Utilizing Transfer Learning Approach for Epidemiological Study of Region-Specific Post-Acute Sequelae of SARS-COV-2

Summary

The proposed work aims to design and implement a novel transfer learning technique to enable better understanding of long-term health outcomes and support tailored epidemiological insights.

Abstract

Post-Acute Sequelae of SARS-CoV-2 (PASC) is posing an extreme risk environment with severe consequences, especially for middle-aged and older adults and those with chronic health conditions such as cardiovascular diseases, cancers, respiratory illnesses, and diabetes. Physical disorders include severe and persistent body pains, fatigue, and difficulty with body movements. Similarly, mood disorders such as depression, mood swings, feelings of hopelessness, and difficulty concentrating are significant predictors of PASC. Individuals, predominantly middle-aged and elderly people, are facing increased physical and mental consequences, including frequent hospitalizations, medically unstable conditions, and in some cases, unexpected fatalities. Timely identification of PASC is essential to mitigating the severity of associated health issues. This research recommends a latent transfer model to integrate patient data from different regions, to extract insights into the data, and to deliver personalized healthcare solutions. This study features the potential of the latent transfer learning model to improve data simplification, generalization, personalization, insight detection, variability, scalability, and adaptability to improve public health outcomes.

Introduction

The Coronavirus Disease of 2019 (COVID-19) was first identified in Wuhan, China. On 31^st December 2019, China reported to the World Health Organization (WHO) about lung inflammation diseases with unknown etiology noticed in the city of Wuhan, province of Hubei, China¹. From 31^st December 2019 to 3^rd January 2020, 44 people were infected by this virus and had a history of contact with the wholesale market- "Huanan Seafood". Initially, patients were witnessed with fever, fatigue, and dry cough. The outbreak of this virus occurred on 30^th January 2020, and it was announced as a Public Health Emergency of International Concern by the WHO¹. By mid-February 2020, more than 25 countries were affected by the novel coronavirus, with 70,635 and 794 cases in China and other countries, respectively. One thousand seven hundred seventy-two deaths (1772) in China and three deaths in all countries were officially declared. On 31^st December 2019, a bunch of pneumonia infection cases were reported, and the WHO reported there were no deaths reported in Wuhan. In Thailand, the first case was announced in Bangkok on 13^th January 2020 for a 61-year-old Chinese female who returned from Wuhan City on 8^th January 2020 after visiting Wuhan City. She was symptomatic with a sore throat, cough, fever, chills, and headache. The person was tested COVID-19 positive in thermal surveillance on 8^th January 2020 and admitted to the hospital that day by Thai officers. The person recovered from the disease. The genetic sequencing structure of the virus was shared by the Chinese government and enabled more countries to efficiently examine and diagnose patients².

In the United States of America, the first case was reported in Washington on 19^th January 2020 for a 35-year-old male who returned from Wuhan City on 15^th January 2020 after his family gathering. He was symptomatic with cough and fever from 15^th January 2020 and was advised to undergo medical tests. He did not have contact with a COVID-19-infected person, but during his visit to Wuhan, he was infected with COVID-19 due to the outbreak in China. He tested COVID positive on 20^th January 2020 and was quarantined in an airborne isolation unit of a medical center. After his admission into the isolation ward, he was normal with a 110/min pulse rate, 134/87 mmHg blood pressure, 37.2 °C body temperature, 16 bpm respiratory rate, and 96% oxygen saturation with normal lung auscultation. Regular COVID monitoring and investigation tests were undertaken, and he was noted with a cough, nausea, and vomiting for two days³. On the second day of admission, he complained about abdominal discomfort, and on the next day, there was no evidence of abnormalities in the chest radiograph. However, on the fifth day, significant evidence of pneumonia was seen in the chest radiograph, and oxygen saturation dropped to 90%. He was advised to undergo oxygen therapy, which was supplemented to improve saturation levels. On the 8^th day, her condition improved, and she was discharged from the clinic. In India, the first case was reported in Kerala on 27^th January 2020 for a 20-year-old female who had returned from Wuhan City due to the COVID-19 outbreak in China. She was asymptomatic from 23^rd January 2020 to 26^th January 2020 and witnessed a sore throat and mild cough from 27^th January 2020. She did not have contact with the COVID-19-infected persons, but during her train journey from Wuhan to Kunming, she noticed several people with respiratory problems, colds, and coughs. Hence, Kerala government authorities instructed her to stay in an isolated place and visit a medical facility for any developed COVID symptoms in the coming days. After her admission into the general hospital, she was normal with an 82/min pulse rate, 130/80 mmHg blood pressure, 98.5 °F body temperature, and 96% oxygen saturation with normal lung auscultation. Regular COVID monitoring and investigation tests were undertaken on the 3^rd, 7^th, and 20^th day from the day of illness. Oropharyngeal swab tests were conducted on the first day and every alternate day for 23 days from the day of illness, and tested negative from the 17^th day. She was cleared with all the tests and was discharged from the hospital on 20^th February 2020⁴.

