Explainable prediction is increasingly required in clinical decision support, especially when models must generalize across institutions. We present a stability-bound binary rule search workflow that operates on fully binarized clinical data and expresses decisions as sparse, human-readable rules. Clinical variables are converted into 0/1 indicators using clinically meaningful thresholds, so that each rule corresponds to a binary mask over a small set of interpretable features. A Binary Rule Search (BRS) engine explores conjunctions of up to four predictors (k=1–4), and candidate rules are evaluated by the Matthews-correlation-coefficient (MCC) on development and validation splits. Robustness is summarized by the Stability-Bound-Rule-Score (SBRS), a geometric-style combination of the lower 95% confidence bounds of MCC in both splits. The workflow was applied to two open-access datasets: a heart attack dataset (303 patients) and a hepatitis C dataset (615 patients). In the heart attack data, a four-feature rule combining age 55–64 years, typical chest pain, absence of angiographically stenosed vessels (CA = 0) and a reversible thallium perfusion defect achieved MCC 0.71 and 0.73 in the development and validation sets, with SBRS = 1.59. In the hepatitis C data, rules built from elevated aspartate aminotransferase together with intermediate or high alkaline phosphatase and increased bilirubin reached MCC 0.75 and 0.84, with SBRS = 1.67. Because all predictors are binarized, the final rules can be displayed as compact binary mask plots or implemented as short checklists and look-up tables. Overall, this stability-bound binary rule search workflow yields sparse, stable and clinically interpretable rule sets for cardiovascular risk stratification and chronic liver disease screening.
Huyut et al. (Sat,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: