Can an AI questionnaire screen for Parkinson’s across different surveys? Two online cohorts show promise—but not primary-care proof
Quest2Dx separated self-reported Parkinson’s cases from controls in two large US online cohorts and transferred between their questionnaires. The peer-reviewed study tests retrospective classification, not undiagnosed primary-care patients, clinical benefit or a safe referral threshold.
The short version
- What happened
- The peer-reviewed model used two large online research resources: PPMI Online, reported as 40,012 participants, and Fox Insight, reported as 54,571. The usable training samples changed with questionnaire-completeness thresholds.
- Who it affects
- People with possible Parkinson’s symptoms could eventually benefit from a shorter, low-cost way to prompt specialist assessment, but this study did not test undiagnosed patients or establish a safe referral threshold.
- What remains unproven
- The retrospective cohorts were online Parkinson’s research communities, not consecutive primary-care patients with unknown diagnosis.
Keep in mind
The retrospective cohorts were online Parkinson’s research communities, not consecutive primary-care patients with unknown diagnosis.
AUROC does not provide the false-positive and false-negative counts at a deployable threshold, and no patient or service outcome was measured.
Question overlap was established with embedding similarity plus visual review, and transfer between two related US-linked cohorts may overstate portability to new countries, languages or clinics.
The short-question subsets were selected and evaluated within the same evidence base and require independent prospective validation.
Multi-source analysis · 2 linked sources. Primary analysed source date: 4 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 11 min The evidence, methods, implications and limitations
The direct answer: the model transferred between two surveys, but that is not the same as a clinical screening test
A language-model-based system called Quest2Dx distinguished people recorded as having Parkinson’s disease from healthy controls in two large online questionnaire datasets, and it retained high discrimination when trained on one survey and tested on the other. The peer-reviewed Nature Aging paper reports mean area under the receiver operating characteristic curve, or AUROC, of 0.977 on PPMI Online and 0.974 on Fox Insight. In cross-questionnaire tests, the reported averages were 0.920 when training on PPMI and testing on Fox Insight, and 0.952 in the other direction. AUROC measures how often a model ranks a randomly chosen case above a randomly chosen control across all possible thresholds; it does not tell a clinic how many people would be missed or referred unnecessarily at the threshold it would actually use.
That distinction is the centre of the finding. Quest2Dx addresses a real technical problem: two questionnaires may ask about similar symptoms with different wording and contain many skipped answers, while conventional models often depend on fixed column positions or imputed values. The new model encodes the meaning of each question-answer pair and learns a representation for missing entries. Its ability to move between datasets is more informative than a result on one randomly split dataset. Yet the participants were already part of Parkinson’s research cohorts and the experiment was retrospective. The paper did not enrol people arriving in primary care with an unknown diagnosis, compare referrals against a clinical work-up or measure whether earlier screening changed health outcomes.[1][2]
What data went into the comparison and why the denominator changes
The paper describes PPMI Online as containing 40,012 participants and Fox Insight as containing 54,571. Both are US-linked online research resources with participant-reported information from people with and without Parkinson’s disease. The full questionnaire pools are much larger than the model inputs: the analysis ultimately used 187 PPMI questions and 378 Fox Insight questions after preprocessing. Participants often left answers blank, so the researchers evaluated several minimum-completeness thresholds rather than pretending every person supplied a complete form. This means the headline cohort sizes are not the denominator for every result. At stricter thresholds, fewer people remained available for training and evaluation.
For the cross-questionnaire experiment, the researchers identified 129 questions judged to overlap between the two surveys using embedding similarity plus visual review. At the 50% completeness threshold, the table reports training sets of 24,347 for the PPMI-to-Fox experiment and 27,060 for the reverse direction. At 90% completeness, those training sets fell to 2,998 and 3,583. The design used five folds: four partitions of the training questionnaire for model fitting, the fifth for validation, and the other questionnaire as the external test set. That is a useful transfer test, but question matching included human judgment and both cohorts share a Parkinson’s-research context. A general-practice form in another language, country or health system could differ more substantially.[1]
The distinction that matters
What the Parkinson’s questionnaire study did—and did not test
The model was evaluated retrospectively in two online research cohorts. AUROC measures ranking across thresholds; it is not the percentage of patients correctly diagnosed in primary care.
