Predicting immune checkpoint inhibitor-related hepatitis using electronic health records of patients.

e13564 Background: Immune checkpoint inhibitor (ICI) therapies have shown impressive results in treating oncology patients. However, some patients exhibit immune-related adverse events (irAE-s), one significant irAE is autoimmune hepatitis. Oncologists routinely screen for hepatic toxicity with a complete metabolic panel prior to each ICI administration. Predictive modeling of irAE-s based on patient factors has the potential to help guide treatment selection and monitoring protocols. We have developed a widely usable model based on patient history and routinely collected standard blood panels that can predict whether a patient will experience hepatitis with ICI administration. Methods: We defined irAE hepatitis as any single value of AST, ALT, or alkaline phosphatase three-times the upper limit of normal (ULN) as following ICI treatment. The goal was to determine whether the level of the biomarkers exceed this threshold within certain, pre-defined time-windows, as determined by medical experts. We used feature engineering to compress the time-series of lab panels into single meaningful statistical descriptors, such as mean or maximum of these vectors. The dataset was highly unbalanced with many more negative cases (out of the 3231 patients, depending on the window length used for feature generation, 100-400 positives were found), which warranted the application of synthetic resampling methods. Finally, we trained various ensemble models ( e.g., random forest, gradient boosting), on both the resampled and original dataset, to obtain the final, predictive model for the likelihood of irAE hepatitis. Models were tuned to favor high recall and lower precision (identification of patients for increased monitoring) or moderate recall and moderate precision (maximizing F1-score). Results: We explored several modelling methods, such as KNN, Logistic Regression, Random Forests (RF), Gradient Boosting (GB), and stacking. GB without resampling produced the best model with the moderate recall and precision. To achieve high recall with low precision, we needed to resample the dataset with random majority class under sampling and then use RF (see table below for exact results). Conclusions: In this study, we show contemporary machine learning methods can be used as a screening tool for patients at risk for irAE hepatitis. This method could be used to identify patients who would benefit from additional laboratory monitoring between ICI administrations or guide clinical decisions about therapy cessation in advance of toxicity. Additionally, these methods may be further developed and adapted to improve clinical trial exclusion criteria for patients most likely to develop irAE hepatitis.[Table: see text]