C24-07 Multi-Task Model for Early Sepsis Prediction: Disentangling Infection and Organ Dysfunction

Rationale Sepsis is defined as life-threatening organ dysfunction caused by infection. Most machine learning models predict sepsis as a composite of organ dysfunction and infection. However, clinical decision making may benefit from disentangling these two distinct but related probabilities. A patient with a high likelihood of infection but low risk of organ dysfunction may be approached differently than one with high risk of organ dysfunction but low likelihood of infection. To address this gap, we developed a multi-task long short-term memory (LSTM) model that simultaneously predicts infection, organ dysfunction, and composite sepsis. Methods All patients presenting in the Emory Healthcare system were included. All patients had at least three hours of continuous monitoring data (i.e., did not have sepsis, transfer, discharge, or death within three hours). Hospitalizations were split into non-overlapping cohorts: training (15,346; balanced 1:1 sepsis vs. non-sepsis), validation (3,278; balanced), and testing (19,593; natural prevalence). The multi-task LSTM model used shared temporal features with three prediction heads for infection, organ dysfunction, and composite sepsis. It analyzed 3 to 36 hours of data, including vital signs, laboratory results, and medication records, to predict each outcome for two to four hours (2-4h) and for four to eight hours (4-8h) in advance. Model performance was evaluated using AUROC and calibration metrics. Results The study included 38,217 hospitalizations across Emory Healthcare hospitals, including 10,955 sepsis cases identified using CDC Adult Sepsis Event criteria. The LSTM model achieved an AUROC of 0.870 in composite sepsis prediction for the test set, with calibration of Brier score of 0.003 and expected calibration error 0.014. On the validation set, discrimination for infection and organ dysfunction prediction was strong (AUROC 0.85 and 0.79, respectively). Conclusions A proposed multi-task LSTM model can disentangle infection and organ dysfunction while maintaining high accuracy for overall sepsis prediction in real-world clinical conditions. By providing three distinct probabilities for infection, organ dysfunction, and composite sepsis, this approach may align better with clinical reasoning and supports more tailored interventions. This multi-output framework offers a more clinically actionable approach to early sepsis detection than conventional composite models.