ShockModes:A Multimodal Model for Prognosticating Intensive Care Outcomes from Physician Notes and Vitals

Abstract

Objective. Shock Index (SI) is widely used for prognosticating outcomes in ICU and emergency settings. We aimed to create a multi-modal early warning system (EWS) for development of abnormal shock index using routinely available vitals and clinical notes. Material and Methods. 17,294 ICU-stays in MIMIC-III data were scored for SI. A new episode of abnormal SI was defined as SI > 0.7 for >30 minutes AND preceded by >=24 hours of normal SI. ICU stays with <24 hours admission, or SI >0.7 within the first 24 hours of admission, or missing SI in >50% in the 24 hour early warning window were excluded, leaving a final cohort of 337 normal and 84 abnormal SI instances. 3117 features from vitals time-series combined with BERT-based features from clinical notes were used to train a battery of machine learning models. The best multimodal pipeline (ShockModes) was assessed for interpretability using SHAP features. Results. Vitals-based, notes-based and multi-modal classifiers achieved the best sensitivity of 0.81, 0.81, and 0.83 with corresponding specificity of 0.92, 0.99, and 0.94 respectively, thus demonstrating the potential of ShockModes for early detection, while preventing false alarms. Global SHAP values revealed Fourier-features of heart rate and heparin sodium prophylaxis as top features. Sensitivity of early detection was highest in acute respiratory failure and chronic kidney disease patients. Conclusion. The multimodal, interpretable early warning system ShockModes can be used for prognosticating SI based outcomes in ICU and emergency settings.Objective. Shock Index (SI) is widely used for prognosticating outcomes in ICU and emergency settings. We aimed to create a multi-modal early warning system (EWS) for development of abnormal shock index using routinely available vitals and clinical notes. Material and Methods. 17,294 ICU-stays in MIMIC-III data were scored for SI. A new episode of abnormal SI was defined as SI > 0.7 for >30 minutes AND preceded by >=24 hours of normal SI. ICU stays with <24 hours admission, or SI >0.7 within the first 24 hours of admission, or missing SI in >50% in the 24 hour early warning window were excluded, leaving a final cohort of 337 normal and 84 abnormal SI instances. 3117 features from vitals time-series combined with BERT-based features from clinical notes were used to train a battery of machine learning models. The best multimodal pipeline (ShockModes) was assessed for interpretability using SHAP features. Results. Vitals-based, notes-based and multi-modal classifiers achieved the best sensitivity of 0.81, 0.81, and 0.83 with corresponding specificity of 0.92, 0.99, and 0.94 respectively, thus demonstrating the potential of ShockModes for early detection, while preventing false alarms. Global SHAP values revealed Fourier-features of heart rate and heparin sodium prophylaxis as top features. Sensitivity of early detection was highest in acute respiratory failure and chronic kidney disease patients. Conclusion. The multimodal, interpretable early warning system ShockModes can be used for prognosticating SI based outcomes in ICU and emergency settings.