以機器學習分析營養素對阿茲海默症影響之研究

Abstract

隨著人類壽命的延長，阿茲海默症已成為重要的文明病之一。若能透過日常營養攝取來預防阿茲海默症，不僅有助於降低醫療成本，亦具備操作簡單、容易落實的優點。本研究旨在運用機器學習技術，探討葉酸（Folate）、膽鹼（Choline）與甜菜鹼（Betaine）之攝取量與認知功能及失智風險的關聯性。方法上，本研究捨棄傳統的硬性填補（Hard Imputation），轉而利用 廣義相加模型(Generalized Additive Models, GAM）之平滑函數（Smoothing Functions） 捕捉營養素與認知功能間的非線性反應曲線。針對多維數據之缺失與交互作用，進一步導入張量積平滑 (Tensor Product Smooths, $te$)，藉由其優異的插值特性與對數據稀疏性（Sparsity）的適應力，在不損失樣本統計代表性的前提下，藉由平滑函數擬合數據之全局趨勢，有效緩解高比例缺失數據對分析結果的偏誤，精確刻畫葉酸、膽鹼與甜菜鹼間的協同效應，並結合空間密度分群演算法 (Density-Based Spatial Clustering of Applications with Noise, DBSCAN)，於 GAM 重建之連續曲面中識別出高認知保護力的營養攝取叢集（Clusters），最終針對不同性別與年齡族群，提出精準的營養攝取建議，以期達成預防神經退化並維護高齡大腦功能的目標。 傳統營養科學在探討營養素對阿茲海默症之影響時，多半採用單一營養素的統計檢定，較少同時綜合多種營養素，因而可能忽略特徵之間的交互作用與潛在的非線性關係，使分析結果較為片面。本研究以 GAM 模型處理營養素與阿茲海默症之間可能存在的非線性關係，並且能以視覺化方式呈現各營養素的邊際與交互效果，彌補了傳統營養科學上的不足。 實驗結果顯示，由本研究方法篩選出的建議攝取範圍內樣本，其認知表現普遍優於其他攝取範圍之樣本。並且本研究之三合一營養素組合的認知功能表現也勝過傳統單一營養素的營養素表現。總體而言，本研究提供了一套兼具可解釋性與實務操作性的範圍篩選模組，對營養素攝取建議與失智預防策略具有高度應用價值。

As human life expectancy increases, Alzheimer’s disease (AD) has emerged as a significant disease of civilization. Utilizing daily nutritional intake to prevent AD not only helps reduce healthcare costs but also offers the advantages of being easy to implement and maintain. This study aims to apply machine learning techniques to investigate the association between the intake of folate, choline, and betaine and their impact on cognitive function and dementia risk. Methodologically, this research eschews traditional hard imputation methods, which often introduce bias. Instead, it leverages the smoothing functions of Generalized Additive Models (GAM) to capture the non-linear response curves between nutrient intake and cognitive performance. To address missing data and complex interactions in multidimensional datasets, we further incorporate tensor product smooths ($te$). By utilizing its superior interpolation properties and adaptability to data sparsity, we fit the global trends of the data without compromising statistical representativeness, thereby mitigating the bias caused by high missing data rates. This approach allows for a precise characterization of the synergistic effects among folate, choline, and betaine. Subsequently, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is applied to the continuous surfaces reconstructed by the GAM to identify nutritional intake clusters associated with high cognitive protection. Ultimately, this study provides tailored nutritional intake recommendations for different genders and age groups, aiming to prevent neurodegeneration and maintain cognitive health in the elderly population. Traditional nutritional epidemiology typically examines the effect of a single nutrient on Alzheimer's disease using classical statistical tests, and rarely analyzes multiple nutrients jointly. As a result, potential interactions between nutrients and nonlinear relationships may be overlooked, filling a significant gap in conventional nutritional research. In contrast, the GAM framework used in this study can flexibly capture possible nonlinear relationships between nutrient intake and Alzheimer's disease, while its interpretability and visualizability allow the marginal and interaction effects of each nutrient to be presented in an intuitive way, facilitating understanding and support from the medical and nutritional sciences. The experimental results indicate that subjects within the recommended intake ranges identified by this methodology generally exhibited superior cognitive performance compared to those outside these ranges. Furthermore, the triple-nutrient combination proposed in this study demonstrated superior cognitive outcomes compared to traditional single-nutrient approaches. Overall, this study presents a range selection framework that is both interpretable and practically viable, offering high application value for formulating nutrient intake recommendations and dementia prevention strategies