A concept-enhanced, knowledge graph-guided framework for interpretable PCOS prediction: a case study in explainable...

Area of Science:

• Biomedical informatics
• Machine learning applications in healthcare
• Genomics and personalized medicine

Background:

• Bridging the gap between machine learning (ML) efficiency and therapeutic use requires evidence-based frameworks.
• Current ML approaches often lack transparency, hindering clinical adoption.

Purpose of the Study:

• To evaluate a novel framework combining biological expertise, knowledge graphs (KG), and probabilistic inference with ML for Polycystic Ovary Syndrome (PCOS) prediction.
• To enhance the interpretability and clinical applicability of ML models in diagnosing complex conditions like PCOS.

Main Methods:

• Utilized PCOS clinical data (541 patients) and integrated clinically relevant concepts via a KG with TransE-based scoring.
• Employed a Bayesian network (BN) for dependency modeling and incorporated knowledge-driven features into ML models.
• Screened models using LazyPredict and constructed an ensemble from the top 10 performers, focusing on accuracy and interpretability.

Main Results:

• The concept-integrated ensemble model achieved high predictive accuracy (93.65% cross-validation accuracy, 0.98 ROC-AUC).
• Incorporating KG and BN features significantly improved model interpretability, with BN causal probabilities being key.
• Generated patient-specific, guideline-compliant explanations for model predictions.

Conclusions:

• Integrating biomedical concepts enhances PCOS prediction interpretability without sacrificing performance.
• The proposed framework supports clinician-aligned PCOS screening tools for transparent risk classification and informed treatment decisions.