Early detection of nasopharyngeal carcinoma through machine-learning-driven prediction model in a population-based healthcare record database

Objective: Early diagnosis and treatment of nasopharyngeal carcinoma (NPC) are vital for a better prognosis. Still, because of obscure anatomical sites and insidious symptoms, nearly 80% of patients with NPC are diagnosed at a late stage. This study aimed to validate a machine learning (ML) model utilizing symptom-related diagnoses and procedures in medical records to predict nasopharyngeal carcinoma (NPC) occurrence and reduce the prediagnostic period. Materials and Methods: Data from a population-based health insurance database (2001–2008) were analyzed, comparing adults with and without newly diagnosed NPC. Medical records from 90 to 360 days before diagnosis were examined. Five ML algorithms (Light Gradient Boosting Machine [LGB], eXtreme Gradient Boosting [XGB], Multivariate Adaptive Regression Splines [MARS], Random Forest [RF], and Logistics Regression [LG]) were evaluated for optimal early NPC detection. We further use a real-world data of 1 million individuals randomly selected for testing the final model. Model performance was assessed using AUROC. Shapley values identified significant contributing variables. Results: LGB showed maximum predictive power using 14 features and 90 days before diagnosis. The LGB models achieved AUROC, specificity, and sensitivity were 0.83, 0.81, and 0.64 for the test dataset, respectively. The LGB-driven NPC predictive tool effectively differentiated patients into high-risk and low-risk groups (hazard ratio: 5.85; 95% CI: 4.75–7.21). The model-layering effect is valid. Conclusions: ML approaches using electronic medical records accurately predicted NPC occurrence. The risk prediction model serves as a low-cost digital screening tool, offering rapid medical decision support to shorten prediagnostic periods. Timely referral is crucial for high-risk patients identified by the model.