Development and Validation of a Machine Learning-Based Risk Prediction Model for Ischemic Stroke-Diabetes Comorbidity

Aims: This study aimed to develop and validate machine learning-based risk prediction models for ischemic stroke-diabetes mellitus (IS-DM) comorbidity using routinely available clinical data, and to compare the performance of traditional logistic regression with backpropagation neural networks (BPNN). ​ Methods Health records of 16,406 community-dwelling adults from Beijing, China, we analyzed. From 41 initial candidate predictors across five categories, seven optimal predictors were selected through ​univariate analysis​ followed by ​multivariate analysis. The dataset was randomly split into training (70%) and validation (30%) sets. We developed prediction models using both logistic regression and BPNN approaches, with model performance evaluated through confusion matrix, AUC, and 10-fold cross-validation. ​ Results The single-hidden-layer BPNN model with three hidden nodes demonstrated superior predictive performance, achieving an AUC of 0.921 (95% CI: 0.92-0.93), outperforming logistic regression. Key predictors included age, marital status, fasting glucose, HbA1c, systolic blood pressure, serum creatinine, and serum sodium. However, the BPNN required significantly more computational resources. ​ Conclusion Machine learning approaches, particularly BPNN, can effectively predict IS-DM comorbidity risk using routine clinical parameters. These models could enhance early comorbidity detection in community settings and inform targeted prevention strategies. Despite it predictive efficacy, the computational demands of BPNN should be considered for clinical implementation.