In the United Kingdom, the first case of COVID-19 was reported on 23^rd January 2020 by a 50-year-old female who had returned from Wuhan city after her visit. Initially, she was asymptomatic and developed symptoms of fever, weakness, sore throat, and dry cough on 26^th January 2020. After her admission into the general hospital, she was normal with an 18/min respiratory rate, 120/80 mmHg blood pressure, 37.6 °C body temperature, and 97% oxygen saturation. Regular COVID monitoring and investigation tests were undertaken, and the patient's condition is clinically stable. Initial examinations reported her with mild lymphopenia and a rise in C-reactive protein. After the 3^rd day of admission into the hospital, she was cleared of all symptoms, and Oropharyngeal swab tests were revealed as COVID-negative from the second day in the hospital. She was isolated for 2 weeks, and her nose and throat swab tests were negative⁵.

Coronaviruses (CoVs) were discovered in the 1960s, and they were classified as follows. Orthornavirae is termed the kingdom/realm of various viruses that contain genomes made of ribonucleic acid (RNA) and that translate with an RNA-dependent RNA polymerase (RdRp). RdRp transcribes the viral RNA genome into messenger RNA (mRNA) and replicates the Genome. The characteristics of the viruses that are in Orthornavirae include evolution, genetic mutations, recombination, and reassortment⁶. Genomes that are made of RNA are of three types, namely, positive-strand RNA viruses, negative-strand RNA viruses (-ssRNA viruses), and double-stranded RNA viruses (dsRNA viruses). Positive-strand RNA viruses (+ssRNA viruses) are a collection of associated viruses that have positive-sense, single-stranded RNA. The Genome of the +ssRNA virus can be used as courier RNA (mRNA) and is directly transformed into virus-related proteins by the ribosomes of the host cells. +ssRNA viruses are associated with the phyla such as Pisuviricota, Kitrinoviricota, and Lenarviricota. Positive-strand RNA viruses contain various pathogens such as Hepacivirus C, Dengue, Ebola, Measles, Rabies, Influenza, Severe acute respiratory syndrome CoV, Middle East respiratory syndrome coronavirus CoV, and Severe acute respiratory syndrome CoV⁷. Negative-strand RNA viruses (-ssRNA viruses) are a collection of associated viruses that have negative-sense, single-stranded RNA. The Genome of the -ssRNA virus can be seen as courier RNA (mRNA) and produced by an enzyme named RNA-dependent RNA polymerase (RdRp). -ssRNA viruses constitute the phylum Negarnaviricota. Negative-strand RNA viruses contain two major branches in the phylum, namely Haploviricotina and Polyploviricotina. Important vertebrate-ssRNA viruses are Ebola, Hanta, Influenza, Lassa fever, and Rabies. Double-stranded RNA viruses (dsRNA viruses) are a collection of polyphyletic viruses that have double-stranded RNA genomes. The Genome of dsRNA transcribes a +ssRNA by the viral RNA-dependent RNA polymerase (RdRp). The Genome of the +ssRNA virus can be used as courier RNA (mRNA) and is openly transformed into virus-related proteins by the ribosomes of the host cells. Whereas the Genome of -ssRNA virus can serve as courier RNA (mRNA) and be produced by an enzyme RNA-dependent RNA polymerase (RdRp). dsRNA viruses are associated with the phyla Duplornaviricota and Pisuviricota. Double-stranded RNA viruses include rotaviruses (which affect young children) and bluetongue virus (which affects cattle and sheep)⁸.

