A collaborative multi-view fusion framework with Deep Forest for multi-class diagnosis of Alzheimer's disease.

Objective: A key aspect of diagnosing Alzheimer's disease (AD) is identifying mild cognitive impairment (MCI), a sensitive transitional stage between normal cognition (NC) and AD. We divided AD disease progression into three stages: NC, MCI, and AD according to severity. We used CoopLearning to integrate multi-view medical data with a Deep Forest (DF) classifier to classify three stages of AD to aid diagnosis. Methods: We used data from the Alzheimer's Disease Neuroimaging Initiative database. The research variables were divided into five views: demographic information, biological sample information, relevant scale information, pixel identification, and magnetic resonance imaging regional information. We compared CoopLearning for feature selection and fusion, least absolute shrinkage and selection operator cross-validation (LassoCV) for single-view feature selection, and LassoCV for multi-view feature selection. The fused data were processed using Synthetic Minority Oversampling Technique (SMOTE), SMOTETomek, and SMOTE Edited Nearest Neighbors (ENN) methods to achieve class balance. The balanced data were then applied to multiple classifiers, such as DF, Random Forest, etc., for performance evaluation. Results: Multi-view feature selection and fusion using CoopLearning, multi-view feature selection, and single-view overall feature selection using LassoCV identified 20 features from five views. Applying these results to the classifier revealed that model performance was optimal when using SMOTEENN in conjunction with CoopLearning multi-view feature selection and the DF classifier (Accuracy = 0.9898, multi-class area under the receiver operating characteristic curve = 0.9997). Conclusion: Collaborative multi-view fusion substantially improved direct multi-class classification across the AD continuum. Integrating clinical and neuroimaging data with Deep Forest may provide a robust and interpretable framework for early auxiliary diagnosis and disease-stage stratification in Alzheimer's disease.