GeneRisk: An Explainable, Privacy-Preserving Framework for Multi-Disease Polygenic Risk Assessment and Evidence-Grounded Clinical Guidance

Abstract

While high-throughput genomic profiling provides unprecedented resolution into inherited polygenic vulnerability, clinical translation remains constrained by three pervasive barriers: high computational barriers to client-side Variant Call Format (VCF) analysis, the black-box nature of complex machine learning classifiers, and severe privacy risks tied to unencrypted cloud transmission of raw genomic sequence data. In this paper, we present GeneRisk, an end-to-end, privacy-preserving computational intelligence platform for explainable multi-disease risk assessment. GeneRisk implements a zero-disk client-side streaming parser capable of extracting clinically validated single-nucleotide polymorphisms (SNPs) across a 250 kb linkage disequilibrium (LD) clumping window at genome-wide significance (p < 5 × 10??). These genomic factors are integrated with patient lifestyle, demographic, and biometric parameters via a calibrated XGBoost late-fusion ensemble. Global and instance-level interpretability is established using Game-Theoretic TreeSHAP, decomposing individual multi-morbidity projections into exact additive log-odds attributions and proportional heritability splits. To bridge quantitative risk scores with clinical decision-making, an evidence-grounded Retrieval-Augmented Generation (RAG) pipeline links dense vector embeddings of clinical guidelines (AHA, ADA, NCCN, CPIC) via Qdrant to an optimized multi-model inference engine. Evaluated across N = 5,000 patient profiles across five major clinical panels, the framework demonstrated robust discriminatory power, achieving 5-fold cross-validation ROC-AUCs of 0.7113 ± 0.0131 for Type 2 Diabetes, 0.6918 ± 0.0361 for Alzheimer\'s Disease, and 0.6916 ± 0.0258 for Coronary Artery Disease. In live clinical benchmark trials, the RAG architecture achieved an empirically measured Overall Groundedness of 0.9053 ± 0.0306 (\'Very Good\' grade), 100% citation accuracy, and 91.25% numerical consistency. By eliminating raw genomic data uploads and providing actionable, auditable explanations, GeneRisk establishes a robust, privacy-compliant foundation for genomic intelligence in primary healthcare.

Introduction

GeneRisk is an explainable and privacy-preserving precision-medicine framework designed to translate genome-wide association study (GWAS) findings into practical disease-risk assessment. It addresses three major challenges in genomic healthcare: the interaction between genetic and lifestyle factors, the lack of transparency in machine-learning predictions, and the privacy risks associated with uploading large genomic files to cloud servers.

The proposed system combines Polygenic Risk Scores (PRS), multimodal machine learning, explainable artificial intelligence, and Retrieval-Augmented Generation (RAG) to estimate disease risk and provide evidence-based clinical guidance.

The system follows a three-tier architecture:

1. Client-side preprocessing: A WebAssembly/TypeScript-based parser processes large Variant Call Format (VCF) files in browser memory and extracts selected genetic variants. According to the proposed design, raw genomic files are not uploaded or persistently stored; only a reduced vector of genetic variant information is sent to the analytical backend.

2. Core analytical engine: The system calculates PRS using linkage disequilibrium clumping and thresholding, then combines genetic scores with demographic and lifestyle factors such as age, body mass index, blood pressure, smoking, physical activity, diet, and sleep. An XGBoost model analyzes these features to estimate disease risk.

3. Explainability and clinical guidance: TreeSHAP identifies the contributions of individual genetic and non-genetic factors to each prediction. A RAG pipeline retrieves relevant clinical guidelines to generate evidence-grounded explanations and recommendations. The proposed architecture also uses temporary in-memory processing for report generation to reduce persistent storage of sensitive information.

The mathematical methodology includes PRS calculation using genetic variant dosages and GWAS effect sizes, population standardization, and regularized gradient-boosted decision trees. SHAP explanations decompose model predictions into individual feature contributions, helping users understand which factors influence estimated risk.

The experimental evaluation uses 5,000 simulated patient profiles per disease panel, divided into 4,000 training samples and 1,000 held-out test samples. Model performance is assessed using accuracy, ROC-AUC, five-fold cross-validation, Brier score, and sensitivity. These metrics evaluate discrimination, prediction quality, and the ability to identify individuals at elevated risk.

Conclusion

GeneRisk provides a comprehensive, privacy-preserving, and explainable computational intelligence architecture that successfully unifies polygenic risk scoring, lifestyle risk factors, machine learning late-fusion, game-theoretic explainability, and guideline-grounded retrieval-augmented generation. The system demonstrates robust discrimination across five major chronic diseases, establishes perfect citation precision, and operates with zero disk leakage of sensitive genomic sequence data. This work bridges statistical genetics with preventive clinical workflows, paving the way for trustworthy, accessible precision healthcare.

References

[1] C. M. Lewis and E. Vassos, \'Polygenic risk scores: from research tools to clinical instruments,\' Genome Medicine, vol. 12, no. 1, p. 44, 2020. [2] A. V. Khera et al., \'Genetic risk, adherence to a healthy lifestyle, and coronary disease,\' New England Journal of Medicine, vol. 375, no. 24, pp. 2349–2358, 2016. [3] A. Ahmed et al., \'A comparative analysis of LIME and SHAP interpreters with explainable ML-based diabetes predictions,\' IEEE Access, vol. 13, pp. 37370–37388, 2025. [4] S. M. Lundberg and S. I. Lee, \'A unified approach to interpreting model predictions,\' in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017. [5] B. Dibaeinia et al., \'PRSformer: Disease prediction from million-scale individual genotypes,\' in Advances in Neural Information Processing Systems (NeurIPS), 2025. [6] A. Raghavendra et al., \'Explainable multimodal fusion of genomic and clinical data for multi-disease prediction: A deep learning approach,\' Journal of Computing and Biomedical Informatics, vol. 10, no. 2, 2026. [7] M. Khater et al., \'Explainable machine learning model for Alzheimer detection using genetic data: A genome-wide association study approach,\' IEEE Access, vol. 12, pp. 95091–95105, 2024. [8] Y. Lim et al., \'Robust SNP-based prediction of rheumatoid arthritis through machine-learning-optimized polygenic risk score,\' Journal of Translational Medicine, vol. 21, no. 92, 2023. [9] A. Alireza et al., \'Enhancing prediction accuracy of coronary artery disease through machine learning-driven genomic variant selection,\' Journal of Translational Medicine, vol. 22, no. 356, 2024. [10] S. Mashmool et al., \'Edge computing in healthcare using machine learning: A systematic literature review,\' WIREs Data Mining and Knowledge Discovery, vol. 16, no. 1, p. e70069, 2026. [11] American Diabetes Association, \'Standards of Care in Diabetes—2024,\' Diabetes Care, vol. 47, no. Suppl 1, pp. S1–S343, 2024. [12] Clinical Pharmacogenetics Implementation Consortium (CPIC), \'CPIC Guidelines for Dosing and Variant Interpretation,\' Clin. Pharmacol. Ther., 2024.

Copyright

Copyright © 2026 Mr. Sathyanarayan S, Rahul K S, Likhith T S, Prekshith K L, Mohammed Saffi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.