Abstract
PURPOSE: The goal of this retrospective analysis was to assess how well predictive models could determine which patients would develop liver chemistry signals during clinical trials based on their pretreatment (baseline) information. PATIENTS AND METHODS: Based on data from 24 late-stage clinical trials, classification models were developed to predict liver chemistry outcomes using baseline information, which included demographics, medical history, concomitant medications, and baseline laboratory results. RESULTS: Predictive models using baseline data predicted which patients would develop liver signals during the trials with average validation accuracy around 80%. Baseline levels of individual liver chemistry tests were most important for predicting their own elevations during the trials. High bilirubin levels at baseline were not uncommon and were associated with a high risk of developing biochemical Hy's law cases. Baseline γ-glutamyltransferase (GGT) level appeared to have some predictive value, but did not increase predictability beyond using established liver chemistry tests. CONCLUSION: It is possible to predict which patients are at a higher risk of developing liver chemistry signals using pretreatment (baseline) data. Derived knowledge from such predictions may allow proactive and targeted risk management, and the type of analysis described here could help determine whether new biomarkers offer improved performance over established ones.