Methodology

Statistics — ORACLE methodology

Cohort

ORACLE is built on N=637 SynCardia 70cc Total Artificial Heart device operations from the STS INTERMACS registry. The unit of analysis is the device operation, not the patient — some patients had more than one implantation, and each is treated as an independent observation. The analytic cohort for the companion manuscript (“Living With a Total Artificial Heart”) is a more stringent N=593 subset after exclusions; ORACLE uses the broader 637 cohort to support its exploratory role. ORACLE's performance metrics are therefore not directly comparable to the manuscript's numbers.

Endpoint

The primary outcome is device success at 12 months. Device success is a composite: a device operation is counted as a success if the patient is alive on the device at 12 months, or if the patient was transplanted within 12 months of the implant. Device failure is on-device death by 12 months. Transplantation is treated as a competing event — at the time of transplant, the device's purpose is fulfilled and the patient leaves the at-risk pool. Patient outcomes after transplantation or any other form of explant are not analysed by ORACLE; this is a device-level analysis.

Model

Three logistic regression models are fitted to the same 13-variable feature set: probability of device success at 12 months (the primary output), probability of on-device death at 12 months (its complement), and probability of transplantation within 12 months (a sub-classification of success). The three are fitted independently; in the user interface, the displayed alive-on-device probability is derived as (device success − transplantation) to ensure the three displayed outcomes sum to unity. Coefficients are sign-constrained based on clinical priors — see the companion manuscript methods for the rationale (small cohort, low events-per-variable, established directional effects of standard predictors).

Validation

Internal validation by optimism-corrected bootstrap (B=200) yields a c-statistic of 0.659 across all three outcomes. Temporal validation in a held-out era yields c=0.65, essentially identical to the bootstrap-corrected estimate, indicating stable performance over time without substantial overfitting. Calibration intercept −0.12, slope 0.72 on the temporal holdout — modest under-fitting at high predicted risk, no formal recalibration applied because external validation in an independent cohort is the appropriate next step. Brier score 0.216–0.220.

Competing risks

Cumulative incidence curves are computed by the Aalen-Johansen estimator, which is the correct treatment for competing risks (death and transplant compete; only death is a failure). Device success over time is plotted as the complement of cumulative incidence of on-device death.

Limitations

ORACLE has not been externally validated against an independent cohort. Discrimination is modest. The model uses pre-implant variables only — no post-implant information leaks. ORACLE is a research prototype and is not a medical device. See the disclaimer on the landing page.

For the full methodology refined for peer review, see the companion manuscript currently in preparation.