Shapley Regression Model Enhances APDS Rare Disease Diagnosis
Researchers have introduced Shapley regression, a novel game-theoretic machine learning model designed to improve the diagnosis of Activated PI3K8 Syndrome (APDS), a rare genetic immune disorder. Current diagnostic methods face challenges due to heterogeneous symptoms and overlapping clinical presentations, while existing AI models often lack interpretability. This new approach replaces traditional linear predictors with a k-additive cooperative game, explicitly modeling symptom co-occurrence while maintaining the transparency of logistic regression. Empirical studies on eight public biomedical datasets demonstrated that a 2-additive model with l2 regularization offers an optimal balance between predictive power and noise robustness. Furthermore, application to a real-world cohort of 222 patients successfully distinguished APDS cases from controls, validating known phenotypes and uncovering pairwise symptom interactions confirmed by clinical experts. This lightweight method bridges the gap between complex deep learning expressiveness and the need for interpretable, data-driven tools in routine electronic health records, potentially reducing diagnostic delays for rare diseases.
Wire timeline
Shapley Regression Model Enhances APDS Rare Disease Diagnosis
Researchers have introduced Shapley regression, a novel game-theoretic machine learning model designed to improve the diagnosis of Activated PI3K8 Syndrome (APDS), a rare genetic immune disorder. Current diagnostic methods face challenges due to heterogeneous symptoms and overlapping clinical presentations, while existing AI models often lack interpretability. This new approach replaces traditional linear predictors with a k-additive cooperative game, explicitly modeling symptom co-occurrence while maintaining the transparency of logistic regression. Empirical studies on eight public biomedical datasets demonstrated that a 2-additive model with l2 regularization offers an optimal balance between predictive power and noise robustness. Furthermore, application to a real-world cohort of 222 patients successfully distinguished APDS cases from controls, validating known phenotypes and uncovering pairwise symptom interactions confirmed by clinical experts. This lightweight method bridges the gap between complex deep learning expressiveness and the need for interpretable, data-driven tools in routine electronic health records, potentially reducing diagnostic delays for rare diseases.
cs.AI updates on arXiv.org