Explainable AI Reveals Clinical Subtypes and Key Predictors of Alzheimer’s Disease: A Systematic Approach Combining Unsupervised Clustering and Machine Learning (Preprint)

<sec> <title>BACKGROUND</title> Alzheimer's Disease (AD) is a neurodegenerative disorder, making the identification of risk factors and patient subtypes crucial for early intervention. However, the complexity of clinical data and the "black-box" nature of many predictive models pose significant challenges. This study introduces a systematic framework to overcome these hurdles by integrating machine learning with explainable AI. The primary objective of this study is to overcome the challenges posed by the complexity of clinical data and the "black-box" nature of predictive models in Alzheimer's Disease (AD) research. To achieve this, the study introduces a systematic framework that integrates unsupervised clustering to identify distinct patient subtypes, a high-performance machine learning model to predict AD diagnosis, and an explainable AI technique (SHAP) to interpret the model's predictions. The ultimate goal is to provide a transparent and reliable method for identifying key risk factors, understanding their impact, and advancing personalized risk assessment and intervention strategies for AD. This study employed a multi-stage analytical approach on a public clinical dataset of 2,149 individuals. First, K-Means clustering, an unsupervised learning algorithm, was to identify distinct patient subtypes based on their comprehensive clinical profiles. Second, a high-performance XGBoost model was trained to predict AD diagnosis. Finally, SHAP (Shapley Additive Explanations), a leading explainable AI technique, was applied to interpret the model's predictions, quantify the importance of each clinical feature, and understand their directional impact. The XGBoost model achieved high predictive accuracy (95.12%) and an AUC of 0.9480 on the test set. SHAP analysis identified the five most influential predictors for AD diagnosis: Functional Assessment, Activities of Daily Living (ADL), Memory Complaints, MMSE score, and Behavioral Problems. Furthermore, our clustering analysis successfully stratified the patient population into seven distinct clinical subtypes. Notably, a "Memory-Complaint Dominant" subtype (Cluster 6) was identified with the highest AD incidence rate (65.1%), while a relatively asymptomatic subtype (Cluster 5) had the lowest rate (29.2%). This integrated framework effectively stratifies AD patients and identifies robust, interpretable predictors of the disease. By combining unsupervised clustering with explainable machine learning, this study provides a paradigm for advancing personalized risk assessment and targeted intervention strategies in Alzheimer's Disease. </sec> <sec> <title>OBJECTIVE</title> The primary objective of this study is to overcome the challenges posed by the complexity of clinical data and the "black-box" nature of predictive models in Alzheimer's Disease (AD) research. To achieve this, the study introduces a systematic framework that integrates unsupervised clustering to identify distinct patient subtypes, a high-performance machine learning model to predict AD diagnosis, and an explainable AI technique (SHAP) to interpret the model's predictions. The ultimate goal is to provide a transparent and reliable method for identifying key risk factors, understanding their impact, and advancing personalized risk assessment and intervention strategies for AD. </sec> <sec> <title>METHODS</title> This study employed a multi-stage analytical approach on a public clinical dataset of 2,149 individuals. First, K-Means clustering, an unsupervised learning algorithm, was to identify distinct patient subtypes based on their comprehensive clinical profiles. Second, a high-performance XGBoost model was trained to predict AD diagnosis. Finally, SHAP (Shapley Additive Explanations), a leading explainable AI technique, was applied to interpret the model's predictions, quantify the importance of each clinical feature, and understand their directional impact. </sec> <sec> <title>RESULTS</title> The XGBoost model achieved high predictive accuracy (95.12%) and an AUC of 0.9480 on the test set. SHAP analysis identified the five most influential predictors for AD diagnosis: Functional Assessment, Activities of Daily Living (ADL), Memory Complaints, MMSE score, and Behavioral Problems. Furthermore, our clustering analysis successfully stratified the patient population into seven distinct clinical subtypes. Notably, a "Memory-Complaint Dominant" subtype (Cluster 6) was identified with the highest AD incidence rate (65.1%), while a relatively asymptomatic subtype (Cluster 5) had the lowest rate (29.2%). </sec> <sec> <title>CONCLUSIONS</title> This integrated framework effectively stratifies AD patients and identifies robust, interpretable predictors of the disease. By combining unsupervised clustering with explainable machine learning, this study provides a paradigm for advancing personalized risk assessment and targeted intervention strategies in Alzheimer's Disease. </sec>