Discover Public Health

Research

Dynamics of COVID‑19 in India using WANFIS (Wavelet Adaptive

Neuro‑Fuzzy Inference System) model
Sarbjit Singh1,3 · Kulwinder Singh Parmar2 · Jatinder Kumar3

Received: 2 December 2024 / Accepted: 7 May 2025

© The Author(s) 2025  OPEN

Abstract
The preset paper discusses the COVID-19 pandemic in India and the development of a data-driven model to predict
COVID-19 confirmed cases, casualties, and recoveries in the country. The Coronavirus Disease 2019 (COVID-19) was first
identified in December 2019 in the Hubei Province of the People’s Republic of China. It quickly spread to 220 countries
worldwide and had a significant impact. In India, the second wave of COVID-19 hit in April 2021, resulting in over 40
million reported cases and three lakh casualties. India ranked second in COVID-19 infections globally, after the United
States of America. To better understand the dynamics of the COVID-19 pandemic in India, a data-driven WANFIS (Wave-
let Adaptive Neuro-Fuzzy Inference System) model was developed. This model uses discrete wavelet decomposition to
extract information from input data and predict the escalation of confirmed cases, casualties, and recoveries in India.
The WANFIS model’s effectiveness was compared to other models like the artificial neural network (ANN) model and
individual ANFIS model, and it proved to be more robust in predicting COVID-19 transmission. The proposed WANFIS
model has the potential to effectively forecast the transmission of infectious diseases, enabling government and health
officials to anticipate and prepare for emergencies more effectively.

Keywords WANFIS model · COVID-19 · Daubechies wavelet · Fuzzy logic · Neural network

1 Introduction

COVID-19 originated in December 2019 in Wuhan, China, and was declared a pandemic by the World Health Organization
in January 2020. The rapid spread of this contagious virus worsened the situation in China and its neighboring countries.
By the end of March 2020, COVID-19 had spread to nearly every corner of the world. The primary cause of COVID-19 is
the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) [1–5].
The coronavirus belongs to the Coronaviridae family and consists of virus particles, or virions, that are approximately
120 nm in size. The virus has a crown-like or corona appearance due to the presence of club-shaped glycoprotein spikes
on its envelope. Figure 1 illustrates the structure of the SARS-CoV-2 virus [6].
During the first wave of the COVID-19 pandemic, the common symptoms of the original strain included fatigue,
cough, and fever in mild cases, while moderate cases involved mild pneumonia and difficulty breathing. Severe
pneumonia could lead to multiple organ failure and death. However, the second wave of COVID-19 brought new

* Kulwinder Singh Parmar, [email protected]; [email protected]; Sarbjit Singh, [email protected]; Jatinder Kumar,
[email protected] | 1Department of Mathematics, Guru Nanak Dev University College, Narot Jaimal Singh, Pathankot,
Punjab, India. 2Department of Mathematical Sciences, I.K. Gujral Punjab Technical University, Jalandhar, Punjab, India. 3Department
of Mathematics, Guru Nanak Dev University, Amritsar, Punjab, India.

Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

Fig. 1  Structure of SARS-

CoV-2

symptoms such as nausea, abdominal pain, joint pain, vomiting, diarrhea, weakness, loss of appetite, skin rashes,
gastrointestinal problems, and discoloration of fingers and toes. It is important to note that these symptoms may
vary from person to person and everyone should be cautious of both old and new strains of COVID-19.
The highly contagious coronavirus has endangered global health, infected a total of 693,282 people and caused
33,106 deaths as of March 30, 2020, according to the World Health Organization (WHO) [7, 8]. Typical symptoms of
COVID-19 included mild to high fever, continuous dry cough, and life-threatening pneumonia. Older people and
those with chronic diseases are more susceptible to this virus [9–12]. According to WHO situation reports, as of July
12, 2020, 12,401,262 confirmed cases of COVID-19, including 559,047 casualties, were reported globally. The death
toll linked to COVID-19 surpasses that of previously identified coronavirus strains SARS-CoV and MERS-CoV.
During the first wave of COVID-19, the virus did not spread as quickly in India as in other countries across the
globe. The first case of COVID-19 in India was reported on January 30, 2020, in Kerala, but it spread to several parts of
the country within two months. As of July 30, 2020, Maharashtra, Delhi, Andhra Pradesh, Karnataka, and Tamil Nadu
were leading in the list of COVID-19 cases in India. The Ministry of Health and Family Welfare (MoHFW) confirmed
1,583,792 COVID cases, 1,020,582 recoveries, and 34,968 casualties in India as of the same date. With this data, India
stood first in Asia and third-highest in the world in terms of COVID-19 infections.
The United States of America (USA) and Brazil were the worst-hit countries during the first wave of COVID-19. In
response to the pandemic, India enforced a complete lockdown of 21 days in the country on March 24, 2020, and
extended it on April 14, May 3, and May 17. The lockdown significantly affected India’s 1.3 billion population and
had adverse impacts on the country’s economy. Considering the challenging economic conditions, the Government
of India decided to relax the lockdown restrictions in a phased manner, with the first phase of lockdown relaxation
(called Unlock 1) starting on June 02, 2020 [13–15].
The second wave of COVID-19 in India hit in April 2021, overwhelming medical facilities and the healthcare system.
This led to shortages of beds, critical drugs, and oxygen cylinders for COVID patients in government and renowned
hospitals. As of May 26, 2021, India reported 26 million cases, ranking second only to the United States. By June 29,
2021, India had surpassed 30 million infections, with approximately 400,000 fatalities and over 29 million recoveries.
The decline in the second wave occurred in India in the third week of May 2021, with major cities witnessing a drop
in new infections. The second wave of COVID-19 varied across countries, with the USA and the U.K. experiencing
their peaks in early January, while Europe saw its second wave during the usual flu season between November and
January. Jain et al. [16] provided a detailed analysis of the variations between the first and second waves of COVID-
19 in India. The Indian Council of Medical Research (ICMR) in New Delhi predicted a third wave of COVID-19 in India
by the end of 2021, potentially driven by new variants arising from repeated mutations.

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

2 Literature review

The outbreak of COVID-19 has raised questions about its transmission, containment, and safety precautions. Math-
ematical modeling and time series analysis have been instrumental in addressing these concerns. These techniques
have led to the implementation of safety measures such as quarantine, social distancing, lockdowns, and the screen-
ing of infected or suspected individuals. Researchers have developed various models to understand the spread of
the virus and its transmission dynamics. For example, Hellewell et al. [17] observed a decline in new cases with travel
restrictions, while Chen et al. [18] assessed the risk of transmission from bats to humans. Other models have been
proposed to predict outbreak characteristics and fatalities in heavily affected countries. The severity of the pandemic
has prompted continuous data analysis of COVID-19 cases by researchers [19–26].
The coronavirus has emerged as a serious global epidemic, claiming nearly 7.01 million lives and causing signifi-
cant economic disruption. The daily rise in confirmed cases heightens the urgency for medical professionals to find
solutions, making it essential to predict future cases for effective planning. Research shows that asymptomatic indi-
viduals are key players in disease transmission, underscoring the need to include them in our forecasts. Time-series
forecasting techniques are valuable for analyzing infection rates and improving decision-support systems [27, 28].
Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) algorithms were employed by Singh et al. [29] for
detecting contagious disease pneumonia using datasets of chest X-rays from both healthy and infected individuals.
The results revealed a remarkable accuracy of Wavelet Transform-based SVM classification over traditional methods
like K-NN in pneumonia detection. Mahanty et al. [30] proposed a novel approach using chest X-ray (CXR) images to
identify COVID-19 patients. The datasets used for analysis included 2,313 images each of COVID-19 positive cases,
pneumonia cases, and normal cases, utilizing advanced transfer learning models like VGG16, GoogleNet, and Xcep-
tion. The hybridizations of VGG16 and Choquet Fuzzy, GoogleNet and Choquet Fuzzy, and Xception and Choquet
Fuzzy achieved significant accuracies of 97.04%, 98.48%, and 99.57%, respectively.
A novel reinforcement learning method was given by Kukker and Sharma [31] to diagnose pneumonia and tuber-
culosis (TB) from X-ray images. The developed classifier used a modified fuzzy Q-learning (MFQL) algorithm and
wavelet-based pre-processing to assess the severity of pneumonia and tuberculosis. The results showed that the
remarkable accuracy of MFQL approach that outperformed traditional classifiers like Support Vector Machine (SVM)
and k-Nearest Neighbor (KNN).
To analyze CT and X-ray findings from numerous COVID-19 pneumonia patients, Mahanty et al. [32] applied a
Support Vector Machine (SVM) classifier and explored deep transfer learning models like VGG16 and Xception using
Keras and TensorFlow for detecting positive cases from X-ray images. The results found that both VGG16 and Xcep-
tion significantly outperformed the SVM classifier, with Xception achieving an accuracy of 97.46% and an F-score of
98%. A new technique for classifying chest CT scans of COVID-19, pneumonia, and normal patients, using a Sugeno
fuzzy integral ensemble with three transfer learning models—SqueezeNet, DenseNet-201, and MobileNetV2 was
proposed by Mahanty et al. [33]. The performance of this method was found to be superior to the individual models
and other ensemble strategies in accuracy, with MobileNetV2 achieving a 99.15% accuracy rate. The analysis given
by Mahanty et al. [34] focused on outbreaks in India, Pakistan, Myanmar, Brazil, Italy, and Germany up to June 4,
2020, with 28-day projections. The Verhulst model outperformed the others, achieving an R-score of 0.9973, while
the proposed model reached an impressive R-score of 0.9981, demonstrating exceptional accuracy.
In another research, Kukker and Sharma [35] developed a Reinforcement Learning method to accurately identify
Coronavirus disease (COVID-19) alongside viral pneumonia and normal conditions using chest X-ray images. The
JAYA-Optimized Fuzzy Reinforcement Learning algorithm with Wavelet Transform was used for feature extraction
and Principal Component Analysis for feature reduction. The prediction accuracy of the developed approach outper-
formed Support Vector Machines, Artificial Neural Networks, Tree Algorithms, and K-Nearest Neighbors. Mahanty et al.
[36] proposed a hybrid model that combines a fuzzy time series (FTS) approach with a non-linear growth model to
predict active COVID-19 cases in countries Italy, Brazil, India, Germany, Pakistan, and Myanmar. The proposed hybrid
approach forecasted out of sample 26 observations in Phase 1 and 14 observations in Phase 2. The hybrid model out-
performed traditional logistic growth and fuzzy time series methods and improved R-squared values in both Phases.
A groundbreaking technique called Enhanced JAYA (EJAYA) assisted Q-Learning for classifying pulmonary dis-
eases like pneumonia and tuberculosis (TB) from chest X-ray images was developed by Kukker et al. [37]. The valida-
tion of proposed approach on 3000 pneumonia images and 600 TB samples yielded over 98% and 95% accuracies

