Optimizing hepatitis C diagnosis through reinforcement learning feature selection and multi-model machine learning evaluation

Abstract

Hepatitis C virus (HCV) infection remains a leading cause of liver cirrhosis and hepatocellular carcinoma globally, affecting approximately 50 million people with chronic infection worldwide. Traditional diagnostic approaches often rely on extensive biomarker panels, resulting in increased healthcare costs and clinical complexity without corresponding improvements in diagnostic accuracy. This study presents a novel multi-agent reinforcement learning (MARL) framework for optimizing HCV diagnosis through intelligent feature selection, aiming to identify minimal yet highly effective biomarker subsets while maintaining superior classification performance. We developed a four-agent MARL system comprising Greedy Q-learning, Exploratory Q-Learning, Parsimonious Deep Q-Network (DQN), and Random agents to dynamically select optimal feature subsets from clinical data. The framework was validated using a real-world dataset of 1,004 patients with 14 clinical features from Government Medical College, Bakshi Nagar, Jammu, India. Selected features were evaluated across four machine learning classifiers: Logistic Regression, Decision Tree, Random Forest, and XGBoost. The MARL framework successfully identified compact feature subsets (3–4 features) achieving exceptional diagnostic performance (accuracy: 0.98–0.99, F1-score: 0.98–0.99). The Parsimonious DQN agent consistently outperformed other approaches across all classifiers. Notably, SGOT emerged as the most critical biomarkers for HCV detection. The framework demonstrated robust stability across six independent experimental runs, with ensemble methods (Random Forest and XGBoost) showing superior performance with RL-selected features. This study demonstrates that reinforcement learning-based adaptive feature selection significantly reduces diagnostic complexity and healthcare costs while maintaining or exceeding the performance of traditional full-feature approaches. The MARL framework offers a scalable, interpretable, and clinically viable solution for early HCV diagnosis, with potential applications across diverse healthcare analytics domains.

Introduction

The Healthcare system in most nations faces severe challenges due to the increase in various diseases. The expansion of diseases has affected the country’s well-being and financial stability. While diseases cannot be completely eradicated, their timely prediction can save the lives of many patients. Previously, healthcare relied on manual systems, which were simpler and easier to implement but had shortcomings such as being time-consuming, error-prone, and inefficient. Currently, the healthcare system has become digital, offering features like online consultations and utilizing technology to the fullest to serve patients remotely. Rapid advancements in Artificial Intelligence (AI) have the potential to revolutionize healthcare by integrating it into clinical practice¹.Reporting AI’s role in clinical practice is crucial for successful implementation by equipping healthcare providers with essential knowledge and tools¹.

AI is a powerful area of computer science with the potential to fundamentally transform the practice of medicine and the delivery of healthcare². There are extensive applications of AI in healthcare. The applications of technology and AI in healthcare have the capability to provide several benefits like quick and accurate diagnosis, customized treatment plans, medical research and drug discovery, remote monitoring and telehealth, robotic surgery, virtual patient care, administrative applications, medical and imaging diagnostics, rehabilitation, patient engagement and compliance etc. AI has proven to be highly powerful in diverse areas of disease, including skin cancer, colon cancer, cardiovascular diseases, diabetes, neurological disorders, respiratory diseases, and eye conditions. In this work, we have chosen to focus on one specific liver disease—Hepatitis C—which can become chronic if not treated at an early stage.

Hepatitis C is a viral infection due to which inflammation occurs in the liver. Hepatitis C virus (HCV) belongs to the Hepacivirus genus of the Flaviviridae family. It spreads majorly via infected blood like sharing syringes, ineffective infection control in healthcare system etc. Hepatitis C can be acute or chronic. If the virus remains in the body for up to six months and is treated during this period, it is termed an acute infection. However, if the virus is not treated and persists in the body for more than six months, it becomes chronic, which can lead to severe consequences such as liver fibrosis, cirrhosis, and liver cancer. Chronic diseases are illnesses that endure more than three months³. Chronic HCV infection is one of the main causes of liver cirrhosis and hepatocellular carcinoma worldwide⁴. Now, two major diagnostic tests are importantly considered for HCV infection. Firstly, Anti-HCV antibody test to detect any previous or current infection. Secondly, HCV Ribonucleic Acid test to detect the active infection. Globally, an estimated 50 million people have chronic HCV infection, with about 1.0 million new infections occurring per year⁵.World Health Organization(WHO) estimated that in 2022, approximately 2,42,000 people died from hepatitis C, mostly from cirrhosis and hepatocellular carcinoma (primary liver cancer)⁵. A significant reduction in morbidity and mortality from HCV is the major aim of WHO by 2030.

The diagnosis of HCV can be confirmed through three major tests: Liver Function test (Aspartate Aminotransferase (AST)/Alanine Transaminase (ALT)), HCV antibody test, HCV RNA PCR test. Previously, HCV detection was performed manually, which required great physical effort and was time consuming. The major drawback is that sometimes most of the patients remain asymptomatic for years which later on leads to serious problems, also it becomes difficult to handle complex datasets. To overcome all the cons of the manual system, AI is introduced for the early prediction of HCV like many other diseases. With the help of AI, the chances of infection, progress of the disease as well as the treatment strategies can be handled with ease providing high accuracy rate, quick-response, personalized treatment plans, handling large and complex datasets.

AI-based research includes several machine learning (ML) algorithms that learn the patterns associated with the disease and perform prediction on the basis of dataset. ML can be classified as supervised ML, unsupervised ML, and reinforcement learning (RL). Deep learning (DL) is also implemented for the prediction of many diseases like HCV. But we have chosen RL for our research as most of the previous research is performed using ML and DL. In RL, the agent learns by experiencing the environment. That is, the agent is provided with the environment to experience it by performing actions which are either positive or negative, and the agent learns by doing to get maximum possible rewards. When the agent explores the environment by performing some actions, it gets feedback which is either positive or negative. Positive feedback is known as reward, whereas negative feedback is known as penalty. The agent continuously performs three actions that is take action, change the state or remain in the same state and gets the feedback. One of the main reasons for using RL is that we can calculate rewards to increase the accuracy of the model. The main of the paper is organized as follows:

While Q-Learning and Deep Q-Networks have been applied to feature selection in medical diagnosis, the proposed framework stands apart due to its groundbreaking conceptual design and distinct methodological contributions. This work moves beyond basic Reinforcement Learning for Hepatitis C Virus (HCV) data, pioneering a comparative analysis of multiple policies within a single unified feature selection environment. The explicit exposition of novelty are explained as follows:

1. Comparative multi-policy assessment- a key contribution: Unlike established methods centered on single algorithm efficacy, our approach pioneers a novel paradigm: comparing multiple, structurally different RL agents simultaneously for feature selection. These agents utilize distinct behavioral strategies (‘Greedy’ focusing solely on immediate rewards), ‘Exploratory’ (balancing discovery and exploitation), ‘Parsimonious’ (a DQN-based approach prioritizing feature sparsity) and a ‘Random’( which takes decision randomly). By conducting a parallel comparison in a shared environment, we can precisely determine how various learning strategies influence feature selection and shape the resulting model. This novel multi-policy approach offers a more comprehensive and detailed understanding of the inherent trade-offs in RL-based feature selection.

2. Accelerating innovation through purposeful diversity: This work progresses beyond simply testing if RL can select features, offering a comprehensive analysis of which RL approaches are best suited to specific clinical objectives and data characteristics. Beyond simply demonstrating that RL works, this methodology provides a robust way to evaluate different behaviors, clarifying how specific learning mechanisms impact model robustness, performance, and interpretability.

3. Divergence from conventional Single-agent RL methods: Existing research in single-agent RL tends to emphasize the performance metrics of a single-tuned algorithm, which while beneficial provides a limited scope. Departing from traditional methods, we establish a dynamic, collaborative arena for multiple RL paradigms to compete, uncovering the most robust solutions. This system provides a more thorough evaluation, featuring a solid random baseline to rigorously assess the actual advancements and capability improvements derived from advanced learning strategies. Understanding this distinction is vital for creating practical tools, as clinical situations frequently require balancing performance, interpretability, and resource efficiency, all of which are inherently represented by our diverse agents.

