# StrokeGuard: Cloud-Deployed Calibrated-MLP with K-Means Boundary Undersampling for Scalable Stroke Risk Prediction

“StrokeGuard: Cloud-Deployed Calibrated-MLP with K-Means Boundary Undersampling for Scalable Stroke Risk Prediction” (2026), by Mohamed Shaban et al. Health Professions research paper. Original abstract below, unchanged. A free PDF is available.

- Authors: Mohamed Shaban, Shahd Ibrahim, Manal M. Ghaith, Ahmed M. Anter
- Published: November 2, 2026
- Field: Health Professions
- DOI: https://doi.org/10.21203/rs.3.rs-11286884/v1
- Free PDF: https://www.researchsquare.com/article/rs-11286884/latest.pdf

## Abstract

Stroke accounts for roughly one in nine deaths worldwide and ranks among the most disabling non-communicable diseases globally. Timely identification of at-risk individuals is clinically actionable. Yet, two persistent obstacles hamper the deployment of automated screening tools: (1)the pronounced scarcity of stroke-positive records in routine clinical databases, where positive cases often constitute less than 5% of the cohort and yield imbalance ratios that can exceed 19:1; and (2)~the near-universal absence of validated, production-grade inference infrastructure in published stroke-ML research. We present StrokeGuard, an end-to-end system that couples a novel K-Means cluster-based undersampling strategy with a threshold-calibrated multi-layer perceptron (MLP) and delivers the trained model to a multi-region cloud platform with fully benchmarked latency, throughput, and cost characteristics. The undersampling algorithm divides the PCA-transformed feature space into four clusters, then draws majority-class records whose proximity to the local minority centroid is smallest, thereby concentrating the training signal on the class boundary while reducing the imbalance ratio from 19.6:1 to 5:1 without generating synthetic data. The MLP is optimized with class-weighted binary cross-entropy, and its decision boundary is shifted to τ=0.3 to maximize G-Mean. Preprocessing comprises Random Forest regression for missing BMI imputation (R²=0.707), z-score normalization using training-set statistics, and Principal Component Analysis retaining 89% of total variance. Evaluated on 45,108 clinical records (2,631 held-out test samples), the system achieves 92.51% accuracy, 72.60% recall, 80.51% precision, 96.49% specificity, 83.70% G-Mean, 76.35% F1 score, and AUC-ROC~=~0.8706, outperforming seven baseline classifiers. Five-fold cross-validation yields 92.36% ± 0.31%, and the train-to-test accuracy gap is 1.61%, confirming generalization. Cloud deployment across the US East, the EU West, and Asia-Pacific produces mean inference latency of 18.72~ms (P95: 24.34~ms; P99: 34.24~ms) and throughput up to 54.7 req/s, with auto-scaling from 1 to 8 instances at linear cost ($0.115 to $0.920/hr). All three regions sustain sub-40 ms P99 latency, satisfying real-time clinical screening requirements.

Canonical page: https://nerdi.pages.dev/paper/W7225118991
Read three real papers a day in Nerdi (iPhone): https://apps.apple.com/app/apple-store/id6813826387?pt=128757794&ct=web-paper&mt=8