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

respectively. Training accuracies using the proposed technique for pneumonia classes were approximately 97%, while
for TB detection cases were 96% as compared to contemporary techniques.
In the realm of time series analysis and forecasting during the COVID-19 pandemic, innovative methodologies such as
Wavelet Adaptive Neuro-Fuzzy Inference System (WANFIS), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Artificial
Neural Networks (ANN) have proven invaluable in modeling the complex, nonlinear, and dynamic systems inherent in
this global crisis.
Time series forecasting serves a critical purpose: it aims to predict future values based on historical data. For COVID-19,
this means accurately forecasting cases, deaths, and other crucial epidemiological metrics. A range of forecasting models,
including ARIMA, SARIMA (Seasonal ARIMA), and advanced machine learning techniques such as ANNs and ANFIS, have
been extensively employed to analyze and predict pandemic trends with growing precision.
Traditional statistical models like ARIMA and SARIMA were initially the go-to choices for forecasting COVID-19 cases
based on historical data. However, these models often fall short in capturing the intricate nonlinearities and irregular
patterns that have characterized the pandemic’s trajectory. In contrast, cutting-edge approaches utilizing ANNs, Support
Vector Machines (SVM), and Genetic Algorithms (GA) have surged in popularity. Their robust ability to model complex,
nonlinear relationships within the data provides a significant advantage, offering the flexibility and adaptability that
traditional time series models lack.
These advancements underline the urgent need for adopting these sophisticated methodologies to navigate the
complexities of COVID-19 and improve our forecasting capabilities, ultimately informing effective responses and inter-
ventions in the face of an ongoing global challenge.
In the present study, the WANFIS (Wavelet Adaptive Neuro-Fuzzy Inference System) model is used to predict the
prevalence of confirmed COVID-19 cases, recovered cases, and casualties in India. The hybrid WANFIS model offers a
novel and innovative approach to forecast COVID-19 trends, combining the strengths of multiresolution analysis (via
wavelet decomposition) with the adaptive and interpretable nature of ANFIS. The novelty stems from our application of
Daubechies discrete wavelet decomposition at various orders to analyze COVID-19 datasets.
This hybrid approach effectively addresses the complex and dynamic nature of COVID-19 data. Its ability to manage
noisy inputs, capture both short-term and long-term trends, and provide clear insights makes it a vital tool for accurate
real-time forecasting. This empowers decision-makers in pandemic management across diverse countries and regions
to respond effectively [38–42].

3 Materials and methods

3.1 Datasets for the proposed study

This study utilizes a combined WANFIS (Wavelet Adaptive Neuro-Fuzzy Inference System) model, incorporating discrete
wavelet decomposition and an ANFIS model to analyze the cumulative COVID-19-confirmed infections, recoveries, and
deaths in India. The data range from May 08, 2020, to June 27, 2021, is considered for the study. The first confirmed case
of COVID-19 in India was reported on January 30, 2020, with the first death occurring on March 12, 2020. During the
period from May 8, 2020, to June 27, 2021, we experienced a significant rise in both confirmed cases and deaths. This
challenging time underscored the importance of proactive measures, public awareness, and community support in order
to effectively manage and overcome the pandemic together. A total of 416 sample values in each time series are divided
into sets. The initial 386 sample values are used for training the model, while the remaining 30 samples are reserved
for testing and model validation. This particular data split, with 386 examples allocated for training and 30 for testing,
effectively represents approximately 90% of the data for training and 10% for evaluation. Using a larger training dataset
significantly enhances the model’s capacity to learn patterns, thereby helping to mitigate underfitting. While a smaller test
set may limit comprehensive evaluation, it still provides a valuable, quick insight into the model’s generalization capabili-
ties. This strategy is particularly beneficial when data is scarce, as it maximizes the amount of data used for training [43].
Data Sources: https://​www.​who.​int/​emerg​encies/​disea​ses/​novel-​coron​avirus-​2019 & https://​www.​covid​19ind​ia.​org.

3.2 Wavelet decomposition

A wavelet is a short, localized pulse in both time and space. It is a compactly supported function with finite energy and zero
mean. Unlike the Fourier method, the wavelet method provides a good balance between time location and frequency detail.

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

Wavelet transforms are helpful in revealing time and space information in a signal and can be used to identify patterns in a
given time series [44]. The Wavelet Transform is a highly regarded tool for analyzing signals and time-series data, yet it faces
challenges in directly addressing non-stationary behavior. Despite this limitation, its true potential shines when integrated
with complementary techniques like ARIMA, ANN, SVM, ANFIS etc. By leveraging the hybrid approach, we can unlock power-
ful capabilities for forecasting intricate and non-stationary time-series datasets, making it an invaluable resource for effective
data analysis [45–49].
Wavelet analysis involves using an original function h(t) called a mother wavelet, and then analyzing scaled and translated
versions of it to extract important characteristics of the data [50, 51]. The continuous wavelet transform (CWT) of a signal X (t)
for the mother wavelet h(t) is defined by the following equation:
∞ � �
t−q
p∫
1
CWT (p, q) = √ X(t)h dt (1)
p
−∞

The representation of a signal X (t) in wavelet form consists of a set of coefficients CWT (p, q) for the mother wavelet h(t).
In the first step of the discrete wavelet decomposition of a signal X(t), the signal is mapped to its wavelet coefficients,
producing a low-frequency smooth component, called approximation, and a high-frequency coarse component, called
detail. At level 1, the signal X (t) is decomposed into its approximation (A1 ) and the detail (D1 ), represented by X = A1 + D1.
This decomposition procedure is then applied to the obtained approximation (A1 ) to yield another approximation and detail
at the next finer level producing A1 = A2 + D2. By recursive applying this procedure, low-frequency components (approxi-
mations) A1 , A2 , ..., An and high-frequency components (details) D1 , D2 , ..., Dn are obtained. The entire process of wavelet
decomposition is clarified in the hierarchical diagram (Fig. 2) [52, 53].
The original time series X (t) in terms of approximation and detail coefficients can be represented as

∑
2J - k−1
∑ ∑
J 2J - k−1
X(t) = AJ,k 𝜑j,k (t) + Dj,k 𝜓j,k (t) (2)
k=-∞ j=1 k = - ∞

where AJ,k and Dj,k denote the approximation and the detail coefficients, respectively, and are defined in terms of the
father wavelet function 𝜑J,k and mother wavelet function 𝜓j,k (t) by their respective Eqs. (3) and (4).
∞

∫ (3)
AJ,k = X(t)𝜑J,k (t) dt
−∞

∫ (4)
Dj,k = X(t)𝜓j,k (t) dt
−∞
j ( ) j ( )
with 𝜑j,k (x) = 2 ∕2 𝜑 2j − x and 𝜓j,k (t) = 2 ∕2 𝜓 2j − x .

Fig. 2  Decomposition of
signal X into low- and high-
frequency components

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

3.3 Artificial neural network (ANN)

An Artificial Neural Network (ANN) is a data-driven soft computing model based on a learning algorithm. ANN works
similarly to the human brain, making it an accurate and reliable modeling technique. Traditional statistical modeling
techniques struggle with nonlinear data. ANN’s supervised learning algorithms can handle nonlinearity more effectively,
making it a valuable tool for forecasting complex real-life problems related to the stock market, environmental pollution,
river flow forecasting, weather, agriculture, and engineering [38, 39].
The basic structure of neural networks includes interconnected processing nodes known as neurons. These neurons
can carry out simple mathematical operations [40]. Typically, an artificial neural network (ANN) consists of three layers:
the input layer, hidden layer, and output layer, each containing multiple interconnected nodes (see Fig. 3). Input data is
initially fed into the input layer. In the hidden layer, the processing unit’s summation component calculates the weighted
sum, and the outputs are then generated from this sum using an activation function. Several activation functions, such
as linear, step, ramp, sigmoid, hyperbolic tangent, and Gaussian function, are commonly utilized to produce activation
values [41].
Various types of artificial neural networks (ANNs) are widely used, depending on the architecture. These include feed-
forward neural networks, time delay feed-forward neural networks, recurrent neural networks, Elman back-propagation
neural networks, and convolutional neural networks. In feed-forward neural networks, information flows in one direction,
while recurrent neural networks are able to recall and reuse output information multiple times until a minimum error
is achieved [42]. This ability to retain and reuse information helps the network remember past data, thus enhancing its
performance. ANNs are adept at handling real-time series forecasting problems. Researchers have also developed algo-
rithms that combine ANNs with well-known statistical models to improve prediction accuracy [54–56].

