Obesity Surgery (2022) 32:2772–2783

https://doi.org/10.1007/s11695-022-06146-1

REVIEW

Artificial Intelligence in Bariatric Surgery: Current Status and Future

Perspectives
Mustafa Bektaş1 · Beata M. M. Reiber1 · Jaime Costa Pereira2 · George L. Burchell3 · Donald L. van der Peet1

Received: 24 February 2022 / Revised: 3 June 2022 / Accepted: 3 June 2022 / Published online: 17 June 2022
© The Author(s) 2022

Abstract
Background Machine learning (ML) has been successful in several fields of healthcare, however the use of ML within
bariatric surgery seems to be limited. In this systematic review, an overview of ML applications within bariatric surgery is
provided.
Methods The databases PubMed, EMBASE, Cochrane, and Web of Science were searched for articles describing ML in
bariatric surgery. The Cochrane risk of bias tool and the PROBAST tool were used to evaluate the methodological quality
of included studies.
Results The majority of applied ML algorithms predicted postoperative complications and weight loss with accuracies up to 98%.
Conclusions In conclusion, ML algorithms have shown promising capabilities in the prediction of surgical outcomes after
bariatric surgery. Nevertheless, the clinical introduction of ML is dependent upon the external validation of ML.

Keywords Artificial intelligence · Machine learning · Deep learning · Bariatric surgery

Introduction Artificial intelligence is defined as computer science

capable of imitating several aspects of human intelligence
Artificial intelligence (AI) is a new field in medicine gaining and behavior [2]. With the use of large datasets, AI
major interest within healthcare, but its development in models can be trained to conduct several complicated
clinical settings is already referred to as a digital revolution tasks [3]. Machine learning, one of the domains of AI, is
for healthcare [1]. a computer system in which models are trained to form
new predictions or decisions by analyzing large quantities
of data [4]. A specific subclass of machine learning known
as deep learning uses multiple layers to analyze imported
Key Points data. In each layer, weights are calculated for several
Machine learning has been used to predict surgical outcomes in
bariatric surgery.
factors from the data. After repeating this process, a final
Surgical outcomes have been predicted with accuracies up to 98%. model is trained and ready to be applied on new data.
External validation is required for the clinical introduction of Examples of both machine and deep learning techniques
machine learning. are presented in Table 1.
* Mustafa Bektaş
Several potentials of AI models have already been dem-
[email protected] onstrated in clinical practice [14, 15]. For example, machine
learning algorithms have been applied to MRI, X-ray, and
1
Department of Gastrointestinal Surgery, Amsterdam UMC CT images to detect tumors in various organs. Addition-
Location Vrije Universiteit Amsterdam, De Boelelaan 1117,
1081 HV Amsterdam, the Netherlands
ally, input from large numbers of electronic health records
2
enabled AI models to identify risk factors for multifactorial
Department of Computer Science, Vrije Universiteit
Amsterdam, De Boelelaan 1105, 1081 HV Amsterdam,
outcomes such as length of stay, mortality, and early hospital
the Netherlands readmission after surgery [16]. Recently, in colorectal sur-
3
Medical Library Department, Amsterdam UMC Location
gery, machine learning was used to predict outcomes such
Vrije Universiteit Amsterdam, De Boelelaan 1117, as lymph node metastasis, response to chemoradiotherapy,
1081 HV Amsterdam, the Netherlands and postoperative complications. For these outcomes,
Obesity Surgery (2022) 32:2772–2783 2773

Table 1  Definitions of subclasses within AI

Subclass Definition

Machine learning (ML) ML involves computer science that is able to perform desired tasks based on input data. When provided
with sufficient data, algorithms can recognize patterns in data and train the model to perform better.
After completion of the final model, the algorithm can be applied to new unknown data [5]
Decision tree (DT) Within a DT model, multiple factors are classified into tree branches. Based on the algorithm, these
branches are divided into nodes, forming several tree pathways. In the end, this model tends to find
the smallest tree that optimally fits the data [6]
Gradient boosting (GBM) In GBM, weights are added to several factors after classification. Afterwards an assessment of weights
occurs, in which weights are modified based on the difficulty to classify the factors. this process is
repeated until a final optimal model is generated [7]
Random forest (RF) RF involves the formation of multiple decision trees with specific values for predictors. This technique
combines all decision trees in order to build an accurate model for predictions [8]
Support vector machine (SVM) SVM models use mapped input data to discover the optimal boundary to separate several classes and
values [9]
Deep learning As a specific branch of machine learning, deep learning can recognize patterns within datasets by using
multiple processing layers. Within each layer, weights are present for several factors within the model.
After the training process, an optimal model is built to perform on new data [10]
Artificial neural networks (ANNs) Similar to our brain system, data is passed through multiple processing layers within ANNs. Each layer
contains weights in order to make decisions for the resulting output. By repeat of this process, this
model can improve results and produce the most accurate model in the end [11]
Convolutional neural networks (CNNs) CNNs are a specific type of neural networks, however no weights are used in the layers. Instead, multi-
ple layers are functioning as filters to register patterns or regions of images [12]
Radiomics A radiomics model analyzes images in order to retrieve specific texture features that are registered as a
0 or 1. By detecting these features, various pathologies could be recognized [13]