PPMI Online
Quest2Dx mean AUROC 0.977
High within-questionnaire discrimination
Fox Insight
Quest2Dx mean AUROC 0.974
Several comparators were also above 0.9
PPMI → Fox
Average AUROC 0.920
Varied from 0.980 to 0.715 by completeness setting
Fox → PPMI
Average AUROC 0.952
Transfer was more stable in the reported settings
Not tested
No prospective unknown-diagnosis cohort
No referral threshold, patient benefit or clinical workflow result
The model beat several tabular baselines, but strong ranking is not a diagnosis
Within the two datasets, Quest2Dx was compared with a support-vector machine, random forest, XGBoost, a multilayer perceptron and a transformer using mode imputation. The closest comparator was the imputation-based transformer, with mean AUROC of 0.966 on PPMI and 0.971 on Fox Insight, against 0.977 and 0.974 for Quest2Dx. XGBoost reached 0.906 and 0.952, and random forest 0.903 and 0.949. The authors report paired statistical tests favouring Quest2Dx in nearly all dataset-completeness settings. The improvement is clear on PPMI and small on Fox Insight, where several methods already scored above 0.9.
Those numbers should not be translated into ‘97.7% accurate’. AUROC is threshold-independent and can remain high even when a chosen operating point creates too many false positives for a low-prevalence setting. Primary-care screening would require a prespecified threshold, sensitivity, specificity, positive predictive value and negative predictive value in the population where the tool is offered. It would also need a plan for people flagged by the model: a screening result is not a diagnosis, and Parkinson’s assessment can involve neurological examination, longitudinal observation and tests that rule out other explanations. The paper calls Quest2Dx a promising route to screening; it does not report regulatory authorization or clinical deployment.[1]
Cross-questionnaire performance is the important advance—and it was not equally stable everywhere
Average transfer performance can hide the difficult corner. When trained on PPMI and tested on Fox Insight, Quest2Dx reported AUROC of 0.980 at both 50% and 60% completeness, 0.973 at 70%, 0.953 at 80% and 0.715 at 90%. The drop coincided with the training set shrinking to 2,998. In the reverse direction, AUROC ranged from 0.943 to 0.973 across the completeness settings shown. This pattern supports the authors’ argument that both data quality and data quantity matter. It also warns against describing the system as uniformly portable: the model struggled in one high-completeness, low-sample configuration.
The paper reports an average PPMI-to-Fox AUROC confidence interval whose upper bound exceeds 1, which is outside the possible AUROC range. That appears to reflect an unbounded interval calculation across a small set of completeness-level estimates rather than the uncertainty around one clinical test. Readers should rely on the setting-specific results and bounded task metric, not treat that average interval as a literal probability range. More generally, the validation is between two datasets available to the same research field, not between independent clinics implementing a frozen model. Prospective external validation should lock the model and threshold before seeing new patients, then report every eligible person and every exclusion.[1]
A shorter questionnaire may be more usable, but the study selected it on the same evidence base
Long surveys are a practical barrier. The authors used the model’s attention weights to rank questions and tested subsets of the top 10, 20 and 30 PPMI items. The paper reports AUROC above 0.9 with 20 questions in the evaluated settings and close to 0.95 for 30 questions at 50% completeness. Questions highlighted by the university included everyday difficulties, rapid-eye-movement sleep behaviour, anxiety, stress and exposure to industrial solvents or degreasers. These factors can be relevant in combination, but none is specific enough to diagnose Parkinson’s on its own.
The shorter form is not yet an independently validated screening instrument. The same modelling process identified the important questions and evaluated the subsets, which can make performance look better than it will in a new population. Attention weight is also not proof that a feature causes disease or that the model’s explanation matches clinical reasoning. Before a short questionnaire is offered to the public, researchers would need to predefine the items, wording, scoring rule and referral threshold, then test them in a separate representative cohort. They should compare completion, anxiety, follow-up burden and missed diagnoses—not only AUROC.[1][2]
What this could mean for patients and primary-care teams
A low-cost questionnaire that works despite missing answers could help primary-care teams notice combinations of non-motor symptoms that are easy to overlook, particularly when specialist access is limited. For a person with emerging symptoms, earlier recognition might shorten the path to assessment and supportive care. But a false reassurance could delay help, while a false alarm could create fear and place more demand on neurology services. Those consequences depend on disease prevalence, the threshold and the availability of confirmatory care. None was measured here.