3.4 Adaptive Neuro‑Fuzzy Inference System (ANFIS) model

The Adaptive Neuro-Fuzzy Inference System (ANFIS), also known as a neuro-fuzzy system, is an approach that uses data
to approximate solutions to real-time problems. This method combines two techniques: the neural network (N.N.) and
the fuzzy inference system (FIS). The fuzzy logic handles uncertainty and imprecision in the input by adjusting the con-
necting weights to fuzzy rules, while the neural network with back-propagation provides adaptability [36]. In the first
step of the ANFIS model, input variables are mapped to input membership functions, which, along with fuzzy if–then
rules, create an initial fuzzy model using the input–output data. This initial model generates output characteristics
that map to output membership functions, resulting in the final system output. The working methodology of a fuzzy
inference system (FIS) consists of several steps: (a) fuzzification, (b) aggregation, (c) activation, (d) accumulation, and
(e) defuzzification. The main types of fuzzy systems include the Mamdani fuzzy system, the Singleton fuzzy system, and
the Takagi–Sugeno fuzzy system [57].

Fig. 3  The basic structure of

the Artificial Neural Network

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

To explain ANFIS architecture, let us use a fuzzy inference system with two inputs x , y , and one output z . The rule of
the system is embedded with fuzzy if–then rules of Sugeno’s type fuzzy system.
Rule: If x is P and y is Q , then z is f (x, y), where P and Q are fuzzy sets in the antecedents and z = f (x, y) is a crisp func-
tion in the consequent.
For input values of x and y , z = f (x, y) is generally a polynomial. The constant polynomial f (x, y) leads to the formation
of a zero-order Sugeno fuzzy model, a particular case of the Mamdani fuzzy inference system. Similarly, the first-order
polynomial f (x, y) forms a first-order Sugeno fuzzy model.
The two-rule first-order Sugeno fuzzy inference system is shown in Fig. 4.
Rule 1: If x is P1 and y is Q1, then f1 = a1 x + b1 y + c1.
Rule 2: If x is P2 and y is Q2, then f2 = a2 x + b2 y + c2.
A detailed description of the individual layers of this ANFIS structure is given below:
Layer 1: Every node i is an adaptive square node with a node function O1, i provided by
O1, i = mPi (x), for i = 1, 2 (5)

Or
O1, i = mQi−2 (y), for i = 3, 4 (6)

where x and y are inputs to the node i , and Pi,Qi denote the linguistic labels, and mA(x) are the membership function
which should preferably be a bell-shaped function with minimum value 0 and maximum value 1 and is given by
1
mA(x) =
| x−𝛽i |2𝛾i (7)
1+| 𝛼 |
| i |
Or Gaussian function
( )2
x−𝛽i

mA(x) = e
− 𝛼i (8)
{ }
where 𝛼i , 𝛽i , 𝛾i represents the set of parameters to be learned, called premise parameters.
Layer 2: In this layer, every node i is fixed, and the output generated is the product of all incoming signals. Each node
output represents the firing strength of a rule. The node function for this layer is
O2, i = wi = mPi (x) mQi (y) ; i = 1, 2 (9)

Layer 3: In this layer, every node i is a fixed node labeled and it calculates the ratio of firing strengths of the rules.

Fig. 4  ANFIS Structure

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

wi
O3, i = wi = ; i = 1, 2 (10)
w1 + w2

The outputs generated by this layer are referred to as normalized firing strengths.
Layer 4: In this layer, every node i is an adaptive node with a node function O4, i , and the consequent of the rules are
performed.
( )
O4, i = wi fi = wi pi x + qi y + ri ; i = 1, 2 (11)
{ }
where pi , qi , ri represents the set of consequent parameters.
Layer 5: In this layer, the single node is fixed and is labelled, which evaluates the overall sum of all incoming signals
referred to as the final output.
∑
wi f i
� i
O5, i = wi f i = ∑ (12)
i
wi
i

The Adaptive Neuro-Fuzzy Inference System (ANFIS) is a powerful paradigm that seamlessly integrates the strengths
of neural networks and fuzzy logic. Within this sophisticated framework, activation functions are pivotal for driving
learning and optimization, enhancing the neural network’s performance. Meanwhile, membership functions (MFs) are
instrumental in defining fuzzy sets and crafting rules, thereby reinforcing the system’s fuzzy logic foundation. Together,
these components dramatically elevate ANFIS’s accuracy, interpretability, and efficiency.
In the ANFIS model, activation functions are critical during the backpropagation step, refining parameters within
the fuzzy inference system (FIS). They effectively tackle non-linearity, ensure robust gradient flow, and promote stabil-
ity throughout the learning process, all of which are essential for optimal performance. Membership functions, on the
other hand, serve a vital role in the fuzzification layer, transforming precise input values into fuzzy representations. They
capture the degree of membership of each input to its respective fuzzy set, enabling nuanced decision-making.
In the Adaptive Neuro-Fuzzy Inference System (ANFIS), two crucial hyperparameters Training Epochs and Tolerance
Levels play a pivotal role in shaping the model’s performance:

1. Training Epochs: This parameter dictates the length of the training process. Increasing the number of epochs allows
for meticulous fine-tuning of the membership functions (MFs) and rule parameters, leading to a more robust model.
However, it’s essential to strike the right balance, as excessive epochs can lead to overfitting, while insufficient epochs
may yield underfitting.
2. Tolerance Levels: This parameter sets the precise stopping criterion for training. A low tolerance guarantees thorough
training, ensuring that the model achieves high accuracy, but may extend the training duration. On the other hand,
a high tolerance accelerates training but risks premature termination before optimal performance is reached. To
maximize effectiveness, incorporate gradient monitoring to dynamically adjust the tolerance, ensuring that your
model trains efficiently while achieving peak performance.

3.5 WANFIS (Wavelet Adaptive Neuro‑Fuzzy Inference System) prediction model

The wavelet decomposition method is an excellent tool for breaking down time series data into different frequency
components. This process helps identify important features of the data. However, there are some issues associated
with the wavelet decomposition method, such as choosing the right mother wavelet, its order, and the level of
decomposition [58, 59]. The choice of the mother wavelet depends on how well its attributes match the features of
the data. There are several families of mother wavelets available for wavelet decomposition, such as Haar, Daube-
chies, Symlet, Morlet, Coiflet, and Biorthogonal. Selecting the right mother wavelet for time series decomposition is
essential, as it depends on the time series characteristics, including smoothness and frequency content, as well as
our analytical goals. Daubechies wavelets are favoured for their balance of time–frequency localization and compu-
tational efficiency. These wavelets are widely acclaimed for their exceptional ability to balance time and frequency
resolution effectively. With their compact support, these wavelets are localized in time while maintaining essential
smoothness, which is crucial in analyzing epidemiological data. COVID-19 datasets are often plagued by noise from
reporting errors, missing data, and inconsistencies in testing. Daubechies wavelets excel at denoising signals, skilfully

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

eliminating high-frequency noise while preserving vital low-frequency patterns, such as infection waves and the
effects of lockdown measures. In contrast, Haar wavelets lack the smoothness of Daubechies wavelets and fall short
in handling gradual trend changes. Likewise, Coiflet and Symlet wavelets tend to demand more computational
resources, making them less practical. Furthermore, Meyer and Morlet wavelets suffer from poor time localization,
rendering them less suitable for analyzing discrete signals. In summary, Daubechies wavelets stand out as a superior
tool for comprehensive and accurate analysis in the realm of epidemiology. For tasks like anomaly detection or tran-
sient signal analysis, simpler wavelets like Haar or specialized ones like Morlet can provide distinct advantages worth
considering for better outcomes [60, 61]. The Daubechies wavelet is considered most suitable for non-stationary,
complex, and spiky data. It is known for producing more accurate forecasts compared to other wavelet families, and
it is widely used for prediction purposes [62, 63]
Combining wavelet decomposition with Adaptive Neuro-Fuzzy Inference System (ANFIS) creates a groundbreak-
ing hybrid model that excels at navigating complex, noisy, and nonstationary data. This innovative approach not only
boosts predictive accuracy but also ensures robust performance and wide-ranging applicability across multiple fields.
By effectively extracting significant features, reducing noise, and enhancing computational efficiency, this model stands
out as a powerful tool for data analysis.
Wavelet decomposition dissects a signal into components spanning various frequency bands and time scales, captur-
ing both time-domain and frequency-domain insights. By generating detailed and approximation coefficients, wavelet
decomposition empowers ANFIS to hone in on essential features. This method adeptly captures local features and
trends, enabling ANFIS to accurately model nonstationary behaviors that may otherwise go unnoticed. Furthermore,
wavelet decomposition simplifies the input data by distilling it into fewer, more informative coefficients. This is crucial
for time-series prediction, as it allows the model to account for both short-term fluctuations and long-term trends. The
multiresolution perspective that wavelet decomposition offers lets ANFIS identify patterns at varying detail levels. By
decomposing data into relevant components, the risk of overfitting in the ANFIS model is significantly reduced, enhanc-
ing its overall reliability and effectiveness.
The computational complexity of the Wavelet Adaptive Neuro-Fuzzy Inference System (WANFIS) is primarily influenced
by two critical factors: the order of Daubechies wavelets and the number of wavelet decomposition levels. Both ele-
ments are pivotal in determining processing time, memory usage, and the overall efficiency of the model. As the order of
Daubechies wavelet increases, the filter length expands, enabling superior feature extraction and the ability to capture
intricate details within the signal. This capability can make a substantial difference in performance.
Daubechies wavelets ranging from order 2 (db2) to order 12 (db12) are commonly used in practice. Choosing the
appropriate Daubechies wavelet orders (e.g., db2–db12) for wavelet decomposition, especially with models like ANFIS,
is vital for effective data analysis. Key considerations include the nature of the data, required resolution, and the balance
between computational complexity and performance. Daubechies wavelets are orthogonal with compact support, where
the "order" reflects the number of coefficients or vanishing moments. Lower-Order (db2–db4) are ideal for smoother
signals, ensuring rapid computation and effective handling of noisy data. Mid-Order (db6–db8) provide a balanced
trade-off between time and frequency resolution, suitable for moderately complex signals. Higher-Order (db10–db12)
are best for capturing intricate details or features in complex or nonstationary data [64–66].
The Wavelet Adaptive Neuro-Fuzzy Inference System (WANFIS) model grapples with a significant challenge: overfit-
ting. This issue leads to excessive learning from the training data, including noise and outliers, which compromises the
model’s performance and reliability. The overfitting in WANFIS model can make it overly sensitive to minor variations that
do not reflect genuine trends. It results in erratic and excessively complex fuzzy rules. Such unpredictability makes the
model’s predictions unreliable, often manifesting as random spikes that bear no relation to actual patterns. To effectively
combat overfitting, it is crucial to select an optimal decomposition level of 3 to 5 levels. This strategic choice ensures
that only relevant features are preserved while filtering out noise and extraneous details, preventing the model from
memorizing irrelevant fluctuations. Lower decomposition levels (1–2) retain excessive noise, whereas excessively high
levels (6–8) risk discarding critical data, leading to unstable learning outcomes.
The smoothness of the time series graph increases with the increasing order of Daubechies wavelet. Typically, the
level of decomposition is set to the order of the highest approximation part that aligns with the trend of the original
time series. On the other hand, the corresponding detail parts reveal sharp peaks [36, 67].
The decomposition of time series X (t) by Daubechies wavelet of order p at level l produces the approximations ( Ap, l )
and detail (Dp, l ) components of the same length as that of the original time series and are used separately as inputs for
the Adaptive Neuro-Fuzzy Inference System (ANFIS) model. These constituent approximations and details are partitioned
further into training and testing segments. The training segments of each approximation and detail series are given to