Abbreviations: ML, machine learning; DT, decision tree; GBM, gradient boosting machine; RF, random forest; SVM, support vector machine;
ANN, artificial neural networks; CNN, convolutional neural networks

predictions were performed with accuracies up to 96%. This systematic review aims to provide an extensive overview of
could emphasize the potential of machine learning to sup- (potential) machine learning applications within bariatric
port risk stratification and facilitate clinical decision-making surgery.
for general surgeons [17–19].
Currently, bariatric surgery has evolved to being a key in
treating the worldwide pandemic of morbid obesity. Optimal Materials and Methods
postoperative weight loss including resolution of obesity-
related comorbidities leads to a decreased burden of disease Search Strategy
and related mortality [20, 21]. Despite an increasing amount
of large data set studies in bariatric surgery, several factors A systematic search was performed in accordance with the
such as short- and long-term complication rates and weight Cochrane Handbook for Systematic Reviews of Interven-
loss remain unpredictable. An example in which AI could tions version 6.0 and PRISMA guidelines. To identify all
benefit bariatric surgery is insufficient weight loss after relevant publications, systematic searches were conducted
surgery. Ten to thirty percent of patients show insufficient in the bibliographic databases PubMed, Embase.com,
weight loss after bariatric surgery [22]. Risk factors for this Clarivate Analytics/Web of Science Core Collection, and
are extremely diverse varying from socio-economic factors the Wiley/Cochrane Library from inception up to the 7­ th
such as insurance policy to a specific type of microbiome of July 2021. The search included keywords and free text
[23, 24]. A complete overview of all risk factors and ideally terms for (synonyms of) ‘machine learning’ combined with
an algorithm to calculate the risk of insufficient weight loss (synonyms of) ‘digestive system surgical procedures’ and
for each patient separately is still missing. Assembling an ‘bariatric surgery’. The full search strategy can be found in
algorithm to identify both patients at major risk of insuf- the Supplementary information (see Appendix).
ficient weight loss and high risk of postoperative complica-
tions would assist the bariatric surgeon as well as the patient Selection Process
to reach a well-informed decision.
Despite the potential benefits of AI, the scope of machine Two reviewers (MB and JCP) conducted the title and
learning applications is rarely reported. Therefore, this abstract screening independently in accordance with the
2774 Obesity Surgery (2022) 32:2772–2783

inclusion and exclusion criteria. Studies were only selected the retrospective design of these studies, a moderate risk
for full-text assessment if both reviewers agreed on inclu- of bias was found in the intervention classification domain.
sion. Controversies between reviewers were resolved by Furthermore, results of the Probast score per domain are
discussions, resulting in consensus. Studies were included demonstrated in Fig. 2b.
if they met the following criteria: (i) describing machine
learning algorithms within bariatric surgery, (ii) clinical Categorization of Machine Learning Techniques
study, (iii) including adults. Studies were excluded if they
(i) did not describe bariatric surgery specifically, (ii) were Purposes of machine learning were prediction of postop-
not written in English, (iii) were certain publication types: erative complications (n = 5), prediction of the amount of
reviews, editorials, letters, legal cases, or interviews. postoperative weight loss (n = 3), aid in decision-making
preoperatively (n = 1), predicting presence of hiatal her-
Risk of Bias Evaluation nias (n = 1), and prediction of quality of life (n = 1). The
frequency at which each form of machine learning technique
The ROBINS-I assessment tool was applied by two review- was used in the included studies is summarized in Fig 3.
ers (MB and JCP) to evaluate the methodological quality
of included non-randomized studies [25]. Additionally, the Postoperative Complications
PROBAST tool was used by two reviewers (MB and JCP) to
assess the quality of machine learning models [26]. Conflicts Five studies demonstrated the use of machine learning algo-
between reviewers were solved by discussions. rithms to predict postoperative complications.
Sheikhtaheri et al. developed a model to predict postop-
Data Synthesis and Outcome Assessment erative complications within 90 days after one anastomosis
gastric bypass surgery (OAGB), by using an ANN algorithm
Following full-text screening, the following data were [27]. These complications included bleeding, anastomotic
extracted from the included studies; first author, year of pub- leakage, obstruction, intraabdominal abscess, and pulmo-
lication, country, number of patients included, mean age of nary embolism. Thirty-two factors ranging from age and
the study population, percentage of female patients, study BMI to smoking and laboratory test results were considered
design, follow-up time, surgical procedure, type of machine important in this prediction model. For the postoperative
learning, external validation, purpose of machine learning, period of 10 days, the highest accuracy of the model was
outcome measurements, and prediction performance. The obtained; an AUC of 0.98 was observed.
categorization of studies was based on machine learning Cao et al. (2019) applied multiple machine learning algo-
purposes and results were demonstrated separately. rithms to detect severe complications within 30 days after
bariatric surgery [28]. Machine learning techniques included
decision tree, random forest, gradient boosting, SVM, and
Results ANN models. Results have revealed the following perfor-
mances for the models (accuracy, AUC): decision tree (92%,
Study Selection and Characteristics 0.5), random forest (95%, 0.51), gradient boosting (96%,
0.58), SVM (96%, 0.5), and ANN (96%, 0.54).
The systematic literature search generated a total of 1821 ref- Consequently, Cao et al. (2020) applied ANN, and CNN
erences after removal of duplicates. After screening of titles models to predict serious complications within 30 days after
and abstracts, 21 studies remained for full-text assessment. bariatric surgery. Serious complications were defined as Cla-
Eleven full texts were included. The flow chart of the search vien–Dindo classification grade 3b and higher (i.e., anasto-
and selection process is presented in Fig. 1. Table 2 sum- motic leakage, organ failure, or death) [29]. For each model,
marizes the general characteristics of the included studies. the predictive performance was described by means of the
accuracy, and AUC. The ANN model showed an accuracy of
Risk of Bias Evaluation 84%, and an AUC of 0.54. For the CNN model, the accuracy
was 95% and the AUC appeared to be 0.57 for predicting
As all included studies were either retrospective (n = 10) or postoperative complications.
prospective (n = 1) cohort studies, the ROBINS-I assessment The authors of the 4­ th study used ANN and GBM models
tool was used for quality assessment of all included studies to predict gastrointestinal leak and venous thromboembo-
(Fig. 2a). Since the primary outcome of this study was the lism in patients undergoing a laparoscopic gastric bypass or
type of machine learning techniques being used, domains laparoscopic sleeve gastrectomy [30]. For gastrointestinal
such as bias due to confounding and bias in outcome meas- leakage, the ANN and GBM model showed the following
urements obtained low risk of bias scores. However, due to predictive capabilities, respectively; an AUC of 0.75 and
Obesity Surgery (2022) 32:2772–2783 2775

