assessed 4650 studies to predict OSA
action analyses was used in a study that identified five repurposed severity, demonstrating 86.9% sensitivity and 99.5% specificity (Fer­
candidate drugs for treating opioid use disorder patients (Zhou et al., nandez et al., 2019).
2021). AI techniques could be applied to consumer sleep technologies which
acquire accurate, sophisticated, and ubiquitous massive robust popula­
3.5. Sleep disorders tion sleep datasets. Photoplethysmography is the most common non-
EEG wearable sensing technology, with sleep staging accuracy over
As a universal function of living species, sleep comprises one-third of 70% (Zhao and Sun, 2021). Radiofrequency, piezoelectric sensor, and
human life. Sleep disorders are frequent and have been associated with a impulse-radio ultra-wideband radar were used to assess sleep architec­
wide variety of dysfunction of body systems, including neurological ture and respiratory events to enable algorithm-based machine learning
disorders, pain symptoms, and metabolic disorders (Krause et al., 2017; (de Zambotti et al., 2020). Metadata and AI technology will provide
Sun et al., 2020). Sleep disturbances have many components and dy­ novel insights into sleep medicine.
namic temporal interactions that are not linear. Not all of these com­
ponents are present in all patients. Complex and heterogeneous features 3.6. Alzheimer’s disease
of sleep disturbances increase the difficulty of implementing precision
medicine in sleep medicine. AI is capable of uncovering clinically spe­ AD is the primary cause of neurodegenerative dementia, character­
cific patterns, predicting prognosis, detecting precision therapeutic ef­ ized by specific abnormal protein misfolding in the brain and an explicit
fect, clustering findings, and providing comprehensible analytics by pattern of deposition that can develop years before any clinical mani­
assessing metadata to obtain beneficial information. festations. Owing to several failures in delivering effective therapies for
Sleep stage classification plays a vital role in the objective detection AD patients, studies have shifted focus from a curative approach to a
of sleep disorders. Polysomnography (PSG) is the gold standard of sleep preventive perspective in preclinical and prodromal stages. AI has pro­
assessment, in which recordings are scored as different variables to vided a new opportunity to extract reliable predictors and automatically
define the continuity, efficiency, and architecture of sleep. Scoring PSG classify different AD phenotypes, but biomarkers that are used for early
results by humans is both time- and labor-intensive, and inter-scorer diagnosis have a relatively limited ability to predict AD at the individual
reliability can be as low as 68% for wake, 23% for stage N1 sleep, level. NeuCube, a new type of AI that implements neuromorphic, brain-
79% for stage N2 sleep, 69% for stage N3 sleep, and 69% for hypopneas like information processing principles, was applied to longitudinal MRI
(Danker-Hopfe et al., 2004). The automated scoring of PSG by super­ datasets to accurately classify and predict mild cognitive impairment
vised AI machine learning and deep recurrent and convolutional neural and dementia 2 years ahead of cognitive deficiency, with 95% and 91%
networks showed high concordance with manual visual scoring (Biswal accuracy, respectively (Doborjeh et al., 2021). Machine learning algo­
et al., 2018). AI models that are used to score PSG are constantly rithms outperformed existing all-cause dementia prediction and inci­
improving, moving closer to truly accurate and efficient representations dence tools, achieving superior accuracy based on only a few key
of sleep and reducing inter- and intra-operator variability, which has variables (James et al., 2021; Park et al., 2020). You et al. developed a
potential to be a new gold standard. Unsupervised principal component dementia risk prediction model based on a large prospective UK Biobank
analysis, cluster analysis, high-throughput algorithms, and K-mean cohort that contained more than 500,000 participants. The model was
clustering iteratively test large numbers of potential PSG data grouping established using readily accessible predictors that can be collected from
to identify a particular configuration of distinct endotypes of sleep dis­ quick questionnaires, physical measures, and blood tests. The model
orders (Schutz et al., 2019). Electroencephalographic microstructures achieved area-under-the-curve of accuracy around 0.85, showing su­
represent novel metrics that are leveraged by AI to predict the presence perior prognostic accuracy compared with previously published models
of sleep disordered breathing (Chervin et al., 2012; Mullins et al., 2021). (You et al., 2022). The fusion of multimodal data can also provide
Six optimized layers that were converted from a single electrocardio­ complementary information for more accurate diagnosis. As a concrete
gram were analyzed by a CNN to provide a precision method to detect example, a previous study proposed a novel deep learning-based
obstructive sleep apnea (OSA) events, demonstrating excellent perfor­ multimodal and multiscale framework to discriminate individuals with
mance (Urtnasan et al., 2018). AD from individuals from mild cognitive impairment individuals (Lu
Sleep disorders, such as insomnia and OSA, affect a wide range of et al., 2018). Performance increased 4% by merging multiple modalities

