Multimodal transformer decoder for the early detection of clinical deterioration in heart failure: Insights from temporal lead-time analysis
K Minjun, J H MoonAbstract
Background
Predicting clinical deterioration including in-hospital mortality, transition to hospice care, and readmission is essential for guiding early interventions and advanced care planning in hospitalized heart failure (HF) patients.
Purpose
This study developed a multimodal transformer decoder model that analyzes sequential chest X-rays (CXRs) and electronic health records (EHRs). This model acts as a comprehensive safety net, predicting these composite adverse outcomes early across varying lead-time windows.
Methods
The model was trained using a dataset of 13,821 admissions comprising sequential CXRs (≥2 AP/PA projections) and paired EHRs. To evaluate the model's capacity as an early safety net, a test cohort of 677 heart failure admissions was utilized. To assess early warning capabilities, the analysis stratified the test data into four lead-time windows prior to the event (death or discharge): < 24h (n=164), 24–48h (n=108), 48–72h (n=88), and > 72h (n=317).
Results
Within the test cohort (mean age 71.2 ± 14.4 years; 54.8% male), the multimodal transformer decoder model demonstrated significant predictive utility in the < 24h window, achieving an area under the curve (AUC) of 0.838 (95% CI, 0.745–0.930), a sensitivity of 60.0%, and a specificity of 91.8% (p < 0.0001). The model maintained robust discriminative ability at 24–48h (AUC 0.808, 95% CI, 0.687–0.929) and > 72h (AUC 0.708, 95% CI, 0.617–0.798) prior to the event (both p < 0.0001). Although discriminative performance dipped at 48–72h (AUC 0.624, 95% CI, 0.450–0.798), specificity remained consistently high across all intervals (85.3%–93.2%). This suggests that the model maintains promising predictive potential even beyond the 72-hour threshold prior to clinical deterioration.
Conclusion
By integrating sequential CXRs and EHRs, the multimodal transformer decoder effectively anticipates clinical deterioration in heart failure patients even beyond 72 hours prior to the event.data Table 1Performance chart 1