The UTC Graduate School is pleased to announce that Megan McCoy will present Doctoral research titled, A COMPARATIVE ANALYSIS OF STATISTICAL AND MACHINE LEARNING MODELS WITH APPLICATION IN AI-POWERED STROKE RISK PREDICTION on 10/10/2025 at 10 AM in Lupton 302. Everyone is invited to attend. 

Computational Science

Chair: Lan Gao

Abstract:

Rapid detection of large vessel occlusion (LVO) in stroke is crucial due to its high mortality and narrow window for intervention. Machine learning (ML) and deep learning-based AI tools show promise for LVO prediction, yet clinical use is limited by the inconsistent pre-hospital data with varying LVO rates, the hard-to-interpret ”black box” nature of many ML algorithms, and high costs of the AI tools. To address this gap, this study proposes a novel hybrid neural network (HNN) model that integrates classical statistical methods with neural networks, combining the structured framework and interpretability of statistical learning with the flexibility and regularization of ML. The HNN was validated through simulation studies using NIHSS scores, demographics, and medical history across diverse LVO prevalence rates and sample sizes. Its performance was benchmarked against logistic regression, Naive Bayes, Random Forest, Decision Tree, and standard neural networks, using metrics including sensitivity, specificity, accuracy, PPV, NPV, AUC, and ROC analysis. The HNN was further applied to a large, multi-center dataset from over 100 hospitals nationwide, employing consistent evaluation metrics. Sampling strategies optimized predictive performance, and SHAP analysis elucidated key predictors influencing model outcomes. In both simulated and real-world settings, the HNN achieved at least a 20% improvement in sensitivity while maintaining robust performance across other metrics. These findings highlight the HNNs potential to address barriers to AI adoption in stroke care by combining interpretability with superior predictive power, offering a practical, scalable tool for pre-hospital LVO detection to enhance faster clinical decision-making and improve patient outcomes.