7
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

compared with a single modality. Because of the availability of large evaluating efficacy. AI can communicate the most appropriate treatment
public neuroimaging datasets (e.g., the Alzheimer’s Disease Neuro­ or coordinate various approaches based on tracking information about a
imaging Initiative database), many studies have focused on the diag­ patient’s diagnosis, propensities, and treatment progression (Fiske et al.,
nosis of dementia and achieved balance accuracy levels well above 90% 2019). Fourth, AI can optimize medical resources and alleviate the
(Feng et al., 2019; Salehi et al., 2020). shortage of mental health services. In developing countries, AI could
Learning convoluted relationships and patterns from empirical data provide specialist care to patients in remote geographical areas where
and extracting predictive data models through AI have established a mental health professionals are scarce (Luxton, 2016). Moreover, pro­
fertile ground in automatically predicting conversion to AD stages and fessional AI-based systems are unaffected by subjective factors, such as
prognosticating progression precisely. Support vector machine methods disruption, stress, and fatigue, and can help reduce overall costs of
(i.e., the earliest AI-based machine-learning models that determine healthcare (Ray et al., 2022). Smartphones and wearable devices can
optimal decision boundaries) have remained the most common appli­ capture longitudinal, intensive, and multimodal mental health data to
cation in the identification and classification of AD (Cuingnet et al., generate operable clinical insights and predictions at the individual
2013). Convolutional neural networks, which extract T1 weighted level. AI robots and virtual psychotherapists are expected to indepen­
image data and employ two or more machine learning algorithms, have dently complete some simple and repetitive work instead of doctors and
gained popularity in AD classification, with promising performance nurses (Lecomte et al., 2020). To a certain extent, this will save medical
(Bashyam et al., 2020; Ghafoorian et al., 2017). Non-convolutional resources, improve work efficiency, and alleviate the scarcity of psy­
artificial neural network algorithms have typically been used with chiatric professional resources and services.
large-scale data in targeting AD classification (Mehdipour Ghazi et al., Although AI has made great progress in the diagnosis and treatment
2019; Suk et al., 2014). Based on the collective output of multiple of psychiatric disorders, there are still many problems to be solved. AI
intramural decision trees, random forest algorithms generate AD pre­ algorithms themselves suffer from inapplicability, vulnerability, and
diction based on MRI and PET data (Huang et al., 2016). Verbal episodic difficulty in interpretation outside the training domain (Kelly et al.,
memory, global cognitive efficiency, verbal fluency, and fluid functions 2019). We should pay special attention to the following potential
are subsets of optimal neuropsychological predictors of a connected task problems (Fig. 3). First, the training results of AI may be influenced by
from which AI can extract linguistic features to enhance classification sample size and whether the sample is biased. The training of an AI
performance (Asgari et al., 2017; Hernandez-Dominguez et al., 2018). model requires a large amount of sample data. The bias-variance
Machine learning might be more sensitive in detecting early stages of AD trade-off in machine learning refers to the delicate balance between
because it captures multivariable patterns as biomarkers. Combined underfitting and overfitting the training data subset. When a flexible
with high-dimensional neuroimaging data, AI has potential to extract model closely fits the training data subset, it exhibits less bias but runs
sensitive biomarkers from clinical features and diagnose early stages of the risk of not generalizing well to the test data subset, resulting in more
AD. variance (Tandon and Tandon, 2019). Achieving the optimal balance
between underfitting and overfitting is a process of iterative experi­
4. Advantages and disadvantages of AI mentation and fine-tuning (Tandon, 2023a), carefully evaluate the
model’s performance on unseen data, closely monitor the training pro­
AI has great potential to redefine how people diagnose and under­ cess, and continually refine the approach to strike the best trade-off are
stand mental illness (Lee et al., 2021; Liu et al., 2020). A person’s unique crucial for building the most accurate model. Thus, it is not allowed to
biological, psychological, and social characteristics can adequately draw conclusions that are conducive to a clear definition based on a
explain their overall mental health, but these biological, psychological, small sample size because the AI training steps and machine learning
and social characteristics need to be better understood. AI-based stra­ algorithms have the risk of overfitting (Walter et al., 2019). Second,
tegies can serve as better pre-diagnostic screening tools and risk models notable shortcomings of AI include a lack of empathy and compassion
to determine an individual’s susceptibility to mental illness and risk of for patients who suffer from trauma and illness (Blease et al., 2020). In
developing it (Shatte et al., 2019). The use of AI in psychiatry has a the long run, patients may become overly reliant on emotional comfort
number of advantages (Fig. 3). First, it provides an intelligent that is provided by AI (Yew, 2021). Additionally, the data processing
self-assessment system to monitor the evolution of symptoms, treatment flow and result reporting of existing studies are different. Replicability
progress, and drug compliance, which helps to understand a patient’s and reproducibility issues plague computational models, hindering their
status and estimate risk in a timely manner. It is easier to disclose reliability. Many studies applying computational methods to psychiatry
symptoms to doctors through AI rather than direct to the doctor for some fail to follow up on their results, either by not translating them into
patients. Second, machine learning has been used to integrate de­ explicit testable hypotheses or by not conducting specific tests to vali­
mographic data, routine testing, blood indicators, and imaging data to date the hypotheses. Consequently, the field lacks meaningful examples
establish models to assist in the diagnosis of schizophrenia, AD, ADHD, of clinical applications in diagnostics or therapeutics, related to psy­
depression, and other mental disorders. Third, AI models have great chosis (Tandon, 2023b). When machine learning is applied to different
clinical potential in choosing the best treatment plan for patients and disciplines, a unified standard for data analysis and result reporting

Monitor the change of symptoms, treatment progress and Small sample results in over fitting of the model
drug compliance

Improve the early diagnosis rate and reduce the Lack of empathy and compassion, and lack of unified
misdiagnosis rate data analysis and results reporting standards
Advantages Disadvantages
Develop the optimal therapeutic strategies, evaluate the Unable to open the "black box" in the AI decision-
efficacy and predict the prognosis making process

Save medical resources and alleviate the shortage of Ethical issues of personal privacy and medical data
mental health services disclosure

Fig. 3. Advantages and disadvantages of artificial intelligence in psychiatry.

8
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

should be established (Riley, 2019). Third, the current deep learning and treatment (Koutsouleris et al., 2022). Optimized machine learning
technology is to fit a nonlinear function composed of neural networks on algorithms are used to build an early warning, disease classification,
big data. This learning behavior has great advantages in the universality disease prediction, and treatment efficacy prediction model that com­
of applications and meets our requirements for making machines “do bines clinical brain disease and quantitative neurobiological indicators
more.” However, the rationality and reliability of its results cannot be to achieve the construction of a big data medical system.
