Forecasting Hospital Mental-Health Admissions with a Novel Hybrid Deep Learning Architecture

Forecasting hospital admissions, especially for mental health patients, is crucial for efficient resource allocation planning. This study used multivariate time-series data to forecast hospital admissions for mental health patients at the Thunder Bay Regional Health Sciences Centre in Ontario, Canada. The main contribution of this work is proposing a novel deep learning architecture, the Stacked-dilated-causal Convolutional Neural Network and Bidirectional Long Short-Term Memory (SCNN-BiLSTM), which outperformed other statistical and machine learning models. The hybrid architecture provides full history-coverage of the input window and maintains the causal structure of the input time series, making it suitable for long-term forecasting. For univariate one-week forecasting, among the commonly used methods, CNN-BiLSTM achieves the minimum Mean Absolute Percentage Error (MAPE) score (13.8). However, for 4-week forecasting, the MAPE score increases to 14.78. Modelling multivariate data reduces this score to 13.33. The multivariate model trained with the proposed architecture performs the best by achieving a MAPE score of 12.96 in 4-weeks forecasts. Training multivariate models using the SCNN-BiLSTM architecture developed in this study may be potentially utilized in other health forecasting domains to achieve improved forecasting accuracy.