Forecasting Psychiatric Inpatient Admissions Using Bi-Directional Long Short-Term Memory (Bi-LSTM) with Dynamic Regularization

Accurate forecasting of psychiatric inpatient admissions is crucial for allocating healthcare resources, scheduling staff, and maintaining continuity of care. However, the time-series data in this domain are frequently multivariate, non-linear, and have long-term relationships, rendering typical forecasting approaches ineffective. Traditional models such as linear regression and autoregressive models have been widely employed for healthcare demand forecasting, but they struggle to capture complicated temporal patterns, particularly in psychiatric settings. Recent techniques, particularly the Bi-directional Long Short-Term Memory (Bi-LSTM) model, have shown promise because of their capacity to represent sequential data. However, ordinary Bi-LSTM models are still prone to gradient instability and error accumulation in long-term projections. Thus, this study proposed an enhancement to the Bi-LSTM model with a dynamic regularization integrated into the forget gate. The regularization adapts dynamically based on the variance of previous gate activations, minimizing reliance on unpredictable patterns and boosting long-term learning stability. By including this adaptive process, the model becomes more resilient to fluctuations and better able to capture important temporal connections. Experimental results show that the proposed model is able to achieve good results ranging from 0.1974 to 101.5289, demonstrating an improvement from the other approaches in the literature.