Fig. 1  PRISMA flow diagram

of the search

0.70. In predicting venous thromboembolisms, the ANN following seven factors appeared to be important for the pre-
algorithm and gradient boosting model achieved the follow- diction of 30-day morbidity and mortality: age, race, BMI,
ing values, respectively; an AUC of 0.65 and 0.67. Out of hypertension, diabetes mellitus, functional status, and previ-
37 variables, the most important factors in predicting both ous surgery.
gastrointestinal leakage and venous thromboembolisms
were age, height, and weight-related measures, hematocrit, Weight Loss
albumin, and assistant training level. A history of deep vein
thrombosis was an additional important variable for predic- All three studies aimed to predict postoperative weight loss
tion of venous thromboembolisms. by applying ANN models.
Wise et al. (2019) aimed to predict the readmission rate Piaggi et al. aimed to predict the percentage excess weight
of 3.1%, the reoperation and reintervention rate of 8.7%, loss (%EWL) in women with severe obesity, 2 years after
and the mortality rate of 0.07% within 30 days after lapa- the laparoscopic adjustable gastric banding procedure [32].
roscopic sleeve gastrectomy in a large cohort [31]. For this %EWL at 2 years postoperatively was 48.2%. The ANN
ANN model, an AUC of 0.59 was detected. Moreover, the model developed was based on preoperative data including
Table 2  General characteristics of included studies
2776

Authors Year Country Patients s Age Female (%) Study design Follow-up Surgical proce- Type of External ML Purpose Study out- Prediction
(mean) dures machine learn- valida- comes performance
ing tion (ACC/AUC)

Sheikhtaheri 2019 Iran 1509 39 NS Retrospective 30 days OAGB Neural network Yes Predict post- Accuracy; 0.98/0.97
et al Cohort operative AUC​
complica-
tions
Cao et al 2019 Sweden 37,811 41 75,9 Retrospective 30 days NS Multiple No Predict post- AUC​ NA
Cohort machine operative
learning complica-
tions
Cao et al 2020 Sweden 44,061 42 NS Retrospective 30 days NS Neural network No Predict post- Accuracy; 0.95/0.57
Cohort operative AUC​
complica-
tions
Nudel et al 2021 USA 436,807 45 79,3 Retrospective 30 days Lap gastric Multiple No Predict post- AUC​ -/0.69
Cohort bypass; LSG machine operative
learning complica-
tions
Wise et al 2020 USA 101,721 44 79,4 Retrospective 30 days LSG Neural network No Predict post- AUC​ -/0.59
Cohort operative
complica-
tions
Piaggi et al 2010 Italy 235 42 100 Retrospective 2 years Gastric Band- Neural network No Predict weight AUC​ -/0.80
Cohort ing loss
Wise et al 2016 USA 647 47 79,6 Retrospective 1 year Lap gastric Neural network No Predict weight AUC​ -/0.83
Cohort bypass loss
Lee et al 2007 Taiwan 249 33 71,1 Prospective 2 years OAGB; Gastric Neural network No Predict weight Accuracy 0.94/-
Cohort Banding loss
Aminian et al 2020 USA 13,722 54 65 Retrospective 4 years Lap gastric Random forest No Assist in AUC​ -/0,71
Cohort bypass; decision-
LSG; Gastric making
Banding;
Duodenal
Switch
Assaf et al 2021 Israel 2482 43 62,7 Retrospective - LSG Decision tree No Predict diagno- Accuracy 0.88/-
Cohort sis of hiatal
hernia
Cao et al 2019 Sweden 6687 43 77 Retrospective 5 years Lap gastric Neural network No Predict postop- Mean squared NA
Cohort bypass erative Qual- error
ity of Life

Abbreviations: LSG, laparoscopic sleeve gastrectomy; Lap gastric bypass, laparoscopic gastric bypass; OAGB, one-anastomosis gastric bypass; NS, not specified; ACC​, accuracy; AUC​, area
under the curve; NA, not applicable
Obesity Surgery (2022) 32:2772–2783
Obesity Surgery (2022) 32:2772–2783 2777

a Wise et al. (2016) used an ANN model to predict the

percentage excess body mass index loss (%EBMIL) 180
and 360 days after laparoscopic Roux-en-Y gastric bypass
surgery based on preoperative variables such as BMI, race,
and gender [33]. The %EBMIL was 73.5% 1 year, postopera-
tively. The AUC for this model was observed to be 0.83. The
variables gender, race, BMI, and diabetes mellitus appeared
to be the key factors for postoperative weight loss.
Lastly, Lee et al. used 17 preoperative factors to pre-
dict successful %EWL 2 years after laparoscopic OAGB
or gastric banding [34]. Success in %EWL was defined
as %EWL > 50% which was accomplished by 84% of the
patients. The ANN model showed an accuracy of 94%
b and the type of operation, HbA1c, and triglyceride levels
appeared to be essential for predicting successful %EWL at
2 years postoperatively.

Decision‑Making

Aminian et al. developed a prediction model using an RF

algorithm to estimate the risk of long-term end-organ com-
plications in patients with type 2 diabetes and obesity when
considering bariatric surgery [35]. The discriminating ability
at 10 years was measured in the area under the curve (AUC)
and resulted in the following for the surgical and non-surgi-
cal groups, respectively; all-cause mortality 0.79 and 0.81,
coronary artery events 0.66 and 0.67, heart failure 0.73 and
Fig. 2  a Methodological quality assessment of the non-randomized 0.75, and nephropathy 0.73 and 0.76. The five most impor-
studies, according to ROBINS-I assessment tool. b Quality of tant variables in the prediction models of all-cause mortality
machine learning models according to the Probast tool were age, BMI at enrollment, history of heart failure, insulin
use, and smoking status.