4. Enhanced therapeutic insight through behavioral analysis: Moving beyond simple performance metrics, evaluating multi-agent systems provides profound clinical insights. Through observation of Parsimonious agents determine minimal attributes, and how exploratory agents identify unconventional combinations, we gain enhanced insights into the foundational data structure essential for diagnosing Hepatitis C. This enables the optimization of feature selection for specific clinical objectives, balancing model performance with practical constraints like sparsity, and cost-effectiveness.

The remainder of the paper is organized as follows: Section II presents a comprehensive literature review of reinforcement learning and machine learning applications in disease diagnosis. Section III introduces the system model and problem formulation, describing the four-layer hierarchical architecture and mathematical framework for multi-agent feature selection, while Section IV details the proposed MARL framework including environment setup, agent configurations, and interaction mechanisms. Section V presents comprehensive experimental results analyzing feature selection performance, classification metrics, and agent-wise comparisons, followed by Section VI providing discussion. Finally, we conclude the work in Section VII.

Literature review

In this section, we present the literature review.

Zohora et al.⁶ explored the diagnosis of Type-II diabetes using Q-Learning, Decision Tree, and KNN using the data set of Indian Women Diabetes, PIMA which was developed by the National Institute of Diabetes and Digestive and Kidney diseases. The dataset consists of 769 subjects having 9 features like age, insulin level, blood pressure, skin thickness, BMI, blood glucose, etc. Type II diabetes is a condition when there is a lack of production of a hormone named insulin in human body which helps in the absorption of glucose that provides energy in the body. When the glucose stays in the bloodstream instead of reaching the cells, it will lead to increase sugar levels in the human body thus causing type II diabetes. Finally, the reinforcement learning technique used here that is Q-learning performed well compared to the supervised learning techniques that work more closely on three key parameters, namely Age, BMI and Glucose.

Kanthavel et al.⁷ come with the study of osteoarthritis using CNN, DNN, ResNet 150, Dense Vox-Net, and the proposed deep reinforcement learning. A dataset of 4000 X-ray images was used for the prediction of the early stageof knee disorders. 65% is used for training and 35% is used for Testing. The proposed model provides 93.79% accuracy, 93.24% sensitivity, 94.12% specificity, 0.938 precision ratio.

Li et al.⁸ worked to predict Coronary artery occlusion using Combined Multitask Progressive Time-Series networks, where the main objective of this research is to replace the expensive, Coronary Angiography with AI-based methods. The dataset of 17,885 patients is taken into consideration from triple A hospitals in Shanghai. Out of which 70% is used for training set, 10% is used for validation set and 20% is used for test set. Coronary Artery is the blood vessel that supply blood to the heart muscles but when there is some hindrance in them, there is lack of blood flow and there by lack of oxygen in the heart, this condition will lead to the disease named Coronary Heart Occlusion.

Huang et al.⁹ introduced RL for feature selection and then Multi-layer perceptron was trained to identify several gene signatures for Renal Cell Carcinoma. Renal cell carcinoma is a type of kidney cancer when there occur some malignant cells inside the kidney within the tiny tubes whose role is to filter waste from the blood to produce urine. The dataset of Renal cell carcinoma is taken from The Cancer Genome Atlas which includes three subtypes: Kidney clear cell carcinoma, Kidney papillary cell carcinoma and Kidney chromophobe. Sequential Reinforcement Active Feature Learning approach is introduced for selecting the gene subsets dynamically.

Songjing et al.¹⁰ explored the dataset of approximately1.3 million elderly individuals for the prediction of lung cancer using Deep Q-Network in comparison with Support Vector Machine and Logistic Regression to reduce the risk factor. The dataset to be collected from Behavioral RiskFactor Surveillance System and US Environmental Protection Agency. When there includes the formation of malignant cells in the tissues of the lungs, whose major role is breathing, this uncontrolled formation of the abnormal cells in the lungs will cause lung cancer.

Ostelio et al.⁵ performed the evaluation of treatment policies for patients having Hepatitis C using RL techniques. The dataset is taken from Veterans Health Administration corporate data warehouse. The algorithms used are Q-learning, DQN, Policy gradient methods, Baseline models. Hepatitis C is a liver problem which includes the inflammation of the liver due to the presence of the hepatitis C virus which belongs from the Flaviviridae family. The major cause of this disease is infected blood, contaminated needles etc.

Kamepalli et al.¹¹ worked for the diagnosis of heart disease using Q-Learning algorithm in comparison with the traditional machine learning classifiers like K-Nearest Neighbour and Decision Tree taking into consideration the dataset from University of California, Irvine ML repository. The dataset is collected from Cleveland heart disease dataset where three main parameters are taken into consideration that is rest blood pressure, cholesterol and age.

Xiaodan et al.¹² performed the research for sepsis, a life- threatening emergency due to some irregular response of the human body to an infection, which may further damage the human organs and can even harm the survival rate. The dataset is included from the multi-parameter intelligent monitoring in intensive care from USA. The models to be used for implementation are DQN, Double DQN, Dueling DQN, Policy Gradient methods, Baseline models. The best part is both evaluation metrics and clinical outcomes are used for evaluation.

Ganesh et al.¹³ introduced Convolutional Long-Short Term Memory using Double Deep Q-Network (ICL-DDQN)for the prediction of lung cancer, leukemia etc. Three different types of Cancer dataset is considered which is based on gene expression microarray and is publicly available for research purposes. Firstly, Grey wolf optimization model is utilizedfor the feature selection, after which deep reinforcement learning is utilized to develop convolutional neural network with long short time memory so as to select the parameters for classification. Convolutional neural network with long short time memory is beneficial for the conversion of features from low-level to higher level.

Barata et al.¹⁴ worked for the AI-based decision support for skin cancer. The dataset from HAM10000 is considered for the training purpose having 10,015 dermatoscopic images, ISIC2018ChallengeTestsetusedfor external validation and reader study which includes 1,511 dermatoscopic images from different countries like Austria, Turkey, Newzealand, Sweden etc. and patient-centered dataset specially foe RL. Supervised Learning as well as Reinforcement Learning models are implemented on the dataset. Due to the formation of the malignant cells in the tissues of the skin, the disease named Skin Cancer come about.

A.S.Sabeena et al.¹⁵ focused on Diabetic Retinopathy using a Deep Q-Network model trained with Gradient Descent Sea Turtle Foraging algorithm taking into consideration the Indian Diabetic Retinopathy Image Dataset to avoid the loss of patient’s vision. When the increase in the blood sugar levels damages the blood vessels in the retina, which is a light-sensitive tissue in the back of the human eye, this eye condition is termed as Diabetic Retinopathy. The Table 1 unveils the difference between the proposed framework and the existing works.

Table 1.

A relative comparison of the proposed approach with the state -of-the-art approaches.

