Reward-based Recurrent Neural Network for Major Depressive Disorder Prediction in Time Series

Predicting depressive symptoms in patients poses challenges due to low accuracy. Depression is a prolonged symptom, and there is heterogeneity among patients’ symptoms, making it difficult for errors in recurrent neural network predictions to be corrected, resulting in lower model generalizability. We propose and construct a reward factor, generated by fitting the gap between the previous stage’s predicted results and actual data, and incorporate it into the recurrent neural network. The neural network extracts sequential features of patients’ depression scores over time series, utilizing the reward factor to adjust subsequent prediction network parameters, thus achieving early warning of severe depressive symptoms in patients over extended periods. Finally, we validate the approach using authentic patient data and compare the results with other classical methods, demonstrating that the reward-based recurrent neural network achieves higher prediction accuracy.