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

an ANFIS model with a predefined tolerance level and some epochs to develop FIS structures. The predicted approxima-
tions ( Â ̂
p,l ) and detail (Dp,l ) coefficients are evaluated by their respective FIS structures over the testing segments of these
coefficients. The sum of the forecasted approximation and detail components gives rise to the final forecast X(t) ̂ of the
time series X (t), which can be expressed mathematically by the following Eq. (13).

X̂
(t) = Â ̂
p, l + Dp, l ; p = 2, 3, ..... , 12 and l = 1, 2, 3, .... (13)

where the hat(∧) symbol indicates the predictions.

A flow chart of the application of the WANFIS model to input time series is shown in Fig. 5.
Many hybrid models rely on raw data or traditional feature extraction techniques such as moving averages and the
Fourier Transform. However, the Discrete Wavelet Transform (DWT) takes analysis to a higher level by decomposing
signals into various frequency bands. This capability not only enhances the identification of short-term spikes but also
uncovers long-term patterns. By effectively filtering out noise while retaining valuable data, DWT significantly improves
prediction accuracy.
The ARIMA-ANN, the wavelet-ANN etc. hybrid models suffer from sensitivity to hyperparameter tuning and lack the
adaptability which is inherent in WANFIS. Numerous studies comparing WANFIS against ANN, LSTM, SVM, and ARIMA in
the realms of COVID-19 forecasting and stock market prediction have consistently demonstrated WANFIS’s superiority.
The WANFIS model offers a powerful solution for tackling complex, noisy, and non-stationary data. While this innovative

Fig. 5  Flow Chart of WANFIS

Model

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

approach can enhance model performance, it does come with challenges, such as increased computational complexity,
the potential for overfitting, and the difficulty of selecting the most suitable wavelet and features. Additionally, it can
diminish model interpretability. Therefore, to unlock the full potential of a WANFIS model, it is crucial to address these
challenges through thoughtful feature selection, and careful hyperparameter tuning. Embracing these strategies can
lead to more effective and reliable outcomes.

4 Application, results, and discussion

Figure 6 displays the time series plots of daily cumulative date-wise cases of COVID-19 confirmation, recoveries, and
deaths in India. The statistics highlight the severity of the second wave in comparison to the first wave of COVID-19 in
India.
Time series plots of cumulative confirmed cases, recoveries, and deaths during India’s COVID-19 pandemic reveal
critical trends from the first and second waves. Initially, cumulative confirmed cases surged, while recoveries lagged
but eventually caught up. Deaths remained steady initially but rose significantly during the second wave. The first wave
began in April 2020, peaking in September and declining until February 2021. By December 2020, recoveries surpassed
new infections, leading to a drop in active cases, with a relatively low mortality rate attributed to effective hospital
management.
In contrast, the second wave was devastating, with cases skyrocketing from March 2021, peaking at over 400,000
daily by April and May—the highest in the world at that time. The healthcare system became overwhelmed, causing a
temporary drop in the recovery rate and a widening gap between cases and recoveries. This situation underscored the
urgent need for effective public health responses in the face of such crises.

4.1 Statistical analysis

The statistical analysis is used to identify trends and patterns in time series data, as well as to show standard statistical
measures and the spread of the data. Table 1 displays the statistical parameters for the cumulative confirmed, recovered,
and death cases of COVID-19 data in India. The minimum and maximum values in Table 1 indicate the wide range of

Fig. 6  Plots of time series of

cumulative daily confirmed
cases, recoveries, and deaths

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

Table 1  Statistical parameters Data Time series No. of sam- Mean Standard Variance Skewness Kurtosis
of COVID-19 dataset ple points (× ­106) Deviation
(× ­105)

Whole Confirmed cases 416 9.80 82.53 6.81 × ­1013 1.01 3.42
Recovered cases 416 8.91 76.71 5.88 × ­1013 1.02 3.64
Death cases 416 0.13 0.96 9.24 × ­109 0.87 3.62
Training Confirmed cases 386 8.28 64.30 4.13 × ­1013 0.85 3.78
Recovered cases 386 7.44 57.62 3.32 × ­1013 0.60 3.27
Death cases 386 0.11 0.72 5.30 × ­109 0.25 2.85
Testing Confirmed cases 30 29.30 7.06 4.99 × ­1011 -0.48 2.08
Recovered cases 30 27.80 11.53 1.33 × ­1012 -0.55 2.13
Death cases 30 36.70 0.23 5.30 × ­108 -0.32 1.66

cumulative confirmations, recoveries, and deaths. High values of standard deviation and variance for each dataset sug-
gest high randomness and scatteredness in this multi-time series data [68, 69].

4.2 Time series analysis

Most real-time series contain non-stationary statistical properties, making it necessary to investigate non-stationarity
for time series analysis. Non-stationarity in time series often results from abrupt changes and randomness. The primary
statistics used to detect non-stationarity in time series are the autocorrelation function (ACF) and partial autocorrelation
function (PACF) plots.
A slowly dwindling ACF pattern indicates randomness in the time series. The PACF plot is commonly used to estimate
the parameters of an autoregressive (A.R.) model. To address non-stationarity, differencing transformation of time series
data is beneficial [63, 70, 71]. Figure 7 shows the ACF and PACF plots of the cumulative confirmed, recovered, and death
cases in India due to COVID-19.
It is evident from the ACF and PACF plots in Fig. 7 that there is randomness in each time series. These plots indicate
that a second-order differencing is needed to eliminate the non-stationarity in the time series of cumulative confirmed
cases, recoveries, and deaths.
Second-order differencing is a powerful technique for transforming non-stationary time series data into a stationary
form by applying the differencing operation twice. This method effectively removes both quadratic trends and seasonal
effects, significantly enhancing the stability of the dataset. By eliminating linear and quadratic trends, second-order
differencing not only stabilizes the mean but also prepares the data for advanced models such as ARIMA and WANFIS,
which rely on the assumption of stationarity. This vital preconditioning step dramatically boosts forecasting accuracy,
making the time series more reliable and minimizing the risk of spurious correlations.

4.3 Artificial neural network

The study uses an artificial neural network (ANN) to make predictions for the time series of daily cumulative confirmed,
recovered, and death cases. It employs a feed-forward neural network with 15 hidden neurons and a single input, and
utilizes the Levemberg-Marquardt training algorithm in the learning process. The sigmoid activation function is used for
learning operations at the hidden layer, while the linear activation function evaluates the final output [72–74]. The Lev-
enberg–Marquardt (LM) algorithm stands out as a premier optimization technique for training Artificial Neural Networks
(ANNs), particularly in addressing complex nonlinear least squares problems. This powerful hybrid approach masterfully
combines the strengths of Gradient Descent (GD) and the Gauss–Newton method, positioning it as one of the fastest
and most efficient training algorithms for moderate-sized networks.
For applications involving highly nonlinear functions, the Levenberg–Marquardt algorithm proves to be remarkably
effective, surpassing simpler methods such as Gradient Descent. While LM is undoubtedly a formidable tool, it is essential
to select optimization methods tailored to the specific complexities and data sizes of a given problem. Overall, the Lev-
enberg–Marquardt algorithm is recognized as one of the fastest and most stable training approaches for ANNs operating
on moderate-sized datasets. However, in the realm of big data or deep learning, employing advanced techniques like

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

Fig. 7  Autocorrelation and partial autocorrelation function plots of time series cumulative confirmed cases, recoveries, and casualties

Adam, RMSprop, or Stochastic Gradient Descent (SGD) becomes imperative, given their superior scalability and reduced
memory demands.
This type of ANN using the Levemberg-Marquardt training algorithm retains the memory of executed operations
during the training process, which helps in producing a more accurate final production of the testing data. Finally, the
performance of the ANN is assessed by comparing the observed samples with its output for the testing time series of
daily cumulative COVID-19 confirmed cases, recoveries, and deaths [75, 76].