Author	Year	Objective	Dataset	Algorithm used	Results	Pros	Cons
6	2020	Type-II diabetes	PIMA Indian Women Diabetes Dataset	Q-Learning, K- Nearest Neighbor, Decision Tree	Accuracy: 84.0% Precision: 100.0% Recall: 59.3% F1-Score: 74.4%	Beneficial to reduce the treatment efficiency time	The dataset is very limited and future research can be performed by including more attributes
7	2022	Osteo -arthritis	Publicly available osteoarthritis dataset with X-ray images	RL-deep neural network, Convolutional Neural Network, Recurrent Neural Network	Accuracy: 93.7% Precision: 93.8% Sensitivity: 93.2% Specificity: 94.1%	Accuracy is obtained with minimum data	Generalizability is affected due to the use of limited data
10	2022	Lung Cancer	United States resident’s data	Deep Q-Network, Support Vector Machine, Logistic Regression	Accuracy: 93.8% Area Under the Receiver Operating Characteristic Curve: 90.3% P-value: 0.3%	A large scale dataset covering over 1.3 million elderly individuals enhances generalizability	Computational resources are required for the implementation of Deep Q-Network
5	2022	Hepatitis C	Veterans Health Administrative corporate data warehouse	Q-Learning, Deep Q-Network, Policy gradient methods, Baseline models	Accuracy: 88.0% Precision: 83.0% Recall: 82.0% F1Score: 84.0%	Proved as a novel approach	Simulated patient data is followed instead of real-world
8	2022	Coronary Heart disease	Triple A hospitals in Shangai	Naïve Bayes, Decision Tree, Single LSTM, Combined Reinforcement Multitask progressive time-series networks	Accuracy: 93.7% F1-Score: 95.4%	Artificial Intelligence based methods have replaced the expensive angiography	High computation power is required for the modules
9	2022	Renal Cell Carcinoma	TCGARCC	SRAFL-RF, Chi2-RF, Lasso- RFChi2- SVM, Lasso-SVM	Accuracy: 97.8% F1-Score: 95.4%	SFARL selects the gene subsets dynamically provides improved accuracy when compared to traditional methods	Limited validation and generalizability
13	2023	Lung Cancer, Leukemia Prediction	GEM dataset	ICL-DDQN	Accuracy: 87.9% Precision: 93.3% Recall: 94.1% F1Score: 92.3%	Dynamic Feature selection approach is implemented	Proposed models may not work well with limited data
12	2023	Sepsis	MIMIC-II	DQN, DDQN, Double DQN, Policy Gradient Methods	Accuracy: 87.0% Precision: 85.0% Recall: 84.0% F1Score: 85.0%	Ml evaluation metrics as well as clinical outcomes are used	Dataset is taken into consideration from only a single hospital system
11	2023	Heart disease	UCI ML Repository	KNN, DT, Q-Learning	Accuracy: 87.9% Precision:99.2% Recall: 58.3% F1Score: 76.5% Area Under the Curve: 80.6%	The prediction can be performed using only few lifestyle data	Prediction can be improved by adding other factors like ECG, the number of major vessels, also including large datasets
18	2023	Type-2 diabetes Mellitus	NHANES	Random Forest, Logistic Regression, Support Vector Machine, RL-based model	Accuracy:93.0% F1-Score: 91.0% Area Under the Curve: 95.0%	Worked on large dataset	Huge computational cost
14	2023	Skin Cancer	HAM10000, ISIC2018	Supervised Learning models, Reinforcement Learning models	Accuracy: 79.2% Sensitivity: 87.1%	The major benefit is the improvement in the sensitivity from supervised learning to reinforcement learning	Seven lesion categories are taken into consideration while their include wider range
15	2024	Diabetic Retinopathy	Indian Diabetic Retinopathy Image dataset	DNN, Deep CNN, GDST-FA based DQN, TPCNN	Accuracy: 98.0% Specificity: 97.1% Sensitivity: 98.4%	Different approaches are combined to get better efficiency of the system as compared to an individual approach	Advanced neural networks can also be utilized for better outcomes
16	2025	Liver Fibrosis stage and Inflammatory activity in chronic Hepatitis C infection	Chronic Hepatitis C patient records from Chinese hospitals	Extreme Learning Machine	Fibrosis Stage: ELM: 69.1% K-Means ELM: 72.3% Inflammatory Activity: ELM: 69.9% K-Means ELM: 66.6%	Streamlined training with low resource usage	Model robustness is heavily influenced by the variance in the training data
17	2026	Liver Fibrosis stage and Inflammatory activity in chronic Hepatitis C infection	Clinical profiles of chronic hepatitis C patients	CR-SCAD	Fibrosis Stage Prediction: 69.9% Inflammation Activity Grading: 65.0%	Reduces multi-collinearity in clinical data	Higher computational costs as compared to basic classifiers

Cai et al.¹⁶ evaluates the performance of a k-means-ELM hybrid against a standard Extreme Learning Machine for the classification of Liver Fibrosis and Inflammatory activity, using a clinical dataset of 123 cases. K-means-ELM provided the best performance in staging fibrosis, while stand-alone ELM proved slightly superior for classifying inflammatory activity.

Cai et al.¹⁷ Using a dataset of 123 clinical cases, this study focuses on the comparison of CR-SCAD classifier to standard baseline methods. CR-SCAD demonstrates similar effectiveness to ELM methods for staging fibrosis, while showing moderate performance in assessing inflammatory activity.

Instead of the traditional single-agent reinforcement learning approach which rely on a single explore-exploit trade-off, this research adopts a multi-policy reinforcement learning approach for investigating feature subset selection performance under different algorithmic approaches. In particular, four reinforcement learning agents- Greedy Q-Learning, Exploratory Q-Learning, Parsimonious Deep Q-Learning (DQN) and a random baseline are employed.

While many RL-based feature selection studies focus on maximizing accuracy with a single mode, they tend to overlook the impact of policy bias, search behavior, and stability on the reliability of medical diagnosis. Conversely, the proposed methodology centers on comparative behavioral policy analysis, facilitating the isolation of stable, minimal and clinically interpretable feature subsets. Each agent performs a unique analytical function: Greedy Q-learning focuses on exploitation, Exploratory Q-learning favors extensive searching to avoid local optima, Parsimonious DQN forces explicit sparsity, and the Random Agent serves as a diagnostic tool for feature redundancy. The innovation here extends beyond just using reinforcement learning for Hepatitis C feature selection, it compares multiple RL policies, providing a deeper understanding of feature importance, trade-offs, and robust performance. This multi-policy formulation shifts the focus from purely maximizing accuracy to ensuring a more interpretable and practical feature selection strategy for clinical settings.

System model and problem formulation

System model

The performance model for the early prediction of Hepatitis C is organized into four hierarchical layers, each performing a specific role in the whole framework- from acquiring data to feature selection using RL and implementation via ML. The layers are detailed as:

Hospital and patient layer

This layer constitutes ’h’ number of hospitals where:

1

and H is the total number of hospitals. Let n_h be the number of patients in hospital h and

2

denotes the total number of patients in all the hospitals.

Let

3

represents the raw data vector for kth patient in hospital h, consisting of m clinical features.

Consider

4

be the patient-specific features, such that:

5

Finally, the data matrix can be written as:

6

7

Data collection layer

The data transmitted to data collection layer from the Hospital and patient layer is now collected, organized and stored securely at this centralized data integration unit. This layer ensures the consistency and quality after converting the raw data into a structured format.

Let the raw data be:

8

which can be transformed into a structured dataset as:

9

where:

10

represents the structured form of input vector i_jafter applying preprocessing function f(.) This will provide consistent dataset as:

11

where N is the number of patients and M is the number of clinical features. In the preprocessing stage, the missing values are imputed using median strategy and Min–Max Scaler is applied for scaling all continuous features and finally, the data is integrated.

Reinforcement layer

The RL layer come up with the task of selecting the optimal features from the original set of features.

Let

12

be the total set of features, where the main aim is to choose the subset of features as X ⊂ Y such that

13

The feature subset is set on dynamically by the multi-agent setup including Greedy agent, Exploratory agent,

Parsimonious agent and Random agent depending on the performance of the classification model, using rewards which then further balances the evaluation metrics.

Let T = total number of features.

A = {A1, A2, A3, A4}: set of four RL agents.

Each agent A_klearns a policyπ_k, where:

14

mapping state s_t to action for each feature at timestep:

15

The action vector for agent A_k is:

16

Feature subsets per agent: For every agent A_k, the subset of features to be selected for sample is:

17

where:

x_j^(k)represents the jth feature of the input sample its processing by the kth agent, which means that the features selected by agent A_k are retained only.

x_j^’: represents the original jth feature vector of a sample.

a_k: be the binary vector representing the feature subset selected by agent A_k. Each element in a^k is either 0 or 1,

where 1 indicates that the particular feature is selected by the agent A_k , and 0 indicates it is not.

: represents the element-wise product. It filters the features of x′_j when applied to x′_j and a_k. If an element in a_k is 1,the corresponding feature from x′_j is kept, but if it is zero, the feature is discarded.

Reward evaluation per agent: On the basis of the performance of a ML model trained using the features to be.

selected, every agent receives a reward R_k, such that:

18

where F denotes a model evaluation metric.

x^(k)is the data set to be transformed using particular features selected by the agent A_k.

