透過機器學習結合多重基因風險指數預測心肌病變

Abstract

本研究旨在運用機器學習技術，結合多基因風險分數，預測心肌病的發生。我們採用臺灣大學醫學院附設醫院（NTUH）和臺灣生物資料庫（TWB）的資料集，首先進行全基因組關聯研究（GWAS），以確定單核苷酸多態性（SNPs）、二元特徵和年齡之間的相關性。隨後在多基因風險分數（PRS）分析中，我們從發現性GWAS中獲取具體權重（連續特徵的β值和二元特徵的對數比率）。計算目標樣本中所有個體的PRS後，這些分數可以應用於邏輯回歸分析中，預測與感興趣特徵有遺傳重疊的特徵。我們使用先進的機器學習模型和交叉驗證技術，評估NTUH和TWB數據集中預測心肌病發展的準確性。在評估中，我們考慮了多種心肌病特徵和預測因素，包括PRS、作為潛在危險因素的臨床參數和ICD-10以及ICD-10-CM。

The main goal of this study is to utilize machine learning techniques to combine polygenic risk scores and predict the occurrence of cardiomyopathy disease. To achieve this, we employ datasets from the National Taiwan University Hospital and Taiwan Biobank and conduct initial genome-wide association studies to identify correlations between single nucleotide polymorphisms and phenotype [4]. Afterwards, for the analysis of polygenic risk scores, specific weights are derived from discovery genome-wide association studies. These weights are then used to calculate the polygenic risk scores for all individuals in the target sample. These scores can be utilized in a firth regression analysis to predict phenotype that are expected to have genetic overlap with the specific trait of interest, i.e., cardiomyopathy [4]. To evaluate the accuracy of predicting cardiomyopathy development, we use cutting-edge machine learning models and cross-validation techniques on both the National Taiwan University Hospital and Taiwan Biobank datasets. In our evaluation, we take into account various cardiomyopathy features and predictors, including polygenic risk scores, clinical parameters as potential risk factors, as well as ICD-10 and ICD-10-CM codes [9].