완전동형암호화를 활용한 질병감수성 예측 기계 학습모델 성능평가

Abstract

The activation of polygenic risk score (PRS) in precision medicine has raised concerns about privacy pres- ervation. Current encryption methods that delete or modify variant information may reduce prediction accuracy. The purpose of this study was to evaluate the feasibility of PRS-based disease susceptibility prediction in a fully homo- morphic encryption (FHE) environment while assessing computational trade-offs. We conducted disease susceptibility prediction using synthetic public genome-wide association study (GWAS) data on coronary artery disease (CAD) and applied inverse probability weighting, k-fold cross-validation, and linkage disequilibrium (LD) shrinkage to enhance polygenic risk score stability and accuracy. Four machine learning models—PRS only logit, ridge regression, linear support vector machine, and radial basis function support vector machine— were trained and tested on fully encrypted genomic and clinical data, with performance evaluated using area under the curve (AUC), mean absolute error (MAE), and turnaround time. A total of 1,401 samples and 404,663 SNPs were used. Ridge regression balanced prediction accuracy and computational efficiency (AUC = 0.7631, MAE = 0.4375, Turnaround Time = 1,343.65s). RBF SVM achieved the highest AUC (0.7792, MAE = 0.3120) but required the longest time (1,983.57s). PRS Only Logit had the shortest time (279.27s) but the lowest AUC (0.6339, MAE = 0.4162). These results suggest ridge regression may have the best balance between prediction accuracy and computational efficiency, making it the most practical choice in an FHE environment.