Machine learning layer

The minimal feature subsets elected with the Multi-agent RL framework is now implemented on the ML models to be selected for the prediction of the Hepatitis C disease. The RL environment rewards agents for selecting features that maximize the F1-score of the particular ML model while penalizes excessive feature selection to ensure sparsity.

Let.

19

represents the dataset after feature selection by RL framework and .

: the presence or absence of hepatitis C.

Y_model : the predicted labels from the ML models.

M : the classification model used.

Model training and prediction

where M_RF denotes Random Forest,M_XGBdenotes XGBoost, M_DT denotes Decision Tree, M_LR denotes Logistic Regression.

DataSet description

This study utilized a retrospective dataset of 1,004 anonymized patient records with 14 features sourced from a single tertiary-care medical institution named as Government Medical College, Bakshi Nagar, Jammu, Jammu and Kashmir, India. The dataset provides routine laboratory values, including liver enzymes, for the assessment of Hepatitis C and the target variable, Status, indicating the presence of HCV RNA (1 for detected, 0 for not detected). After preprocessing, 938 samples along with 8 features are finally considered.

To ensure patient confidentiality, all personally identifiable information was redacted before analysis. Ethical restrictions and institutional policies prevent public release of raw clinical data. Nevertheless, full transparency is ensured through the provided anonymized metadata and comprehensive preprocessing procedures. Data access is available via reasonable request and institutional clearance.

As a single-center study, our results reflect local demographic and laboratory characteristics, which may limit their applicability to other patient populations without further validation. In terms of empirical justification, we have excluded Age and Gender in our research taking into consideration the following points:

1. Bias mitigation: The datasets related to healthcare includes attributes like age, gender etc. These attributes cause biasness in the results due to the reason that the model may produce results learning some patterns without actually relying on the actual biochemical markers.

2. Data minimization: According to General Data Protection Regulation (GDPR), the data collected must be exactly as per the requirement of the study.

3. Anonymization: Removing personal attributes retains data privacy.

4. Data sparsity and noise: In most of the datasets, 70–80% patients avoid to share these details due to which there are chances of overfitting. Due to this, using wrong data will lead to reduce the model’s accuracy.

Data preprocessing

To ensure data quality and stability, an initial preprocessing step was carried out prior to the reinforcement learning- based feature selection. The raw data underwent data cleansing and transformation to ensure it was accurate, consistent, and ready for modeling, including reinforcement learning-based feature selection and machine learning classification. Originally containing 14 attributes, the dataset was reduced to 8 features during preprocessing by removing patient name, contact information, gender, age, status and viral load. To comply with ethical standards and ensure confidentiality, all personally identifiable information was removed from the dataset. The demographic variables are removed to minimize potential bias, allowing the model to concentrate exclusively on biochemical markers. Viral load measurements were omitted to prevent data leakage and artificially high performance results, as these tests serve as a confirmatory diagnostic step for HCV. Consequently, the finalized model is trained and validated using eight specific liver function parameters.

Handling missing values

The dataset included quantitative results from common Liver Function Tests specifically: Albumin, Bilirubin, Total Protein, Alkaline Phosphatase, Aspartate Aminotransferase, Gamma Glutamyl Transferase and Globulin. We performed median imputation to handle the missing data, substituting all absent entries with their feature’s median value.

The method was selected for the following reasons:

• The Median is less sensitive to extreme values, which are frequently observed in clinical liver biomarkers.

• It maintains the original data distribution shape more effectively than imputation using the mean.

• It avoids introducing systematic error that could affect the rewards in reinforcement learning during the feature selection process.

Mathematically,

20

Feature scaling: The Min–Max Scaler was implemented to normalize all features because the variables possess different ranges (eg: Albumin and SGOT):

21

Target label encoding: The Conversion of multi-class categories (such as healthy, Fibrosis and Cirrhosis) into two groups for simplified diagnostic modeling is a standard modeling technique. The HCV status was assigned a binary value: 0 for non-HCV cases and 1 for HCV positives.

Split of training set: The data were divided into training (80%) and testing (20%) while preserving the original class proportions and preventing information leakage from the feature selection process into the classifier evaluation.

Reinforcement learning: Each RL agent (Parsimonious, Random, Greedy, Exploratory) engaged with this preprocessed data set to sequentially determine the most informative subset of features.

Compatability with machine learning: The final refined data collection was utilized for model training: Logistic Regression, Decision Tree, Random Forest, XGBoost. All models were trained on consistently preprocessed data, which guaranteed reliable and comparable performance across different set of features.

Problem formulations

Problem context

It is generally known that the delayed diagnosis of any disease can lead to critical situations. Similarly, Hepatitis C, if not detected at an initial stage, will lead to dire liver problems. Prior investigations on HCV using ML typically depend on traditional feature selection approaches such as filter method, wrapper method, etc. or the features are applied unfiltered, which led to unrecorded acquisition cost, clinical work ability, feature repetition. Due to this, the problems to be faced are discussed as:

Feature selection: The Feature selection procedure to be followed previously is static and myopic, as well as cost-ineffective.

Lack of clinical synergy: The major focus is on accuracy instead of early detection, cost for acquiring features.

This approach uses an optimization framework to select the smallest subset of markers that yields the highest diagnostic efficacy. Assume ‘F’ is the complete set of features and ‘S’ is a selected subset such that S ⊂ F. The objective is to maximize a utility function that balances classification accuracy, represented by the F1-score, against a penalty for model complexity or feature set size. This can be represented as:

22

The formulation helps agents achieve a balance between predictive accuracy and practical clinical utility by selecting concise and reliable features.

To ensure a fair comparison, the weighting parameters α and β were chosen empirically to prioritize diagnostic accuracy and ensure model sparsity, remaining constant across all agents. This approach explains that in scenarios with sparsity constraints, agents prefer stable functional markers to high-variance enzyme markers.

The proposed framework

The initial preprocessing phase guarantees data quality and feature engineering before applying MARL-based feature selection. A multi-agent reinforcement learning ((MARL) framework composed of four autonomous learning agents- Greedy Q-Learning Agent, Exploratory Q-Learning Agent, Parsimonious DQN Agent and Random Agent to select a minimal feature subset (targeting 3–4 features) for HCV diagnosis, aiming for high performance (accuracy/F10.99). After feature selection using the agents, the ML models are implemented on the selected features. The proposed framework is shown in Fig. 1. The RL environment rewards agents for selecting features that maximizes the F1-score of the ML model while penalizing the excessive feature selection to ensure sparsity.

Fig. 1.

Layers in the system model.

Environment (feature selection Env): The Feature Selection Environment is a class for selecting the features in a reinforcement learning environment. The Features are selected for ML models, where an agent’s actions for including or excluding the features results in rewards depending on the model’s performance:

• State: Binary feature mask representing selected features.

• Actions: Each action determines which feature to modify and whether to include the particular feature or exclude it.

• Rewards: The reward is calculated on the basis of the F1-score of the models on the selected features and penalty is included if too many features are selected.

• Model: ML models are trained at each step with the number of features selected.

Agents: Four RL agents were used:

• Greedy agent: A Q-Learning agent with low exploration (ϵ = 0.2) and fast learning (α = 0.1), focusing on quick convergence.

• Exploratory agent: A Q learning agent with high exploration (ϵ = 0.5) and slow learning (α = 0.005), designed to explore various combinations of characteristics.

• Parsimonious agent: A Deep Q-Network (DQN) agent with moderate exploration (ϵ = 0.4) and a sparsity penalty (0.04), balancing performance and feature count.

• Random agent: A baseline agent that selects features randomly for comparison.

  • Interaction: The environment provides cyclical interaction where the agent learns about the current state of the feature.

Reinforcement learning training and stopping criteria

Training was concluded for each agent once the prescribed termination conditions were satisfied. Specifically, the agents underwent training for a predetermined maximum number of episodes, established to guarantee thorough exploration of the feature selection space. Furthermore, convergence was verified by observing the stabilization of cumulative rewards over successive training episodes. Empirical evidence showed that the reward stabilized well before the maximum episode limit, indicating that both the policy and the selected features had converged. The final subset of features for each agent was determined by the policy derived at the conclusion of training. The model ceased learning when cumulative reward improvements became negligible over consecutive episodes, or when the episode cap was reached.

