Development of a deep neural network model for ultra-early neurological deterioration in ischemic stroke and analysis of associated risk factors

Abstract

Background

In this study, we established a deep neural network (DNN)-based predictive model, aiming to provide a basis for improving the treatment prognosis of early neurological deterioration (END) in patients with ultra-early ischemic stroke after intravenous thrombolysis with Alteplase.

Methods

A total of 1747 patients with ultra-early ischemic stroke who received intravenous thrombolysis with Alteplase were retrospectively included into the Yidu cloud database. These patients were assigned into the END group (234 cases) and the (No-END) group (1513 cases) based on whether END occurred within 48 h of admission. Tensorflow module in Python software platform was utilized to establish a DNN-based predictive model. Based on the training set (1397 cases) and a testing set (350 cases), the final DNN-based predictive model was generated for the prediction of END-associated factors. The discriminant performance of the model was evaluated by calculating the area under the receiver operating characteristic (ROC) curve (AUC).

Results

The incidence of END in the participants was 13.39 % after intravenous thrombolysis with alteplase. The END group exhibited higher incidences of hemorrhage, all-cause deaths during hospitalization, and poor prognosis versus the No-END group (all p < 0.001). The AUC of DNN model prediction was 0.853, with a sensitivity of 70.53 % and specificity of 100.00 %. The DNN-based predictive model could effectively predict the END-associated factors.

Conclusion

The DNN-based predictive model in this study has a high predicative accuracy, good generalization ability, and robustness, without overfitting. This model is available for the prediction of END in ultra-early ischemic stroke.