Multimodal digital biomarkers detect early cognitive decline
PrimaryAltoida's theory is that brief, neuroscience-driven digital tasks can reveal early cognitive and functional impairment because motion, speech, and touch behavior during complex tasks reflects subtle neurocognitive dysfunction before it is captured by conventional clinical assessment. Machine-learning models trained on these sensor-derived patterns should identify mild cognitive impairment and Alzheimer's-related risk earlier and more objectively than standard screening alone. Testable predictions include higher sensitivity for early or subthreshold MCI than clinician classification or traditional scales, reproducible sensor signatures across visits, and association of Altoida-derived classifications with later clinical progression or Alzheimer's biomarker status.
Popperian evaluation
The biological foundation is real: motor slowing, speech degradation, and fine-motor control loss are documented correlates of early neurocognitive dysfunction, and the insensitivity of traditional scales like MMSE to early-stage changes is widely acknowledged in the Alzheimer's literature. Where the premise weakens is the inferential leap from 'sensor data contains signal' to 'ML models trained on that signal outperform clinician assessment.' The pilot study (n=21 PD patients) reported 100% sensitivity but 26.7% specificity and Cohen's κ of 0.17, which is near-chance agreement. That pattern is consistent with an over-sensitive threshold, not a validated biomarker. The premise that multimodal sensor streams reflect dysfunction before clinical detection is plausible but unproven at the level claimed.
Supporting evidence: Motor, speech, and touch behavioral changes in early MCI are documented across multiple neuropsychological literatures, independent of Altoida's work.; The digital biomarkers review (fc9cd264) confirms that traditional scales lack sensitivity to early-stage cognitive changes and are affected by rater variability.; The pilot (d9aaae07) did capture all six clinician-classified PD-MCI cases, consistent with the premise that sensor data contains at least some signal.
Counter evidence: The pilot's specificity was 26.7% (95% CI: 7.8–55.1%) and accuracy was 47.6%, worse than a coin flip, with κ=0.17 indicating near-chance agreement with clinical classification.; Wide confidence intervals on sensitivity (54.1–100%) mean the 100% figure is statistically uninformative at n=21.; No evidence that the ML model's internal features map onto identifiable neurocognitive substrates, leaving the 'neuroscience-driven' framing unsupported by mechanistic validation.
The theory's central explanatory claim is that the 11 additional MCI cases flagged beyond clinician classification represent true early or subthreshold impairment detected by superior sensor sensitivity. But the competing explanation is simpler and at least as consistent with the data: the classification threshold is set too low, producing false positives. Without longitudinal follow-up showing that those 11 patients actually progressed to clinical MCI or dementia, the theory cannot distinguish its own signal from noise. No published data links Altoida-derived classifications to later clinical progression or Alzheimer's biomarker status (amyloid PET, CSF tau, etc.). The DSAD/ADAD conference proceedings (8c47177d) discuss biomarker-based progression models but provide no direct evidence for Altoida's digital biomarker validity. The theory explains its own pilot result, but the pilot result is equally well explained by over-classification.
Supporting evidence: The theory accounts for why 11 patients were flagged beyond the six clinician-classified cases, framing them as subthreshold MCI.; The digital biomarkers review acknowledges that continuous digital monitoring could detect changes missed by periodic clinical visits.
Counter evidence: Low specificity (26.7%) and PPV (35.3%) are equally consistent with over-classification rather than superior early detection.; No longitudinal data confirms that the additionally flagged patients went on to develop clinical MCI or dementia.; No direct comparison with Alzheimer's biomarker status (amyloid, tau, neurodegeneration markers) validates the digital biomarker against a ground truth.; The pilot was in Parkinson's disease, not Alzheimer's, so generalization to the primary claimed domain is itself untested.
