Federated Learning for Mental Health Disorder Detection: A CNN-LSTM Approach with Non-IID Data

Marron G.A.C.; Mamani-Aliaga A.H.; Bustincio R.W.C. · Proceedings - International Conference of the Chilean Computer Science Society, SCCC (2025) · DOI: 10.1109/sccc67219.2025.11420615

Resumen

Traditional centralized approaches for NLP-based (Natural Language Processing) mental-health monitoring pose significant privacy risks because sensitive user data must be stored on central servers. To address this gap, we propose a Federated Learning (FL) framework applied to NLP sentiment-analysis models - i.e., training across decentralized data silos where only model updates, not raw data, are shared. The model combines a CNN-LSTM architecture with BERT embeddings and is trained on clinically annotated tweets distributed in a non-IID manner across clients. Each client performs local training of the hybrid model, initialized with pre-trained BERT weights, while a central server aggregates updates using Federated Averaging (FedAvg), aiming to identify indicators (not diagnoses) of depression, stress, bipolar disorder, and anxiety without centralizing raw data. Evaluated with accuracy, macro F1, and AUROC, the federated model achieves competitive performance relative to centralized baselines, reflecting a justified privacy-utility trade-off. Overall, this work integrates advanced NLP techniques with privacy-preserving training and ethical data handling, offering a privacy-conscious solution for monitoring mental-health signals on social media.

Instituciones: Universidad Nacional de San Agustín de Arequipa