A New Deep Learning Algorithm for Detecting the Lag Effect of Fine Particles on Hospital Emergency Visits for Respiratory Diseases

There exists a time lag between short-term exposure to fine particulate matter (PM2.5) and incidence of respiratory diseases. The quantification of length of the time lag is significant for preparation and allocation of relevant medical resources. Several classic lag analysis methods have been applied to determine this length. However, different models often lead to distinct results and which one is better is subtle. The prerequisite of obtaining the reliable length is that the model can truly reveal the underlying pattern hidden in the above relationship. Long short-term memory (LSTM) is an artificial recurrent neural network (RNN) architecture used in the field of deep learning, whose strong capability makes it widely applied in many fields. In this study, we manage to exploit it to acquire the time-lag length in the exposure-response relationship. The relationship between exposure and response is assumed as linear and non-linear, and models with and without confounding factors are performed under these two assumptions. Results of DLNM model show that the best hospital emergency visit prediction appears in 3 lag days, with the maximum RR value of 1.004357 (95% CI: 1.000938-1.009563). Then, a vary of LSTM models with different time steps are performed, which are evaluated by mean absolute error (MAE), the mean absolute percentage error (MAPE), the root of mean square error (RMSE) and R square (R
²
). The results show that LSTM of time step 3 achieves the lowest MAE (33), MAPE (9.86), RMSE (42) and the highest R
²
(0.78), consistent with the result of DLNM model. Also, the proposed model is compared with ARIMA model, one of the commonly used forecasting models, showing better accuracy. This demonstrates that LSTM can be used as a new method to detect the lag effect of PM2.5 on respiratory diseases.