Fig 1.
Diagrammatic representation of the study.
Table 1.
Comparison of characteristics between patients who experienced in-hospital mortality and those who were discharged alive.
Fig 2.
Visual timelines of two patients, one who was discharged alive (left) and one died during hospital stay (right). Time from admission (0–48 hours) is on the x-axis and variables grouped by category are on the y-axis. Continuous variables (e.g. temperature) are normalized so that black indicates minimum and white indicates maximum values. Binary variables (e.g. vasopressor medication) are represented so that black indicates absence and white indicates presence of the variable. As shown, the visual timeline of the patient who died depicts increased activity in the intervention and medication areas as well as multiple changes to vital sign readings as compared to the visual timeline of the patient who was survived the hospitalization.
Table 2.
Model discrimination for predicting mortality on the test dataset (n = 34,747 admissions).
Fig 3.
Original (left) and Grad-CAM attention-based heatmap (right) derived for the CNN-RL model for a min-max normalized standard ordering visual timeline of a test patient who died in-hospital. As can be seen, the area of interventions and diagnostic testing have been highlighted as the most contributing areas for predicting a high probability of mortality for this patient.