Multistage machine learning model for automated referral triage in pain medicine

Effective referral triage in pain medicine is essential to ensure that patients receive timely and appropriate care. This study presents a multistage machine learning framework to better identify patients who may benefit from one of five specialised pain procedures. Two years of clinic data were used for training and validation, consisting of 231 features for 3,552 patients. The proposed approach started with the baseline model selection, followed by stage iterations acquiring the elbow method for stage decision. The Easy Ensemble was selected among methods and applied for each stage. Results showed that the approach improves prediction accuracy in the true positive rate (TPR) and area under the curve (AUC). The final-stage models achieved the improvement over single-stage model by as much as 34.8% and 23% on TPR and AUC. This multistage framework can enhance triage accuracy and hold potential for broader application in other clinical settings.