〔Abstract〕 To construct a deep learning model based on convolutional neural network ( CNN)-long short term memory network (LSTM)-Attention to explore the correlation between meteorological and clinical factors and the incidence of ischemic stroke . Methods A fusion model CNN-LSTM-Attention based on CNN , LSTM , and Attention was constructed by incorporating clinical data and meteorological data of ischemic stroke inpatients . The predictive performance of the model was evaluated by maximum prediction error and root mean square error (RMSE) . The impact of different lag days on prediction performance was investigated by selecting lag periods ran- ging from 1 to 7 days . Results In both short-term and long-term predictions , the CNN-LSTM-Attention fusion model (short-term : 1 . 5 and 0. 6 ; long-term : 8. 3 and 2. 5) showed superior maximum prediction bias and RMSE compared to the LSTM model ( short-term : 2. 8 and 1 . 2 ; long-term : 19. 5 and 5 . 5) and the CNN-LSTM model (short-term : 2. 0 and 0. 8 ; long-term : 11 . 2 and 3 . 3) . After incorporating lag days , the maximum prediction devi- ation and RMSE for lags of 3 days (short-term : 0. 7 and 0. 4 ; long-term : 5 . 5 and 1 . 9) and 5 days (short-term : 0. 8 and 0. 3 ; long-term : 6. 5 and 2. 0) in both short-term and long-term forecasts were smaller than lags of 0 days (short-term : 1 . 5 and 0. 6 ; long-term : 8. 3 and 2. 5) . The maximum prediction deviation and RMSE with a lag of 1 day (short-term : 1 . 5 and 0. 8 ; long-term : 6. 8 and 2. 4) and 7 days (short-term : 1 . 9 and 0. 9 ; long-term : 7. 5 and 2. 7) were both greater than those with a lag of 0 days . Conclusion The established CNN-LSTM-Attention model demonstrates significant predictive value for the onset of ischemic stroke and can provide reference for the ra- tional allocation of medical resources .