fully verified because we cannot fully understand the logic of machine Our life has been turned upside down since the onset of COVID-19.
learning results generation. We cannot open the “black box” of the AI Traditional healthcare reliance on outdated technologies has been
decision-making process and thus do not know how or why a particular challenged throughout this crisis. AI applications target timely, accu­
decision was made (Castelvecchi, 2016). Fourth, in terms of patient rate, contextualized care and empower clinicians to make better de­
privacy, the confidentiality of information, the security management of cisions, showing promising abilities to build a new future (Cosic et al.,
data that are collected by virtual or AI robot therapists, and wearable 2020b). Analyzing digital data that are acquired from sensor-based
devices, there are ethical issues that threaten personal privacy and surveillance that are orthogonal and complementary to traditional
confidentiality can result in medical data disclosure. Therefore, clear data streams by AI algorithms have the potential to improve lead times
guidelines need to be formulated with regard to the clinical application and predictions (Rabii et al., 2022; Radin et al., 2022). The ongoing
of AI in psychiatry. COVID-19 situation has had a huge impact on mental health, both
AI can currently provide technical assistance, but its capabilities are indirectly via disruptive societal changes and directly via neuropsychi­
still limited. To improve the performance of AI algorithms, extensive atric sequelae after infection (Sun et al., 2022; Zheng et al., 2021). The
and high-quality databases need to be built (Su et al., 2020). AI should demand for mental health remedies have consequently strongly
also gradually build a lifelong learning framework to avoid catastrophic increased. Because the psychiatric workforce is shrinking during the
forgetting (Sodhani et al., 2020). Additionally, data scientists and cli­ COVID-19 pandemic, now is an optimal time to deploy AI to broaden the
nicians need to increase collaboration and communication to bring AI availability of diagnostic tools and inform prognostic decisions with
closer to clinical applications and improve the interpretability of algo­ algorithm models that accurately match clinicians’ assessments. AI
rithms (Couckuyt et al., 2022). Finally, it is crucial to engage the public techniques will facilitate monitoring long-term mental health effects of
and patients in the research and development of AI algorithms and COVID-19 and future crises that will shape mental health (Cosic et al.,
dynamically and continuously increase public awareness and under­ 2021). With the ongoing digitalization of healthcare records across
standing of its pros and cons. healthcare systems, analyzing large clinical datasets with AI algorithms
is becoming more clinically feasible. Technology giants, such as Apple
5. Conclusion and future perspective and Facebook, have already used models to detect depression and pre­
vent suicide. People often cannot access adequate mental health care
Relative to most diseases, the basis of diagnosis and treatment effi­ during a pandemic. AI implementation can intelligently match re­
cacy in the field of psychiatry are still entirely limited to subjective sources, such as auxiliary support systems, for preclinical emotional and
conscious symptoms and observable clinical symptoms. With the informational support. Machine learning-based clinical workflows also
development of biotechnology and arrival of the age of big data, AI can offer accessible and scalable interventions that can augment traditional
become a bridge to transform medicine, connecting basic research, drug care, presenting a feasible route to augment traditional mental health­
development, and clinical practice. AI approaches in psychiatric care and bridge the gap between treatment and the capacity to deliver it
research combine the clinical, psychometric, and biological validation of (Torous et al., 2021).
digital phenotyping into the implementation of diagnostic and prog­ Current AI technology has made great strides in psychiatry, but its
nostic models in real-world clinical practice (Hauser et al., 2022). Using current functioning is limited. Further advancements and applications
a broad set of NLP techniques, such as statistical analyses of feature are expected in the near future. To better achieve the success of AI in
trends, supervised learning, and unsupervised learning, researchers psychiatry, several aspects should be overcome to bridge the gap. First,
have uncovered patterns of how specific mental health problems man­ the use of AI has historically been focused on unimodal data. Multimodal
ifest in language and characterized changes in mental health (Low et al., AI, although more robust in its estimations over unimodal methods, has
2020). ChatGPT holds significant potential as a promising tool in digital drawbacks in its scalability and time-consuming nature of information
psychiatry. ChatGPT can potentially contribute to digital psychiatry by concatenation (Kline et al., 2022). Both uni- and multimodal AI have
providing information, psychoeducation, initial assessment, symptom advantages and disadvantages. More flexible choices between uni- and
monitoring, emotional support, behavioral tracking, and research in­ multimodal AI should be applied in future detection based on specific
sights, complementing existing mental health services and extending conditions. Second, different psychiatric disorders generally exhibit
their reach. ChatGPT can provide accessible mental health support to similar symptoms, and it can be difficult to distinguish between different
individuals who may have limited access to traditional mental health diagnoses. Researchers are beginning to untangle the common biology
services. By engaging in regular conversations, it can detect changes in that links supposedly distinct psychiatric conditions. More precise AI
language patterns, behavior, or emotional states that might indicate the methodology should be developed to distinguish different diagnoses of
need for further assessment or intervention. The ability of ChatGPT to psychiatric disorders. Third, the real-world application of AI in psychi­
engage in conversations with multiple users simultaneously may help atry still faces many challenges, such as personal privacy, medical data
alleviate the strain on mental health professionals who face heavy disclosure, sample selection bias, and others. More high-quality data are
workloads. The application of ChatGPT in digital psychiatry offers a needed to increase the adaptability of developed models. Fourth, AI
range of benefits but also acknowledges several potential ethical and technicians and experienced psychiatrists are required to jointly build a
scientific integrity challenges (Thornton et al., 2023), including con­ mature data-sharing platform and establish a data standard system.
cerns related to privacy, confidentiality, informed consent, biases in In particular, explainable AI is crucial for the clinical applications in
training data, algorithmic transparency and explainability, maintaining psychiatry, and it plays a vital role in strengthening mutual trust be­
therapeutic boundaries, ensuring the accuracy and reliability of infor­ tween patients and therapists. When AI algorithms are used in health­
mation provided, and addressing potential harm or misuse of the tech­ care, particularly in sensitive domains like psychiatry, it becomes crucial
nology. AI could provide researchers with a large amount of to provide transparent and interpretable explanations for the decisions
corresponding objective features and crucial information by sharing and recommendations made by the AI systems, it should be ensuring all
data on highly engaged populations. Various debiasing approaches in AI stakeholders are aware of its utilization (Zhong et al., 2023). Explainable
models are advocated to alleviate societal inequities and offer help to AI could provide transparency by revealing the underlying factors,
people who are marginalized in terms of access to mental health care features, or patterns that contribute to the decision-making process. This

