Enhancing Risk Prediction in Mental Health Using Ensemble Hybrid Models and Administrative Healthcare Data with Irregular Intervals

Risk prediction estimates the probability of future adverse outcomes for high-risk individuals to enable early intervention. Among all machine learning models, deep learning models can enhance risk predictions with respect to mental health (MH), where administrative healthcare data (Admin-HD) contain complex, irregular temporal information that a single model may not fully capture. This study aims to develop an ensemble hybrid deep-learning model (EnH-DL) to improve performance and reduce errors on unseen data compared to available deep-learning models. The model was evaluated in a cohort of individuals diagnosed with addiction or mental illness (AMH) to predict the risk of outcomes, such as the first healthcare encounter indicating homelessness (FHE-H) and the first healthcare encounters involving police (FHE-P). A sliding window and matrix-based representation were used in the preprocessing phase. Gated recurrent units (GRUs) with time decay, a one-dimensional (1D) convolutional neural network (CNN), and an attention function were integrated to build the EnH-DL. The results showed an average improvement in the area under the curve (AUC) by 2.5% and sensitivity by over 1.5%. An error reduction of over 0.22 indicates improved model reliability for unseen data. Tested in a clinical demand window to simulate real-world settings, the model achieved an 83% AUC and 79% sensitivity for FHE-H on a highly imbalanced test set. In conclusion, EnH-DL outperformed existing models for multiple MH outcomes using Admin-HD.