基於人工智慧與基因資訊之疾病風險預測 － 以紅斑性狼瘡為例

Abstract

精準醫療的發展促進電子健康紀錄與基因資訊的整合，核心目標在於提升疾病風險預測的準確性。然而，傳統全基因組關聯研究多採單變量分析，難以捕捉基因變異間的複雜交互作用，且現有預測模型亦無法充分解釋個體間疾病進展的差異。本研究聚焦於系統性紅斑性狼瘡及其併發之狼瘡性腎炎，創新性地結合人工智慧與統計方法，建構整合性的疾病進展風險預測模型。研究整合電子健康紀錄與全基因組關聯研究資料，提出多種多基因風險評分建構策略，並應用機器學習技術以篩選具預測潛力之關鍵特徵。模型核心採用集成式學習演算法，能有效處理基因變異間的連鎖不平衡問題，並納入非線性關係與交互作用的考量，以提升預測表現。研究結果顯示，納入基因變異間交互作用有助提升模型準確性，抗核抗體滴度作為系統性紅斑性狼瘡風險分層之重要臨床指標，所建模型亦辨識出具潛在意義之關鍵基因特徵，對個體化風險評估具高度貢獻。結合可解釋性人工智慧技術，有效提升模型在臨床決策中的可信度，進一步強化其於個人化醫療上的應用潛力。本研究突破傳統多基因風險評分建構框架，展現機器學習於高維度基因變異資料處理與模型解釋層面的優勢，提供自體免疫疾病風險預測之創新方法，並突顯臨床與基因資料整合於早期預防與個人化醫療中的實務價值。

The advancement of precision medicine has facilitated the integration of electronic health records (EHRs) and genomic data to improve disease risk prediction. However, traditional genome-wide association studies (GWAS) often rely on univariate analyses, making capturing complex interactions among genetic variants challenging. Existing models also fall short of explaining individual differences in disease progression. This study focuses on systemic lupus erythematosus (SLE) and its complication, lupus nephritis. This innovative approach integrates artificial intelligence (AI) and statistical methods to construct an integrated model for predicting disease progression risk. By integrating EHRs with GWAS data, we propose several strategies for constructing polygenic risk scores (PRS) and applying machine learning (ML) to identify key features with predictive potential. The core model uses an ensemble algorithm that effectively handles linkage disequilibrium and captures non-linear relationships and interaction effects to enhance prediction. The results indicate that incorporating gene-gene interactions may improve model accuracy. Antinuclear antibody titers serve as key clinical indicators for risk stratification in SLE. The proposed model also identifies potentially significant genetic features contributing to individualized risk assessment. Incorporating explainable AI techniques further enhances the model’s credibility in clinical decision-making and the potential for personalized medicine. This study goes beyond traditional PRS frameworks, highlighting the advantages of ML in handling high-dimensional genomic data and enhancing interpretability. It provides an innovative approach for predicting autoimmune disease risk and highlights the value of integrating clinical and genomic data for early prevention and personalized healthcare.