Extended Data Fig. 2: Individual Patient Survival Probability.
From: Arrhythmic sudden death survival prediction using deep learning analysis of scarring in the heart
a. Survival probability curves are shown for an example patient in the external test set who experienced sudden cardiac death from arrhythmia (SCDA, left display), and for one who did not (No SCDA, right display). Survival probability curves are plotted over time for Survival Study of Cardiac Arrhythmia Risk (SSCAR) (solid, blue), Cox proportional hazards (Cox PH) model (dashed, green) on the clinical covariates, Kaplan-Meier estimator (dot-dashed, purple), together with the indicator ground truth (dotted, black). For the patient with SCDA, SSCAR crosses the 50% survival probability threshold considerably closer to the SCDA time, as compared to the alternative curves, highlighting the model’s high calibration. For the censored patient (no SCDA), SSCAR estimates higher survival probability at the time of non-SCDA event compared to the other models. b. Examples of SSCAR’s predicted probability distributions for the time to SCDA (shaded areas, pdf(TSCDA)) for two patients in the external test set who experienced SCDA (P1, blue and P2, orange). The predicted times to event (Predicted TSCDA) are depicted as solid vertical lines (peaks of distributions); actual times (Actual TSCDA are depicted by dotted vertical lines. Note that SSCAR has a larger prediction error in P2 compared to P1, seen on the graph as the distance between the respective solid and dotted lines. However, SSCAR ‘recognizes’ the inaccurate TSCDA prediction and compensates for that by also predicting a more spread out distribution (larger scale parameter) for P2. This direct relationship between the prediction error and predicted scale parameter holds more generally for the entire dataset, suggesting SSCAR learns to quantify the degree of inaccuracy in the TSCDA prediction.