Pisuviricota is termed a phylum of RNA viruses, which includes all + ssRNA viruses and dsRNA. Its lower classification contains three classes, namely, Duplopiviricetes, Pisoniviricetes, and Stelpaviricetes. Pisoniviricetes is a class of +ssRNA viruses that infect and contaminate eukaryotes. Its lower classification contains three sub-classes, namely Nidovirales, Picornavirales, and Sobelivirales. Nidovirales is an order of enveloped, +ssRNA that infects and contaminates vertebrates and invertebrates⁹. This contains a suborder of viruses, namely, Abnidovirineae, Cornidovirineae, Nanidovirinae, Arnidovirineae, Mesnidovirineae, and Tornidovirineae. Various families are recognized under Nidovirales, such as Abyssoviridae, Coronaviridae, Nanghoshaviridae, Nanhypoviridae, Arteriviridae, Medioniviridae, Cremegaviridae, Mesoniviridae, Euroniviridae, Gresnaviridae, Roniviridae, Tobaniviridae, Mononiviridae, and Olifoviridae⁸. Coronaviridae is a family of enveloped, +ssRNA viruses that infects mammals, amphibians, and birds. Its lower classification contains two subfamilies, Letovirinae and Orthocoronavirinae. The members of Orthocoronavirinae are identified as coronaviruses. Coronaviruses are a subfamily of Coronaviridae, which is related to RNA viruses that cause infections in mammals and birds. In mammals and birds, these viruses cause mild to severe respiratory swab infections, such as upper respiratory swab infections and lower respiratory swab infections¹⁰. These viruses cause SARS, MERS, and COVID-19 in humans and some mammals, diarrhea in pigs and cows, hepatitis, and encephalomyelitis. Coronaviruses are enveloped +ssRNA that belong to the Kingdom of Orthornavirae, Phylum of Pisuviricota, Class of Pisoniviricetes, Subclass of Nidovirales, Family Coronaviridae, and Subfamily of Orthocoronavirinae. Coronaviruses are categorized into four genera, namely Alpha coronavirus, Beta coronavirus, Gamma coronavirus, and Delta coronavirus¹¹. Based on this research, the following hypotheses are formed.

Severe Acute Respiratory Syndrome Coronavirus-2 (SARS-CoV-2) spike mutation is extremely glycosylated. It is critically significant to examine the biological consequences of these spike mutations. More than one mutation can originate at the same position in the spike protein sequence. The spike protein sequences provided a chance to estimate the wide spectrum of mutations detected across the globe, and they are represented in Table 1. The mutations considered were a consequence of human-to-human transmission of the virus. Vaccination prompted immune reactions capable of powerfully neutralizing SARS-CoV-2¹². Though scientific investigations discovered the advent of SARS-CoV-2 variants concealing mutations in the spike, the core objective remained the neutralization of host antibodies. There was an evolving indication of reduced neutralization of some SARS-CoV-2 variants following vaccination. Vaccine manufacturers formulated various platforms for a potential update of vaccine structures, and surveys of genetic and antigenic variations in the global virus population were conducted with experiments to illuminate the phenotypic influences of mutations¹³.

The National Center for Biotechnology Information (NCBI) virus database contains 10333 human SARS-CoV-2 spike proteins from Africa, Asia, Europe, North America, South America, and Oceania. The dissemination of several mutations in spike proteins examined from different geographic locations is shown in the above table. Details of the Alpha, Beta, Gamma, and Delta coronavirus with their genera, features, and descriptions are provided in the above table. Each Genera is described by its features, such as fundamental, variants, significance, shape, measurement, structure, membrane fusion, release, contamination, and synthesis¹⁴.

As mentioned in Table 1, virus strains such as HCoV-229E, HCoV-OC43, HCoV-NL63, and HCoV-HKU1 observed mild to moderate symptoms of illness, whereas SARS-CoV, MERS-CoV, and SARS-CoV-2 observed severe critical symptoms of illness in humans. Virus microorganisms pass through the body and infect the host cell. Depending on the category of coronavirus, it infects cells in different portions of the host body. As soon as it has attached itself to the strong and healthy cell, it goes into it¹⁵. After the virus is settled inside the host cell, it will unclutter, and its DNA and RNA will emanate out and travel towards the nucleus. They will move in a fragment, which will result in various reproductions of the virus. These copies will come out of the nucleus to be assembled and receive protein, which protects their DNA and RNA. These new-fangled reproductions of the virus will vacate the previously infected and diseased cells to contaminate other healthy host cells again and multiply. If the host's immune system succeeds in fighting off the coronavirus, the individual will not become sick; otherwise, unhealthy complications may occur. If the immune system fails to control the coronavirus, it results in pathogenesis. With this, the virus enters the body and infects different organs. Depending on how severe the signs or symptoms are, the individual will take to rest or pursue medical assistance¹⁶. The life cycle and various strains of coronavirus are depicted in Figure 1.