4.4 Adaptive Neuro‑Fuzzy Inference System (ANFIS) model

The Sugeno fuzzy inference system with back-propagation utilizes input data from 386 sample points of cumulative
confirmed, recovered, and death cases for training, with 1000 epochs and a tolerance level of 0.0001 [77]. The member-
ship functions in an Adaptive Neuro-Fuzzy Inference System (ANFIS) model are fine-tuned through a powerful hybrid
learning algorithm that seamlessly integrates gradient descent and least-squares estimation. This vital tuning process
effectively optimizes the parameters of the membership functions, significantly reducing the error between the model’s
predictions and the actual target outputs. The essential Steps for tuning and achieving Optimal performance are: (a)

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

selection of membership function type (b) initial parameter settings (c) hyperparameter tuning (d) testing on unseen
data. By meticulously following these steps, we can enhance the model’s accuracy and reliability [78].
The tolerance level and the number of training epochs for a Fuzzy Inference System (FIS) are critical factors that directly
impact the success of model. Striking the right balance between achieving high accuracy and preventing overfitting or
underfitting is essential. By carefully tuning these parameters, we can significantly enhance the convergence speed and
overall performance of the training process, leading to more reliable outcomes. A smaller tolerance level (­ 10–5) and high
epochs are used for high precision.
The predefined tolerance level for Fuzzy Inference System (FIS) structures in ANFIS model is essential for optimizing the
system’s training process. This level significantly impacts the model’s precision, the efficiency of training, and its capability
to generalize to new data. Choosing the right tolerance level for FIS structures requires careful consideration of several
key factors, including model complexity, the desired accuracy, computational limitations, and the characteristics of the
data being used. A well-justified tolerance level can lead to more robust and effective models.
The FIS structure obtained is then used to predict 30 sample testing points. The observed 30 sample testing points are
compared with the outcome sample points estimated by the ANFIS method using different error measures [79]. Figure 5
illustrates the working architecture of an ANFIS model.
To maximize the ANFIS model’s potential, focus on improving its generalization. Begin by normalizing data, address-
ing missing values, and removing outliers. Simplify the model by reducing rules and membership functions to prevent
overfitting. Use k-fold or leave-one-out cross-validation to ensure robust assessment. Employ techniques like bagging
and boosting to enhance model performance. Lastly, test on out-of-sample data to evaluate true generalization [80, 81].

4.5 WANFIS (Wavelet Adaptive Neuro‑Fuzzy Inference System) model

In this study, the WANFIS model is used to analyze the time series data of COVID-19 cases in India, including confirmed
infections, recoveries, and fatalities. The WANFIS model combines two techniques: Wavelet Decomposition and ANFIS.
By employing wavelet decomposition as an initial step, this model skilfully dissects the original signal or time series into
essential components, such as low-frequency trends and high-frequency fluctuations. This decomposition empowers the
ANFIS model to hone in on the significant patterns in the data, leading to enhanced accuracy and a marked reduction
in noise. ANFIS, which integrates the strengths of fuzzy logic and neural networks, creates a powerful framework that
learns from data. This unique combination facilitates a deeper understanding of data relationships while leveraging the
adaptability of neural networks [82, 83].
ANFIS often struggles with memorizing training data instead of truly learning general patterns, which can hinder its
effectiveness. When faced with an excessive number of fuzzy rules, it tends to overfit the noise in the data, significantly
diminishing its predictive power on unseen datasets. Additionally, ANFIS faces challenges with highly nonlinear time
series due to its inflexible membership functions. It also fails to efficiently capture both short-term fluctuations and
long-term trends, leading to suboptimal performance. Moreover, ANFIS is particularly sensitive to noise and does poorly
when data distributions change over time, a phenomenon known as non-stationarity.
In contrast, wavelet decomposition serves as a powerful preprocessing tool that extracts meaningful features from
input data while effectively filtering out noise. The WANFIS model, which integrates wavelet analysis, addresses these
limitations by significantly reducing the number of fuzzy rules needed. This simplification prevents unnecessary complex-
ity and enhances generalization by breaking down intricate patterns into more manageable components. The Wavelet
Transform decomposes signals into multiple frequency components, both low- and high-frequency. This division enables
high-frequency details to capture short-term fluctuations, while low-frequency components effectively model long-term
trends. As a result, WANFIS achieves superior feature extraction and improved generalization on unseen data.
In WANFIS model, we break down the input time series data into low-frequency (approximation) and high-frequency
(detail) components using the Daubechies mother wavelet. We considered Daubechies wavelets of order 2 to 10 and
chose to decompose the time series at level 6 using a Daubechies wavelet. The decomposition level is a crucial factor
that dictates how many times the original signal can be divided into approximation (low-frequency) and detail (high-
frequency) components. By opting for a higher decomposition level, we can extract more intricate features, empowering
the WANFIS model to uncover hidden patterns, trends, and noise variations in the COVID-19 dataset. If the decomposi-
tion level is set too low, significant short-term fluctuations—such as the effects of lockdowns, shifts in policy, or sud-
den outbreak spikes—risk being overlooked. Conversely, an excessively high decomposition level might result in over-
decomposition, causing the loss of vital trend information. The resulting approximation and detail components at this
level are depicted in Figs. 8, 9, and 10 for the confirmed cases, recoveries, and deaths. The low-frequency approximation

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

Fig. 8  Six-level approxima-

tion and detail components
using Daubechies wavelet of
order 10 for the time series of
confirmed cases

Fig. 9  Six-level approxima-

tion and detail components
using Daubechies wavelet of
order 10 for the time series of
Recovered cases

components capture the overall trend of the time series, while the high-frequency detail components reflect its fluctua-
tions. We found that the third order approximation component ( A3 in Figs. 8, 9 and 10) closely aligns with the original
time series trends, as shown in Fig. 6a. Additionally, the localized variations in the time series are reflected in the detail
components D1 to D6. Based on our findings, the wavelet decomposition at level 3 produced accurate forecasts. Our analy-
sis also indicates that the Daubechies wavelet of order 10 outperformed other orders in predicting the time series data.

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

Fig. 10  Six-level approxima-

tion and detail components
using Daubechies wavelet of
order 10 for the time series of
Death cases

After decomposing each time series by Daubechies wavelet of order ten at level 3, the ANFIS model is applied
individually to the highest approximation part A3 and the intricate parts D1,D2,D3 , to yield the respective forecasts
̂1 , D
Â3 and D ̂2 , D
̂3 . The conclusive forecast is the sum of the individual estimates for each time series of confirmed,
recovered, and death cases and is expressed as:

X̂ ̂1 + D
(t) = Â3 + D ̂2 + D
̂3 (14)

The effectiveness of the proposed model can be evaluated by comparing the observed sample values with the
model’s forecasted values using the mean absolute deviation (MAD) and the mean absolute percentage error (MAPE)
values. These values are determined by the following equation:

1 ��
n
⎫
MAD = O − Fi �� ⎪
n i=1 � i ⎪
n � ⎬ (15)
�
100 � �Oi − Fi � ⎪
MAPE =
n i=1 Oi ⎪ ⎭

where Oi and Fi denote the observed and the predicted sample values respectively.
Among the most effective and widely recognized metrics are Mean Absolute Deviation (MAD) and Mean Absolute
Percentage Error (MAPE). Assessing these values across various models not only uncovers which models excel under
particular conditions but also empowers informed decision-making. MAD quantifies the average absolute error in the
same unit as the original data. This metric is invaluable for grasping the absolute magnitude of errors. MAPE, on the
other hand, offers a powerful advantage by expressing errors as a percentage. This scale-independent feature makes
MAPE exceptionally useful for comparing models across diverse datasets and industries. It provides a relative error
metric that not only facilitates comparisons but is also essential for real-world decision-making, where percentage-
based accuracy is often more intuitive. To achieve a truly comprehensive performance evaluation, there are additional
metrics such as Root Mean Square Error (RMSE), Mean Squared Error (MSE), Symmetric Mean Absolute Percentage
Error (sMAPE), and R-squared. By leveraging these metrics together, we can make more robust and informed decisions.