9
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

transparency holds AI systems accountable for their recommendations, Asgari, M., Kaye, J., Dodge, H., 2017. Predicting mild cognitive impairment from
spontaneous spoken utterances. Alzheimers Dement (N. Y) 3, 219–228. 〈https://
fostering trust and ensuring that the outcomes are aligned with best
www.ncbi.nlm.nih.gov/pubmed/29067328〉.
clinical practices. By providing clear explanations, explainable AI helps Bahado-Singh, R.O., Vishweswaraiah, S., Aydas, B., et al., 2019. Artificial intelligence
patients understand the rationale behind diagnostic or treatment rec­ analysis of newborn leucocyte epigenomic markers for the prediction of autism.
ommendations. Explainable AI could also help identify and address Brain Res 1724, 146457. 〈https://www.ncbi.nlm.nih.gov/pubmed/31521637〉.
Barnett, I., Torous, J., Staples, P., et al., 2018. Relapse prediction in schizophrenia
these biases or errors by providing visibility into the decision-making through digital phenotyping: a pilot study. Neuropsychopharmacology 43,
process. This allows for critical examination of the AI system’s 1660–1666. 〈https://www.ncbi.nlm.nih.gov/pubmed/29511333〉.
behavior and enables corrections or improvements to be made. Bashyam, V.M., Erus, G., Doshi, J., et al., 2020. MRI signatures of brain age and disease
over the lifespan based on a deep brain network and 14 468 individuals worldwide.
In general, AI implementation in psychiatry continues along the path Brain 143, 2312–2324. 〈https://www.ncbi.nlm.nih.gov/pubmed/32591831〉.
of becoming the most transformative technology that humanity has Ben-Zeev, D., Brian, R., Wang, R., et al., 2017. CrossCheck: Integrating self-report,
developed. Future psychiatric research needs to consider the imple­ behavioral sensing, and smartphone use to identify digital indicators of psychotic
relapse. Psychiatr. Rehabil. J. 40, 266–275. 〈https://www.ncbi.nlm.nih.gov/
mentation of AI to improve diagnostic, prognostic, and therapeutic de­ pubmed/28368138〉.
cision making. Big data that track patient symptoms over time in mental Biswal, S., Sun, H., Goparaju, B., et al., 2018. Expert-level sleep scoring with deep neural
healthcare could fuel the integration of AI in psychiatric research and networks. J. Am. Med Inf. Assoc. 25, 1643–1650. 〈https://www.ncbi.nlm.nih.gov/
pubmed/30445569〉.
service delivery. Blease, C., Locher, C., Leon-Carlyle, M., et al., 2020. Artificial intelligence and the future
of psychiatry: Qualitative findings from a global physician survey. Digit Health 6,
CRediT authorship contribution statement 2055207620968355. 〈https://www.ncbi.nlm.nih.gov/pubmed/33194219〉.
Brinker, T.J., Hekler, A., Hauschild, A., et al., 2019. Comparing artificial intelligence
algorithms to 157 German dermatologists: the melanoma classification benchmark.
Jie Sun, Qun-Xi Dong, San-Wang Wang, Yong-Bo Zheng, and Ying Eur. J. Cancer 111, 30–37. 〈https://www.ncbi.nlm.nih.gov/pubmed/30802784〉.
Han wrote the first draft of the manuscript. Ying Han, Xiao-Xing Liu, and Bzdok, D., Meyer-Lindenberg, A., 2018. Machine learning for precision psychiatry:
opportunities and challenges. Biol. Psychiatry Cogn. Neurosci. Neuroimaging 3,
San-Wang Wang constructed the figures. Tang-Sheng Lu, Kai Yuan, and
223–230. 〈https://www.ncbi.nlm.nih.gov/pubmed/29486863〉.
Jie Shi revised the manuscript. Lin Lu, Bin Hu, and Ying Han supervised Carpenter, K.L., Sprechmann, P., Calderbank, R., et al., 2016. Quantifying risk for anxiety
the review and revised the manuscript. All authors contributed to the disorders in preschool children: a machine learning approach. PLoS ONE 11,
article and approved the final version of the manuscript. e0165524. 〈https://www.ncbi.nlm.nih.gov/pubmed/27880812〉.
Casanova, R., Barnard, R.T., Gaussoin, S.A., et al., 2018. Using high-dimensional
machine learning methods to estimate an anatomical risk factor for Alzheimer’s
Declaration of Competing Interest disease across imaging databases. Neuroimage 183, 401–411. 〈https://www.ncbi.
nlm.nih.gov/pubmed/30130645〉.
Castelvecchi, D., 2016. Can we open the black box of AI? Nature 538, 20–23. 〈https://
The authors declare that they have no conflict of interest. www.ncbi.nlm.nih.gov/pubmed/27708329〉.
Chan, C.C., Norel, R., Agurto, C., et al., 2023. Emergence of language related to self-
experience and agency in autobiographical narratives of individuals with
Acknowledgements schizophrenia. Schizophr. Bull. 49, 444–453. 〈https://www.ncbi.nlm.nih.gov/
pubmed/36184074〉.
Images in Figs. 1 and 2 were created with BioRender.com. Chand, G.B., Dwyer, D.B., Erus, G., et al., 2020. Two distinct neuroanatomical subtypes
of schizophrenia revealed using machine learning. Brain 143, 1027–1038. 〈https://
www.ncbi.nlm.nih.gov/pubmed/32103250〉.
Financial Disclosure Chang, M., Womer, F.Y., Gong, X., et al., 2021. Identifying and validating subtypes
within major psychiatric disorders based on frontal–posterior functional imbalance
via deep learning. Mol. Psychiatry 26, 2991–3002. 〈https://www.ncbi.nlm.nih.
This work was supported in part by STI2030-Major Projects (no. gov/pmc/articles/PMC8505253/〉.
2022ZD0212900 and 2021ZD0200800). Che St, Z., Sauver, J., Liu, H., et al., 2018. Deep learning solutions for classifying patients
on opioid use. AMIA Annu Symp. Proc. 2017, 525–534. 〈https://www.ncbi.nlm.nih.
gov/pubmed/29854117〉.
Search strategy and selection criteria Chekroud, A.M., Zotti, R.J., Shehzad, Z., et al., 2016. Cross-trial prediction of treatment
outcome in depression: a machine learning approach. Lancet Psychiatry 3, 243–250.
〈https://www.ncbi.nlm.nih.gov/pubmed/26803397〉.
References for this review were identified by searching PubMed with Chen, Y., Storrs, J., Tan, L., et al., 2014. Detecting brain structural changes as biomarker
the search terms “artificial intelligence,” “machine learning,” “deep from magnetic resonance images using a local feature based SVM approach.
learning,” “mental health,” “psychiatry,” “diagnosis,” “prognosis,” and J. Neurosci. Methods 221, 22–31. 〈https://www.ncbi.nlm.nih.gov/pubmed/2
4041480〉.
“treatment.” Additional articles were identified by searching the refer­ Chervin, R.D., Shelgikar, A.V., Burns, J.W., 2012. Respiratory cycle-related EEG changes:
ences of relevant articles. Only papers that were published in English response to CPAP. Sleep 35, 203–209. 〈https://www.ncbi.nlm.nih.gov/pubmed/222
were reviewed. The final list of included articles was generated based on 94810〉.
Colombo, F., Calesella, F., Mazza, M.G., et al., 2022. Machine learning approaches for
relevance and originality with regard to the broad scope of this review. prediction of bipolar disorder based on biological, clinical and neuropsychological
markers: A systematic review and meta-analysis. Neurosci. Biobehav Rev. 135,
References 104552. 〈https://www.ncbi.nlm.nih.gov/pubmed/35120970〉.
Corchado, J.M., Bajo, J., Abraham, A., 2008. GerAmi: Improving healthcare delivery in
geriatric residences. IEEE Intell. Syst. 23, 19–25. 〈https://www.computer.org/csdl/
Abbas, H., Garberson, F., Glover, E., et al., 2018. Machine learning approach for early
magazine/ex/2008/02/mex2008020019/13rRUEgarxg〉.
detection of autism by combining questionnaire and home video screening. J. Am.
Cosic, K., Popovic, S., Kukolja, D., et al., 2016. Multimodal analysis of startle type
Med Inf. Assoc. 25, 1000–1007. 〈https://www.ncbi.nlm.nih.gov/pubmed/2974163
responses. Comput. Methods Prog. Biomed. 129, 186–202. 〈https://www.ncbi.nlm.
0〉.
nih.gov/pubmed/26826902〉.
Acion, L., Kelmansky, D., van der Laan, M., et al., 2017. Use of a machine learning
Cosic, K., Popovic, S., Sarlija, M., et al., 2020a. Artificial intelligence in prediction of
framework to predict substance use disorder treatment success. PLoS ONE 12,
mental health disorders induced by the COVID-19 pandemic among health care
e0175383. 〈https://www.ncbi.nlm.nih.gov/pubmed/28394905〉.
workers. Croat. Med J. 61, 279–288. 〈https://www.ncbi.nlm.nih.gov/pubmed/3
Afshar, M., Sharma, B., Dligach, D., et al., 2022. Development and multimodal validation
2643346〉.
of a substance misuse algorithm for referral to treatment using artificial intelligence
Cosic, K., Popovic, S., Sarlija, M., et al., 2020b. Impact of human disasters and COVID-19
(SMART-AI): a retrospective deep learning study. Lancet Digit Health 4, e426–e435.
pandemic on mental health: potential of digital psychiatry. Psychiatr. Danub 32,
〈https://www.ncbi.nlm.nih.gov/pubmed/35623797〉.
25–31. 〈https://www.ncbi.nlm.nih.gov/pubmed/32303026〉.
Albert Haque, M.G., Adam S.Miner, Li Fei-Fei, 2018, Measuring depression symptom
Cosic, K., Popovic, S., Sarlija, M., et al., 2021. AI-based prediction and prevention of
severity from spoken language and 3D facial expressions. arXiv preprint arXiv,
psychological and behavioral changes in ex-COVID-19 patients. Front Psychol. 12,
1811.08592. 〈https://arxiv.org/abs/1811.08592〉.
782866. 〈https://www.ncbi.nlm.nih.gov/pubmed/35027902〉.
Allesoe, R.L., Nudel, R., Thompson, W.K., et al., 2022. Deep learning-based integration of
Couckuyt, A., Seurinck, R., Emmaneel, A., et al., 2022. Challenges in translational
genetics with registry data for stratification of schizophrenia and depression. Sci.
machine learning. Hum. Genet. 141, 1451–1466. 〈https://www.ncbi.nlm.nih.gov/
Adv. 8, eabi7293. 〈https://www.ncbi.nlm.nih.gov/pubmed/35767618〉.
pubmed/35246744〉.
Alshaer, H., Hummel, R., Mendelson, M., et al., 2019. Objective relationship between
Cuingnet, R., Glaunes, J.A., Chupin, M., et al., 2013. Spatial and anatomical
sleep apnea and frequency of snoring assessed by machine learning. J. Clin. Sleep.
regularization of SVM: a general framework for neuroimaging data. IEEE Trans.
Med 15, 463–470. 〈https://www.ncbi.nlm.nih.gov/pubmed/30853041〉.

