Prior-informed machine learning for clinically interpretable medical diagnosis

Abstract

Traditional machine learning (ML) in medical diagnostics often suffers from limited clinical interpretability and poor generalization, owing to its black-box, data-driven nature. Here, we present Prior-Informed Machine Learning (PIML), a framework that integrates clinically grounded generative reasoning with data-driven, ML-enabled discriminative modeling through Bayesian inference to provide clinically interpretable, patient-specific diagnostic outcomes. By embedding the optimization of sensitivity and specificity in data-driven diagnosis—requirements for subsequent Bayesian inference—directly into the ML model development process, PIML enables a synergistic integration of machine learning and clinical expertise that cannot be achieved through either data-driven or clinically grounded approaches alone. PIML was validated across three clinical domains—cervical neoplasia detection, high-risk oral lesion detection, and preterm birth detection—where Bayesian inference was systematically employed to integrate electrical impedance spectroscopy (EIS)-based machine learning with domain-specific clinical expertise, thereby enabling more comprehensive and enhanced diagnostic decision-making. Our results demonstrate that PIML generally improves diagnostic performance compared with standard ML approaches and provides competitive or superior performance relative to knowledge-based and feature-level prior integration strategies. This work presents an innovative, systematic framework with strong potential for widespread adoption in addressing medical diagnostic challenges across heterogeneous clinical environments.