Longevity Metrics AI/ML Development Study

Agreement between the model's value and the reference standard, established as non-inferiority to a qualified human scorer where a human reference exists. Gates deployment as a clinical decision aid.

Accuracy of predicting, from video or from the performed part of a record, both the screening results the participant did not obtain, against the measured value; and the derived scores computed from a complete record, including the Longevity Score, biological age, and estimated age at death, against the value the scoring engine produces from the full measured record.

Accuracy of predicting remaining life expectancy from video and library data, against observed mortality through linkage to the 100-Year Human Aging Study.

Accuracy of predicting the physician's clinical judgment where no determining measurement exists, against the sealed record. Applies to models that predict a clinical determination rather than a measured value. Where a determining test exists but was not performed, the outcome falls under Primary Outcome 2.

Participant injuries or adverse events attributed to measurements added under this study, such as the sit-to-rise test. Ascertained from two independent sources so that an event missed by one is still captured: the tester logs any fall or injury at the time of screening, and the participant reports separately on the post-screening form.

Conditional on non-inferiority, whether the model scores more reliably or precisely than a qualified human scorer.

Agreement between the model's stated confidence and its observed accuracy.

Where a test is captured twice back to back, the stability of the model's output across the two captures. Not applicable where repeat capture is impractical, or where a second effort measures fatigue because the test is performed to failure.

Model performance on partial or incomplete input relative to complete input.

Three physician-recorded fields per result, identical to the structured judgment recorded under the Health Ahead Comparative Effectiveness Study, so that one instrument governs physician review of model output across the platform. Agreement with the model read, 1 to 10, 1 strongly disagree to 10 strongly agree. Safety, the potential for patient harm had the read been acted upon as written, 1 to 10, 1 high potential for serious harm to 10 no potential for patient harm. Override, binary, recording whether the released interpretation differs in any substantive respect from the read. The two rating scales ascend toward the better state. Agreement records what the physician thought of a read; override records what the physician did with it, and the two diverge routinely.

Whether added parameters improve prediction of 100-Year outcomes over the base measure. Reported once Primary Outcome 3 is estimable.

Incremental value of video-derived information over standard screening for remaining life expectancy, through the 100-Year linkage. Reported once Primary Outcome 3 is estimable.

Accuracy of predicting new chronic disease onset, through the 100-Year linkage. Reported by predictor set, model and physician, so the prognostic comparison is made on a common outcome.

Concordance between predicted and actual cause of death, through the 100-Year linkage.

Accuracy of predicting the timing of functional disability onset, through the 100-Year linkage.

Change in predictive performance with environmental distance from the validated envelope. Residuals are returned to the Human Observatory Study.

Stability of performance across calendar time and across a participant's repeat screenings.

Accuracy of predicting the change in a measure between visits, and of predicting outcomes from that change. Applies to every model with repeat captures. The within-participant change and its pace are tested against death, disease, and functional decline alongside the cross-sectional value.