Diagnosis

Assaf et al. developed a decision tree model for the preop-

erative prediction of the presence of hiatal hernias (HH)
in patients undergoing a laparoscopic sleeve gastrectomy
procedure [36]. This is relevant as the presence of a hiatal
hernia may impose per-operative technical challenges
which is why foreknowledge is beneficial. The model
showed an accuracy of 88.2% for the prediction of hiatal
hernias. Thirteen variables were observed to be influencing
the prediction of hiatal hernias, in which reflux symptoms,
higher age, and BMI were discovered to be associated with
Fig. 3  Applied forms of machine learning a higher risk of hiatal hernias. Additionally, lower age and
BMI have been discovered to be related to shorter opera-
tion lengths.
the comprehensive test of psychopathology Minnesota Mul-
tiphasic Personality Inventory-2. The model showed an AUC Postoperative Quality of Life
of 0.80 for this prediction. Age, paranoia, antisocial prac-
tices, and Type A behavior were independent predictors of Cao et al. (2019) built a CNN model to predict the postop-
%EWL. erative health-related quality of life 1, 2, and 5 years after
2778 Obesity Surgery (2022) 32:2772–2783

a primary gastric bypass procedure [37]. The postoperative However, other factors such as race, inflammatory bowel
quality of life was measured by the RAND-SF-36 question- disease, laboratory results, and functional status are more
naire and the obesity-related problems scale (OP). Perfor- controversial. Not all clinical variables were included in a
mance of the machine learning algorithm was presented as similar or homogeneous manner across the included studies.
the mean squared error, indicating the discrepancy between This is despite the hypothesis that inclusion of previously
the observed value and predicted value. The mean squared excluded variables may improve the accuracy of machine
error for the CNN model was 0.035 in predicting the post- learning models to predict postoperative complications and
operative quality of life. related risk factors. In the field of breast cancer surgery,
the exclusion of variables in machine learning models was
prevented by determining many variables based on pre-oper-
Discussion ative, intra-operative, and post-operative means [39]. These
findings could suggest that guidelines are needed to secure a
From this systematic review, it can be concluded that arti- comprehensive list of clinical factors that can be used for an
ficial intelligence has potentials in several fields within optimal training process of machine learning models.
bariatric surgery. Various models have been created to pre- Three studies have attempted to predict postoperative
dict severe complications with AUCs up to 0.98. Secondly, weight loss. Neural networks demonstrated the highest AUC
weight loss was predicted by AUCs ranging from 0.80 to of 0.94 in predicting postoperative weight loss. For decades
0.83. Lastly, an AUC up to 0.81 was observed in predicting now, researchers in the bariatric field have attempted to iden-
the postoperative quality of life, diagnosis, and end-organ tify all risk factors for insufficient weight loss after bariat-
complications of patients with morbid obesity. ric surgery. Multiple studies have shown that postoperative
Five studies have applied machine learning models to pre- weight loss is dependent on multiple factors, both objec-
dict postoperative complications for patients undergoing bar- tive measures such as BMI and subjective measures such as
iatric surgery. Among several models, neural networks have patient-related measures. It could therefore be specifically
shown the highest accuracy of 98% in predicting postopera- beneficial and interesting for bariatric surgeons to implement
tive complications. Ideally, by using machine learning mod- AI as a means of identifying risk factors for, for example,
els, bariatric surgeons will be able to better predict (severe) insufficient WL. However, as Nudel et al. noted [30], exter-
postoperative complications for each unique patient. These nal validation of the machine learning model was missing
predictions can, in theory, influence the decision towards a due to insufficient data. Therefore, more large datasets are
different type of bariatric operation or different timing of the needed before accurate and valid models can be developed.
operation, more specific prophylactic measures to prevent For predicting the risk of long-term end-organ compli-
a certain type of complication, or a shared decision with cations, such as coronary artery events, heart failure, and
complete informed consent. nephropathy in patients suffering from type 2 diabetes and
In a recent study, the “low-risk bariatric patient” was morbid obesity, a random forest model showed an AUC
defined by the absence of factors such as a medical history of 0.66, 0.73, and 0.73, respectively. According to Amin-
of thromboembolic events, diabetes mellitus, and kidney or ian et al. [35], this random forest model may support and
pulmonary disease [38]. In this review, overlapping risk fac- accelerate the process of decision-making toward bariatric
tors have been identified in the included studies predicting surgery. This is desirable as the duration of obesity itself
postoperative complications and weight loss (Table 3). It and the presence of its related comorbidities have repeat-
is of no surprise that age, BMI, previous intra-abdominal edly been reported to lead to less postoperative weight
surgery, diabetes, and cardiovascular disease were identi- loss and higher comorbidity-related mortality [40–42]. As
fied as risk factors for postoperative severe complications. weight loss after bariatric surgery is not always associated

Table 3  Summary of Postoperative complications Postoperative weight loss

overlapping factors for
postoperative complications and Protective factors Risk factors Helping factors Inhibiting factors
weight loss
Low BMI Non-White race Female gender Older age
Diabetes mellitus* Diabetes mellitus*
Older age High BMI
Previous bariatric surgery

