Shapley Regression for Rare Disease Diagnosis Support: a case study on APDS

Researchers propose Shapley regression, a game-theoretic machine learning method for diagnosing APDS, a rare genetic immune disorder. The approach combines interpretability with predictive power by modeling symptom interactions while maintaining transparency, validated on both public datasets and a real-world cohort of 222 patients.

This research addresses a critical gap in rare disease diagnosis by developing an interpretable machine learning approach that outperforms traditional linear methods while remaining transparent to clinical practitioners. APDS diagnosis has historically suffered from delayed recognition due to heterogeneous symptoms and overlapping clinical presentations, creating a real clinical need for systematic detection frameworks. The proposed Shapley regression method innovates by replacing linear predictors with k-additive cooperative games, explicitly capturing how symptoms co-occur and interact rather than treating them independently.

The technical contribution bridges two competing demands in medical AI: predictive accuracy and clinical interpretability. While deep learning models excel at pattern recognition, their black-box nature creates adoption barriers in healthcare settings where clinicians must understand and validate model decisions. Shapley regression maintains the convexity and transparency of logistic regression while incorporating symptom interactions through game theory, a principled mathematical framework that physicians can conceptually grasp.

The empirical validation strengthens the work's credibility. Testing across eight public biomedical datasets established that 2-additive models with L2 regularization achieve optimal performance-robustness trade-offs. Application to the real-world APDS cohort demonstrated that the method accurately distinguished cases from matched controls and confirmed known phenotypic associations while enabling exploration of previously undocumented symptom pairs. Clinical expert validation of these pairwise interactions provides crucial evidence that the model captures clinically meaningful patterns rather than statistical artifacts.

The broader implications extend beyond APDS diagnosis. This methodology provides a template for interpreting rare disease detection systems generally, supporting adoption of AI tools in clinical practice where regulatory approval and physician trust remain paramount considerations.

• →Shapley regression combines game theory with logistic regression to model symptom interactions while maintaining clinical interpretability
• →A 2-additive model with L2 regularization achieved optimal balance between predictive accuracy and noise robustness across validation datasets
• →Real-world validation on 222 APDS patients confirmed known phenotypes and revealed new clinically validated symptom interactions
• →The method bridges deep learning expressiveness and traditional model transparency, addressing a critical barrier to clinical AI adoption
• →This approach provides a replicable framework for developing interpretable diagnostic tools for other rare genetic disorders