The theory generates three concrete, testable predictions that could genuinely refute it: (1) higher sensitivity for early/subthreshold MCI than clinician classification or traditional scales, (2) reproducible sensor signatures across repeated visits, and (3) association of Altoida classifications with later clinical progression or Alzheimer's biomarker positivity. Each prediction specifies what would count as failure. A well-powered study showing no sensitivity advantage over MoCA or ACE-III would damage prediction 1. Poor test-retest reliability would kill prediction 2. A longitudinal cohort where Altoida-flagged subjects do not progress at higher rates would falsify prediction 3. The predictions are specific enough that negative results would be informative, not just reinterpreted. The reason this does not score higher: prediction 2 (reproducibility) has zero supporting publications and no data at all, and the existing pilot data on prediction 1 is so underpowered that it neither confirms nor refutes. The theory is well-structured for falsification but largely untested.
Supporting evidence: The pilot study explicitly acknowledges its own limitations (wide CIs, small n, no longitudinal confirmation) and calls for the exact studies that would test these predictions.; Sensitivity, specificity, and κ against clinician classification are clearly operationalized metrics that a larger trial could confirm or refute.; Association with Alzheimer's biomarker status is a binary, measurable endpoint.
Counter evidence: Test-retest reproducibility (prediction 2) has no supporting data whatsoever, listed at low confidence in the reasoning graph.; The theory does not specify what level of specificity would count as adequate, leaving room for post-hoc threshold adjustment that could insulate it from falsification on the specificity dimension.
Reasoning tree
Public endorsements
Emmanuel Streel is listed as a co-author on a 2025 peer-reviewed paper in Alzheimer's & Dementia (Translational Research & Clinical Interventions), published under Altoida Inc. affiliation, arguing that digital biomarkers redefine clinical outcomes and the concept of meaningful change in Alzheimer's disease. That authorship is a direct, public endorsement of the core theory: that digital measurement captures clinically relevant signal beyond conventional assessment. The pilot study on Parkinson's MCI (publication d9aaae07) demonstrates exactly the testable prediction in the theory — ML/AR-based NeuroMarker classification compared against clinician-led classification, with sensitivity and specificity reported. Streel's co-authorship on the 2025 paper (PMID 40463636) places him on record as an active scientific advocate for digital biomarkers in cognitive decline, not merely an employee executing someone else's agenda.
Evidence publication IDs: d9aaae07-e634-40d8-af25-e23a0ca5ffec
Tarnanas is Altoida's founder and chief scientific officer, which makes him the originator of the theory, not merely a third-party endorser. His 19-year VR/AR research program explicitly targets neurodegenerative disease onset detection via sensor-driven tasks (TechCrunch, 2019). His TEDx talk anchors the spatial-cognition-in-early-Alzheimer's claim to the 2014 Nobel in Medicine (O'Keefe/Mosers). He made prepared public remarks at the $6.3M Series A announcing Altoida as an AI Alzheimer's-risk predictor. Most directly, he is a named inventor on PCT/IB2022/055333, a patent for measuring cognitive performance via a gamified device environment specifically targeting MCI-phase and Alzheimer's prediction — a direct technical instantiation of the theory's core mechanism. No contradicting statements found in any evidence item.
Evidence publication IDs: cf880163-f2f3-4164-8df7-d5bc9d0d25ea
None of the four records contain a statement attributable to John Harrison. Record 1 is a Frontiers Health Facebook post about the Altoida–Click Therapeutics partnership, authored by Frontiers Health, not Harrison. Record 2 is an anonymous Facebook group post describing Altoida's app. Record 3 is a 2013 Interaxon patent whose inventor list includes 'Paul Harrison Baranowski', a different person at a different company. Record 4 is a PMWC conference directory that lists Altoida Inc among hundreds of companies with no speaker or quote attributed to Harrison. There are zero direct quotes from John Harrison in the evidence set, and no record where he addresses the theory's core claims — sensor-derived biomarkers, MCI sensitivity, or Alzheimer's-risk classification. As CSO his role is consistent with the theory's scientific direction, but that inference alone does not constitute a public endorsement.