Introduction: Machine learning (ML) involves computational methods and learning mechanisms that can help generate new knowledge from large databases. These techniques are uniquely suited to analyses of paediatric intestinal failure (IF) where classification, prognostication, identification of risk factors, and knowledge extraction can be challenging. Heterogeneous patient populations, small sample sizes, and variations in treatments may complicate our ability to manually identify trends in data. Factors that predict enteral autonomy (EA) and patient management (optimal intravenous lipid emulsion [ILE] solutions; central venous catheter [CVC] complications) have been described, yet a systematic approach that may be prospectively applied to large datasets has never been performed in this patient population. Our study aims to identify predictors of EA through machine learning (ML) and assess their impact on TPN duration. For patients on TPN, assessments of lipid emulsion (ILE) and central venous catheter (CVC) type were also performed. Methods: Demographics, procedures, and outcomes for patients (N=17) from McMaster Children’s Hospital’s (MCH) Intestinal Rehabilitation Program (1999-2021) were collected for a retrospective chart review. Predictors of EA status were obtained using a decision tree ML model. Performance was evaluated as area under the receiver operator characteristics (AUC) and precision-recall (PR) curves. These predictors were assessed for ability to impact TPN duration with Kaplan-Meier (KM) estimates and Cox proportional-hazards (CPH) models. ILE treatments were compared using the Mann-Whitney U test, and CVC types were compared with KM, CPH, and analysis of variance (ANOVA). Results: The model retrieved 10 predictors of EA status. AUC and PR curves were 91% and 71% respectively. Of these 10 predictors: sex, serial transverse enteroplasty (STEP) procedure, ileocecal valve (ICV) presence, and remnant small bowel length were significant for predicting EA status under KM, but not CPH. For ILE treatment, patients on SMOF Lipid® had lower total bilirubin values compared to Intralipid® at six months (4.0 versus 10.0 uMol/L). Longevity of peripherally inserted central catheters (PICC) was 847 days versus 268 days for tunnelled CVCs, and 181 days for ports (95% CI: 0-2044, 93-443, 144-218). Types of complications were not significantly different across CVC. Conclusions: Our study supports the role for ML frameworks in advanced pattern recognition, and promoting adaptive clinical management. Validation of this model using our single-centre dataset supports similar conclusions derived from larger, multi-center registries. Ongoing use of our ML framework offers a novel contribution to intestinal failure analyses by helping identify unique risk factors and protective factors for EA, and approaches to clinical management in the paediatric IF population.
Read full abstract