The research. Honest numbers.
Two kinds of work: original work aimed at Alzheimer’s and ALS, and validation studies where a public benchmark grades the method. Every figure is script-generated from checksum-locked data — never typed by hand — and every caveat is kept in the record alongside the result.
← The MissionBy Michael Key · ORCID
Original, mission-driven work — the contribution itself. Each study aims at something not already established: a calibrated model that knows where it’s blind, an honest negative result, or a tool a family or clinician could actually use.
Molecular Aggregation & Variant Effect
Twelve investigations on public deep-mutational-scanning atlases (Aβ and TDP-43). A calibrated Aβ variant-effect model that knows where it’s blind; two clean negatives on cross-disease generalization and scaling.
Speech-Based ALS Prognosis
Calibrated forecasts of how ALS progresses, function by function. The speech signal comes from DementiaBank and the Speech Accessibility Project; longitudinal ALSFRS (ALS Functional Rating Scale) trajectories come from PRO-ACT, with the Alzheimer’s clinical track grounded in ADNI. Data-access applications are in progress; the molecular study used only public data, and the speech and clinical tracks begin only once access is granted. Raw cohort data stays local and never enters a chat or AI context.
Validation studies on public neurodegeneration cohorts, where a published benchmark is the answer key. They don’t claim new discoveries — they show that the discipline (locked input → versioned script → audited number, with calibration as the headline) holds up on real patient-level data, and they’re how this work earns access to the gated cohorts the contributions need. Every result is reported as a reproduction graded against a public baseline, with its honest delta — never as a finding of my own.
Forecasting Cognitive Decline & Incident Dementia
Can current-wave predictors forecast a later dementia diagnosis with trustworthy probabilities? Built on the Health and Retirement Study with Langa-Weir classifications — public data, already in hand. Calibration is the headline; discrimination is reported against a range from comparable studies in other cohorts, never claimed as beating an HRS benchmark.
How Much Does Leaky Evaluation Inflate Accuracy?
A methods study on small public neuro datasets (Parkinson’s voice, gait). It measures the optimism gap between record-wise evaluation — which leaks the same subject across train and test — and honest subject-level evaluation. That leakage-inflation gap is the result, and the protocol every gated-cohort study below reuses.
Study page will appear when the first investigation is complete and independently reproducible.
Predicting MCI→AD Conversion
Can baseline imaging and cognition forecast conversion to symptomatic Alzheimer’s within a fixed window, with calibrated probabilities? A pre-registered “simple matches complex” test, with a guardrail to report the complex model if it wins. Built against stand-in (synthetic) data until access is granted.
Study page will appear when the first investigation is complete and independently reproducible.
Speech-Based Alzheimer’s Detection
Can speech alone distinguish Alzheimer’s from healthy controls, with calibrated confidence? Reproduced against the published ADReSS benchmark as a cited target, reported with an honest delta — not a claim of my own. Built against stand-in (synthetic) data until access is granted.
Study page will appear when the first investigation is complete and independently reproducible.
Longitudinal Cognitive Forecasting
Can cognitive trajectories be forecast with calibrated intervals — not just point predictions? Benchmarked against the TADPOLE challenge winners on interval calibration, reproduced rather than out-competed. Built against stand-in (synthetic) data until access is granted.
Study page will appear when the first investigation is complete and independently reproducible.
Every reported number is regenerated by a versioned script, recorded in a script-written results.json, and locked to a sha256 checksum of the source data. If a number drifts between runs, the build gate catches it before the site updates. Honest caveats are recorded next to every claim, not appended as disclaimers.