Identification of Prognostic Clinical Features in Grade 4 Immune-Related Adverse Events: A Triangulation Study.

BACKGROUND

Grade 4 immune-related adverse events (irAEs) is life-threatening complications of immune checkpoint inhibitor therapy. Due to its rarity and scarcity of data, there is a lack of systematic research on key factors influencing poor prognosis. This exploratory study aimed to identify clinical features robustly associated with mortality in patients with grade 4 irAEs. Given the extremely small sample size and high‑dimensional data, a triangulation approach integrating traditional univariate statistics and machine learning to maximize the reliability of feature selection.

METHODS

This study included 26 cancer patients admitted to the ICU for grade 4 irAEs. To maximize robustness from limited data, a "triangulation" approach was employed. Prognostic features were independently identified through two parallel approaches: (1) traditional univariate statistical analysis, and (2) multiple machine learning algorithms evaluated by Leave-One-out Cross-Validation. Features consistently highlighted as significant by both independent methodologies were integrated to form a final high-confidence feature set.

RESULTS

Univariate analysis identified 21 features significantly associated with mortality. Machine learning analysis refined this to 11 important features. Through "triangulation", 8 features were consistently validated: body mass index and VEGF-inhibitors were inversely associated with mortality, while vasopressor therapy, oxygen therapy, lactate levels at day 1 and 2, pneumonia and percentage of neutrophils, exhibited a positive correlation with the mortality.

CONCLUSION

This small-sample exploratory study identified 8 routinely available early ICU clinical features robustly associated with mortality in grade 4 irAEs patients using a "triangulation" framework. These characteristics highlight the pivotal roles of shock, respiratory failure, and inflammation. While not directly constructing a clinical prediction model, they may facilitate early risk stratification and provide hypotheses for prioritized validation in future large-sample studies.