Computational cost and runtime analysis

Because reinforcement learning with function approximation is both iterative and stochastic, calculating a precise, closed-form computational complexity is challenging. To evaluate the practical computational cost of the proposed framework, we report on its empirical runtime performance. Throughout several independent training sessions, the agent’s runtime was monitored based on actual elapsed wall-clock hours. Experiments were executed on CPU, using the Google Colab framework. Each training session lasted roughly 180 s, reaching convergence within about 100 episodes. The duration increases proportionally with agent and episode counts, whereas the expense per episode is mainly driven by reward computation and classifier training. Experimental results verify that the proposed framework is computationally viable for mid-sized clinical datasets.

Data splitting and evaluation protocol

A rigorous training–testing split was implemented to guarantee unbiased evaluation and prevent data leakage. For each experimental trial, the entire dataset was divided into an analysis set (80%) and a retained validation set (20%) using proportional sampling to maintain group ratios. To enhance statistic reliability, this partitioning process was repeated across multiple independent trials using varied random seeds. Notably, the reinforcement learning-based feature selection process completely excluded the test set. The RL agents were trained solely on the training set, with feature selection, guided exclusively by performance metrics derived from the same data. Once the RL agents finalized their feature selection and downstream models were trained, the holdout test set was applied only once per run. By avoiding information leakage during feature selection, this evaluation approach guarantees that performance metrics represent true generalization to new, unseen data. A separate validation set was not explicitly defined.

Justification of multi-agent design choices

The proposed framework utilizes a comparative multi-agent reinforcement learning (MARL) approach, where each agent learns autonomously without sharing policies, optimizing joint rewards, or communicating with others. Multiple agents are employed not to cooperate, but to systematically analyze different HCV diagnostic strategies under identical settings. The Greedy agent acts as a fixed, exploitation–oriented benchmark, reflecting a rule-based diagnostic approach. By prioritizing stochastic actions, the exploratory agent investigates the effect of discovery-based learning on feature exploitation and model generalization. The Parsimonious DQN agent aims for concise, clinically interpretable feature sets by penalizing redundancy, thereby fulfilling practical deployment equipments. The random agent serves as a control baseline, enabling comparison with uninformed actions and providing context for the effectiveness of the learned policies. These agents were deliberately chosen to represent a diverse range of learning behaviors, rather than to cooperate with one another. Comparing agent performance illuminates the accuracy-stability-efficiency trade-offs, strengthening the validation of our diagnostic framework.

Reward function design and sensitivity analysis

The framework is designed to select the features using Reinforcement Learning agents, a reward function is optimized which balances the predictive performance with feature sparsity. For a particular selected feature subset, the reward function was defined as:

23

where F1score depicts the F1-Score of the particular model trained on the selected features, Nfeatures depicts the subset of features to be selected, and Nthreshold depicts the maximum number of features to be acceptable and are predefined as 4.

α and β are the two parameters to act as penalty coefficients where the importance provided to the classification performance is controlled by ‘α’ and ‘β’ controls the penalty which is applied when too many features are selected.

Particularly, for Greedy and exploratory agent α is 0.02, where as for Parsimonious agent α is 0.04and the penaly β is.

set to 0.1 for all agents. A sensitivity analysis was performed to assess the sensitivity and robustness of the feature selection.

Statistical analysis

A statistical analysis was performed to critically examine the performance of several RL agents and to validate whether the observed performance differences were statistically meaningful. This statistics was performed depending on the independent experimental runs which is N = 6 in this study, which indicates the variability and average performance of the models. Particularly, the paired t-tests are applied for comparing the results of F1-score between pair of agents for each classifier. In this study, the paired t-tests are selected as the performance scores are acquired from the same set of experimental runs, making the observations dependent.

Pairwise comparisons were performed for all the four classifiers, using the F1-scores generated by different agents across all the six runs. A significance level of p < 0.05 was adopted for all primary analyses. This level of significance suggests that the performance gaps identified are statistically robust and unlikely to be incidental.

Experimental results

This section describes the results obtained from the different classifiers implemented in the dataset.

Feature selection performance of RL-based agents

The research evaluated the efficacy of four reinforcement learning agents—Parsimonious agent, Random agent, Exploratory agent, and Greedy agent in determining the best feature sets for the detection of Hepatitis C. Each agent exhibited a different degree of variability regarding both the number of features chosen and the consistency of those choices over multiple independent executions. The agent’s interaction with the feature space varied; consequently, each exhibited a different selection pattern. The aim of this multi-agent approach is to simulate varied behaviors and identify the most important factors. Each agent chooses features during every episode as shown in Fig. 2 and the corresponding selection frequency was then computed.

Fig. 2.

Performance Curves of various agents across training episodes. The horizontal axis shows training progress in episodes, while the vertical axis represents the episodic return, computed as the classification performance of the chosen feature subset.

Figure 3 displays a heatmap where the normalized frequency of feature selection for each agent is visualized (0indicates no selection, 1 indicates constant selection).This visualization offers a granular look at how individual agents prioritize particular liver function biomarkers during the optimization of the HCV detection subset feature. Across all agents, the features most reliably chosen are Albumin, Bilirubin, GGT and SGOT highlighting their robust ability to predict HCV classification. In Contrast, the infrequent selection of liver enzyme markers (SGOT, SGPT, and GGT) by the Greedy and Parsimonious agents implies their limited impact on classification performance.

Fig. 3.

Heatmap of feature selection frequency across agents.

Agent-wise feature selection behavior

1. Greedy agent: The heatmap delineates how the acquisitive agent variably chooses only Total Protein.

2. Parsimonious agent: The analysis preferentially included core protein markers such as Albumin, and SGOT while excluding all other indicators.

3. Random agent: The agent established a baseline by picking features at random with equal probability and the selected features includes Albumin, Bilirubin, GGT Globulin, SGOT, Total Protein. The resulting selection data did not fixate on a narrow choice, as evidenced by the heatmap’s varied shades, which reflect a thorough examination.

4. Exploratory agent: This agent explores mainly Bilirubin and GGT. The overall selection rate of each feature as shown in Fig. 4, aggregated across all agents and runs, was computed to shed light on its role in the RL-based feature selection process. Albumin, Bilirubin, GGT and SGOT represent the top selected features in Fig. 4, underscoring their robust predictive significance for classifying liver conditions. Their frequent selection highlights their strong utility in determining liver health status.

Fig. 4.

Global feature importance based on selection frequency.

Classification performance across machine learning models

The performance of the selected feature subsets was assessed using four machine learning models—Logistic Regression, Decision Tree, Random Forest and XGBoost. The effectiveness of the model was measured on a held-out-test set, with Accuracy, Precision, AUC, recall and F1-Score being the chosen metrics. Reinforcement Learning- driven feature selection demonstrably improved classification accuracy compared to models that use all available features.

1. Accuracy: The accuracy of every possible agent and classifier pairing is shown in Fig. 5.The Parsimonious agent consistently demonstrated superior accuracy, ranging from 0.97 to 0.99, relative to other models; this suggests that utilizing a moderate search strategy facilitates the identification and selection of more effective, discriminatory liver biomarkers, for the detection of HCV. The Exploratory agent also performed reliably well (0.88–0.98), with high exploration rate. The Random agent which acts as a baseline, performs better than Greedy agent with the results as (0.91–0.93). On the contrary, the low performing agent is Greedy agent (0.62–0.78). Random forest performs consistently high for Parsimonious agent, Random agent and Greedy agent where as XGBoost performs high with Exploratory agent.

   Accuracy is heavily influenced by the chosen feature-selection strategy: specifically, methods that encourage exploration (such as ’Exploratory’ and ’Parsimonious’) yielded Feature subsets that results in superior classification performance when detecting HCV.