The above figure depicts the life cycle and stages of coronavirus in a detailed manner. There are seven stages, namely (1) attachment and admission, (2) fusion and endocytosis, (3) proteolysis, (4) translation and cleavage, (5) replication, (6) exocytosis, and (7) matured virus release. In the attachment and admission phase, Coronavirus microorganisms are admitted into the host, and the aim is for host-resistant reconnaissance and intervention strategies. To enter host cells, these viruses first attach to a cell's surface receptor for binding, then enter endosomes, and ultimately begin fusion. Later, the coronavirus fuses with the virus-related and lysosomal sheaths into the host cell. A virus surface-fastened spike protein interferes with these viruses' entry. Endocytosis is a cellular process in which a substance is carried into the host cell¹⁷. The material to be adopted is enclosed by a range of cell sheaths, which then shoot up inside the host cell to generate a vesicle encompassing the ingested substance. Endocytosis comprises pinocytosis and phagocytosis; the former is cell drinking, and the latter is cell eating. Proteolysis is the itemization of virus-related proteins into minor polypeptides/amino acids. Structured mRNA conversion is a post-transcriptional process that reins in gene communication and promptly diverges protein profusion. It has a main part in copious biological procedures such as cell growth, cell progress, synaptic plasticity, stress retorts, and productive viral development¹⁸. The viral translation procedure can be sectioned into three phases, namely, instigation, elongation, and dissolution.

Viral replication is the process of generating genetic viruses through the contamination of infected host cells. For viral reproduction to occur, viruses must first enter the host cell. By repeatedly replicating their viral genome and encapsulating these copies, viruses survive and continue spreading by infecting new host cells¹⁹. Exocytosis is the mechanism through which vesicles containing viral particles are secreted and released from the infected cell, enabling the newly formed viruses to enter the host body and infect additional healthy cells. Post-acute sequelae of SARS-CoV-2 (PASC) is recognized as a syndrome that primarily affects the lungs, although it can impact multiple organs, resulting in organ damage or even organ failure²⁰. Such complications may lead to death or long-term health problems. For example, the heart may develop an increased risk of heart failure or other complications; the lungs may sustain long-term breathing difficulties following pneumonia; the brain may be affected by strokes, seizures, or Guillain-Barré syndrome; and the kidneys may suffer damage that can result in septic shock. Liver injury can occur due to enzyme imbalance, while rhabdomyolysis leads to the breakdown of muscle tissue, releasing myoglobin into the bloodstream, which can become fatal if the kidneys cannot filter it quickly. Additional complications include the formation of abnormal blood clots that cause internal bleeding and organ failure, secondary infections from bacteria such as strep or staph, disseminated intravascular coagulation due to dysfunctional clotting responses, and chronic fatigue, often accompanied by brain fog, severe tiredness, dizziness, or impaired thinking. Based on these challenges, the following hypotheses are proposed: H1: Utilizing a transfer learning algorithm can considerably enhance the prediction accuracy and transferability of models by applying epidemiological data to identify region-specific PASC. H1a: Deep learning models that are fine-tuned using epidemiological data perform better than other models. H1b: Combining region-specific PASC determinants with transfer learning models improves the accuracy and efficiency of classification and clustering. H1c: This combination can also reveal previously undisclosed or overlooked region-specific patterns. H1d: Transfer learning enables reduced training time and faster convergence compared to other models.

Protocol

This study was conducted in accordance with ethical guidelines for research involving human data. All patient data used in this study were fully anonymized prior to access and analysis. No personally identifiable information was used. The research protocol was reviewed and approved by the Institutional Review Board (IRB), in compliance with the Declaration of Helsinki. Where applicable, data access permissions were obtained from relevant authorities or data repositories. As this study involves secondary analysis of de-identified data, informed consent was waived by the IRB. The proposed latent transfer model is a machine learning based semi-supervised learning technique designed to find the hidden patterns and transfer them from the source dataset to the target dataset, ensuring it is highly appropriate for analyzing epidemiological problems of PASC¹⁶. Post-COVID conditions often manifest with diverse symptoms across different populations, leading to inconsistent and sparse data. Proposed work addresses the problem of sparse data due to diverse symptoms across different regions by learning improved latent representations to represent hidden health patterns from the datasets, and by representing them as emerging datasets. To perform this conversion, these latent transfer techniques use auto or variational encoders to represent both source and target data in a common latent space. And this common space safeguards the obtained hidden patterns as clusters of post-COVID symptoms are well-preserved across diverse regions²¹. The prototype can further advance through improved generalization with domain adaptation. Using this approach, the model can help recognize and predict at-risk clusters and support the surveillance of public health in remote areas. This work also enhances learning by reducing the dependence on completely labeled data. With this, the proposed approach becomes a powerful tool for analyzing the gaps due to its generalized approach towards various populations.