10
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

Pattern Anal. Mach. Intell. 35, 682–696. 〈https://www.ncbi.nlm.nih.gov/ J. Autism Dev. Disord. 49, 11–21. 〈https://www.ncbi.nlm.nih.gov/pubmed/3
pubmed/22732664〉. 0019273〉.
Danker-Hopfe, H., Kunz, D., Gruber, G., et al., 2004. Interrater reliability between scorers Insel, T., Cuthbert, B., Garvey, M., et al., 2010. Research domain criteria (RDoC): toward
from eight European sleep laboratories in subjects with different sleep disorders. a new classification framework for research on mental disorders. Am. J. Psychiatry
J. Sleep. Res 13, 63–69. 〈https://www.ncbi.nlm.nih.gov/pubmed/14996037〉. 167, 748–751. 〈https://pubmed.ncbi.nlm.nih.gov/20595427/〉.
Dedoncker, J., Vanderhasselt, M.A., Ottaviani, C., et al., 2021. Mental health during the Jaganathan, K., Kyriazopoulou Panagiotopoulou, S., McRae, J.F., et al., 2019. Predicting
COVID-19 pandemic and beyond: The importance of the vagus nerve for splicing from primary sequence with deep learning. Cell 176, 535–548 e524.
biopsychosocial resilience. Neurosci. Biobehav Rev. 125, 1–10. 〈https://www.ncbi. 〈https://www.ncbi.nlm.nih.gov/pubmed/30661751〉.
nlm.nih.gov/pubmed/33582230〉. James, C., Ranson, J.M., Everson, R., et al., 2021. Performance of machine learning
Dipnall, J.F., Pasco, J.A., Berk, M., et al., 2016. Fusing data mining, machine learning algorithms for predicting progression to dementia in memory clinic patients. JAMA
and traditional statistics to detect biomarkers associated with depression. PLoS ONE Netw. Open 4, e2136553. 〈https://www.ncbi.nlm.nih.gov/pubmed/34913981〉.
11, e0148195. 〈https://www.ncbi.nlm.nih.gov/pubmed/26848571〉. Joyce, J.B., Grant, C.W., Liu, D., et al., 2021. Multi-omics driven predictions of response
Disease, G.B.D., Injury, I., Prevalence, C., 2018. Global, regional, and national incidence, to acute phase combination antidepressant therapy: a machine learning approach
prevalence, and years lived with disability for 354 diseases and injuries for 195 with cross-trial replication. Transl. Psychiatry 11, 513. 〈https://www.ncbi.nlm.nih.
countries and territories, 1990-2017: a systematic analysis for the Global Burden of gov/pubmed/34620827〉.
Disease Study 2017. Lancet 392, 1789–1858. 〈https://www.ncbi.nlm.nih.gov/ Kelly, C.J., Karthikesalingam, A., Suleyman, M., et al., 2019. Key challenges for
pubmed/30496104〉. delivering clinical impact with artificial intelligence. BMC Med 17, 195. 〈https://
Doborjeh, M., Doborjeh, Z., Merkin, A., et al., 2021. Personalised predictive modelling www.ncbi.nlm.nih.gov/pubmed/31665002〉.
with brain-inspired spiking neural networks of longitudinal MRI neuroimaging data Kim, E.S., Berkovits, L.D., Bernier, E.P., et al., 2013. Social robots as embedded
and the case study of dementia. Neural Netw. 144, 522–539. 〈https://www.ncbi. reinforcers of social behavior in children with autism. J. Autism Dev. Disord. 43,
nlm.nih.gov/pubmed/34619582〉. 1038–1049. 〈https://www.ncbi.nlm.nih.gov/pubmed/23111617〉.
Dwyer, D.B., Falkai, P., Koutsouleris, N., 2018. Machine learning approaches for clinical Kline, A., Wang, H., Li, Y., et al., 2022. Multimodal machine learning in precision health:
psychology and psychiatry. Annu Rev. Clin. Psychol. 14, 91–118. 〈https://www. A scoping review. NPJ Digit Med 5, 171. 〈https://www.ncbi.nlm.nih.gov/pubmed/3
ncbi.nlm.nih.gov/pubmed/29401044〉. 6344814〉.
Esteva, A., Robicquet, A., Ramsundar, B., et al., 2019. A guide to deep learning in Koutsouleris, N., Hauser, T.U., Skvortsova, V., et al., 2022. From promise to practice:
healthcare. Nat. Med. 25, 24–29. 〈https://www.ncbi.nlm.nih.gov/pubmed/3061733 towards the realisation of AI-informed mental health care. Lancet Digit Health 4,
5〉. e829–e840. 〈https://www.ncbi.nlm.nih.gov/pubmed/36229346〉.
Feng, C., Elazab, A., Yang, P., et al., 2019. Deep learning framework for Alzheimer’s Krause, A.J., Simon, E.B., Mander, B.A., et al., 2017. The sleep-deprived human brain.
disease diagnosis via 3D-CNN and FSBi-LSTM. IEEE Access 7, 63605–63618. Nat. Rev. Neurosci. 18, 404–418. 〈https://www.ncbi.nlm.nih.gov/pubmed/2
〈https://ieeexplore.ieee.org/document/8703035〉. 8515433〉.
Fernandez, C., Rusk, S., Glattard, N., et al., 2019. Using novel EEG phenotypes and Lecomte, T., Potvin, S., Corbiere, M., et al., 2020. Mobile Apps for mental health issues:
artificial intelligence to estimate OSA severity. A375-A375 Sleep 42. https://doi.org/ meta-review of meta-analyses. JMIR Mhealth Uhealth 8, e17458. 〈https://www.
10.1093/sleep/zsz067.930. ncbi.nlm.nih.gov/pubmed/32348289〉.
Feyaerts, J., Henriksen, M.G., Vanheule, S., et al., 2021. Delusions beyond beliefs: a LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521, 436–444. 〈https://
critical overview of diagnostic, aetiological, and therapeutic schizophrenia research www.ncbi.nlm.nih.gov/pubmed/26017442〉.
from a clinical-phenomenological perspective. Lancet Psychiatry 8, 237–249. Lee, E.E., Torous, J., De Choudhury, M., et al., 2021. Artificial intelligence for mental
〈https://www.ncbi.nlm.nih.gov/pubmed/33485408〉. health care: clinical applications, barriers, facilitators, and artificial wisdom. Biol.
Fischbach, G.D., Lord, C., 2010. The Simons Simplex Collection: a resource for Psychiatry Cogn. Neurosci. Neuroimaging 6, 856–864. 〈https://www.ncbi.nlm.nih.
identification of autism genetic risk factors. Neuron 68, 192–195. 〈https://www. gov/pubmed/33571718〉.
ncbi.nlm.nih.gov/pubmed/20955926〉. Li, A., Zalesky, A., Yue, W., et al., 2020. A neuroimaging biomarker for striatal
Fiske, A., Henningsen, P., Buyx, A., 2019. Your robot therapist will see you now: ethical dysfunction in schizophrenia. Nat. Med 26, 558–565. 〈https://www.ncbi.nlm.nih.go
implications of embodied artificial intelligence in psychiatry, psychology, and v/pubmed/32251404〉.
psychotherapy. J. Med Internet Res 21, e13216. 〈https://www.ncbi.nlm.nih.gov/ Libin, E., Libin, A., 2003. New diagnostic tool for robotic psychology and robotherapy
