Towards Remote Continuous Monitoring of Cytokine Release Syndrome

Area of Science:

• Oncology
• Immunotherapy
• Machine Learning
• Critical Care Medicine

Background:

• Cytokine release syndrome (CRS) is a severe adverse event in cancer immunotherapy.
• Accurate monitoring and prediction of CRS severity are challenging.
• Early intervention is crucial for managing CRS and maximizing immunotherapy benefits.

Purpose of the Study:

• To develop and evaluate a machine learning algorithm for forecasting CRS severity.
• To predict CRS (no CRS, mild, severe) within 24 hours using vital signs and Glasgow Coma Scale (GCS).
• To assess the impact of different data availability scenarios on prediction accuracy.

Main Methods:

• An XGBoost-based machine learning algorithm was developed.
• The algorithm utilized vital signs and GCS questionnaire inputs.
• Models were trained and evaluated on a cohort of 1,139 oncology patients in ICU settings.

Main Results:

• The algorithm achieved a micro-average AUC of 0.94 for predicting 3 CRS grades when using all time-series features.
• Models using data from the preceding 24 hours showed a micro-average AUROC of 0.88.
• Removing blood pressure or GCS inputs significantly decreased model performance (p<0.05).

Conclusions:

• Machine learning, particularly XGBoost, can accurately predict CRS severity in oncology patients.
• Vital signs and GCS are critical predictors of CRS.
• Timely CRS prediction facilitates early intervention, improving patient outcomes with immunotherapies.