BMI, body mass index

* = type not specified
Obesity Surgery (2022) 32:2772–2783 2779

with health-related quality of life, predicting the increase models may have a significant impact on decision-making
in quality of life after bariatric surgery is a welcome algo- within bariatric surgery. As machine learning models are
rithm in the process of expectation management and shared improved and validated, surgeons could be one step closer to
decision-making, preoperatively [43, 44]. Neural networks achieving personalized decision-making for patients under-
have shown a mean squared error of 0.035 in predicting the going bariatric procedures.
postoperative health-related quality of life 1, 2, and 5 years To use machine learning models for the prediction of
after bariatric surgery, indicating an accurate estimation, surgical outcomes in bariatric surgery, data from laparoscopic
since the mean squared error was close to 0. This neural bariatric surgery should be accessible [50]. Laparoscopic
network model might provide the opportunity to improve videos of bariatric procedures could be collected to serve as a
postoperative care and rehabilitation for patients undergoing training database for machine learning models. By providing
bariatric surgery. However, due to missing patient informa- accurate image navigation during surgery, anatomical
tion, the generalizability of this model might be uncertain. landmarks and unexpected intraoperative findings such as
Missing data could be solved by imputation, as this was done adhesions and abdominal wall hernias could be identified
in the study of Tseng et al. [45], in which machine learn- efficiently by machine learning models [51]. In addition,
ing models were used to predict acute kidney injury after perioperative data collected from anesthesiologists could
cardiac surgery. One study predicted the presence of hiatal be collected such as continuous blood pressure measures or
hernias. The importance of hiatal hernia (HH) present at oxygen saturation as factors possibly predicting postoperative
the time of bariatric surgery remains controversial but is complications. Furthermore, as robotic surgery is often
increasingly recommended to be corrected simultaneously performed in bariatric surgery, machine learning models
with the laparoscopic sleeve gastrectomy [46]. Neverthe- could also improve the performance of robotic surgery by
less, gastroesophageal reflux symptoms may worsen or per- providing 3D mapping during surgery and evaluating surgical
sist, and a secondary operation with conversion from sleeve skills afterward [52].
gastrectomy to LRYGB may be necessary [47]. The fore- This review has several limitations. External validation
knowledge of the presence of HH may both influence the cohorts seem to be missing for most studies, indicating the
patient and surgeon in decision-making towards LRYGB and uncertainty of machine learning models. Therefore, big data
predict a longer operation time. However, as the authors of from clinical settings are required to achieve appropriate
this study mention, the accuracy of the models developed is generalizability and accuracy for machine learning models
not impressive and the study should be regarded as proof of [53]. Additionally, due to the presence of inconsistencies in
concept, exploring the possibilities with AI. reported accuracies and AUCs, a meta-analysis could not
Due to the missing external validation in most studies, the be conducted.
first step for future studies in bariatric surgery should be the
inclusion of external validation cohorts to gain more gener-
alizability of machine learning models. Afterwards, clinical Conclusion
trials should be conducted to facilitate the implementation
of ML models within bariatric surgery. For both steps, large In this review, promising predictive capabilities of machine
amounts of data are required for the training process of these learning have been discovered within bariatric surgery.
models. This data could be retrieved from available patient Machine learning has predominantly been used for predic-
databases or robotic surgery, eventually facilitating the train- tion of postoperative complications and weight loss. How-
ing process of machine learning [48, 49]. ever, ML algorithms have mainly been applied to datasets
This review has revealed that machine learning mod- without external validation. To overcome this problem,
els have potentials to predict postoperative complications, additional data from large patient databases, laparoscopic
weight loss, end-organ complications, quality of life, and surgery, or robotic surgery should be used. By validating
preoperative diagnosis. After the necessary steps to improve ML models, the clinical implementation of ML will be
generalizability and clinical validation, machine learning facilitated.
2780 Obesity Surgery (2022) 32:2772–2783

Appendix

Table 4

Table 4  Search strategy in PubMed

Search Query Results

#3 #1 AND #2 1062
#2 “Digestive System Surgical Procedures”[Mesh] OR “Bariatric Surgery”[Mesh] OR “Laparotomy”[Mesh] OR 489,138
“Roux-en-Y”[Tiab] OR “Cholecystostom*”[Tiab] OR “Choledochostom*”[Tiab] OR “Gastroenterostom*”[Tiab]
OR “Jejunoileal Bypass*”[Tiab] OR “Pancreaticojejunostom*”[Tiab] OR “Peritoneovenous Shunt”[Tiab] OR
“Portoenterostom*”[Tiab] OR “Gastric Bypass*”[Tiab] OR “Appendectom*”[Tiab] OR “Cholecystectom*”[Tiab]
OR “Sphincterotom*”[Tiab] OR “Colectom*”[Tiab] OR “Cecostom*”[Tiab] OR “Colostom*”[Tiab] OR
“Duodenostom*”[Tiab] OR “Ileostom*”[Tiab] OR “Jejunostom*”[Tiab] OR “Esophagectom*”[Tiab]
OR “Hemorrhoidectom*”[Tiab] OR “Hepatectom*”[Tiab] OR “Liver Transplant*”[Tiab] OR “Pancreas
Transplant*”[Tiab] OR “Pancreatectom*”[Tiab] OR “Pancreaticoduodenectom*”[Tiab] OR “Proctectom*”[Tiab]
OR “gastrectom*”[tiab] OR “Gastrostom*”[tiab] OR “Esophagoplast*”[tiab] OR “Esophagostom*”[tiab] OR
“Hepatectom*”[tiab]
#1 “Machine Learning”[Mesh] OR “Machine Learning”[Tiab] OR “machine intelligen*”[tiab] OR “machine 154,754
vision*”[tiab] OR “machine learning”[tiab] OR “transfer learning”[tiab] OR “deep learning”[tiab] OR “neural
network*”[tiab] OR “support vector machine*”[tiab] OR “automatic segmentation*”[tiab] OR “Long short term
memory”[tiab] OR “LSTM”[tiab] OR “supervised learning”[tiab] OR “unsupervised learning”[tiab] OR “rein-
forcement learning*”[tiab] OR “hierarchical learning*” [tiab] OR “Image Interpretation*”[tiab] OR “Prediction
model*”[tiab] OR “image recognition”[tiab] OR “perceptron”[tiab]

Table 5

Table 5  Search strategy in Embase.com

Search Query Results

#5 #3 NOT #4 1227
#4 #3 AND (‘chapter’/it OR ‘conference abstract’/it OR ‘conference paper’/it OR ‘conference review’/it OR ‘editorial’/it OR 857
‘erratum’/it OR ‘letter’/it OR ‘note’/it OR ‘short survey’/it OR ‘tombstone’/it)
#3 #1 AND #2 2084
#2 ‘gastrointestinal surgery’/exp OR ‘laparotomy’/exp OR ‘biliary tract surgery’/exp OR (‘Roux-en-Y’ OR ‘Cholecystostom*’ 720,511
OR ‘Choledochostom*’ OR ‘Gastroenterostom*’ OR ‘Jejunoileal Bypass*’ OR ‘Pancreaticojejunostom*’ OR ‘Peritoneo-
venous Shunt’ OR ‘ Portoenterostom*’ OR ‘Gastric Bypass*’ OR ‘Appendectom*’ OR ‘Cholecystectom*’ OR ‘Sphincter-
otom*’ OR ‘Colectom*’ OR ‘Cecostom*’ OR ‘Colostom*’ OR ‘Duodenostom*’ OR ‘Ileostom*’ OR ‘Jejunostom*’ OR
‘Esophagectom*’ OR ‘Hemorrhoidectom*’ OR ‘Hepatectom*’ OR ‘Liver Transplant*’ OR ‘Pancreas Transplant*’ OR
‘Pancreatectom*’ OR ‘Pancreaticoduodenectom*’ OR ‘Proctectom*’ OR ‘gastrectom*’ OR ‘Gastrostom*’ OR ‘Esophago-
plast*’ OR ‘Esophagostom*’ OR ‘Hepatectom*’):ti,ab,kw
#1 ‘machine learning’/exp OR (‘Machine Learning’ OR ‘machine intelligen*’ OR ‘machine vision*’ OR ‘machine learning’ OR 335,846
‘transfer learning’ OR ‘deep learning’ OR ‘neural network*’ OR ‘support vector machine*’ OR ‘automatic segmentation*’
OR ‘Long short term memory’ OR ‘LSTM’ OR ‘supervised learning’ OR ‘unsupervised learning’ OR ‘reinforcement
learning*’ OR ‘hierarchical learning*’ OR ‘Image Interpretation*’ OR ‘Prediction model*’ OR ‘image recognition’ OR
‘perceptron’):ti,ab,kw
Obesity Surgery (2022) 32:2772–2783 2781