pubmed/31094356〉. studies. Cyber Behav. 6, 369–374. 〈https://www.ncbi.nlm.nih.gov/pubmed/14511
Fitzpatrick, K.K., Darcy, A., Vierhile, M., 2017. Delivering cognitive behavior therapy to 448〉.
young adults with symptoms of depression and anxiety using a fully automated Linardon, J., Cuijpers, P., Carlbring, P., et al., 2019. The efficacy of app-supported
conversational agent (Woebot): a randomized controlled trial. JMIR Ment. Health 4, smartphone interventions for mental health problems: a meta-analysis of
e7785. 〈http://mental.jmir.org/2017/2/e19/〉. randomized controlled trials. World Psychiatry 18, 325–336. 〈https://www.ncbi.
Fu, J., Yang, S., He, F., et al., 2021. Sch-net: a deep learning architecture for automatic nlm.nih.gov/pubmed/31496095〉.
detection of schizophrenia. Biomed. Eng. Online 20, 75. 〈https://www.ncbi.nlm.nih. Liu, G.-D., Li, Y.-C., Zhang, W., et al., 2020. A brief review of artificial intelligence
gov/pubmed/34344372〉. applications and algorithms for psychiatric disorders. Engineering 6, 462–467.
GBD 2019 Mental Disorders Collaborators, 2022. Global, regional, and national burden https://doi.org/10.1016/j.eng.2019.06.008.
of 12 mental disorders in 204 countries and territories, 1990-2019: a systematic Liu, Y., Qu, H.Q., Mentch, F.D., et al., 2022a. Application of deep learning algorithm on
analysis for the Global Burden of Disease Study 2019. Lancet Psychiatry 9, 137–150. whole genome sequencing data uncovers structural variants associated with multiple
〈https://www.ncbi.nlm.nih.gov/pubmed/35026139〉. mental disorders in African American patients. Mol. Psychiatry 27, 1469–1478.
Ghafoorian, M., Karssemeijer, N., Heskes, T., et al., 2017. Deep multi-scale location- 〈https://www.ncbi.nlm.nih.gov/pubmed/34997195〉.
aware 3D convolutional neural networks for automated detection of lacunes of Liu, W., Fang, P., Guo, F., et al., 2022b. Graph-theory-based degree centrality combined
presumed vascular origin. Neuroimage Clin. 14, 391–399. 〈https://www.ncbi.nlm. with machine learning algorithms can predict response to treatment with
nih.gov/pubmed/28271039〉. antipsychotic medications in patients with first-episode schizophrenia. Dis. Markers
Graham, S., Depp, C., Lee, E.E., et al., 2019. Artificial intelligence for mental health and 2022, 1853002. 〈https://www.ncbi.nlm.nih.gov/pubmed/36277973〉.
mental illnesses: an overview. Curr. Psychiatry Rep. 21, 116. 〈https://www.ncbi. Low, D.M., Rumker, L., Talkar, T., et al., 2020. Natural language processing reveals
nlm.nih.gov/pubmed/31701320〉. vulnerable mental health support groups and heightened health anxiety on reddit
Guo, S., Palaniyappan, L., Liddle, P.F., et al., 2016. Dynamic cerebral reorganization in during COVID-19: observational study. J. Med Internet Res 22, e22635. 〈https://
the pathophysiology of schizophrenia: a MRI-derived cortical thickness study. www.ncbi.nlm.nih.gov/pubmed/32936777〉.
Psychol. Med 46, 2201–2214. 〈https://www.ncbi.nlm.nih.gov/pubmed/27228263〉. Lu, D., Popuri, K., Ding, G.W., et al., 2018. Multimodal and Multiscale Deep Neural
Hauser, T.U., Skvortsova, V., De Choudhury, M., et al., 2022. The promise of a model- Networks for the Early Diagnosis of Alzheimer’s Disease using structural MR and
based psychiatry: building computational models of mental ill health. Lancet Digit FDG-PET images. Sci. Rep. 8, 5697. 〈https://www.ncbi.nlm.nih.gov/pubmed/2
Health 4, e816–e828. 〈https://www.ncbi.nlm.nih.gov/pubmed/36229345〉. 9632364〉.
Henson, P., D’Mello, R., Vaidyam, A., et al., 2021. Anomaly detection to predict relapse Luxton, D.D., 2016. An introduction to artificial intelligence in behavioral and mental
risk in schizophrenia. Transl. Psychiatry 11, 28. 〈https://www.ncbi.nlm.nih.gov/ health care, Artificial intelligence in behavioral and mental health care. Elsevier,,
pubmed/33431818〉. CA, USA, pp. 1–26.
Hernandez-Dominguez, L., Ratte, S., Sierra-Martinez, G., et al., 2018. Computer-based Mehdipour Ghazi, M., Nielsen, M., Pai, A., et al., 2019. Training recurrent neural
evaluation of Alzheimer’s disease and mild cognitive impairment patients during a networks robust to incomplete data: Application to Alzheimer’s disease progression
picture description task. Alzheimers Dement (Amst. ) 10, 260–268. 〈https://www. modeling. Med. Image Anal. 53, 39–46. 〈https://www.ncbi.nlm.nih.gov/pubmed/3
ncbi.nlm.nih.gov/pubmed/29780871〉. 0682584〉.
Hopfield, J.J., 1982. Neural networks and physical systems with emergent collective Mengoni, S.E., Irvine, K., Thakur, D., et al., 2017. Feasibility study of a randomised
computational abilities. Proc. Natl. Acad. Sci. USA 79, 2554–2558. 〈https://www. controlled trial to investigate the effectiveness of using a humanoid robot to improve
ncbi.nlm.nih.gov/pubmed/6953413〉. the social skills of children with autism spectrum disorder (Kaspar RCT): a study
Huang, J., Zhao, Y., Qu, W., et al., 2022. Automatic recognition of schizophrenia from protocol. BMJ Open 7, e017376. 〈https://www.ncbi.nlm.nih.gov/pubmed/2
facial videos using 3D convolutional neural network. Asian J. Psychiatr. 77, 103263. 8645986〉.
〈https://www.ncbi.nlm.nih.gov/pubmed/36152565〉. Mijderwijk, H.J., Steiger, H.J., 2022. Predictive analytics in clinical practice: advantages
Huang, L., Jin, Y., Gao, Y., et al., 2016. Longitudinal clinical score prediction in and disadvantages. Acta Neurochir. Suppl. 134, 263–268. 〈https://www.ncbi.nlm.
Alzheimer’s disease with soft-split sparse regression based random forest. Neurobiol. nih.gov/pubmed/34862550〉.
Aging 46, 180–191. 〈https://www.ncbi.nlm.nih.gov/pubmed/27500865〉. Mohammadi, M., Al-Azab, F., Raahemi, B., et al., 2015. Data mining EEG signals in
Huijnen, C., Lexis, M.A.S., Jansens, R., et al., 2019. Roles, strengths and challenges of depression for their diagnostic value. BMC Med Inf. Decis. Mak. 15, 1–14. 〈https:
using robots in interventions for children with autism spectrum disorder (ASD). //www.ncbi.nlm.nih.gov/pmc/articles/PMC4690290/〉.