2. Precision: Precision as shown in Fig. 6 highlights the model’s ability to minimize false positives, a vital aspect of clinical decision support systems for identifying positive HCV cases. The high precision scores (0.989–0.996) attained by the Parsimonious agent indicates that a comprehensive investigation of the data features allows the model to distinguish the Hepatitis C patients reliably and with minimal misclassification. The strong, reliable precision (0.905–0.985) of the Exploratory agent suggests that the data possesses numerous valuable characteristics, allowing even chance selection to capture relevant clinical insights. The Random agent (0.917–0.937) even performs better than the Greedy agent. On the contrary, the low performing agent is the Greedy agent(0.602–0.760).

   Among the different algorithms, precision was most stable for Logistic Regression in case of Parsimonious agent, Random Forest in case of Greedy agent and random agent ,and Random Forest and XGBoost in case of Exploratory agent.

3. F1-Score: Fig. 7 displays the average F1-Scores for all possible pairing of agents and classifiers. Across all models, the Parsimonious agent consistently yielded the highest F1-scores(0.966–0.991), demonstrating that moderate exploration of the state-action leads to more informative feature subsets for HCV detection. Similarly ,the F1-scores for Exploratory agents range from 0.871 to 0.980. After this, Random agent performs better than Greedy agent as(0.909–0.934). By comparison, the Greedy agent yielded significantly poorer F1-scores, falling within the range of 0.620–0.794. Among classifiers, Random Forest consistently exhibited stable performance when utilized by high performing agents, while the highest improvement for a classifier paired with exploratory search was achieved by XGBoost.

4. Recall: The heatmap shown in Fig. 8 shows the recall values. The Exploratory (0.871–0.980) and Parsimonious (0.966–0.991)agents provides the highest recall values among all classifiers. On the other hand, Random agent (0.909–0.934) performs better performance than the Greedy agent where the Greedy agent performs very low with (0.620–0.794) values.

Fig. 5.

Heatmap of accuracy for different agent-classifier combinations.

Fig. 6.

Heatmap of precision for different agent-classifier combinations.

Fig. 7.

Heatmap of F1-Score for different agent-classifier combinations.

Fig. 8.

Heatmap of Recall for different agent-classifier combinations.

Comparative analysis of RL agents

Parsimonious agent: The Parsimonious agent achieves overall best performance as shown in Table 2 as:

Table 2.

Performance comparison of machine learning classifiers of parsimonious agent on the independent test set.

Classifiers	Accuracy	Precision	Recall	F1-Score
XGBoost	0.991	0.994	0.987	0.990
Random forest	0.992	0.992	0.991	0.991
Logistic regression	0.971	0.996	0.941	0.966
Decision tree	0.987	0.989	0.983	0.986

Greedy agent: The Greedy agent is demonstrated as the lowest performing agent as shown in Table 3.

Table 3.

Performance comparison of machine learning classifiers of greedy agent on the independent test set.

Classifiers	Accuracy	Precision	Recall	F1-Score
XGBoost	0.775	0.759	0.771	0.761
Random forest	0.781	0.760	0.841	0.794
Logistic regression	0.626	0.602	0.642	0.620
Decision tree	0.779	0.760	0.826	0.788

Random agent: The Random agent achieves surprisingly competitive performance as shown in Table 4.

Table 4.

Performance comparison of machine learning classifiers of random agent on the independent test set.

Classifiers	Accuracy	Precision	Recall	F1-Score
XGBoost	0.935	0.930	0.934	0.932
Random forest	0.938	0.937	0.932	0.934
Logistic regression	0.911	0.917	0.903	0.909
Decision tree	0.925	0.920	0.930	0.924

Exploratory agent: The Exploratory Agent shows very strong performance as shown in Table 5.

Table 5.

Performance comparison of machine learning classifiers of exploratory agent on the independent test set.

Classifiers	Accuracy	Precision	Recall	F1-Score
XGBoost	0.981	0.985	0.975	0.980
Random forest	0.979	0.985	0.970	0.977
Logistic regression	0.880	0.905	0.845	0.871
Decision tree	0.978	0.983	0.970	0.976

The Parsimonious agent among all classifiers stands out as the most consistently high performing agent with a very compact feature set. The Exploratory agent proves to be very strong in case of tree-based models and XGBoost. The Random agent proves to be a robust baseline, surpassing Greedy agent and showing comparable performance with higher feature counts. The performance of the Greedy agent is lowest with Logistic Regression.

Generally, machine learning agents based on learning principles performed much better than the Random agent except Greedy agent, which underscores the utility of RL for optimizing features in medical classification tasks.

To evaluate the reliability of each RL agent, the F1-score was tracked across six independent experiment runs. Figure 9 highlights the performance evolution of each agent. The Exploratory and Parsimonious agents perform very well, demonstrating a high F1-score that is consistent across all tests, while the Random agent performs less effective than Parsimonious and Exploratory but better than the Greedy agent. These observations demonstrate the enhanced durability of agents with capabilities for a wider and moderate range of exploration.

Fig. 9.

Mean F1-Score across multiple experiment trends.

Baseline strategy

In the experimental setup, a random agent was included to establish an internal performance baseline. In random agent, the selection of features is performed randomly without depending on any policy which is totally different than the selection process of the RL agents that depends on a particular policy for feature selection. This provides an opportunity to compare the informed agents as well as the uninformed random agent. To make sure significant internal evaluation, this random agent worked as a baseline. The comparison of the RL agents is performed with this baseline agent instead of the conventional filter, wrapper and embedded methods which are actually out of the scope of this study.

Ablation study on reinforcement learning policies

An ablation-style evaluation was conducted to compare the performance of the three RL-based agents: Greedy, Parsimonious and Exploratory. Each of these policies follow a different strategy to select the subset of features. The Greedy agent focuses on the improvement of the performance immediately, the exploratory agent promotes extensive search across the feature space, and the parsimonious agent highlights feature sparsity while preserving the classification performance. Table 6 depicts the comparative results of the different policies along with their learning strategies, accuracy, F1-score and the number of selected features. Among all agents, Parsimonious agent performs well with better accuracy and F1-score as shown in the table 7, 8 and 9.

Table 6.

Ablation analysis of RL policies for feature selection.

S. no	Agent	Best accuracy	Best F1-Score	Mean features selected	Most frequent features selected
1	Greedy	78.1%	79.4%	1	Total Protein
2	Exploratory	98.1%	98.0%	2	Bilirubin, GGT
3	Parsimonious	99.2%	99.1%	2	Albumin, SGOT
4	Random	93.8%	99.1%	5	Albumin, Bilirubin, GGT, Globulin, SGOT

Table 7.

Paired t-test analysis comparing the mean performance of different RL agents across repeated experiments for Logistic Regression classifier.

Agent 1	Agent 2	T-statistic	P-value	Significance (p < 0.05)
Greedy	Exploratory	− 2.818	0.037	Yes
Greedy	Parsimonious	− 4.564	0.006	Yes
Greedy	Random	− 3.480	0.018	Yes
Exploratory	Parsimonious	− 1.191	0.287	No
Exploratory	Random	− 0.324	0.759	No
Parsimonious	Random	0.743	0.491	No

Table 8.

Paired t-test analysis comparing the mean performance of different RL agents across repeated experiments for decision tree classifier.

Agent 1	Agent 2	T-statistic	P-value	Significance (p < 0.05)
Greedy	Exploratory	− 2.166	0.083	No
Greedy	Parsimonious	− 2.221	0.077	No
Greedy	Random	− 1.703	0.149	No
Exploratory	Parsimonious	− 0.973	0.375	No
Exploratory	Random	0.829	0.445	No
Parsimonious	Random	1.052	0.341	No

Table 9.

Paired t-test analysis comparing the mean performance of different RL agents across repeated experiments for random forest classifier.

Agent 1	Agent 2	T-statistic	P-value	Significance (p < 0.05)
Greedy	Exploratory	− 2.143	0.085	No
Greedy	Parsimonious	− 2.308	0.069	No
Greedy	Random	− 1.821	0.128	No
Exploratory	Parsimonious	− 1.818	0.129	No
Exploratory	Random	0.737	0.494	No
Parsimonious	Random	1.057	0.339	No

Impact of reward function penalty on agent performance and feature selection

The selection of feature subsets and predictive performance are confirmed by the specific values of α and β used in the reward function which is shown in the respective sensitivity analysis. Figure 10 provides a visual summary of the aggregated results across various penalty configurations. Under default penalty, Parsimonious and random agent provides high F1-score while selecting 4 or 2 features on an average whereas Greedy and Exploratory agents comes up with low F1-score while selecting only very few features, with the Greedy often selecting only one. Under low feature penalty, the agents are encouraged for selecting more features, still Parsimonious agent provides best results using compact features giving preference for sparse solutions. For high feature penalty, agents were suggested to select even smaller feature subsets, for this Parsimonious and Exploratory agents performs high F1-score.

