Multimodal Early Septic Shock Prediction Model using Lasso Regression with Decaying Response

Sepsis is one of the leading causes of in-hospital deaths. In this paper, we propose a multimodal early prediction system for septic shock, which is a serious complication for patients with sepsis. Our system utilizes patients vital and laboratory time series data in combination with medical notes to predict patients at risk of septic shock in real-time. Our model achieves 0.89 ROC AUC and a median detection time of 30.64 hours before the onset of the septic shock event at a specificity of 0.67. Our proposed method uses lasso regression with modified response decay to model the risk as we go further away in time from the septic shock event. We compare the performance of our model to baseline and state-of-the-art models used to predict early deterioration and septic shock in ICU patients. Our model shows improvement in both ROC AUC and median detection time at all specificities. This translates into an earlier prediction time at a higher true positive rate for the same false positive rate. Moreover, we investigate the ROC AUC and median detection time trade-off and propose an automated way for model selection under such trade-off. We also demonstrate how the model selection step could also be modified to optimize for a predefined user utility.