Table 6  Search strategy in Clarivate Analytics/Web of Science Core Collection

Search Query Results

#4 #1 AND #2 667
Refined by: [excluding] DOCUMENT TYPES: (LETTER OR MEETING ABSTRACT OR EDITORIAL MATERIAL OR
CORRECTION)
#3 #1 AND #2 747
#2 TS = (“Roux-en-Y” OR “Cholecystostom*” OR “Choledochostom*” OR “Gastroenterostom*” OR “Jejunoileal Bypass*” OR 294,577
“Pancreaticojejunostom*” OR “Peritoneovenous Shunt” OR “Portoenterostom*” OR “Gastric Bypass*” OR “Appendec-
tom*” OR “Cholecystectom*” OR “Sphincterotom*” OR “Colectom*” OR “Cecostom*” OR “Colostom*” OR “Duodenos-
tom*” OR “Ileostom*” OR “Jejunostom*” OR “Esophagectom*” OR “Hemorrhoidectom*” OR “Hepatectom*” OR “Liver
Transplant*” OR “Pancreas Transplant*” OR “Pancreatectom*” OR “Pancreaticoduodenectom*” OR “Proctectom*” OR
“gastrectom*” OR “Gastrostom*” OR “Esophagoplast*” OR “Esophagostom*” OR “Hepatectom*”)
#1 TS = (“Machine Learning” OR “machine intelligen*” OR “machine vision*” OR “machine learning” OR “transfer learning” 477,557
OR “deep learning” OR “neural network*” OR “support vector machine*” OR “automatic segmentation*” OR “Long short
term memory” OR “LSTM” OR “supervised learning” OR “unsupervised learning” OR “reinforcement learning*” OR
“hierarchical learning*” OR “Image Interpretation*” OR “Prediction model*” OR “image recognition” OR “perceptron”)

Table 7  Search strategy in Wiley/Cochrane Library

Search Query Results

#3 #1 AND #2 7
#2 (“Roux en Y” OR “Cholecystostom*” OR “Choledochostom*” OR “Gastroenteros- 4113
tom*” OR “Jejunoileal Bypass*” OR “Pancreaticojejunostom*” OR “Peritoneo-
venous Shunt” OR “Portoenterostom*” OR “Gastric Bypass*” OR “Appendectom*”
OR “Cholecystectom*” OR “Sphincterotom*” OR “Colectom*” OR “Cecostom*”
OR “Colostom*” OR “Duodenostom*” OR “Ileostom*” OR “Jejunostom*” OR
“Esophagectom*” OR “Hemorrhoidectom*” OR “Hepatectom*” OR “Liver Trans-
plant*” OR “Pancreas Transplant*” OR “Pancreatectom*” OR “Pancreaticoduo-
denectom*” OR “Proctectom*” OR “gastrectom*” OR “Gastrostom*” OR “Esopha-
goplast*” OR “Esophagostom*” OR “Hepatectom*”):ti,ab,kw
#1 (“Machine Learning” OR “machine intelligen*” OR “machine vision*” OR “machine 4733
learning” OR “transfer learning” OR “deep learning” OR “neural network*” OR
“support vector machine*” OR “automatic segmentation*” OR “Long short term
memory” OR “LSTM” OR “supervised learning” OR “unsupervised learning” OR
“reinforcement learning*” OR “hierarchical learning*” OR “Image Interpretation*”
OR “Prediction model*” OR “image recognition” OR “perceptron”):ti,ab,kw
2782 Obesity Surgery (2022) 32:2772–2783

