Machine learning for enhanced prognostication: predicting 30-day outcomes following posterior fossa decompression surgery for Chiari malformation type I in a pediatric cohort.

OBJECTIVE

Chiari malformation type I (CM-I) is a congenital disorder occurring in 0.1% of the population. In symptomatic cases, surgery with posterior fossa decompression (PFD) is the treatment of choice. Surgery is, however, associated with peri- and postoperative complications that may require readmission or renewed surgical intervention. Given the associated financial costs and the impact on patients' well-being, there is a need for predictive tools that can assess the likelihood of such adverse events. The aim of this study was therefore to leverage machine learning algorithms to develop a predictive model for 30-day readmissions and reoperations after PFD in pediatric patients with CM-I.

METHODS

This was a retrospective study based on data from the National Surgical Quality Improvement Program-Pediatric database. Eligible patients were those undergoing PFD (Current Procedural Terminology code 61343) for CM-I between 2012 and 2021. Patients undergoing surgery for tumors or vascular lesions were excluded. Unplanned 30-day readmission and unplanned 30-day reoperation were the main study outcomes. Additional outcome data considered included the length of hospital stay, 30-day complications, discharge disposition, and 30-day mortality. Training and testing samples were randomly generated (80:20) to study the 30-day readmission and reoperation using logistic regression, decision tree, random forest (RF), K-nearest neighbors, and Gaussian naive Bayes algorithms.

RESULTS

A total of 7106 pediatric patients undergoing PFD were included. The median age was 9.2 years (IQR 4.7, 14.2 years). Most of the patients were female (56%). The 30-day readmission and reoperation rates were 7.5% and 3.4%, respectively. Headaches (32%) and wound-related complications (30%) were the most common reasons for 30-day readmission, while wound revisions and evacuation of fluid or blood (62%), followed by CSF diversion-related procedures (28%), were the most common reasons for 30-day reoperation. RF classifiers had the highest predictive accuracy for both 30-day readmissions (area under the curve [AUC] 0.960) and reoperations (AUC 0.990) compared with the other models. On feature importance analysis, sex, developmental delay, ethnicity, respiratory disease, premature birth, hydrocephalus, and congenital/genetic anomaly were some of the variables contributing the most to both RF models.

CONCLUSIONS

Using a large-scale nationwide dataset, machine learning models for the prediction of both 30-day readmissions and reoperations were developed and achieved high accuracy. This highlights the utility of machine learning in risk stratification and surgical decision-making for pediatric CM-I.