Development and validation of a machine learning model predicting post-tonsillectomy hemorrhage.

PURPOSE: To develop and validate machine learning models to predict post-tonsillectomy hemorrhage.

METHODS

This was a machine learning analysis of a cohort of patients included in the Norwegian tonsil registry in Norway from 03.01.2017 to 01.05.2025. A perioperative assessment was used to describe the type of surgery performed, the surgical technique used, and the methods used to achieve hemostasis. Postoperative outcomes were assessed in questionnaires 30 days after surgery. Unsupervised models were used to explore the data. Supervised models were developed to predict post-tonsillectomy hemorrhage, defined as any postoperative bleeding necessitating admission. Predictors included in the model were age, sex, type of surgery, surgical technique and means to achieve hemostasis. Model performance was evaluated with the area under the receiver operating characteristics curve (AUC). The best model was evaluated in a held-out test set. The model was explained with a SHAP plot. A decision curve analysis was conducted to assess the potential clinical utility of the model.

RESULTS

A total of 32,037 patients (mean [SD] age 17.84 [12.33]; 18,949 [59.15%] women) were included, with a mean bleeding rate of 6.17%. Unsupervised learning identified sub-groups with differences in bleeding rates. The best performing predictive model was the Adaboost classifier, achieving a test set AUC of 0.71 (95% CI 0.68-0.73). The most important predictors were middle or old age, bipolar diathermy for hemostasis and male sex. The predictive model was superior to alternative strategies in the decision curve analysis.

CONCLUSION

Post-tonsillectomy hemorrhage could be predicted with moderate accuracy. More research is needed to assess if it has potential utility as a clinical decision-support tool.