整合基因多型性和大腦磁振造影進行阿茲海默症的分類和高風險族群的識別

Abstract

阿茲海默症 (Alzheimer’s Disease, AD) 是老年人最常見的神經退化性疾病之一，影響全世界約 65 歲以上人口的 4%。在臨床診斷過程中，遺傳因子與腦部醫學影像皆為關鍵參考依據。本研究提出一個結合遺傳資訊與大腦的結構性磁振造影 (structural MRI, sMR) 的雙階段多模態分類模型，分別利用卷積神經網路(Convolutional Neural Network, CNN) 對大腦影像進行特徵萃取，並採用eXtreme Gradient Boosting (XGBoost) 進行分類。 本研究使用Alzheimer’s Disease Neuroimaging Initiative Phase 1 (ADNI 1) 資料，共納入 421 位受試者 (其中 AD 病患 192 位，認知正常對照組 229 位)。分析資料包括 620,901 個單核苷酸多型性 (Single Nucleotide Polymorphism, SNP) 標記、1,450 張標準化的 1.5 T 腦部 MRI 影像，以及人口統計變項 (年齡、性別與家族病史)。模型效能的評估是基於5-fold cross validation和以受試者為依據的資料切分 (Subject-level data split)，並以 AUROC作為主要評估指標。實驗的結果顯示，整合遺傳與腦影像特徵的模型達到最佳 AUROC 為 0.91，優於僅使用遺傳資訊 (AUROC = 0.86) 或僅使用影像特徵 (AUROC = 0.88) 之模型。 此外，我們分別以多基因風險分數 (Polygenic Risk Score, PRS) 與整合風險分數 (Integrated Risk Score, IRS) 進行高風險子群的識別，成功辨識AD高風險之子族群。最後，針對資料洩漏問題的分析顯示，當洩漏程度增加時模型效能亦會被高估，顯示在醫學影像相關研究中防止資料洩漏的重要性。 本研究的結果顯示，結合基因體與腦部影像資料進行整合式分析，對於阿茲海默症之高風險族群之識別與精準健康之推動具有高度應用潛力。

Alzheimer’s Disease (AD) affects approximately 4% of the population above 65 years old. Both genetic factors and medical images play a major role in providing information during the diagnosis process for the patient. In this study, we propose a two-stage multimodal classification model based on the genetics and structural MRI data using the Convolutional Neural Network (CNN) and eXtreme Gradient Boosting (XGBoost) models. The study was conducted based on 421 subjects (192 AD cases and 229 non-AD controls) from the Alzheimer’s Disease Neuroimaging Initiative Phase 1 (ADNI 1). Our analysis included 620,901 single nucleotide polymorphism (SNP) markers from the Illumina Human610-Quad BeadChip genotyping platform, 1,450 1.5T brain structure magnetic resonance imaging (MRI) from ADNI 1 Standardized MRI Collections, and demographic variables, including age, sex, and AD family history. The performance was evaluated using AUROC (Area Under the Receiver Operating Characteristic Curve) on a 5-fold cross-validation with a subject-level data split. The result showed that the model integrating genetics and brain MRI features achieved a best AUROC of 0.91, outperforming the models considering only genetics (AUROC = 0.81) or brain MRI features (AUROC = 0.88). Furthermore, the high-risk subgroup was identified using the Polygenic Risk Score (PRS) and the Integrated Risk Score (IRS) calculated from SNP data and MRI-extracted features. Last, the analysis of the data leakage problem reveals an inflation of model performance as the severity of leakage increases, highlighting the importance of data leakage prevention in medical image-related studies. The integrative analysis of genomics and brain MRI has great potential to facilitate the identification of high-risk populations and accelerate the advancement of precision health in AD.