Abstract Rationale Despite decades of research there has been little success in randomized clinical trials targeting mortality in sepsis. This failure is presumed to be due to the underlying heterogeneity of sepsis. In the setting of heterogeneity, there is emerging evidence that the causal link between sepsis and outcomes is mediated by specific narrow organ failures and the interpretation of prognosis by clinicians, patients and their surrogates. However, this dynamic interplay has not been described at scale. Large language models (LLMs) allow for the systematic analysis of natural language and may provide insight into these processes. Methods We did a retrospective cohort study using the MIMIC-IV database. Sepsis patients were identified using Sepsis-III criteria, in-hospital deaths were noted and deidentified discharge summaries were extracted. Using the 27 billion parameter open-source model, MedGemma2, a sequential, multi-agent LLM pipeline was used to systematically analyze unstructured clinical narratives from each discharge summary. A chronological step-by-step causal chain from insult to outcome was generated and a primary reason for death was applied. Predefined options included: refractory shock (RS), refractory hypoxemia (RH), comorbid withdrawal of care (CWC), neurologic withdrawal of care (CWC) and sudden cardiac arrest (SCA). Results 820 patients with sepsis died during their index hospitalization in this cohort. CWC N = 282 (34%) was the most common LLM labeled reason for death followed by RS N = 257 (31%), NWC N = 189 (23%), RH N = 66 (8%) and SCA N = 26 (3%). The baseline demographics, comorbidities, infectious sources, and day one SOFA scores are shown in the Table. Although there were numeric differences, the distribution of SOFA scores on day one was not statistically significant. Moreover, the baseline SOFA score utilizing each sub score independently adequately predicted all-cause mortality AUC 0.75 but was a poor predictor of any specific reason for death, AUC for CWC = 0.57, RS = 0.55, NWC = 0.64, RH = 0.53, SCA = 0.54. Conclusions Local, open-source large language models can label the reasons for death amongst patients with sepsis. Comorbid withdrawal of care is the most common reason for death identified through this analysis. The baseline distribution of organ dysfunction poorly predicts the eventual reason for death. This work highlights that reasons for death in sepsis are varied and not fully explained by organ dysfunction. More research is needed to validate this approach, and to define and predict actionable reasons for death with the goal of precision phenotyping and predictive enrichment in our clinical trials. This abstract is funded by: NHLBI
Schenck et al. (Fri,) studied this question.