Development and Validation of an Ensemble Machine Learning Model for Early Detection of Hepatopancreatobiliary Diseases

Hepatopancreatobiliary (HPB) cancers have the highest death rates in oncology due in part to late diagnosis. Non-specific symptoms and the lack of standardized evaluation instruments in primary care lead to diagnostic delays. We developed and validated a machine learning-based clinical decision support model for timely identification and referral of patients at risk for HPB diseases. 

Methods 

We trained a structured Symptomatology Index and an ensemble of Gradient Boosting, AdaBoost, and Support Vector Machines in a retrospective-prospective multicentre cohort study across three tertiary referral centres in Northern India (n=2296). The model was internally validated in 2108 patients and independently validated in a multicentre cohort of 781 patients. 

Findings 

The ensemble model had AUC 0·96 (95% CI 0·94–0·98), sensitivity 0·76, and specificity 0·94. This model demonstrated a 23% relative improvement in sensitivity compared with specialist clinical assessment (p<0·01). Disease-specific sub-models achieved AUCs of 0·87-0·89 for biliary and periampullary malignancies. Early diagnosis (≤4 months) was associated with an absolute increase of 9·6% in the rates of curative resection (p<0·001). A bilingual mobile app achieved 92·6% cross-validated accuracy for real-time symptom triage. 

Interpretation 

A structured-symptom-based ensemble learning framework can detect HPB diseases with high accuracy. Integration via a digital app provides a scalable, low-cost strategy to standardise symptom evaluation, expedite specialist referral, and improve surgical eligibility in resourcelimited settings.