Federated Learning for Cognitive Impairment Detection Using Speech Data

Josep Blazquez-Folch¹María Limones Andrade²Berta Calm¹Juan Miguel Auñón García²MONTSERRAT ALEGRET^1,3Nathalia Muñoz¹Amanda Cano¹Victoria Fernández¹Fernando García-Gutiérrez¹Itziar De Rojas^1,3Pablo Garcia Gonzalez¹Clàudia Olivé¹Raquel Puerta¹María Capdevila-Bayo¹Álvaro Muñoz-Morales¹Paula Bayón-Buján¹Andrea Miguel Romero¹Laura Montrreal Navarro¹Ana Espinosa^1,3Pilar Sanz-Cartagena¹Maitee Rosende-Roca Rodriguez¹Carla Zaldua⁴Peru Gabirondo⁴Yahveth Cantero-Fortiz^1,3Miren Jone Gurruchaga¹Lluís Tárraga Mestre^1,3Merce Boada Rovira^1,3Agustín Ruiz Laza^1,3,5,6Marta Marquié^1,3Sergi Valero^1,3*

• ¹Ace Alzheimer Center Barcelona - Universitat Internacional de Catalunya, Barcelona, Spain
• ²Department of Artificial Intelligence and Big Data, GMV, Madrid, Spain
• ³Networking Research Center on Neurodegenerative Diseases (CIBERNED), Instituto de Salud Carlos III, Madrid, Spain
• ⁴Accexible Impacto s.l., Urduliz, Spain
• ⁵Glenn Biggs Institute for Alzheimer’s & Neurodegenerative Diseases, University of Texas Health Science Center, San Antonio, Spain
• ⁶Department of Microbiology, Immunology and Molecular Genetics., Long School of Medicine. University of Texas Health Science Center, San Antonio, Spain

ABSTRACT Introduction: In Alzheimer's disease (AD) research, clinical, neuroimaging, genetic, and biomarker data are vital for advancing its understanding and treatment. However, privacy concerns and limited datasets complicate data sharing. Federated learning (FL) offers a solution by enabling collaborative research while preserving data privacy. Methods: This study analyzed data from patients assessed at the Memory Unit of the Ace Alzheimer Center Barcelona who completed a standardized digital speech protocol. Acoustic features extracted from these recordings were used to distinguish between cognitively unimpaired (CU) and cognitively impaired (CI) individuals. The aim was to evaluate how data heterogeneity impacted the FL model performance across three scenarios: (1) equal contributions and class ratios, (2) unequal contributions, and (3) imbalanced class ratios. In each scenario, the performance of local models trained using an MLP feed-forward neural network on institutional data was analyzed and compared to a global model created by aggregating these local models using Federated Averaging (FedAvg) and Iterative Data Aggregation (IDA). Results: The cohort included 2,239 participants: 221 CU individuals (mean age 66.8, 64.7% female) and 2,018 CI subjects, comprising 1,219 with mild cognitive impairment (mean age 74.3, 61.9% female) and 799 with mild AD dementia (mean age 80.8, 64.8% female). In scenarios 1 and 3, FL provided modest gains in accuracy and AUC. In scenario 2, FL markedly improved performance for the smaller dataset (balanced accuracy rising from 0.51 to 0.80) while preserving 0.86 accuracy in the larger dataset, highlighting scalability across heterogeneous conditions.In scenarios 1 and 3, FL application resulted in slight accuracy improvements. In scenario 2, FL enhanced model accuracy for the institution with a smaller dataset (improving from 0.61 to 0.86) while maintaining 0.86 accuracy for the institution with a larger dataset, demonstrating scalability for detecting cognitive impairment in different scenarios. Conclusion: These findings demonstrate the potential of FL to enable collaborative modeling of speech-based biomarkers for cognitive impairment detection, even under conditions of data imbalance and institutional disparity. This work highlights FL as a scalable and privacy-preserving approach for advancing digital health research in neurodegenerative diseases.