1. Data sources and collection

This study used publicly available Kaggle datasets (Cardiovascular Disease Dataset, Cancer Prediction Dataset, Early-Stage Diabetics Prediction Dataset, and Mental Health Prediction Dataset). Therefore, IRB approval in this submission is not required.

For datasets involving image data, preprocessing was conducted using Python with TensorFlow and OpenCV libraries. Preprocessing steps included image resizing (e.g., 224 × 224 pixels), grayscale or RGB conversion, normalization of pixel values to [0, 1] range, and data augmentation techniques (rotation, flipping, zooming) to enhance model generalization and reduce overfitting. The image arrays were then converted into NumPy arrays for further processing and model input.

For feature selection, Recursive Feature Elimination (RFE) was implemented using Scikit-learn's RFE class, typically in conjunction with a Logistic Regression or Random Forest estimator. The number of features selected varied per dataset, but a common configuration was to reduce dimensionality to the top 10-15 most influential features based on recursive scoring. This method improved classification accuracy while reducing computational complexity and model overfitting.

This study utilized four publicly available datasets from Kaggle. The Cardiovascular Disease Dataset, Breast Cancer Prediction Dataset, Early-Stage Diabetes Prediction Dataset, and the Mental Health Prediction Dataset. These datasets were downloaded directly from Kaggle and consist of structured tabular data containing clinical and survey-based attributes. Data preprocessing was conducted using Python with key libraries such as Pandas for data manipulation, NumPy for numerical operations, and Scikit-learn for data preprocessing, feature selection, and modeling. Missing values were imputed using mean or mode strategies, categorical variables were encoded via One-Hot or Label Encoding, and numerical features were scaled using either StandardScaler or MinMaxScaler. For imbalanced classes, SMOTE (Synthetic Minority Oversampling Technique) from the imblearn package was applied. Recursive Feature Elimination (RFE) was implemented using Scikit-learn's RFE module with estimators such as Logistic Regression and Random Forest, and the number of features was typically reduced to the top 10-15 most influential ones based on model performance metrics such as ROC-AUC and accuracy.

2. Model architecture

This research proposes a variant of the neural network algorithm called the latent transfer model to train input images from the COVID-19 image database. This work is effective in overcoming the vanishing gradient problem when building a deep neural network. Although various neural network variants are available with different layer configurations, this work primarily focuses on implementing residual learning techniques during the training process of deep neural networks. The proposed latent transfer model uses an improved algorithm of ResNet-50 with a 7 x 7 convolutional layer framework, along with flattened and dense layered networks. These layers are utilized in image transformation and multi-class classification of PASC patients, region-wise, to extract patterns²². This work considers input images of size 256 x 256 x 3, utilizing efficient hyperparameters, learning behavior, optimization procedures, and batch volumes. The proposed latent transfer model considers a dataset of 12,946 images to classify patients at active risk of developing cardiovascular issues, cancer, chronic respiratory diseases, diabetes, and mental health problems. Dataset details and hyperparameters are given in Table 2 and Table 3, respectively.

To ensure reproducibility, all model implementation, preprocessing, and evaluation steps were conducted in a controlled environment using Python within Jupyter Notebook on a system running Windows 11, equipped with an Intel Core i7 (11th Gen) processor, 16GB RAM, and NVIDIA GeForce RTX 3060 GPU (for any accelerated processing, although the GPU was not essential for these datasets). Software libraries used include Pandas 1.5.3, NumPy 1.24, Scikit-learn 1.3.0, Imbalanced-learn 0.11.0 for SMOTE, Matplotlib 3.7.1, and Seaborn 0.12.2 for visualization. Preprocessing involved handling missing values via mean/mode imputation, encoding categorical variables using One-Hot or Label Encoding, and feature scaling using StandardScaler. Model development utilized classifiers such as Logistic Regression, Random Forest, and Support Vector Machines, with Recursive Feature Elimination (RFE) for dimensionality reduction. Models were evaluated using 5-fold cross-validation, and performance was measured using metrics including accuracy, precision, recall, and ROC-AUC score, ensuring a consistent and reproducible pipeline.

