Deep Learning Neural Network based Survival Analysis: Application in Brain Stroke

Abstract

Background: Determining subclinical Brain stroke (BS) risk factors may allow for early and more operative BS prevention measures. To find the main risk factors and moderating effects of survival in patients with BS. For enhancing the accuracy of diagnosis and the quality of patient care, the artificial intelligence filed is obtaining ever-increasing interest. So, the Deep learning based Neural Networks (DLNN) method was approached in patients with BS to predict the outcome by the risk factors. Methods: In this prospective study, a total of 332 patients with BS (mean age 77.4 (SD 10.4) years, 50.6% male), from the Imam Khomeini hospital, Ardabil, Iran, were recruited from 2004 up to 2018. The data of major predictors of survival were gathered from the available documents from the BS registry of the hospital, and the definitive diagnosis of BS was considered based on computerized tomography scan and magnetic resonance imaging. Cox's proportional hazard regressions and DLNN were used to analyze the predictors of survival and the moderating effect by introducing the interaction effects. Results: The survival probability (95% Confidence Interval (CI)) of patients with BS was 39% (34 – 45). The 1-, 5- and 10-year death rates were 0.254, 0.053, and 0. 023, respectively. The independent risk factors associated with BS were age category (Adjusted Hazard Ratio (AHR)=1.81; 90% CI: 1.61 to 2.02), male (AHR=1.41; 90% CI: 1.11 to 1.78), history of blood pressure (AHR=1.39; 90% CI: 1.09 to 1.78) , history of diabetes (AHR=1.37; 90% CI: 1.02 to 1.82) , history of hyperlipoproteinemia (AHR=0.64; 90% CI: 0.46 to 0.9) , oral contraceptive pill use (AHR=0.65; 90% CI: 0.45 to 0.95) , hemorrhagic cerebrovascular accident (AHR=1.49; 90% CI: 1.12 to 1.97). Also, the age category and education level, smoking, and using oral contraceptive pill moderates the relationship between history of cerebrovascular accident, history of heart disease, and history of blood pressure with the hazard of BS, respectively (All p<0.05). The most important predictors for BS mortality based on the optimal model of DLNN method results were time interval after ten years with 92.2% accuracy Age category with 75.6% accuracy, history of hyperlipoproteinemia with 66.88% accuracy, education level with 66.86% accuracy. The other independent variables sex, employment, place of residence, former smoking, waterpipe smoking, history of heart disease, diabetes, oral contraceptive pill use, physical activates, history of cerebrovascular accident type, history of blood pressure history, were at moderate importance about 66.55% accuracy. Conclusion: Instead of considerable advances for the treatment of a patient with BS, effective BS prevention remains the best means for dropping the BS load regarding the related factors which have been found in this study. This study demonstrates that the DLNN strategy shows a satisfying presentation in the prediction of BS mortality with higher diagnostic accuracy based on the main risk factors in comparison to the Cox regression model.