Parkinson's Screening AI Reads Health Questionnaires

A Parkinson’s screening AI built at Rensselaer Polytechnic Institute can read long health questionnaires and pick out people who may need testing for the disease, years before the familiar movement symptoms appear. The model, called Quest2Dx, is described in a Nature Aging paper published on 4 November and announced by RPI the same day.

The point is not to diagnose Parkinson’s but to get the right people referred sooner. Early signs such as loss of smell, sleep disorders, handwriting difficulties and mood changes are easy to brush off, and the release says that can delay referrals to neurologists for years. Quest2Dx is meant to catch the pattern in answers patients already give in primary care questionnaires. So far it has been tested only on research datasets, not in clinics.

“People with Parkinson’s disease get referred to specialists pretty late,” said Pingkun Yan, head of RPI’s Department of Biomedical Engineering, who led the team. “We built this AI model to screen for potential cases earlier.”

Why questionnaires trip up ordinary AI models

Health questionnaires are long. The questionnaire from the Parkinson’s Precision Medicine Initiative (PPMI) database that the team studied has 187 questions, and the Fox Insight study’s has more than 300. Patients often skip questions, and according to RPI typical machine learning tries to fill those gaps by guessing, a step that can introduce errors. Conventional models can also fail when questions are reworded or reordered, because they learn where a question sits rather than what it means.

Quest2Dx uses language modelling to read the meaning of each question and answer. “We used the large language model to understand the meaning and context of each question and answer,” Yan said. That lets it work around missing answers, handle questionnaires worded differently from the one it was trained on, and show which questions and answers mattered most for each person.

Parkinson’s screening AI scored highly on research data

The team trained the model on PPMI data and validated it on the Fox Insight dataset, both supported by the Michael J. Fox Foundation. In the paper, Quest2Dx reached an AUROC, a standard 0-to-1 accuracy measure for telling two groups apart, of 0.977 on PPMI Online and 0.974 on Fox Insight. Trained on one questionnaire and tested on the other, it scored 0.920 one way and 0.952 the other. RPI says it beat earlier models, including an approach built on the Llama 2 language model.

The questions the model found most useful included difficulty getting out of bed and eating, exposure to industrial solvents and degreasers, REM sleep behaviour disorder, anxiety and feelings of tension or stress. None is diagnostic alone. A shortened survey of the 30 most useful questions was “nearly as accurate” as the full questionnaires, RPI says.

“These findings will help design simplified questionnaires for PD screening, paving the way for a tool that can be adopted in future studies and used in primary care settings to enable early diagnosis and timely treatment for patients,” the authors wrote.

A cheap first filter, if it survives real clinics

The most useful result here may be the 30-question survey rather than the headline accuracy. A GP surgery cannot ask every older patient to complete 300 items, but a short form scored by software is the kind of first filter that could fit into a routine check-up. That is what makes this different from tools that need a scan, a blood test or a specialist’s time: the input is something patients can already fill in at home.

The numbers need careful reading, though. High AUROC scores on curated research cohorts, where people with Parkinson’s and healthy volunteers have been carefully recruited, often fall when a tool meets the messier mix of patients in a real waiting room. RPI itself says Quest2Dx can produce both false positives and false negatives and still needs refinement. In screening, both errors carry a cost: false alarms mean worried patients and extra specialist referrals, while misses give false reassurance.

That puts Quest2Dx at an earlier stage than Empatica’s FDA-cleared WellPD wearable, which tracks symptoms in people already diagnosed. It belongs with research such as the AI heart MRI survival study and the NUS DUSP4 CAR-T work: promising results on existing data that still need testing forward in time on real patients. The team has published its code openly, which makes independent checking easier.

Ageing populations raise the stakes

The researchers frame early detection as a public health issue, because age is a key risk factor for Parkinson’s. They write of a “critical need” for non-invasive, easily accessible early methods that fit into primary care before extensive motor impairment occurs. They also believe the approach could be adapted to screen for other conditions, including cardiovascular diseases.

The test that matters next is a prospective study in primary care, showing whether a short Quest2Dx questionnaire flags the right patients when nobody has picked them in advance.