Machine learning to predict immune-related adverse events in cancer patients treated with ICIs.

Abstract

e13620 Background: Immune-related adverse events (irAEs) secondary to treatment with immune checkpoint inhibitors (ICIs) pose a major challenge in cancer treatment. Although irAEs result in substantial morbidity, there is still no reliable tool for predicting these events based on pre-treatment clinical and laboratory data. Machine learning algorithms (MLA) have emerged as a powerful tool in cancer research, offering new opportunities to analyze complex, multidimensional data and uncover patterns that can enhance prediction and treatment strategies. In this study we aimed to develop a personalized prediction model to predict irAEs using different MLA based on an individual’s features prior to ICIs treatment. Methods: We extracted data on patients' cancer diagnoses and treatments from the Cedars-Sinai Medical Center database for 2012–2024, focusing on those treated with ICIs and those with irAEs. We identified demographic and clinical data including sex, age, race, ethnicity, BMI (Body Mass Index), smoking history, alcohol consumption, and co-morbidities, in addition to laboratory measurements, including neutrophil, lymphocyte, and eosinophil absolute counts, as well as the neutrophil-to-lymphocyte ratio. We utilized the TPOT package and scikit-learn to develop a machine learning classification model, exploring sampling strategies like under sampling to balance the dataset. The most generalizable model was trained on a balanced dataset and tested on an unbalanced one to reflect real-world scenarios. The dataset was split 80/20 for training and testing, with five-fold cross-validation to prevent overfitting. We optimized TPOT parameters, including runtime and epochs, to improve model accuracy and performance. Results: Our database includes 4 million patients. Among them approximately 15,000 patients had been diagnosed with cancer (excluding non-melanoma skin cancer), 4,302 patients treated with ICIs and 1,089 were identified as having irAEs based on irAE-directed treatments (steroids and infliximab) in combination with laboratory measurements and oncologists' notes. The model with the best performance was Random Forest classifier with an AUC 0.64, specificity of 0.63 and sensitivity of 0.70. Patients ranked in the top quartile by the model's score were found to be twice as likely (2.14, SE 0.079) to develop irAEs compared to those in the bottom quartile. The 5 most predictive features for the model were as following, age, sex, ethnicity, race and alcohol consumption. Conclusions: This study demonstrates the potential of MLA in predicting irAEs among patients treated with ICIs that could lead to early intervention. These findings highlight the feasibility of utilizing pre-treatment clinical and laboratory data for irAEs prediction. Future efforts should focus on refining models and incorporating additional variables to improve predictive accuracy and clinical applicability.