Explainable AI Outperforms Classical Statistics in Predicting Pediatric Achondroplasia Surgery

For children born with achondroplasia, the most common form of short-limbed dwarfism, some of the most consequential medical decisions revolve around whether and when to operate on the skull and spine. Narrowing of the foramen magnum, the opening at the base of the skull through which the brainstem passes, can compress vital neural structures and prove life-threatening in infancy. Further down the spinal column, stenosis of the lumbar canal can cause progressive neurological injury and crippling pain. Yet deciding when the risks of surgery outweigh the risks of watchful waiting remains one of the most difficult judgment calls in pediatric neurosurgery, made harder by the fact that achondroplasia is rare, its complications vary enormously from child to child, and most centers never accumulate enough patients to support conventional statistical modeling.

A new multicenter study led by Seifollah Gholampour of the University of Chicago Medicine and David F. Bauer of Texas Children’s Hospital, published in the Annals of Biomedical Engineering, tackles precisely this problem. Drawing on 150 pediatric patients treated at four major U.S. hospitals—Boston Children’s Hospital, Cedars-Sinai Medical Center, Oregon Health & Science University’s Doernbecher Children’s Hospital, and Texas Children’s Hospital—the research team assembled a dataset of 34 clinical, demographic, and imaging variables for each child. They then ran a head-to-head competition: traditional penalized statistical inference, in the form of ridge regression and generalized additive models, against nine cost-sensitive machine learning classifiers, with the best performers combined into a stacked ensemble. Crucially, they wrapped the winning model in an explainable artificial intelligence framework so that its reasoning could be inspected rather than taken on faith.

The results were striking. The stacked ensemble achieved an accuracy and macro-F1 score of 0.77 on held-out test data, with stable generalization across validation folds. More telling was the comparison of discrimination, the model’s ability to separate children who would need surgery from those who would not. For cranial surgery, the ensemble’s area under the receiver operating characteristic curve reached 0.78, compared with just 0.68 for the generalized additive model baseline. For spinal surgery the gap was similar: 0.76 versus 0.66. Calibration, meaning the alignment between predicted probabilities and observed outcomes, was acceptable, and decision curve analysis showed positive net benefit across clinically relevant thresholds for both outcomes. In practical terms, the machine learning approach made meaningfully fewer errors in the scenarios that matter most for surgical planning.

Why should a modestly larger cohort than a traditional statistician might demand still favor machine learning? The answer lies in the structure of the data itself. Rare diseases produce small, class-imbalanced datasets in which only a fraction of patients undergo surgery, and the relationships between predictors and outcomes are rarely linear or independent. A foramen magnum that is critically narrow may matter far more in an infant with hydrocephalus than in an older child without it. Achondroplasia’s phenotype, driven by a well-characterized mutation in the FGFR3 gene, produces a constellation of interacting features—frontal bossing, sleep-disordered breathing, Chiari malformation, back pain—whose joint effects defy simple additive models. Ensemble classifiers, which combine many weak learners, are built to capture such nonlinear dependencies and interactions, provided their outputs can be made transparent enough for clinicians to trust.

Transparency came through SHAP, or SHapley Additive exPlanations, a technique rooted in cooperative game theory that assigns each predictor a contribution value for every individual prediction. When the researchers applied SHAP to the best-performing model, the resulting explanations mapped closely onto clinical intuition while also revealing subtleties that conventional inference had missed. For cranial surgery, the dominant drivers were foramen magnum stenosis, hydrocephalus, family history, frontal bossing, sleep disturbance, and age. For spinal surgery, the leading contributors were spinal stenosis, Chiari malformation, family history, back pain, foramen magnum stenosis, and age. These class-specific profiles matter because they tell surgeons not merely what the model got right, but which features of a given child’s presentation pushed the prediction toward or away from intervention.

Perhaps the most scientifically interesting finding concerns age. In the classical inferential baselines, age did not reach statistical significance as a standalone predictor. Yet SHAP dependence patterns showed that the model attributed importance to age in a context-dependent way: the effect of age on surgical risk depended on whether a child had foramen magnum or spinal stenosis, rather than acting as an independent risk factor. This is exactly the kind of interaction a linear or additive model is designed to smooth away, and its recovery illustrates why the authors argue that explainable machine learning captures clinically contextualized predictive structure that penalized inference can obscure.