Fig. 10.

Sensitivity analysis of the reward function under different feature sparsity penalties.

Statistical validation of agent performance differences

The results of our statistical validation indicate that specific performance gains among agents are significant, whereas others do not meet the p < 0.05 threshold.

Logistic regression: It is observed that the Greedy agent performs statistically significantly worse when compared to Parsimonious, Random and Exploratory agents in case of Logistic Regression. This proves that Greedy is less efficient in this context and the validation proved a significant difference as shown in table 10.

Table 10.

Paired t-test analysis comparing the mean performance of different RL agents across repeated experiments for XGBoost classifier.

Agent 1	Agent 2	T-statistic	P-value	Significance (p < 0.05)
Greedy	Exploratory	− 2.148	0.084	No
Greedy	Parsimonious	− 2.238	0.075	No
Greedy	Random	− 2.076	0.093	No
Exploratory	Parsimonious	− 1.338	0.239	No
Exploratory	Random	0.789	0.466	No
Parsimonious	Random	0.994	0.366	No

Decision tree, random forest and XGBoost: The validation shows no significant difference s among any of the agent pairs. This shows that the observed differences in the average F1-scores were not consistent enough to confidentally declare one agent statistically superior to other.

So, the validation proves to be beneficial because it allows to differentiate the performance of agents for logistc regression.

Computational cost and runtime analysis

The average computational time for several reinforcement learning agents is computed for different experimental runs and shown in the Tables 11 and 12 , which provides a quantitative comparison corresponding to each agent.

Table 11.

Agent training runtime summary.

S. no	Agent	Mean training time	Std dev training time
1	Exploratory	6.454523	0.353609
2	Greedy	6.531288	0.493046
3	Parsimonious	12.539150	0.918941
4	Random	6.168792	0.463537

Table 12.

Classifier evaluation runtime summary per agent.

S. no	Agent	Mean evaluation time	Std dev evaluation time
1	Exploratory	0.057707	0.019848
2	Exploratory	0.057501	0.010600
3	Exploratory	1.406285	0.395911
4	Exploratory	0.173569	0.063140
5	Greedy	0.061257	0.037611
6	Greedy	0.693680	1.570682
7	Greedy	1.281895	0.216519
8	Greedy	0.255928	0.141561
9	Parsimonious	0.051042	0.013214
10	Parsimonious	0.057310	0.015791
11	Parsimonious	1.252024	0.238203
12	Parsimonious	0.168939	0.075189
13	Random	0.050939	0.014219
14	Random	0.056054	0.017276
15	Random	1.297854	0.199096
16	Random	0.171289	0.050965

Discussion

The purpose of this research is to evaluate the utility of a reinforcement learning framework in selecting features for the diagnosis of hepatitis C virus and to compare its results with those of traditional machine learning models. The study reveals that the RL feature selection framework successfully identified a compact and powerful feature subset, thereby reducing model complexity while maintaining and sometimes enhancing predictive performance. Models utilizing features chosen via reinforcement learning achieved comparable performance metrics, including accuracy, precision, recall and f1- score, to models trained on the entire feature set across diverse classification algorithms such as Random Forest, XGBoost, Logistic Regression, and Decision Tree. A key advantage of RL agents is their ability to adaptively search for optimal feature subsets and uncover intricate, non-linear patterns which are often overlooked by traditional statistical or filter-based techniques. RL’s capacity to consistently choose a stable set of important features across multiple trials validates the dependability of the feature selection approach to identify relevant HCV-related biomarkers. By reducing dimensionality, RL improved computational performance, resulting in faster training speeds and lower resource usage compared to full feature models. This is crucial for practical use cases, where there is a strong demand for fast, inexpensive diagnostic systems. Combining reinforcement and other machine learning methods provides a powerful evidence-based strategy for enhancing diagnostic tools for Hepatitis C Virus (HCV). The performance could be further enhanced employing more sophisticated reinforcement learning agents, more extensive agent’s datasets, or specialized reward mechanisms.

A notable discrepancy exists between individual biomarker importance and agent-level selection consistency within the proposed RL model. SGOT acts as a powerful, independent predictor in model subsets, reflecting its importance in clinical assessments of liver injury. Nevertheless, SGOT, SGPT, Bilirubin and Total Protein were chosen far more frequently across various agents and experimental trials. The agents prioritize stable, generalizable biomarkers under sparsity constraints over those with high yet inconsistent, individual importance. The agent’s reward function balances high predictive accuracy with minimal feature usage, causing them to favor reliable biomarkers that result in compact, high performing subsets across different runs. This finding suggests prioritizing stable functional indicators over fluctuating enzyme levels in Hepatitis C diagnostics, resulting in more reliable, interpretable, and applicable models in real-world settings. Potential future works involves exploring deep reinforcement learning methods, multi-agent RL systems, or hybrid feature selection techniques to enhance the robustness and applicability of the predictive model.

Clinical interpretability and practical deployment

Beyond just diagnostic accuracy, clinical interpretability is crucial for the practical adoption of machine learning – based diagnostic systems. This reinforcement learning approach enhances clinical usability by identifying compact sets of 3–4 common liver markers, thus lowering costs and simplifying testing procedures. Because they directly represent liver synthetic activity, consistent monitoring of stable indicators- specifically SGOT, SGPT, Bilirubin and Total Protein enhances the ease of clinical assessment. Since the reinforcement learning process is an offline development step, the optimized feature subset can be directly applied in standard classification pipelines, ensuring no additional computational burden during deployment. By decoupling the training process from the inference phase, this framework remains practical for clinical adoption while ensuring decisions are both transparent and trustworthy. The framework bridges the divide between sophisticated machine learning and functional clinical decision support by offering both superior diagnostic performance and transparent biomarker selection.

Agent behavior analysis

The surprisingly high accuracy of random agent is due to the inherent properties of the dataset, rather than the superiority of the random search. The study’s clinical features exhibit substantial redundancy, as multiple liver function markers reflect overlapping pathological data related to Hepatitis C progression. As a result, subsets of features chosen at random frequently possess enough discriminative information to yield high classification accuracy. Due to strong disease-related patterns, the modeling task is relatively straightforward, despite not using optimal features. Although it demonstrates high average performance, the random agent falls short in stability, consistency, and interpretability, largely because its chosen feature sets are inconsistent and not optimized for simplicity. Conversely, the proposed agent-based system systematically leverages redundant data to select concise, stable, and clinically relevant feature sets, thereby delivering higher robustness, and interpretability than random exploration can provide.

Comparative analysis of agent performance

The tests reveal that the agents perform at distinctly different levels. The parsimonious agent consistently outperforms others by proving its efficacy in identifying a concise yet highly informative subset of features. By optimizing the trade-off between predictive accuracy and feature sparsity, this algorithm retains key biomarkers while discarding redundant variables. The Exploratory agent takes second place, utilizing its capacity to survey a broader feature space and identify powerful feature combinations. The intermediate performance of the Random agent indicates that despite intermittent discovery of useful feature subsets, the absence of a structured learning framework prevents consistent long-term optimization. Conversely, Greedy agent underperforms likely because it prioritizes immediate rewards over sufficient exploration, resulting in inferior feature sets. In conclusion, effective RL-based feature hinges on balancing exploration with sparse optimization.

Future research directions

For image-based deep learning, spectral-based singular pooling acts as an effective method for preserving dominant structural information when handling high-dimensional inputs¹⁹. This work focuses on Reinforcement Learning and classical classifiers for tabular clinical data; however, incorporating spectral information preservation is an interesting direction for subsequent research. Leveraging singular pooling within advanced deep models-such as reinforcement learning frameworks- offers a mechanism to enhance future learning for complex, time-series biochemical patient data. Furthermore, incorporating multimodal data-such as pairing laboratory biomarkers with imaging or longitudinal clinical signals-could leverage singular pooling to preserve key spectral features while minimizing dimensionality. These directions are reserved for future research, as they necessitate significant architectural changes beyond the current study’s scope.