Author contribution Mustafa Bektaş: participated in the design of 13. Kumar V, Gu Y, Basu S, Berglund A, Eschrich SA, Schabath MB,
the study, data collection and interpretation. Wrote and submitted the et al. Radiomics: the process and the challenges. Magn Reson
manuscript. Imaging. 2012;30(9):1234–48.
Beata M.M. Reiber: participated in the design of the study, wrote 14. Mintz Y, Brodie R. Introduction to artificial intelligence in medi-
parts of the manuscript, and revised the manuscript critically. cine. Minim Invasive Ther Allied Technol. 2019;28(2):73–81.
Jaime Costa Pereira: participated in the design and data collection 15. Topol EJ. High-performance medicine: the convergence of human
of the study. and artificial intelligence. Nat Med. 2019;25(1):44–56.
George L. Burchell: performed the literature search and wrote parts 16. Hashimoto DA, Rosman G, Rus D, Meireles OR. Artificial intelli-
of the manuscript. gence in surgery: promises and perils. Ann Surg. 2018;268(1):70–6.
Donald L. van der Peet: participated in the design of the study, 17. Eresen A, Li Y, Yang J, Shangguan J, Velichko Y, Yaghmai V,
wrote parts of the manuscript, and revised the manuscript critically. et al. Preoperative assessment of lymph node metastasis in colon
All authors approved the final version of the manuscript. cancer patients using machine learning: a pilot study. Cancer
imaging: the official publication of the International Cancer Imag-
ing Society. 2020;20(1):30.
Declarations 18. Shaish H, Aukerman A, Vanguri R, Spinelli A, Armenta P, Jam-
bawalikar S, et al. Radiomics of MRI for pretreatment prediction
Ethics approval and consent to participate This article does not contain of pathologic complete response, tumor regression grade, and
any studies with human participants or animals performed by any of neoadjuvant rectal score in patients with locally advanced rectal
the authors. Informed consent does not apply. cancer undergoing neoadjuvant chemoradiation: an international
multicenter study. Eur Radiol. 2020;30(11):6263–73.
Conflict of interest The authors declare no competing interests. 19. Bunn C, Kulshrestha S, Boyda J, Balasubramanian N, Birch S,
Karabayir I, et al. Application of machine learning to the pre-
diction of postoperative sepsis after appendectomy. Surgery.
Open Access This article is licensed under a Creative Commons
2021;169(3):671–7.
Attribution 4.0 International License, which permits use, sharing,
20. Schauer PR, Bhatt DL, Kirwan JP, Wolski K, Aminian A, Brethauer
adaptation, distribution and reproduction in any medium or format, as
SA, et al. Bariatric surgery versus intensive medical therapy for
long as you give appropriate credit to the original author(s) and the source,
diabetes - 5-year outcomes. N Engl J Med. 2017;376(7):641–51.
provide a link to the Creative Commons licence, and indicate if changes
21. Syn NL, Cummings DE, Wang LZ, Lin DJ, Zhao JJ, Loh M,
were made. The images or other third party material in this article are
et al. Association of metabolic-bariatric surgery with long-term
included in the article’s Creative Commons licence, unless indicated
survival in adults with and without diabetes: a one-stage meta-
otherwise in a credit line to the material. If material is not included in the
analysis of matched cohort and prospective controlled studies with
article’s Creative Commons licence and your intended use is not permitted
174 772 participants. Lancet. 2021;397(10287):1830–41.
by statutory regulation or exceeds the permitted use, you will need to
22. Linke K, Schneider R, Gebhart M, Ngo T, Slawik M, Peters T,
obtain permission directly from the copyright holder. To view a copy of
et al. Outcome of revisional bariatric surgery for insufficient
this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.
weight loss after laparoscopic Roux-en-Y gastric bypass: an obser-
vational study. Surg Obes Relat Dis. 2020;16(8):1052–9.
23. Karmali S, Brar B, Shi X, Sharma AM, de Gara C, Birch DW.
Weight recidivism post-bariatric surgery: a systematic review.
References Obes Surg. 2013;23(11):1922–33.
24. Faria SL, Santos A, Magro DO, Cazzo E, Assalin HB, Guadagnini
1. Briganti G, Le Moine O. Artificial intelligence in medicine: today D, et al. Gut microbiota modifications and weight regain in mor-
and tomorrow. Front Med. 2020;7:27. bidly obese women after Roux-en-Y gastric bypass. Obes Surg.
2. Chen M, Decary M. Artificial intelligence in healthcare: an 2020;30(12):4958–66.
essential guide for health leaders. Healthc Manage Forum. 25. Sterne JA, Hernán MA, Reeves BC, Savović J, Berkman ND,
2020;33(1):10–8. Viswanathan M, et al. ROBINS-I: a tool for assessing risk of bias
3. Kulkarni S, Seneviratne N, Baig MS, Khan AHA. Artificial in non-randomised studies of interventions. BMJ. 2016;355: i4919.
intelligence in medicine: where are we now? Acad Radiol. 26. Moons KGM, Wolff RF, Riley RD, Whiting PF, Westwood M,
2020;27(1):62–70. Collins GS, et al. PROBAST: a tool to assess risk of bias and
4. Davenport T, Kalakota R. The potential for artificial intelligence applicability of prediction model studies: explanation and elabora-
in healthcare. Future Healthc J. 2019;6(2):94–8. tion. Ann Intern Med. 2019;170(1):W1–33.
5. El Naqa I, Murphy MJ. What Is Machine Learning? In: El Naqa I, 27. Sheikhtaheri A, Orooji A, Pazouki A, Beitollahi M. A clinical
Li R, Murphy M, editors. Machine learning in radiation oncology. decision support system for predicting the early complications
Cham: Springer; 2015. p. 3–11. of one-anastomosis gastric bypass surgery. Obes Surg.
6. Stiglic G, Kocbek S, Pernek I, Kokol P. Comprehensive decision 2019;29(7):2276–86.
tree models in bioinformatics. PLoS ONE. 2012;7(3): e33812. 28. Cao Y, Fang X, Ottosson J, Näslund E, Stenberg E. A compara-
7. Friedman JH. (2002). Stochastic gradient boosting. Computational tive study of machine learning algorithms in predicting severe
Statistics & Data Analysis. 2002;38(4):367–378. complications after bariatric surgery. J Clin Med. 2019;8(5):668.
8. Breiman L. Random Forests. Mach Learn. 2001;45:5–32. 29. Cao Y, Montgomery S, Ottosson J, Näslund E, Stenberg E. Deep
9. Zhang L, Zhou W, Jiao L. Wavelet support vector machine. IEEE learning neural networks to predict serious complications after
Trans Syst Man Cybern B Cybern. 2004;34(1):34–9. bariatric surgery: analysis of Scandinavian obesity surgery regis-
10. Rusk N. Deep learning. Nat Methods. 2016;13:35. try data. JMIR Med Inform. 2020;8(5): e15992.
11. Abraham A. Artificial neural networks. Handbook of measuring 30. Nudel J, Bishara AM, de Geus SWL, Patil P, Srinivasan J, Hess
system design. New Jersey: John Wiley & Sons; 2005:901–908. DT, et al. Development and validation of machine learning models
12. Albawi S, Mohammed TA, Al-Zawi S. Understanding of a convo- to predict gastrointestinal leak and venous thromboembolism after
lutional neural network. 2017 International Conference on Engi- weight loss surgery: an analysis of the MBSAQIP database. Surg
neering and Technology (ICET); 2017 Aug 1–6. Endosc. 2021;35(1):182–91.
Obesity Surgery (2022) 32:2772–2783 2783