The comparison also exposed a provocative asymmetry. Several predictors highlighted prominently by SHAP—sleep disturbance, back pain, and age—were statistically nonsignificant in the inferential baselines. To a traditional biostatistician, that discrepancy might look like a red flag; to the study’s authors, it is evidence that the two frameworks answer different questions. Penalized inference asks which variables show a population-average marginal association, subject to strict regularization and significance testing. The ensemble-plus-SHAP pipeline asks which variables shift individualized risk estimates, including through interactions. Both perspectives carry information, and the study’s dual-stream benchmark was deliberately designed to make that contrast explicit and reproducible rather than to declare a single winner.

The stakes of getting these predictions right are high. Achondroplasia affects roughly one in every 20,000 to 30,000 births worldwide, and its neurosurgical complications have historically driven much of its excess mortality. Compressed brainstem and upper cervical cord in infants can cause central apnea and sudden death, while progressive spinal stenosis in older children can lead to irreversible neurological deficits. Kaplan–Meier analyses in the new study showed that children the model classified as high-risk underwent surgery significantly earlier than those in lower-risk strata, suggesting that the model’s risk scores align with real clinical timelines. A validated preoperative prediction tool could therefore help standardize referral patterns across the handful of specialized skeletal dysplasia centers that handle these cases, and could help families and clinicians structure surveillance schedules around quantified risk rather than intuition alone.

The study is also notable for how honestly it confronts the limitations of machine learning in rare-disease settings. The literature on medical AI is littered with models that report glittering cross-validation numbers but fail on external data, a problem amplified when event counts per variable fall below classical thresholds. The authors addressed class imbalance with cost-sensitive learning rather than naive synthetic oversampling, an approach consistent with their own prior work questioning the clinical validity of techniques like SMOTE on medical data. They evaluated not only discrimination but also calibration and clinical decision curves, and they compared nine candidate classifiers against a proper inferential baseline rather than a straw man. The centralized ethics approval across all four participating institutions, with de-identified retrospective data, allowed a cohort size that no single center could have assembled.

What emerges is less a finished clinical tool than a template for how predictive modeling should be done when data are scarce, imbalanced, and phenotypically messy. By benchmarking statistical inference against explainable AI under explicit rare-disease constraints, and by insisting that every prediction come with a decomposition of its drivers, the team has produced a reproducible framework that other rare-disease researchers can adapt. For children with achondroplasia, the near-term hope is that risk stratification for foramen magnum decompression and spinal surgery becomes earlier, more consistent, and better communicated. For the broader field of clinical machine learning, the message is sharper still: in rare disease, the models that earn clinical trust will not be the most accurate black boxes, but the ones whose reasoning surgeons can read, question, and act upon.

Subject of Research: Explainable AI versus statistical inference for predicting craniospinal surgery in pediatric achondroplasia

Article Title: Predicting Craniospinal Surgery in Pediatric Achondroplasia: Benchmarking Statistical Inference and Explainable AI Under Rare-Disease Constraints

Article References: Gholampour, S., Huang, J., Keusch, D., Lopes, C., Masarwy, A., Obayashi, J., Ruppert-Gomez, M., Sayama, C. M., Baird, L. C., Danielpour, M., & Bauer, D. F. (2026). Predicting Craniospinal Surgery in Pediatric Achondroplasia: Benchmarking Statistical Inference and Explainable AI Under Rare-Disease Constraints. Annals of Biomedical Engineering. https://doi.org/10.1007/s10439-026-04352-x

Image Credits: AI Generated

DOI: 10.1007/s10439-026-04352-x

Keywords: achondroplasia, pediatric neurosurgery, explainable AI, SHAP, machine learning, foramen magnum stenosis, spinal stenosis, surgical risk prediction, clinical decision support, rare disease, generalized additive models, Annals of Biomedical Engineering