e13695 Background: While immune checkpoint inhibitors (ICIs) are increasingly used for cancer treatment due to favorable tolerability compared with chemotherapy (reference), immune-related adverse events (irAEs) have emerged as a major clinical challenge, as they may lead to treatment interruption and reduced antitumor efficacy. Early identification of patients at high risk for moderate-to-severe irAEs requiring treatment is critical. Although prior studies have shown the utility of machine-learning (ML) models for predicting irAE occurrence, accurate prediction of clinically significant irAEs using pre-treatment data remains limited. Methods: We analyzed data from the All of Us Research Program and identified 998 cancer patients treated with ICIs, including anti-PD-1, anti-PD-L1 and combination regimens. Immune-related adverse events were identified using diagnosis concept sets derived from prior literature, yielding 310 patients with documented irAEs. There were 68 moderate-to-severe irAEs, defined as events requiring systemic corticosteroid initiation within 30 days of irAE onset. Patients without any irAE diagnosis codes served as controls. Baseline features included demographics (age and sex), comorbidities associated with irAE risk (autoimmune disease and others), lifestyle factors (alcohol and smoking history), and available pre-treatment laboratory measurements. We evaluated three supervised machine-learning models: elastic-net regularized logistic regression, XGBoost, and LightGBM. The dataset was split 80/20 for training and testing and trained with five-fold cross-validation. Results: For prediction of any irAE occurrence, elastic-net logistic regression using baseline features achieved the highest performance with AUROC of 0.75, sensitivity of 0.64 and specificity of 0.77. Consistent with literature, the most influential predictors included autoimmune disease, chronic liver disease, ICI combined with chemotherapy, and alcohol consumption. Model performance decreased when predicting moderate-to-severe irAEs using baseline features alone. Including pre-treatment laboratory measurements modestly improved prediction, with the XGBoost being the best-performing model achieving an AUROC of 0.65 (95% CI: 0.48–0.82). While overall discrimination remained modest, SHAP analysis identified metabolic and inflammatory markers, including BMI (body mass index), glucose, and white blood cell count as well as common comorbidities as key contributors. Conclusions: While baseline risk factors predict irAE occurrence, identifying patients at risk for moderate-to-severe irAEs requires personalized immune profiling. Future work will integrate multimodal data, multi-institutional cohorts and biologically informed ML models to enable early risk stratification and guide clinical management.
Gong et al. (Thu,) studied this question.