Interpretability and trustworthiness of selected biomarkers

The feature subsets derived by multiple reinforcement learning agents over six independent runs are summarized in Table 13. The findings highlight significant variations in how the agents select features. The Greedy agent consistently prioritizes a minimal subset of features, typically identifying Total Protein as the primary variable early on, while occasionally incorporating liver biomarkers like Albumin, Bilirubin, GGT and SGOT in subsequent iterations. Conversely, the Exploratory agent adopts a comprehensive approach to feature selection, incorporating key markers, like Albumin, Bilirubin, SGOT, SGPT and ALP to survey a wider range of the available data space. The Parsimonious model selects minimal yet crucial features, frequently identifying key hepatic biomarkers, like SGOT, SGPT, ALP and Total Protein. Meanwhile, the Random agent’s selection of larger, highly variable feature sets typically including Total Protein, Albumin, SGOT,GGT, Bilirubin and SGPT- highlights its failure to adopt a systematic optimization strategy. Significantly, the frequent selection of various biochemical markers related to liver across multiple runs providing the evidence of their stability and clinical applicability of these features keeping up the interpretability of the proposed framework at feature-level. This stability recommends that the reinforcement learning agents are efficient for recognizing the clinically meaningful parameters linked with the Hepatitis C classification thereby enhancing the reliability of the selected biochemical parameters.

Table 13.

Feature selection per run.

S. no	Run	Greedy	Exploratory	Parsimonious	Random
1	Run 1	Total Protein	GGT, Bilirubin	Albumin, SGOT	Albumin, Globulin, Bilirubin, SGOT, GGT
2	Run 2	Total Protein	Albumin, SGOT	GGT, Bilirubin, SGOT, SGPT	Total Protein, Globulin, GGT
3	Run 3	Total Protein	Albumin, Bilirubin, SGOT, SGPT, ALP	SGOT, SGPT, ALP	Total Protein, Albumin, SGOT, GGT, Bilirubin
4	Run 4	Albumin, Bilirubin	Total Protein, SGPT, ALP	SGPT	Total Protein, Albumin, SGOT, GGT, Bilirubin, SGPT
5	Run 5	SGOT, Bilirubin	SGOT	Total Protein, Globulin, SGOT	Total Protein, ALP, SGOT, GGT, Bilirubin, SGPT
6	Run 6	GGT, Bilirubin	Bilirubin	ALP, SGOT, SGPT, Bilirubin	Total Protein, SGPT, GGT

Feature analysis section

Inspite of the fact that definite biochemical parameters are not picked for every experimental run, still their frequency of selection is high among multiple runs indicating strong applicability to the classification task. Due to the stochastic exploration phenomenon applied by the reinforcement learning agents, negligible variations are expected in feature selection. Reinforcement learning unlike deterministic approaches investigates different combinations of parameters to avoid local optima so as to find out multiple informative subsets. Various biomarkers contribute in the complementary diagnostic information in case of clinical datasets where unlike feature combinations achieve comparable predictive performance. Consequently, the selected frequency (0.8) for some biomarkers still depicts stable subset of clinically relevant variables for the assessment of Hepatitis C.

Limitations and future work

This research is subject to several constraints that warrant discussion.

a. Comparison with classical feature selection methods

One of the limitations of the current experimental study is that the conventional feature selection techniques like filter method, wrapper method and embedded method are not compared with the proposed RL agents. The major objective of this study is to understand the internal dynamics of the RL – driven feature selection approach rather than providing their comparison with the conventional feature selection methods. While these methods proved to be a well established baseline, taking into consideration the scope of our research, we have considered random agent as the baseline method in our research. In spite of this, to evaluate the absolute performance of MARL-based agents, conventional methods would provide an additional context and is acknowledged as a limitation of the current study.

b. Single-center dataset and generalizability

The scope of this study confined to one institution, which potentially limits the external validity of the findings.

Variations in population demographics, laboratory standards, and clinical workflows across different locations can impact biomarker and model performance. Consequently, the findings may be limited to the specific characteristics of the analyzed dataset. Therefore, the findings should be viewed as preliminary, requiring additional validation before the proposed model can be applied to larger, diverse patient populations. Future efforts will involve testing the framework on multi-site, heterogeneous populations to verify its stability for real-world healthcare deployment.

c. Experimental Runs

As our evaluation relies on a limited number of training runs, future studies should increase the number of independent runs to 30 or more.

d. Lack of shared reward mechanism in multi-agent learning

The introduced multi-agent reinforcement learning system utilizes individual agents, devoid of direct communication or shared rewards, thereby restricting the extent of cooperative decision-making. Future research will advance beyond independent agent operations by establishing genuine multi-agent collaboration, utilizing shared reward structures and explicit inter-agent communication to facilitate cooperative policy learning.

The model inadequately addresses the cost and availability of data, treating all laboratory parameters as uniformly accessible. In conclusion, despite encouraging preliminary results, the model lacks testing in real-world clinical settings, a critical step for fully assessing its operational readiness.

While our reinforcement learning approach shows promise for HCV diagnosis, we plan to extend the framework to overcome its current limitations.

Conclusion

For the detection of the Hepatitis C virus, this research has demonstrated that reinforcement learning is an efficient and streamlined methodology. By judiciously selecting the most impactful variables, the RL-driven system increased machine learning performance while simultaneously cutting computational costs. When using features selected by reinforcement learning, ensemble methods like Random Forest and XGBoost demonstrated superior performance, achieving strong accuracy, precision, F1-score. These results stress the promise of using RL-driven feature selection alongside machine learning algorithms to create reliable, scalable, interpretable diagnostic platforms. By reducing the dimension of the feature set, the model’s complexity is decreased, leading to more efficient training, better performance on new data, and a lower chance of overfitting. This highlights reinforcement learning as a valuable framework for healthcare analytics, especially in areas where relevant features are intricate and challenging to model by hand.

In conclusion, the RL-ML pipeline provides a highly effective, robust, and clear methodology for detecting HCV at an early stage. Future investigations are poised to incorporate more voluminous datasets, deep reinforcement learning, or hybrid optimization strategies to further boost the reliability and practical clinical use of such automated diagnostic systems.

Author contributions

Kanica Anand Arora conceived the research idea, supervised the study, and contributed to the overall project administration. Anuj Sharma designed the methodology, performed key analyses, and contributed to data interpretation. Deepak Prashar contributed to software development, data processing, and validation of experimental results. Amr Yousef assisted in model implementation, visualization, and technical refinement of the manuscript. Arfat Ahmad Khan contributed to formal analysis, writing the original draft, and enhancing the technical and scientific narrative of the manuscript. Zaffar Ahmed Shaikh provided critical review, editing, and contributed to improving the scholarly presentation and coherence of the final version.

Funding

The paper received no external funding.

Data availability

With appropriate institutional authorization for doctoral research, this study utilized datasets acquired from the Department of Microbiology at Government Medical College, Jammu (Ref No, GMCJ./Micro/2023–24/2105).The dataset is available on Zenodo. The Source code comprising preprocessing scripts, reinforcement learning environment definitions, agent configurations, and evaluation pipeline is publicly available at Github (https://github.com/kanicaanand/HCV-MARL-Feature-Selection/tree/main).

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

The Study was approved by Institutional Ethics Committee, Government Medical College (GMC), Bakshi Nagar, Jammu. This investigation utilized a retrospective analysis of hospital record, supplemented by telephonic patient follow-up data obtained during routine clinical care. Before analysis, all identifying information was erased, rendering the data completely anonymous and non-identifiable. The study was classified as minimal risk, involving no experimental treatments, procedures, or behavioral interventions. In accordance with institutional guidelines for de-identified retrospective data, the institutional Ethics Committee waived the requirement for informed consent. Data usage was strictly confined to research purposes, with no attempts made to re-identify participants.