Abstract Body (Do not enter title and authors here): Introduction: Traditional risk models predicting mortality in cardiogenic shock (CS) rely on static data snapshots, failing to capture the temporal dynamics represented in electronic health records. Conventional machine learning (ML) models can incorporate some temporal information but often struggle to model the complex, non-linear, and long-range relationships present in clinical time series data. Foundation models, pretrained on large, diverse time series datasets, offer a powerful alternative by capturing richer temporal patterns and making predictions in dynamic clinical settings.
Hypothesis: We hypothesized that a ML approach using a pretrained multi-task time series model (UniTS) could be fine-tuned to predict mortality in patients with CS admitted to the cardiac intensive care unit (CICU) using high-resolution, multivariable clinical time series data.
Methods: We performed a retrospective analysis of patients with CS admitted to the Brigham and Women’s Hospital CICU (2015-2024). Patients were split into training (80%) and validation (20%) cohorts. For each observation, 24h of clinical data were used to fine-tune a pretrained UniTS model to develop: (1) a dynamic model generating rolling 24h mortality predictions every 6h, and (2) a static model predicting overall in-hospital mortality after the first 24h of CICU admission. Validation performance was evaluated using area under the curve (AUC). The static model was compared to the IABP-Shock II risk score and SCAI shock stages for the same task using the DeLong test.
Results: Among 2,109 admissions with CS (median age 68 years, 38% women), 25% were AMI-related, and 39% received temporary mechanical circulatory support. In-hospital mortality was 37%, with a median time to death of 3.0 days following CICU admission. The final model included 31 clinical variables (Fig A). The dynamic model achieved a per-prediction AUC of 0.88 for 24-hour mortality. The static model achieved an AUC of 0.83 for in-hospital mortality, significantly outperforming the IABP-Shock II risk score (AUC 0.74; p = 0.03) and SCAI Shock stage (AUC 0.61; p < 0.001) (Fig B).
Conclusion: Leveraging the full temporal and multivariable complexity of CICU time series clinical data, fine-tuned ML foundation models accurately predict both very early and in-hospital mortality. By substantially outperforming traditional risk stratification methods, this time series modeling approach offers a promising tool for dynamic risk assessment in CS.