Predicting the need for syrinx shunting after posterior fossa decompression in Chiari I malformation: a real-world machine learning analysis.

OBJECTIVE: The aim of this study was to develop and validate a machine learning (ML) algorithm to predict the delayed need for syrinx shunt placement following posterior fossa decompression (PFD) for Chiari malformation type I (CM-I) with concurrent syringomyelia. METHODS: This multicenter retrospective cohort study utilized the TriNetX network to identify patients undergoing index PFD for CM-I and syringomyelia (2010-2020) with at least 2 years of continuous follow-up. The primary outcome was unplanned syringosubarachnoid, syringopleural, or syringoperitoneal shunt placement. Five supervised ML classifiers were trained on 30 preoperative clinical and demographic variables using the Synthetic Minority Over-Sampling Technique to address class imbalance. RESULTS: Of 3112 patients (62.3% female, median age 28.4 years) with a median follow-up of 5.1 years, 271 (8.7%) experienced refractory syringomyelia requiring a shunt. The CatBoost classifier achieved the highest discriminative performance on the independent validation set, yielding an area under the curve of 0.87, accuracy of 83%, sensitivity of 0.78, and specificity of 0.82. Shapley Additive Explanations analysis identified adolescent idiopathic scoliosis, prolonged symptom duration, age > 40 years, and preoperative opioid dependency as the strongest predictors of delayed shunt placement. Notably, partial dependence analysis revealed that every 1-month delay in surgical intervention increased the absolute probability of shunting by approximately 0.8%-1.0%. CONCLUSIONS: The CatBoost ML algorithm accurately predicted the delayed need for syrinx shunting after PFD. The prominent risk associated with concomitant scoliotic deformity and surgical delay provides actionable intelligence, strongly supporting early operative intervention and guiding highly individualized long-term postoperative surveillance.