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

The WANFIS model is used to analyze time series data by considering the previous three inputs to predict daily cumula-
tive confirmed cases, recoveries, and deaths. In this analysis, Daubechies wavelets of orders 2, 3, 5, 8, and 10 are employed.
The model is trained on 386 samples of each input time series and then tested on 30 future values for each time series.
The results of training and testing errors for the hybrid prediction model using Db2, Db3, Db5, Db8, and Db10 wavelets
with a decomposition level of 3 are provided in Table 2. It is noted that the predictions using Db10 are more accurate than
those using Daubechies wavelets of the remaining orders. The WANFIS model achieves the lowest Mean Absolute Per-
centage Error (MAPE) values when using Db10, with training and testing MAPE values of 0.3670 and 0.4339, respectively.
Choosing the Daubechies wavelet of order 10 (db10) for wavelet decomposition in a hybrid Wavelet-ANFIS model
significantly enhances prediction accuracy compared to other wavelet families. The advantages of using db10 include:
(a)Precise signal decomposition, capturing both high-frequency and low-frequency components (b)Improved prediction
accuracy for time series data with subtle patterns or noise (c)A balanced approach that fosters strong learning in the
ANFIS model while reducing the risk of overfitting. This makes db10 an excellent choice for effective predictive modeling.
The diagrammatic representation of the predicted approximation part A3 and the intricate details D1,D2,D3 of each
time series of COVID-confirmed cases, recoveries, and deaths for training data using the WANFIS model are shown in
Figs. 11, 12 and 13, respectively. The plots of training vs. trained and testing vs. tested time series of COVID-confirmed
cases, recoveries, and casualties generated by WANFIS are shown in Fig. 14.
The WANFIS model is used to analyze multiple time series data for COVID-19 cases in India, including confirmations,
recoveries, and deaths. The model’s prediction performance is compared with the ANFIS and ANN models using MAD and
MAPE values for training and testing periods, as shown in Table 3. The joint WANFIS model significantly reduces training
and testing errors for each time series. However, the independent ANFIS model demonstrates slightly better outcomes
for both training and testing sets compared to the ANN model. The high training and testing errors of the independent
ANFIS model may be due to its limited ability to generalize. Combining the ANFIS model with wavelet decomposition
is suggested as a way to improve its performance [84]. The accuracy of the proposed WANFIS model is attributed to its
ability to predict wavelet-decomposed data for each time series separately.
In comparing data-driven and soft computing models, it is evident that the WANFIS model outperforms other predic-
tion models (refer to Table 3). The MAPE values of WANFIS, ANFIS and ANN models for COVID-19 datasets of confirmed,
recovered and death cases are respectively 0.8562, 1.1541, 3.2068; 1.2613, 1.9797, 3.8883; and 0.4339, 1.6601, 3.7303.
Specifically, the individual ANFIS model surpasses ANN and reduces percentage errors by approximately 70%, 50%, and
60% for the corresponding time series of COVID-19 confirmations, recoveries, and fatalities.
Additionally, the joint WANFIS model’s predictive efficiency is approximately 80% higher than that of ANN for con-
firmed cases. Furthermore, the proposed model demonstrates around 60% and 90% more effectiveness than ANN in
forecasting COVID-19 recoveries and casualties in India, respectively. Overall, the results of ANFIS and the WANFIS model

Table 2  Training and testing Daubechies wave- Time series Training Testing
results of the WANFIS model let order 4
using Daubechies wavelet of MAD (× ­10 ) MAPE MAD (× ­105) MAPE
orders 2, 3, 5, 8, and 10
Order 2 Confirmed cases 2.5122 0.6654 3.0808 1.0448
(Db2) Recovered cases 2.0670 0.7546 4.4142 1.5689
Death cases 0.0250 0.4626 0.0263 0.7192
Order 3 Confirmed cases 1.7621 0.5843 2.5710 0.8697
(Db3) Recovered cases 1.5586 0.7109 3.2632 1.1539
Death cases 0.0190 0.4358 0.0179 0.4864
Order 5 Confirmed cases 1.6963 0.5894 2.5579 0.8658
(Db5) Recovered cases 1.4254 0.6931 3.3230 1.1753
Death cases 0.0186 0.4314 0.0204 0.5557
Order 8 Confirmed cases 1.7058 0.5537 2.5491 0.8627
(Db8) Recovered cases 1.3950 0.6730 3.6218 1.2835
Death cases 0.0181 0.3794 0.0184 0.4947
Order 10 Confirmed cases 1.7027 0.5438 2.5303 0.8562
(Db10) Recovered cases 1.4217 0.6459 3.5594 1.2613
Death cases 0.0179 0.3670 0.0161 0.4339

Bold values mean best results by Db10 (order 10) as compared to other orders

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

Fig. 11  Plots of training and

trained approximation part A3
and the detailed parts D1,D2,D3
of the time series of confirmed
cases

are superior to those of ANN, making the joint WANFIS model the best performer when compared to ANFIS, ANN, and
regression models.
WANFIS stands out for its innovative application of Discrete Wavelet Transform (DWT) prior to inputting data into
ANFIS. This strategic approach decomposes time series data into a spectrum of frequency components, effectively cap-
turing both low and high-frequency details. By filtering out noise while preserving critical trends, WANFIS significantly
enhances the learning process, leading to superior model performance and a marked reduction in overfitting compared
to conventional ANFIS and ANN models that rely on unprocessed data. Consequently, this noise reduction results in more
accurate predictions with lower errors. The wavelet decomposition utilized by WANFIS dynamically adjusts to fluctuations
in the data, showcasing its remarkable robustness.
A notable downside of ANN is its black-box nature; while it can uncover hidden patterns, it often sacrifices interpret-
ability and demands large datasets for optimal performance. In comparison, WANFIS demonstrates impressive accuracy
even on smaller datasets. This is largely due to the wavelet decomposition that expertly extracts significant patterns,
minimizes noise, and enhances the overall learning process—resulting in a substantially lower Mean Absolute Percent-
age Error (MAPE) compared to both ANFIS and ANN.
Both ANN and ANFIS demand large datasets, resulting in intricate structures and a greater number of parameters. This
complexity significantly slows the training process. ANN, with its multiple hidden layers, can become highly computation-
ally intensive. Furthermore, ANFIS tends to generate an overwhelming number of fuzzy rules, which can compromise its
efficiency. Such challenges underscore the limitations of these models in practical applications. Unlike traditional ANFIS,
which navigates the complexities of raw, unprocessed data, WANFIS streamlines and filters information, paving the way
for faster convergence and minimizing computational overhead.

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

Fig. 12  Plots of training and

trained approximation part A3
and the detailed parts D1,D2,D3
of the time series of recovered
cases

The WANFIS model offers flexible adaptations for diverse time series datasets, enabling enhanced performance
through strategic parameter tuning, targeted feature extraction, and focused optimization.
WANFIS harnesses historical time-series data to forecast future trends, yet it recognizes that real-world events—
such as vaccination rollouts, lockdowns, economic policies, and shifts in mobility—significantly impact disease trans-
mission, economic conditions, and stock market fluctuations in ways that historical data alone cannot capture. The
waves of COVID-19 are not merely biological phenomena but are profoundly shaped by policy decisions and changes
in human behaviour. Traditional ANFIS/ANN models often assume stable relationships; however, WANFIS possesses
the capability to detect policy-induced trend shifts through wavelet decomposition and can adjust its fuzzy inference
system in real-time, ensuring that it remains responsive to our ever-evolving landscape. This adaptability positions
WANFIS as a superior tool for accurate and timely forecasts in uncertain times [85, 86].
The WANFIS model presents significant advantages for forecasting COVID-19 trends across different countries.
By enhancing forecasting accuracy, effectively managing noisy and inconsistent data, and adapting to the unique
conditions of each nation, it emerges as a powerful tool in this endeavor. As pandemic dynamics vary greatly from
one region to another, therefor the model’s success hinges on access to high-quality data and its ability to address
the distinct characteristics and challenges faced by each target country or region.

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

Fig. 13  Plots of training and

trained approximation part A3
and the detailed parts D1,D2
,D3 of the time series of death
cases

5 Conclusion

The objective of this study is to comprehensively understand the dynamics of COVID-19 transmission in India using
the Wavelet Adaptive Neuro-Fuzzy Inference System (WANFIS) model. The study aims to provide a clear picture of the
transmission dynamics of in the country and involves the development of a joint model of discrete wavelet decom-
position and a neuro-fuzzy ANFIS model to predict daily COVID-19 cases, including confirmations, recoveries, and
casualties. The study covers the period from May 8, 2020, to June 27, 2021, with 416 sample points, of which 386 are
used for training and the remaining 30 for testing the model.
The first step involves employing discrete wavelet decomposition to extract important information from the data
by separating it into low and high-frequency components (approximation and detail parts). These components are
then input into the ANFIS model, which is a combination of fuzzy logic and neural networks with back-propagation.
When the ANFIS model is applied to the wavelet decomposed data during the training period, it generates compo-
nent-wise FIS structures, which are then used to predict future components during testing. The summation of these
predicted components produces the final forecast of the input time series. Performance indicators such as Mean
Absolute Deviation (MAD) and Mean Absolute Percentage Error (MAPE) are used to compare observed and predicted
values obtained by WANFIS, ANFIS, and ANN models.

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

Fig. 14  Plots of the outcomes

generated by WANFIS model
for trained and tested values
of confirmed cases, recoveries,
and deaths

The study found that the MAPE values for the WANFIS model are 0.8562, 1.2613, and 0.4339 for confirmed, recovered,
and death cases, respectively. This indicates that the WANFIS model outperforms the ANFIS and ANN models in terms of
forecasting confirmed and death cases. The study concludes that the joint approach of wavelets and adaptive neuro-fuzzy

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

Table 3  A comparison of Model Time series Training Testing

training and testing results
4
of WANFIS, ANFIS and ANN MAD (× ­10 ) MAPE MAD (× ­105) MAPE
models
WANFIS Confirmed cases 1.7027 0.5438 2.5303 0.8562
Recovered cases 1.4217 0.6459 3.5594 1.2613
Death cases 0.0179 0.3670 0.0161 0.4339
ANFIS Confirmed cases 3.9681 1.1526 3.4038 1.1541
Recovered cases 3.2288 1.1179 5.5778 1.9797
Death cases 0.0362 0.5606 0.0620 1.6601
ANN Confirmed cases 7.1605 1.6581 9.5316 3.2068
Recovered cases 6.5243 2.2153 11.1073 3.8883
Death cases 0.0817 1.3256 0.1430 3.7303

inference system leads to lower forecasting MAD and MAPE error values, demonstrating its effectiveness in predicting
the input time series data. The results of the study can provide valuable insights into the COVID-19 situation in India.

Acknowledgements The authors are grateful to the VC, I K Gujral Punjab Technical University (Government of Punjab) and VC, Guru Nanak Dev
University (Government of Punjab), for providing a research facility. The authors are also thankful to the editor for their valuable suggestions
and comments in improving the manuscript.