11
J. Sun et al. Asian Journal of Psychiatry 87 (2023) 103705

Mullins, A.E., Kim, J.W., Wong, K.K.H., et al., 2021. Sleep EEG microstructure is Tenev, A., Markovska-Simoska, S., Kocarev, L., et al., 2014. Machine learning approach
associated with neurobehavioural impairment after extended wakefulness in for classification of ADHD adults. Int J. Psychophysiol. 93, 162–166. 〈https://www.
obstructive sleep apnea. Sleep. Breath. 25, 347–354. 〈https://www.ncbi.nlm.nih.go ncbi.nlm.nih.gov/pubmed/23361114〉.
v/pubmed/32772308〉. Thornton, J., D’Souza, R., Tandon, R., 2023. Artificial intelligence and psychiatry
Park, J.H., Cho, H.E., Kim, J.H., et al., 2020. Machine learning prediction of incidence of research and practice. Asian J. Psychiatr. 81, 103509. 〈https://www.ncbi.nlm.nih.go
Alzheimer’s disease using large-scale administrative health data. NPJ Digit Med 3, v/pubmed/36806373〉.
46. 〈https://www.ncbi.nlm.nih.gov/pubmed/32258428〉. Topol, E.J., 2019. High-performance medicine: the convergence of human and artificial
Pennisi, P., Tonacci, A., Tartarisco, G., et al., 2016. Autism and social robotics: A intelligence. Nat. Med 25, 44–56. 〈https://www.ncbi.nlm.nih.gov/pubmed/3
systematic review. Autism Res 9, 165–183. 〈https://www.ncbi.nlm.nih.gov/ 0617339〉.
pubmed/26483270〉. Torous, J., Wisniewski, H., Bird, B., et al., 2019. Creating a digital health smartphone app
Perlis, M.L., Posner, D., Riemann, D., et al., 2022. Insomnia. Lancet 400, 1047–1060. and digital phenotyping platform for mental health and diverse healthcare needs: an
〈https://www.ncbi.nlm.nih.gov/pubmed/36115372〉. interdisciplinary and collaborative approach. J. Technol. Behav. Sci. 4, 73–85.
Plana-Ripoll, O., Pedersen, C.B., Agerbo, E., et al., 2019. A comprehensive analysis of https://doi.org/10.1007/s41347-019-00095-w.
mortality-related health metrics associated with mental disorders: a nationwide, Torous, J., Bucci, S., Bell, I.H., et al., 2021. The growing field of digital psychiatry:
register-based cohort study. Lancet 394, 1827–1835. 〈https://www.ncbi.nlm.nih.go current evidence and the future of apps, social media, chatbots, and virtual reality.
v/pubmed/31668728〉. World Psychiatry 20, 318–335. 〈https://www.ncbi.nlm.nih.gov/pubmed/345053
Qu, W., Kao, C.H., Hong, H., et al., 2021. Single-channel EEG based insomnia detection 69〉.
with domain adaptation. Comput. Biol. Med 139, 104989. 〈https://www.ncbi.nlm. Turing, A.M., 1950. Computing machinery and intelligence. Mind N. Ser. 59 (236),
nih.gov/pubmed/34739969〉. 433–460.
Qureshi, M.N.I., Oh, J., Lee, B., 2019. 3D-CNN based discrimination of schizophrenia Urtnasan, E., Park, J.U., Joo, E.Y., et al., 2018. Automated detection of obstructive sleep
using resting-state fMRI. Artif. Intell. Med 98, 10–17. 〈https://www.ncbi.nlm.nih.go apnea events from a single-lead electrocardiogram using a convolutional neural
v/pubmed/31521248〉. network. J. Med Syst. 42, 104. 〈https://www.ncbi.nlm.nih.gov/pubmed/29687192〉.
Rabii, K.B., Javaid, W., Nabeel, I., 2022. Development and implementation of Vaidyam, A.N., Linggonegoro, D., Torous, J., 2021. Changes to the Psychiatric Chatbot
centralised, cloud-based, employee health contact tracing database and predictive Landscape: A Systematic Review of Conversational Agents in Serious Mental Illness:
modelling framework in the COVID-19 pandemic. Lancet Digit Health 4, e770–e772. Changements du paysage psychiatrique des chatbots: une revue systematique des
〈https://www.ncbi.nlm.nih.gov/pubmed/36307190〉. agents conversationnels dans la maladie mentale serieuse. Can. J. Psychiatry 66,
Radin, J.M., Quer, G., Pandit, J.A., et al., 2022. Sensor-based surveillance for digitising 339–348. 〈https://www.ncbi.nlm.nih.gov/pubmed/33063526〉.
real-time COVID-19 tracking in the USA (DETECT): a multivariable, population- Walter, M., Alizadeh, S., Jamalabadi, H., et al., 2019. Translational machine learning for
based, modelling study. Lancet Digit Health 4, e777–e786. 〈https://www.ncbi.nlm. psychiatric neuroimaging. Prog. Neuropsychopharmacol. Biol. Psychiatry 91,
nih.gov/pubmed/36154810〉. 113–121. 〈https://www.ncbi.nlm.nih.gov/pubmed/30290208〉.
Ray, A., Bhardwaj, A., Malik, Y.K., et al., 2022. Artificial intelligence and psychiatry: an Wang, M., Hu, K., Fan, L., et al., 2022. Predicting treatment response in schizophrenia
overview. Asian J. Psychiatr. 70, 103021. 〈https://www.ncbi.nlm.nih.gov/ with magnetic resonance imaging and polygenic risk score. Front. Genet. 13,
pubmed/35219978〉. 848205. 〈https://www.ncbi.nlm.nih.gov/pubmed/35186051〉.
Richards, B., Tsao, D., Zador, A., 2022. The application of artificial intelligence to Wei, Q., Cao, H., Shi, Y., et al., 2023. Machine learning based on eye-tracking data to
biology and neuroscience. Cell 185, 2640–2643. 〈https://www.ncbi.nlm.nih.gov/ identify Autism Spectrum Disorder: A systematic review and meta-analysis.
pubmed/35868269〉. J. Biomed. Inf. 137, 104254. 〈https://www.ncbi.nlm.nih.gov/pubmed/36509416〉.
Riley, P., 2019. Three pitfalls to avoid in machine learning. Nature 572, 27–29. 〈https:// Winterburn, J.L., Voineskos, A.N., Devenyi, G.A., et al., 2019. Can we accurately classify
www.ncbi.nlm.nih.gov/pubmed/31363197〉. schizophrenia patients from healthy controls using magnetic resonance imaging and
Rubin-Falcone, H., Zanderigo, F., Thapa-Chhetry, B., et al., 2018. Pattern recognition of machine learning? A multi-method and multi-dataset study. Schizophr. Res 214,
magnetic resonance imaging-based gray matter volume measurements classifies 3–10. 〈https://www.ncbi.nlm.nih.gov/pubmed/29274736〉.
bipolar disorder and major depressive disorder. J. Affect Disord. 227, 498–505. Wu, W., Zhang, Y., Jiang, J., et al., 2020. An electroencephalographic signature predicts