The online cohorts may also differ from the people most likely to be missed by existing services. Joining a long-term Parkinson’s research programme requires awareness, connectivity and willingness to complete repeated surveys. People with limited digital access, lower literacy, different languages or atypical symptoms may be underrepresented. A model built from self-report can learn how diagnosed participants describe themselves as well as early disease signals. Health systems should therefore treat it as a research classifier until prospective testing shows how it behaves before diagnosis, across demographic groups and in routine workflow.[1][2]
Funding, preprint history and what the peer-reviewed date means
The peer-reviewed version was published on 4 November 2026, but the research is not a same-day experiment: a preprint was posted on 24 September 2024 and the authors say the two main datasets were obtained on 13 November 2023. The current source date is legitimate because peer review and formal publication are material changes in evidence status. It would be misleading to imply that the participant data were collected this week or that the result reflects the newest questionnaire populations. The report therefore preserves both the publication date and the older data history.
The work was funded by National Institutes of Health training and research grants and a National Science Foundation CAREER award. The authors declared no competing interests. PPMI and Fox Insight themselves are supported by the Michael J. Fox Foundation and many pharmaceutical, biotechnology and health-technology partners listed in the paper. Public-private funding does not invalidate the analysis, but it matters for transparency and for independent replication. The code is public, while access to the participant-level PPMI and Fox Insight data requires registration and agreement to their terms.[1]
What would change the assessment
The assessment would strengthen if a frozen Quest2Dx model and short questionnaire were tested prospectively among consecutive primary-care patients whose Parkinson’s status was unknown at enrolment. A useful study would report the full flow of invited, completed, referred and diagnosed participants; clinically chosen thresholds; sensitivity and specificity with confidence intervals; subgroup performance; and the number of specialist appointments generated per additional confirmed case. Comparison with a simple clinical score would show whether the language model adds enough value to justify its complexity.
It would weaken if performance falls after language, demographic or health-system shifts, if false positives overwhelm referral capacity, or if the model mainly recognizes features disclosed after diagnosis. For now, the peer-reviewed cross-questionnaire validation is a substantive research result and a better test than a single-dataset split. It is not Breaking because it does not announce an approved test, an immediate change in clinical guidance or an urgent new hazard. The practical message is measured optimism: semantic modelling can make questionnaires more portable, but clinical usefulness remains unproven.[1][2]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring screening: compare the finding and its evidence limits.
Can an LLM screen students for depression reliably? One setup changed its answer in 23.6% of cases
In 6,395 Chinese undergraduates, some language-model prompts matched conventional models at one threshold. But performance changed with the cutoff, XGBoost gave better probability estimates, and repeated GPT calls disagreed for 532 of 2,258 students.
Read the next reportExplore the full source and evidence record
What this means for people
- People with possible Parkinson’s symptoms could eventually benefit from a shorter, low-cost way to prompt specialist assessment, but this study did not test undiagnosed patients or establish a safe referral threshold.
- Primary-care teams would need sensitivity, specificity, predictive values and referral workload in their own population before using a ranking model as a screening service.
- Patients should not treat a questionnaire score as a diagnosis or delay clinical assessment because a model returns a reassuring result.
Global context
Questionnaire-based screening could be attractive where neurologists and advanced diagnostics are scarce, but portability across two US-linked online cohorts is not global validation. Symptom language, environmental exposures, literacy, access to follow-up and Parkinson’s prevalence vary across settings. A system that increases detection in one health service could produce unmanageable false referrals in another. Translation and local validation must preserve meaning rather than merely map words.
What the evidence does not yet show
- The retrospective cohorts were online Parkinson’s research communities, not consecutive primary-care patients with unknown diagnosis.
- AUROC does not provide the false-positive and false-negative counts at a deployable threshold, and no patient or service outcome was measured.
- Question overlap was established with embedding similarity plus visual review, and transfer between two related US-linked cohorts may overstate portability to new countries, languages or clinics.
- The short-question subsets were selected and evaluated within the same evidence base and require independent prospective validation.
What to watch next
- Prospective evaluation in people without a known Parkinson’s diagnosis at enrolment.
- A locked short questionnaire and referral threshold with sensitivity, specificity and predictive values.
- Independent tests across languages, countries, demographic groups and lower-connectivity populations.
- Evidence that screening shortens diagnostic delay or improves care without excessive anxiety and referral burden.
Living evidence record
Impact record IAI-0ZOF4A7
Evidence stage
Studied
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
4 November 2026
Source trail
2 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 4 November 2026
Research question, topics and regions
Whether a language-model-based classifier can distinguish self-reported Parkinson’s disease cases from controls and transfer between two differently worded online health questionnaires
Topics: Parkinson’s disease · Screening · Questionnaires · Language models · Clinical validation · Primary care · Missing data
Regions: United States · North America · Global
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.