Author contributions K.S.P and S.S. formulate the problem, did the analysis, computational work, and write the manuscript. J.K. reviewed the
manuscript.

Funding No funding is availed for this study.

Availability of data and materials Data is available online at WHO website.

Declarations
Ethics approval and consent to participate We further confirm that no aspect of the work covered in this manuscript involving human patients
has been conducted.

Consent for publication Authors give their consent for the publication of this manuscript.

Competing interests The authors declare no competing interests.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which
permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to
the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You
do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party
material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If
material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds
the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://​creat​iveco​
mmons.​org/​licen​ses/​by-​nc-​nd/4.​0/.

References
1. Shen M, Peng Z, Xiao Y. Modeling the epidemic trend of the 2019 novel coronavirus outbreak in China. bioRxiv. 2020.
2. Zhu N, Zhang D, Wang W, Li X, Yang B, Song J, et al. A novel coronavirus from patients with pneumonia in China, 2019. N Engl J Med. 2020.
https://​doi.​org/​10.​1056/​NEJMo​a2001​017.
3. Lai CC, Shih TP, Ko WC, Tang HJ, Hsueh PR. Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) and corona virus disease-2019
(COVID-19): the epidemic and the challenges. Int J Antimicrob Agents. 2020;55:105924.
4. Yang P, Wang X. COVID-19: a new challenge for human beings. Cell Mol Immunol. 2020. https://​doi.​org/​10.​1038/​s41423-​020-​0407-x.
5. Marzouk M, Elshaboury N, Abdel-Latif A, Azab S. Deep learning model for forecasting COVID-19 outbreak in Egypt. Process Saf Environ
Prot. 2021;153:363–75.
6. Britannica. The Editors of Encyclopaedia. “Coronavirus”. Encyclopedia Britannica. 2021. https://​www.​brita​nnica.​com/​scien​ce/​coron​avirus-​
virus-​group. Accessed 29 June 2021.

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

7. Wang LS, Wang YR, Ye DW, Liu QQ. A review of the 2019 Novel Coronavirus (COVID-19) based on current evidence. Int J Antimicrob Agents.
2020;55:105948.
8. World Health Organization (WHO). Coronavirus disease 2019 (COVID-19) Situation Report–70”. WHO. 2020.
9. Baldwin R, di Mauro BW. Mitigating the COVID economic crisis: act fast and do whatever it takes. London: CEPR Press; 2020.
10. Daniels JP. Venezuelan migrants “struggling to survive” amid COVID-19. The Lancet. 2020;395(10229):1023. https://d ​ oi.o
​ rg/1
​ 0.1
​ 016/S​ 0140-​
6736(20)​30718-2.
11. Ma Z. Spatiotemporal fluctuation scaling law and metapopulation modeling of the novel coronavirus (COVID-19) and SARS outbreaks.
arXiv. 2003.03714. 2020.
12. Mandal S, Bhatnagar T, Arinaminpathy N, Agarwal A, Chowdhury A, Murhekar M, Gangakhedkar RR, Sarkar S. Prudent public health
intervention strategies to control the coronavirus disease 2019 transmission in India: a mathematical model-based approach. Indian J
Med Res. 2020. https://​doi.​org/​10.​4103/​ijmr.​IJMR_​504_​20.
13. Huang NE, Qiao F. A data driven time-dependent transmission rate for tracking an epidemic: a case study of 2019-nCoV. Sci Bull.
2020;65:425–7.
14. Lohani AK, Goel NK, Bhatia KKS. Improving real time flood forecasting using Fuzzy Inference system. J Hydrol. 2014;509:25–41.
15. Jeddi S, Sharifian S. A hybrid wavelet decomposer and GMDH-ELM ensemble model for Network function virtualization workload fore-
casting in cloud computing. Appl Soft Comput. 2020;88: 105940.
16. Jain VK, Iyengar KP, Vaishya R. Differences between First wave and Second wave of COVID-19 in India. Diabetes Metab Syndr. 2021. https://​
doi.​org/​10.​1016/j.​dsx.​2021.​05.​009.
17. Hellewell J, Abbott S, Gimma A, Bosse NI, Jarvis CI, Russell TW, Munday JD, Kucharski AJ, Edmunds WJ, Sun F, Flasche S. Feasibility of
controlling COVID-19 outbreaks by isolation of cases and contacts. Lancet Glob Health. 2020;8:e488–96.
18. Chen TM, Rui J, Wang QP, Zhao ZY, Cui JA, Yin L. A mathematical model for simulating the phase-based transmissibility of a novel coro-
navirus. Infect Dis Poverty. 2020;9(1):1–8.
19. Prem K, Liu Y, Russell TW, Kucharski AJ, Eggo RM, Davies N. The effect of control strategies to reduce social mixing on outcomes of the
COVID-19 epidemic in Wuhan, China: a modelling study. Lancet Public Health. 2020;5(5):e261–70.
20. Choi SC, Ki M. Estimating the reproductive number and the outbreak size of Novel Coronavirus disease (COVID-19) using mathematical
model in Republic of Korea. Epidemiol Health. 2020;42: e2020011.
21. Yuan Z, Xiao Y, Dai Z, Huang J, Chen Y. A simple model to assess Wuhan lock-down effect and region efforts during COVID-19 epidemic
in China Mainland. Medrxiv. 2020-02. 2020.
22. Kucharski AJ, Russell TW, Diamond C, Liu Y, Edmunds J, Funk S, Eggo RM. Early dynamics of transmission and control of COVID-19: a
mathematical modelling study. Lancet Infect Dis. 2020;20(5):553–8.
23. Li Q, Guan X, Wu P, Wang X, Zhou L, Tong Y, Ren R, Leung KS, Lau EH, Wong JY, Xing X. Early transmission dynamics in Wuhan, China, of
novel coronavirus–infected pneumonia. N Engl J Med. 2020;382:1199–207.
24. Wu JT, Leung K, Leung GM. Nowcasting and forecasting the potential domestic and international spread of the 2019-nCoV outbreak
originating in Wuhan, China: a modelling study. Lancet. 2020;395:689–97.
25. Hamzah FB, Lau C, Nazri H, Ligot DV, Lee G, Tan CL, Chung MH. CoronaTracker: worldwide COVID-19 outbreak data analysis and predic-
tion. Bull World Health Organ. 2020;1(32):1–32.
26. Lin Q, Zhao S, Gao D, Lou Y, Yang S, Musa SS, Wang MH, Cai Y, Wang W, Yang L, He D. A conceptual model for the coronavirus disease 2019
(COVID-19) outbreak in Wuhan, China with individual reaction and governmental action. Int J Infect Dis. 2020;93:211–6.
27. Singh S, Parmar KS, Makkhan SJS, Kaur J, Peshoria S, Kumar J. Study of ARIMA and least square support vector machine (LS-SVM) models
for the prediction of SARS-CoV-2 confirmed cases in the most affected countries. Chaos Solitons Fractals. 2020;139: 110086.
28. Singh S, Parmar KS, Kumar J. Development of multi-forecasting model using Monte Carlo simulation coupled with wavelet denoising-
ARIMA model. Math Comput Simul. 2024;230:517–40.
29. Singh N, Sharma R, Kukker A. Wavelet transform based pneumonia classification of chest X-ray images. In: 2019 International Conference
on Computing, Power and Communication Technologies (GUCON). EEE; 2019. pp. 540–5.
30. Mahanty C, Kumar R, Asteris PG, Gandomi AH. COVID-19 patient detection based on fusion of transfer learning and fuzzy ensemble
models using CXR images. Appl Sci. 2021;11(23):11423.
31. Kukker A, Sharma R. Modified fuzzy Q learning based classifier for Pneumonia and tuberculosis. Irbm. 2021;42(5):369–77.
32. Mahanty C, Kumar R, Mishra BK, Barna C. COVID-19 detection with X-ray images by using transfer learning. J Intell Fuzzy Syst.
2022;43(2):1717–26.
33. Mahanty C, Kumar R, Patro SGK. Internet of medical things-based COVID-19 detection in CT images fused with fuzzy ensemble and
transfer learning models. N Gener Comput. 2022;40(4):1125–41.
34. Mahanty C, Kumar R, Mishra BK, Hemanth DJ, Gupta D, Khanna A. Prediction of COVID-19 active cases using exponential and non-linear
growth models. Expert Syst. 2022;39(3): e12648.
35. Kukker A, Sharma R. JAYA-optimized fuzzy reinforcement learning classifier for COVID-19. IETE J Res. 2023;69(10):7314–25.
36. Mahanty C, Patro SGK, Rathor S, Rachapudi V, Muzammil K, Islam S, Khan WA. Forecasting of coronavirus active cases by utilizing logistic
growth model and fuzzy time series techniques. Sci Rep. 2024;14(1):18039.
37. Kukker A, Sharma R, Pandey G, Faseehuddin M. Fuzzy lattices assisted EJAYA Q-learning for automated pulmonary diseases classification.
Biomed Phys Eng Express. 2024;10(6): 065001.
38. Zhao S, Musa SS, Lin Q, Ran J, Yang G, Wang W, et al. Estimating the unreported number of novel coronavirus (2019-nCoV) cases in China
in the first half of January 2020: a data-driven Modelling analysis of the early outbreak. J Clin Med. 2020. https://​doi.​org/​10.​3390/​jcm90​
20388.
39. Al-Qaness MA, Saba AI, Elsheikh AH, Abd Elaziz M, Ibrahim RA, Lu S, Ewees AA. Efficient artificial intelligence forecasting models for
COVID-19 outbreak in Russia and Brazil. Process Saf Environ Protect. 2021;149:399–409.
40. Yan S, Zheng D. A deep neural network for predicting synergistic drug combinations on cancer. Interdiscip Sci Comput Life Sci. 2024.
https://​doi.​org/​10.​1007/​s12539-​023-​00596-6.
41. Davidson R, Labys WC, Lesourd JB. Wavelet analysis of commodity price behaviour. Comput Econ. 1998;11(1–2):103–28.

