Predicting Craniospinal Surgery in Pediatric Achondroplasia: Benchmarking Statistical Inference and Explainable AI
Seifollah Gholampour1, Jesse Huang1, Dylan Keusch2
1Department of Neurological Surgery, University of Chicago Medicine, Chicago, IL, USA.
Purpose:
To develop a framework that benchmarks penalized statistical inference against explainable artificial intelligence (XAI) methods for predicting cranial and spinal surgical need in pediatric achondroplasia under small-cohort, class-imbalanced, and phenotypically heterogeneous conditions.
Methods:
We analyzed 34 clinical, demographic, and imaging variables in a multicenter cohort (four U.S. hospitals) of 150 pediatric patients with achondroplasia. We benchmarked penalized statistical inference (ridge regression and generalized additive models [GAMs]) against nine cost-sensitive classifiers, and applied post hoc SHapley Additive exPlanations (SHAP) to interpret the best-performing classifier.
Results:
The stacked ensemble achieved superior test-set performance (accuracy and macro-F1 = 0.77) with stable generalization and significantly higher discrimination than the GAM baseline (cranial AUROC 0.78 vs. 0.68; spinal AUROC 0.76 vs. 0.66). Calibration was acceptable, and decision curve analysis showed positive net benefit across relevant thresholds for both outcomes. SHAP highlighted class-specific drivers: foramen magnum (FM) stenosis, hydrocephalus, family history, frontal bossing, sleep disturbance, and age for cranial surgery, and spinal stenosis, Chiari malformation, family history, back pain, FM stenosis, and age for spinal surgery. SHAP dependence patterns suggested context-dependent age attributions in relation to FM or spinal stenosis, rather than a consistent standalone age effect. Several SHAP-highlighted predictors (e.g., sleep disturbance, back pain, and age) were nonsignificant in the inferential baselines. Kaplan-Meier curves indicated earlier intervention in high-risk phenotypes.
Conclusion:
In a reproducible dual-stream benchmark, XAI improved discrimination over conventional inference by capturing clinically contextualized, complex nonlinear predictive dependencies and interactions associated with cranial and spinal surgical risk in achondroplasia using preoperative variables.