Abstract
Mental health diagnostics, an evolving domain in healthcare, aims to enhance early detection and intervention strategies. The integration of artificial intelligence (AI) and machine learning has facilitated advanced predictive models, but concerns related to data privacy and security remain significant challenges. Traditional centralised data collection methods expose sensitive user information, limiting large-scale adoption in real-world applications. This study addresses these challenges by utilising a multimodal dataset comprising physiological signals (heart rate variability, sleep patterns) and behavioural data (online activity, social media interactions). The proposed framework incorporates a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network within a federated learning environment, ensuring that raw user data remains decentralised and privacy is preserved using differential privacy and encryption techniques. The novelty of this research lies in achieving real-time mental health diagnostics while maintaining stringent privacy safeguards. Experimental results demonstrate that the proposed model achieves an accuracy of 92%, an F1-score of 0.905, and an AUC-ROC score of 0.90, substantially outperforming baseline federated learning models. Furthermore, the model shows resilience against data heterogeneity and privacy threats, making it a promising solution for secure, scalable, and ethical remote mental health assessments.