Vol.:(0123456789)
Research
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y

42. Soni K, Parmar KS, Agrawal S. Modeling of air pollution in residential and industrial sites by integrating statistical and Daubechies
wavelet (Level 5) analysis. Model Earth Syst Environ. 2017;3:1187–98.
43. Rustam F, Reshi AA, Mehmood A, Ullah S, On B-W, Aslam W, Choi GS. COVID-19 future forecasting using supervised machine learning
models. IEEE Access. 2020;8:101489–99.
44. Parmar KS, Soni K, Singh S. Prediction of river water quality parameters using soft computing techniques. In: Deo RC, Samui P, Kisi O,
Yaseen ZM, editors. Intelligent data analytics for decision-support systems in hazard mitigation. Singapore: Springer; 2021. p. 429–40.
45. Valenzuela O, Rojas I, Rojas F, Pomares H, Herrera LJ, Guillen A, et al. Hybridization of intelligent techniques and ARIMA models for
time series prediction. Fuzzy Sets Syst. 2008;159(7):821–45.
46. Box GEP, Jenkins GM. Time series analysis, forecasting and control. San Francisco: Holden Day; 1976.
47. Kantz H, Schreider T. Nonlinear time series analysis. Cambridge: Cambridge University Press; 1997.
48. Hsu K, Gupta HV, Sorooshian S. Artificial neural network modeling of the rainfall runoff process. Water Resour Res. 1995;31:2517–30.
49. Hung NQ, Babel HS, Weesakul S, Tripathi NK. An artificial neural network model for rainfall forecasting in Bangkok Thailand. Hydrol
Earth Syst Sci. 2009;13:1413–25.
50. Yousefi S, Weinreich I, Reinarz D. Wavelet-based prediction of oil prices. Chaos Solitons Fractals. 2005;25(2):265–75.
51. Rasheed J, Jamil A, Hameed AA, Al-Turjman F, Rasheed A. COVID-19 in the age of artificial intelligence: a comprehensive review.
Interdiscip Sci Comput Life Sci. 2021;13:153–75.
52. Wu P, Hao X, Lau EHY, Wong JY, Leung KSM, Wu JT, et al. Real time tentative assessment of the epidemiological characteristics of novel
coronavirus infections in Wuhan, China, as at January 22 2020. Euro Surveill. 2020;25(3):2000044.
53. Zhao S, Lin Q, Ran J, Musa SS, Yang G, Wang W, et al. Preliminary estimation of the basic reproduction number of novel coronavirus
(2019-nCoV) in China, from 2019 to 2020: a data-driven analysis in the early phase of the outbreak. Int J Infect Dis. 2020. https://​doi.​
org/​10.​1016/j.​ijid.​2020.​01.​050.
54. Freire PKDMM, Santos CAG, da Silva GBL. Analysis of the use of discrete wavelet transforms coupled with ANN for short-term stream-
flow forecasting. Appl Soft Comput. 2019;80:494–505.
55. Basheer IA, Hajmeer M. Artificial neural networks: fundamentals, computing, design, and application. J Microbiol Methods.
2000;43(1):3–31.
56. Khashei M, Bijari M. An artificial neural network (p, d, q) model for time series forecasting. Expert Syst Appl. 2010;37(1):479–89.
57. Zhang GP, Qi M. Neural network forecasting for seasonal and trend time series. Eur J Oper Res. 2005;160(2):501–14.
58. Singh S, Parmar KS, Kumar J, Makkhan SJS. Development of new hybrid model of discrete wavelet decomposition and autoregressive
integrated moving average (ARIMA) models in application to one month forecast the casualties cases of COVID-19. Chaos Solitons Fractals.
2020;135: 109866.
59. Mallat S. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans Pattern Anal Mach Intell.
1989;11(7):674–93.
60. Daubechies I. Ten lectures on wavelets. Philadelphia: SIAM; 1992.
61. Meyer Y, Coifman R. Wavelets. Cambridge: Cambridge University Press; 1997.
62. Percival DB, Walden AT. Wavelet methods for time series analysis. Cambridge: Cambridge University Press; 2000.
63. Parmar KS, Bhardwaj R. Wavelet and statistical analysis of river water quality parameters. Appl Math Comput. 2013;219(20):10172–82.
64. Pang Z, Niu F, O’Neill Z. Solar radiation prediction using recurrent neural network and artificial neural network: a case study with com-
parisons. Renew Energy. 2020;156:279–89.
65. Pal R, Sekh AA, Kar S, Prasad DK. Neural network-based country wise risk prediction of COVID-19. Appl Sci. 2020;10(18):6448.
66. Singh S, Parmar KS, Kumar J. Soft computing model coupled with statistical models to estimate future of stock market. Neural Comput
Appl. 2021. https://​doi.​org/​10.​1007/​s00521-​020-​05506-1.
67. Chen AS, Leung MT, Daouk H. Application of neural networks to an emerging financial market: forecasting and trading the Taiwan Stock
Index. Comput Oper Res. 2003;30(6):901–23.
68. Singh S, Parmar KS, Kaur J, Kumar J, Makkhan SJS. Prediction of COVID-19 pervasiveness in six major affected states of India and two-stage
variation with temperature. Air Qual Atmos Health. 2021. https://​doi.​org/​10.​1007/​s11869-​021-​01075-x.
69. Wu S, Wang X, Su J. Statistical analysis of the community lockdown for COVID-19 pandemic. Appl Intell. 2022;52:3465–82. https://​doi.​
org/​10.​1007/​s10489-​021-​02615-9.
70. Melard G, Pasteels JM. Automatic ARIMA modeling including interventions, using time series expert software. Int J Forecast.
2000;16(4):497–508.
71. Parmar KS, Bhardwaj R. Water quality management using statistical and time seriesprediction model. Appl Water Sci. 2014;4(4):425–34.
72. Babu CN, Reddy BE. A moving-average filter-based hybrid ARIMA–ANN model for forecasting time series data. Appl Soft Comput.
2014;23:27–38.
73. Mary SR, Kumar V, Venkatesan KP, Kumar RS, Jagini NP, Srinivas A. Vulture-based AdaBoost-feedforward neural frame work for COVID-19
prediction and severity analysis system. Interdiscip Sci Comput Life Sci. 2022;14(2):582–95.
74. Cumbane SP, Gidófalvi G. Deep learning-based approach for COVID-19 spread prediction. Int J Data Sci Anal. 2024. https://​doi.​org/​10.​
1007/​s41060-​024-​00558-1.
75. Ravenda F, Cesarini M, Peluso S, et al. A probabilistic spatio-temporal neural network to forecast COVID-19 counts. Int J Data Sci Anal.
2024. https://​doi.​org/​10.​1007/​s41060-​024-​00525-w.
76. Hamilton MA, Hamilton D, Soneye O, et al. An analysis of the impact of policies and political affiliation on racial disparities in COVID-19
infections and deaths in the USA. Int J Data Sci Anal. 2022;13:63–76. https://​doi.​org/​10.​1007/​s41060-​021-​00284-y.
77. Jang JSR. ANFIS: adaptive network based fuzzy inference system. IEEE Trans Syst Manag Cybernet. 1993;23:665–85.
78. Chakraborty T, Ghosh I. Real-time forecasts and risk assessment of novel coronavirus (COVID-19) cases: a data-driven analysis. Chaos
Solitons Fractals. 2020;135(109850):1–10.
79. Daubechies I. Orthonormal bases of compactly supported wavelets. Commun Pure Appl Math. 1988;41:909–96.
80. Nayak J, Naik B, Dinesh P, et al. Intelligent system for COVID-19 prognosis: a state-of-the-art survey. Appl Intell. 2021;51:2908–38. https://​
doi.​org/​10.​1007/​s10489-​020-​02102-7.

Vol:.(1234567890)
Discover Public Health (2025) 22:270 | https://doi.org/10.1186/s12982-025-00670-y
Research

81. Chadaga K, Chakraborty C, Prabhu S, Umakanth S, Bhat V, Sampathila N. Clinical and laboratory approach to diagnose COVID-19 using
machine learning. Interdiscip Sci Comput Life Sci. 2022;14(2):452–70.
82. Parmar KS, Bhardwaj R. Statistical, time series and fractal analysis of full stretch of river Yamuna (India) for water quality management.
Environ Sci Pollut Res. 2015;22(1):397–414.
83. Parmar KS, Bhardwaj R. River water prediction modelling using neural networks, fuzzy and wavelet coupled model. Water Resour Manag.
2015;29(1):17–33.
84. Strömbergsson D, Marklund P, Berglund K, Saari J, Thomson A. Mother wavelet selection in the discrete wavelet transform for condition
monitoring of wind turbine drivetrain bearings. Wind Energy. 2019;22(11):1581–92.
85. Cheng C, Jiang WM, Fan B, Cheng YC, Hsu YT, Wu HY, Tsou HH. Real-time forecasting of COVID-19 spread according to protective behavior
and vaccination: autoregressive integrated moving average models. BMC Public Health. 2023;23(1):1500.
86. Altarawneh L, Wang H, Jin Y. COVID-19 vaccine prediction based on an interpretable CNN-LSTM model with three-stage feature engineer-
ing. Health Technol. 2024. https://​doi.​org/​10.​1007/​s12553-​023-​00797-5.

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Vol.:(0123456789)