Federated learning for cognitive impairment detection using speech data

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Alzheimer's disease (AD) research requires diverse data, but privacy concerns hinder sharing.
• Federated learning (FL) offers a privacy-preserving solution for collaborative data analysis.

Purpose of the Study:

• To evaluate the impact of data heterogeneity on FL model performance for cognitive impairment detection using speech biomarkers.
• To compare FL performance against local models under various data distribution scenarios.

Main Methods:

• Acoustic features from digital speech recordings of 2,239 participants (cognitively unimpaired and impaired) were analyzed.
• Federated Averaging (FedAvg) and Iterative Data Aggregation (IDA) were used to aggregate local models trained with MLP feed-forward neural networks.
• Model performance was assessed across scenarios with equal/unequal contributions and imbalanced class ratios.

Main Results:

• FL showed modest accuracy and AUC gains in scenarios with equal contributions and imbalanced classes.
• In scenarios with unequal contributions, FL significantly improved performance on smaller datasets (balanced accuracy from 0.51 to 0.80) while maintaining high accuracy on larger datasets.
• FL demonstrated scalability and effectiveness across heterogeneous institutional data.

Conclusions:

• Federated learning enables collaborative development of speech-based biomarkers for cognitive impairment detection.
• FL is a scalable, privacy-preserving approach suitable for digital health research in neurodegenerative diseases, addressing data imbalance and institutional disparities.