Predicting Wilson Disease Progression Using Machine Learning with Real-world Electronic Health Records

Background: Wilson Disease (WD) is a rare genetic disorder that causes excessive copper accumulation. Research on the natural history of WD is limited. Understanding WD progression over time will facilitate design of clinical trials and personalized medicine approaches.

Objectives: To identify predictors for WD progression from time of diagnosis to cirrhosis, liver failure, and death (all-cause) and to predict patient-level risk of progression to these endpoints at 1, 3, and 5 years after WD diagnosis.

Methods: A retrospective natural history cohort study of adult patients with incident WD diagnosis (first recorded in the study period and no WD diagnosis in the preceding year) was conducted using the US Optum EHR data between 1/1/2007 and 6/30/2020. LASSO Cox regression, Random Survival Forest (RSF), and XGBoosting (XGB) models were used to identify important predictors associated with progression to cirrhosis, liver failure, and death. An ensemble method used weighted averages to rank predictors across models and find the strongest predictors for each outcome. The resulting predictors were reviewed by clinical experts for biological plausibility. RSF and XGB models were developed for patient-level prediction using the top ensemble predictors ( < 15 predictors). The resulting models were validated with an independent sample cohort. C-index and dynamic AUCs were used to evaluate model performance.

Results: In a total of 3,559 patients with WD, 310 developed cirrhosis, 255 developed liver failure, and 604 died during the study period. Increased age at WD diagnosis, history of alcoholism, elevated AST and bilirubin levels within 3 months of WD diagnosis, and presence of both neurologic and hepatic comorbid conditions were the most common predictors for progression to the study endpoints. XGB had superior predictive performance compared with RSF by C-index (cirrhosis: 0.73 for XGB vs 0.71 for RSF; liver failure: 0.78 for XGB vs 0.77 for RSF; death: 0.80 for both XGB and RSF). Individual patient risks were assessed for progression to cirrhosis, liver failure, and death using the XGB model with the top ensemble predictors. The dynamic AUC was 0.74 at Year 1, 0.68 at Year 3, and 0.69 at Year 5 for cirrhosis; 0.82 at Year 1 and 0.77 at both Year 3 and Year 5 for liver failure; 0.81 at Year 1 and 0.82 at both Year 3 and Year 5 for death.

Conclusions: This study identified the most influential clinical predictors and assessed patient-level risk of WD progression using machine learning. As clinical evidence evolves to include more real-world data, results from machine learning prognostic models will increase understanding of disease natural history and etiology and may help improve clinical trial design and guide individualized clinical care.