31. Wise ES, Amateau SK, Ikramuddin S, Leslie DB. Prediction of 43. Waljee JF, Ghaferi A, Cassidy R, Varban O, Finks J, Chung KC,
thirty-day morbidity and mortality after laparoscopic sleeve gas- et al. Are patient-reported outcomes correlated with clinical
trectomy: data from an artificial neural network. Surg Endosc. outcomes after surgery?: A population-based study. Ann Surg.
2020;34(8):3590–6. 2016;264(4):682–9.
32. Piaggi P, Lippi C, Fierabracci P, Maffei M, Calderone A, Mauri 44. Doll HA, Petersen SE, Stewart-Brown SL. Obesity and physical
M, et al. Artificial neural networks in the outcome prediction and emotional well-being: associations between body mass index,
of adjustable gastric banding in obese women. PLoS ONE. chronic illness, and the physical and mental components of the
2010;5(10): e13624. SF-36 questionnaire. Obes Res. 2000;8(2):160–70.
33. Wise ES, Hocking KM, Kavic SM. Prediction of excess weight 45. Tseng PY, Chen YT, Wang CH, Chiu KM, Peng YS, Hsu SP,
loss after laparoscopic Roux-en-Y gastric bypass: data from an et al. Prediction of the development of acute kidney injury
artificial neural network. Surg Endosc. 2016;30(2):480–8. following cardiac surgery by machine learning. Crit Care.
34. Lee YC, Lee WJ, Lee TS, Lin YC, Wang W, Liew PL, et al. Pre- 2020;24(1):478.
diction of successful weight reduction after bariatric surgery by 46. Chen W, Feng J, Wang C, Wang Y, Yang W, Dong Z, et al. Effect
data mining technologies. Obes Surg. 2007;17(9):1235–41. of concomitant laparoscopic sleeve gastrectomy and hiatal her-
35. Aminian A, Zajichek A, Arterburn DE, Wolski KE, Brethauer SA, nia repair on gastroesophageal reflux disease in patients with
Schauer PR, et al. Predicting 10-year risk of end-organ complica- obesity: a systematic review and meta-analysis. Obes Surg.
tions of type 2 diabetes with and without metabolic surgery: a 2021;31(9):3905–18.
machine learning approach. Diabetes Care. 2020;43(4):852–9. 47. Peng BQ, Zhang GX, Chen G, Cheng Z, Hu JK, Du X. Gas-
36. Assaf D, Rayman S, Segev L, Neuman Y, Zippel D, Goitein D. troesophageal reflux disease complicating laparoscopic sleeve
Improving pre-bariatric surgery diagnosis of hiatal hernia using gastrectomy: current knowledge and surgical therapies. Surgery
machine learning models. Minimally invasive therapy & allied for obesity and related diseases: official journal of the American
technologies: MITAT: official journal of the Society for Minimally Society for Bariatric Surgery. 2020;16(8):1145–55.
Invasive Therapy. 2021;1–7. 48. Chand M, Ramachandran N, Stoyanov D, Lovat L. Robotics, arti-
37. Cao Y, Raoof M, Montgomery S, Ottosson J, Näslund I. Predict- ficial intelligence and distributed ledgers in surgery: data is key!
ing long-term health-related quality of life after bariatric surgery Tech Coloproctol. 2018;22(9):645–8.
using a conventional neural network: a study based on the Scan- 49. Balla A, Batista Rodríguez G, Corradetti S, Balagué C, Fernán-
dinavian obesity surgery registry. J Clin Med. 2019;8(12):2149. dez-Ananín S, Targarona EM. Outcomes after bariatric surgery
38. Gero D, Raptis DA, Vleeschouwers W, van Veldhuisen SL, Mar- according to large databases: a systematic review. Langenbecks
tin AS, Xiao Y, et al. Defining global benchmarks in bariatric Arch Surg. 2017;402(6):885–99.
surgery: a retrospective multicenter analysis of minimally inva- 50. Johnston SS, Morton JM, Kalsekar I, Ammann EM, Hsiao
sive Roux-en-Y gastric bypass and sleeve gastrectomy. Ann Surg. CW, Reps J. Using machine learning applied to real-world
2019;270(5):859–67. healthcare data for predictive analytics: an applied example in
39. Juwara L, Arora N, Gornitsky M, Saha-Chaudhuri P, Velly AM. bariatric surgery. Value in health: the journal of the International
Identifying predictive factors for neuropathic pain after breast cancer Society for Pharmacoeconomics and Outcomes Research.
surgery using machine learning. Int J Med Inform. 2020;141: 104170. 2019;22(5):580–6.
40. Guerreiro V, Neves JS, Salazar D, Ferreira MJ, Oliveira SC, 51. Kitaguchi D, Takeshita N, Hasegawa H, Ito M. Artificial intel-
Souteiro P, et al. Long-term weight loss and metabolic syndrome ligence-based computer vision in surgery: recent advances
remission after bariatric surgery: the effect of sex, age, metabolic and future perspectives. Annals of gastroenterological surgery.
parameters and surgical technique - a 4-year follow-up study. Obes 2021;6(1):29–36.
Facts. 2019;12(6):639–52. 52. Bhandari M, Zeffiro T, Reddiboina M. Artificial intelligence and
41. Heller S, Lingvay I, Marso SP, Tsimikas AP, Pieber TR, Poulter robotic surgery: current perspective and future directions. Curr
NR, et al. Development of a hypoglycaemia risk score to identify Opin Urol. 2020;30(1):48–54.
high-risk individuals with advanced type 2 diabetes in DEVOTE. 53. Ngiam KY, Khor IW. Big data and machine learning algorithms
Diabetes Obes Metab. 2020;22(12):2248–56. for health-care delivery. Lancet Oncol. 2019;20(6):293.
42. Anand SS, Islam S, Rosengren A, Franzosi MG, Steyn K, Yusu-
fali AH, et al. Risk factors for myocardial infarction in women Publisher’s Note Springer Nature remains neutral with regard to
and men: insights from the INTERHEART study. Eur Heart J. jurisdictional claims in published maps and institutional affiliations.
2008;29(7):932–40.