Work

Explainable AI for Conditional Stroke Probability Estimation: A Hybrid Bayesian Network Approach

Year: 2026

Type: article

Abstract: Accurate stroke prediction is vital for early intervention, yet most high-performing AI/ML models remain “black boxes” with little intrinsic interpretability. Bayesian net-works (BNs) are probabilistic graphical models that balance predictive performance with interpretable, auditable reasoning. We assessed whether optimizing data preprocessing could enhance BNs diagnostic performance on a highly imbalanced stroke dataset. Using a 2×2×2 factorial design, we compared eight configurations varying three choices: dis-cretized versus continuous (conditional Gaussian) topology, median versus multiple imputation by chained equations (MICE), and no balancing versus synthetic over-sampling (SMOTE). Preserving continuous variables significantly improved discrimina-tion over discretization (AUC 0.812 vs 0.718; p = 0.011), an advantage that persisted across 30 stratified resamples (paired p < 0.001); the hybrid BN matched XGBoost (0.822) and logistic regression (0.819). MICE gave no advantage over median substitution (p = 0.708), and SMOTE did not improve performance while tripling the Brier score, indicating posterior miscalibration. The optimal BN remained interpretable and could simulate illustrative absolute risk reductions (e.g., 43.18% to 31.55% under modelled glycemic control). Continuous BNs offer an interpretable, well-calibrated, and robust framework for clinical decision support, warranting prospective validation in real-world settings.

Source: Algorithms

DOI: 10.3390/a19090736

Authors: Pranil GC, Ravinder-Jeet Singh, Ratvinder Grewal

Institutions: Health Sciences North, Laurentian University, NOSM University

Language: English

Topic: Bayesian Modeling and Causal Inference

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: gold

Estimated APC paid: $1,930