When this chapter was written, the molecular track had closed pointing at a clear next destination: calibrated clinical prognosis for the diseases this mission is about, on the cohorts built for it — PRO-ACT and ADNI for longitudinal ALS and Alzheimer’s, and the Speech Accessibility Project for bulbar-onset speech. Those datasets were access-gated, and the applications took time.
Rather than wait, I did the rigorous thing on data already in hand: the Health and Retirement Study — a large longitudinal cohort available through a registered public release, with a usable research classification. Eight investigations. Every number script-generated and checksum-locked. All eight reproduce byte-identically.
That choice was deliberate, not a detour. Doing clinical-style prognosis end-to-end on public data kept the research moving while access was pending, exercised the ADM process on ground the molecular track never touched — a human longitudinal cohort, competing risks, subgroup fairness, a silver-standard label — and demonstrated the calibrated-honesty discipline later clinical work would require. The destination had not changed; this was the on-ramp to it, built on data anyone could check.
What I was building toward
The practical question is simple to ask and difficult to answer: two years from now, is this person likely to have crossed into dementia—and how much should anyone trust that number? That made calibration the headline metric rather than accuracy alone.
The outcome is the Langa-Weir classification, an algorithmic research proxy rather than a clinical diagnosis. The study could test whether the workflow was calibrated to that label. It could not establish clinical truth or support advice about an individual.
The decisions that shaped the build
Use public longitudinal data rather than wait
The HRS releases made it possible to exercise an end-to-end prognosis workflow while access to disease-specific cohorts was pending. Inputs were checksum-locked, the label was reproduced from documented constants, and all eight investigations were required to regenerate byte-identically.
Treat sample selection as part of the model
The primary analysis keeps people who remain self-respondents at the next interview. Because decline drives the switch to a proxy, the highest-risk people leave that sample earliest. I therefore ran an imputed-inclusive analysis in parallel and kept the transport limit beside every result it controls.
Set the complexity bar before comparing models
The minimal model used age, cognitive score, and prior cognitive status. Penalization and gradient boosting produced small, statistically detectable gains, but their point estimates stayed below the study’s operating threshold and their intervals crossed it. Formal equivalence was not established. Recalibration did not add decision-relevant benefit.
The evidence supported a provisional, decision-specific stop—not a claim that complexity never helps. Added fidelity had not earned the additional burden for this benchmark, and the uncertainty around that judgment remains part of the result.
Test the failure modes that aggregate scores hide
Competing-risk analysis showed that treating death as ordinary censoring materially overstates absolute incidence. Subgroup checks showed that good aggregate calibration did not establish fairness. Continuous-score forecasting improved mainly in the cognitively normal group, not where the clinical motivation was strongest. Those findings changed the interpretation more than another leaderboard comparison would have.
What I will not claim
- The model is calibrated to a research proxy, not clinical ground truth.
- Self-respondent selection limits transport to the full at-risk population.
- Subgroup gaps cannot be separated into label bias, measurement, sampling, case mix, and biology with these data.
- This is not a clinical tool, individual advice, external validation, or a statement about any person.
Where it points
The HRS study demonstrated the required discipline on a human longitudinal cohort before disease-specific work could proceed: calibrated uncertainty, competing risks, subgroup checks, and explicit refusal to treat a proxy label as clinical truth. That is what a validation study was meant to establish.
For the formal results, intervals, sources, and complete limitations, read the HRS study → Current program status is maintained on Research. — Michael