FLAME: Federated Learning and Aggregated Multi-Model Ensemble for Multi-Class Alzheimer's Disease Stage...

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Medical Informatics

Background:

• Accurate Alzheimer's disease (AD) staging is vital for timely intervention and treatment.
• Clinical classification is challenging due to overlapping cognitive profiles across disease stages.
• Existing methods lack privacy-preserving, multi-institutional capabilities for robust AD staging.

Purpose of the Study:

• To develop and evaluate FLAME, an advanced diagnostic system using federated learning for privacy-preserving AD stage classification.
• To systematically review machine learning (ML) and deep learning (DL) models for classifying five AD stages: cognitively normal (CN), subjective memory complaints (SMC), early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI), and AD.
• To optimize a hybrid ensemble model using a Genetic Algorithm (GA) for enhanced classification performance.

Main Methods:

• Evaluated 16 ML and 11 DL models on a 362-feature clinical dataset.
• Developed a GA-optimized hybrid ensemble combining LightGBM and a 5-layer DNN.
• Implemented a federated learning architecture with four aggregation algorithms (FedAvg, FedProx, FedNova, SCAFFOLD) for privacy-preserving multi-institutional deployment.

Main Results:

• FT-Transformer achieved the highest standalone DL performance (accuracy=0.7810, κ=0.7081).
• LightGBM showed superior ML performance (accuracy=0.8156, κ=0.7537).
• The GA-optimized hybrid ensemble (w=0.685) outperformed individual models. FedNova hybrid design achieved the best federated performance (accuracy=0.8213, κ=0.7614).

Conclusions:

• Federated learning combined with evolutionary ensemble optimization offers a scalable, privacy-preserving solution for AD stage classification.
• FLAME provides a clinically viable framework for multi-institutional decision-support systems.
• Severe under-recall of the AD class (F1 ≤ 0.21) remains the primary challenge for clinical translation.