3. Preprocessing and feature engineering

To retrieve statistically significant features from these datasets, a preprocessing technique named "Feature Selection using Recursive Feature Elimination (RFE)" and a modelling pipeline named "Transfer Learning with Multi-Layer Perceptron (MLP)" are used.

4. Training and validation

An organized training and validation approach was used to assess the proposed approach for PASC symptom analysis and prediction. At first, a hierarchical train-test split was applied to get the cumulative system of class labels with an 80:20 ratio of the training and testing data. Followed by k-fold cross-validation with k = 5, was implemented on the training data to evaluate the robustness of the model across various splits of the data. An MLP classifier with two hidden layers (64 and 32 neurons) and 300 iterations was applied. Further feature representation and transfer learning techniques were implemented.

5. Evaluation metrics

Accuracy, F-measure, G-mean, MAPE, Sensitivity, and Specificity were calculated on the test data to compare the proposed model's performance. Outcomes illustrate the improvement of evaluation metrics, Accuracy, F-measure, G-mean, MAPE, Sensitivity, and Specificity on the test data to compare the proposed model's performance with other state-of-the-art models, namely VGG-16 CNN, VGG-19 CNN, DenseNet-121 CNN, InceptionV3 CNN, ResNet-101 CNN, MobileNetV2 CNN, and the proposed Latent Transfer model. The above dataset contains training, validation, and testing samples of cardiovascular, cancer, chronic respiratory, diabetes, and mental illness-related data.

All experiments were conducted in a Windows 11 environment using Python 3.10 on a system with an Intel Core i7 processor, 16GB RAM, and an NVIDIA RTX 3060 GPU. Publicly available Kaggle datasets were preprocessed using Pandas, NumPy, and Scikit-learn for missing value imputation, encoding, scaling, and feature selection via Recursive Feature Elimination (RFE). Models such as Logistic Regression, Random Forest, and SVM were trained and evaluated using 5-fold cross-validation with metrics including accuracy, precision, recall, and ROC-AUC to ensure reproducibility and performance consistency.

The above table shows the details of hyperparameters and the technique or algorithm that uses them. Learning rate, early stopping technique, batch size, epoch size, callbacks, loss function, and metrics to measure the efficiency are given for the Adam optimizer algorithm.

Results

The proposed model is compared with existing transfer models, including VGG-16 CNN, VGG-19 CNN, DenseNet-121 CNN, InceptionV3 CNN, ResNet-101 CNN, and MobileNetV2 CNN. The parameter details and outcomes of the comparison are shown in Table 4, Table 5, and Figure 2, respectively. Results illustrate the improvement of evaluation metrics by using the proposed model over the other models, namely VGG-16 CNN, VGG-19 CNN, DenseNet-121 CNN, InceptionV3 CNN, ResNet-1...

Discussion

The proposal centers on building a latent transfer learning model to handle the impacts of the PASC. This structure leads to severe physical and mental health problems, particularly in the middle-aged and elderly, and those subject to previous chronic diseases, such as cardiovascular diseases, cancer, respiratory diseases, and diabetes. The eventual goal is to unify diverse patient data from different populations with the help of a latent transfer model to allow personalized care, early detection, and e...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Cloud Computing Environment (e.g., Google Colab / AWS EC2)	Google / Amazon	N/A	Used for training deep learning models with GPU acceleration
Epidemiological Dataset (Post-Acute Sequelae of SARS-CoV-2, region-specific)	Public health data repositories (e.g., WHO, CDC, ICMR, regional hospitals)	N/A	Dataset for transfer learning experiments
Jupyter Notebook	Project Jupyter	Open-source	Interactive environment for coding and documentation
Keras	Open-source	Integrated with TensorFlow	High-level API for building and training deep learning models
Matplotlib / Seaborn	Open-source	N/A	Visualization libraries used for graphs and epidemiological trend analysis
NumPy	Open-source	N/A	Numerical computation library
Pandas	Open-source	N/A	Data manipulation and analysis library
Pre-trained CNN/Transformer models (e.g., EfficientNet, BERT-based models)	TensorFlow Hub / HuggingFace	Model-specific	Used for transfer learning and fine-tuning for epidemiological prediction
Python (v3.8 or above)	Python Software Foundation	Open-source	Programming language used for data preprocessing, model training, and evaluation
Scikit-learn	Open-source	N/A	Machine learning library for preprocessing, evaluation metrics, and baseline models
TensorFlow (v2.x)	Google	Open-source	Deep learning framework used for transfer learning and model deployment