〈https://www.ncbi.nlm.nih.gov/pubmed/29156364〉. antidepressant response in major depression. Nat. Biotechnol. 38, 439–447. 〈https://
Salehi, A.W., Baglat, P., Sharma, B.B., et al., 2020, A CNN model: earlier diagnosis and www.ncbi.nlm.nih.gov/pubmed/32042166〉.
classification of Alzheimer disease using MRI, 2020 International Conference on Yew, G.C.K., 2021. Trust in and ethical design of carebots: the case for ethics of care. Int.
Smart Electronics and Communication (ICOSEC). IEEE, pp. 156–161. 〈https:// J. Soc. Robot 13, 629–645. 〈https://www.ncbi.nlm.nih.gov/pubmed/32837630〉.
ieeexplore.ieee.org/document/9215402〉. You, J., Zhang, Y.R., Wang, H.F., et al., 2022. Development of a novel dementia risk
Scangos, K.W., Khambhati, A.N., Daly, P.M., et al., 2021. Closed-loop neuromodulation prediction model in the general population: A large, longitudinal, population-based
in an individual with treatment-resistant depression. Nat. Med 27, 1696–1700. machine-learning study. EClinicalMedicine 53, 101665. 〈https://www.ncbi.nlm.nih.
〈https://www.ncbi.nlm.nih.gov/pubmed/34608328〉. gov/pubmed/36187723〉.
Schutz, S.G., Lisabeth, L.D., Shafie-Khorassani, F., et al., 2019. Clinical phenotypes of de Zambotti, M., Cellini, N., Menghini, L., et al., 2020. Sensors capabilities, performance,
obstructive sleep apnea after ischemic stroke: a cluster analysis. Sleep. Med 60, and use of consumer sleep technology. Sleep. Med Clin. 15, 1–30. 〈https://www.
178–181. 〈https://www.ncbi.nlm.nih.gov/pubmed/31186211〉. ncbi.nlm.nih.gov/pubmed/32005346〉.
Shatte, A.B.R., Hutchinson, D.M., Teague, S.J., 2019. Machine learning in mental health: Zarogianni, E., Storkey, A.J., Johnstone, E.C., et al., 2017. Improved individualized
a scoping review of methods and applications. Psychol. Med 49, 1426–1448. prediction of schizophrenia in subjects at familial high risk, based on
〈https://www.ncbi.nlm.nih.gov/pubmed/30744717〉. neuroanatomical data, schizotypal and neurocognitive features. Schizophr. Res 181,
Sodhani, S., Chandar, S., Bengio, Y., 2020. Toward training recurrent neural networks for 6–12. 〈https://www.ncbi.nlm.nih.gov/pubmed/27613509〉.
lifelong learning. Neural Comput. 32, 1–35. 〈https://www.ncbi.nlm.nih.gov/ Zhang, X., Wang, R., Sharma, A., et al., 2021. Artificial intelligence in cognitive
pubmed/31703175〉. psychology—Influence of literature based on artificial intelligence on children’s
Su, C., Xu, Z., Pathak, J., et al., 2020. Deep learning in mental health outcome research: a mental disorders. Aggress. Violent Behav., 101590 https://doi.org/10.1016/j.
scoping review. Transl. Psychiatry 10, 116. 〈https://www.ncbi.nlm.nih.gov/ avb.2021.101590.
pubmed/32532967〉. Zhao, J., Huang, J., Zhi, D., et al., 2020. Functional network connectivity (FNC)-based
Suk, H.I., Lee, S.W., Shen, D., et al., 2014. Hierarchical feature representation and generative adversarial network (GAN) and its applications in classification of mental
multimodal fusion with deep learning for AD/MCI diagnosis. Neuroimage 101, disorders. J. Neurosci. Methods 341, 108756. 〈https://www.ncbi.nlm.nih.gov/
569–582. 〈https://www.ncbi.nlm.nih.gov/pubmed/25042445〉. pubmed/32380227〉.
Sun, J., Yan, W., Zhang, X.N., et al., 2020. Polygenic evidence and overlapped brain Zhao, X., Sun, G., 2021. A multi-class automatic sleep staging method based on
functional connectivities for the association between chronic pain and sleep photoplethysmography signals. Entropy (Basel) 23. 〈https://www.ncbi.nlm.nih.gov/
disturbance. Transl. Psychiatry 10, 252. 〈https://www.ncbi.nlm.nih.gov/pubmed/3 pubmed/33477468〉.
2709872〉. Zheng, Y.B., Sun, J., Liu, L., et al., 2021. COVID-19 vaccine-related psychological stress
Sun, J., Zheng, Y.B., Liu, L., et al., 2022. The impact of quarantine on pain sensation among general public in China. Front Psychiatry 12, 774504. 〈https://www.ncbi.
among the general population in China during the COVID-19 pandemic. Brain Sci. nlm.nih.gov/pubmed/34950070〉.
12. 〈https://www.ncbi.nlm.nih.gov/pubmed/35053822〉. Zhong, Y., Lyu, Y.A., Yu, S., et al., 2023. The issue of evidence-based medicine and
Taliaz, D., Spinrad, A., Barzilay, R., et al., 2021. Optimizing prediction of response to artificial intelligence. Asian J. Psychiatr. 85, 103627. 〈https://www.ncbi.nlm.nih.go
antidepressant medications using machine learning and integrated genetic, clinical, v/pubmed/37201383〉.
and demographic data. Transl. Psychiatry 11, 381. 〈https://www.ncbi.nlm.nih.gov/ Zhou, J., Park, C.Y., Theesfeld, C.L., et al., 2019. Whole-genome deep-learning analysis
pubmed/34238923〉. identifies contribution of noncoding mutations to autism risk. Nat. Genet 51,
Tandon, N., Tandon, R., 2019. Using machine learning to explain the heterogeneity of 973–980. 〈https://www.ncbi.nlm.nih.gov/pubmed/31133750〉.
schizophrenia. Realizing the promise and avoiding the hype. Schizophr. Res 214, Zhou, M., Wang, Q., Zheng, C., et al., 2021. Drug repurposing for opioid use disorders:
70–75. 〈https://www.ncbi.nlm.nih.gov/pubmed/31500998〉. integration of computational prediction, clinical corroboration, and mechanism of
Tandon, R., 2023a. Controversies in psychiatry practice and research. Asian J. Psychiatr. action analyses. Mol. Psychiatry 26, 5286–5296. 〈https://www.ncbi.nlm.nih.gov/
81, 103508. 〈https://www.ncbi.nlm.nih.gov/pubmed/36806374〉. pubmed/33432189〉.
Tandon, R., 2023b. Computational psychiatry and the psychopathology of psychosis: Zinchuk, A.V., Gentry, M.J., Concato, J., et al., 2017. Phenotypes in obstructive sleep
promising leads and blind alleys. Schizophr. Res 254, 143–145. 〈https://www.ncbi. apnea: a definition, examples and evolution of approaches. Sleep. Med Rev. 35,
nlm.nih.gov/pubmed/36889180〉. 113–123. 〈https://www.ncbi.nlm.nih.gov/pubmed/27815038〉.

12