RPI researchers find 30-question AI screening nearly matches full Parkinson’s surveys

Earlier Parkinson’s referrals are the goal; better questionnaire screening is the finding. An RPI-led team reports that its Quest2Dx AI outperformed earlier models, while a targeted 30-question survey performed nearly as accurately as full questionnaires. Rensselaer Polytechnic Institute described the Nature Aging research November 4. The system is intended to flag people for further testing, not diagnose the disease.

The team, led by RPI biomedical engineering department head Pingkun Yan, wants to help doctors recognize potential cases before obvious movement problems appear. The researchers point to loss of smell, sleep disorders, handwriting difficulties and mood changes as possible early signs. Patients and caregivers may dismiss those changes as minor, they say, delaying referrals to neurologists for years.

Reading answers without filling every blank

The researchers trained Quest2Dx using the PPMI database and validated it with a second dataset from the Fox Insight Study. Both contain lengthy health questionnaires, and patients frequently leave answers blank. Primary care questionnaires can also be complicated enough that clinicians overlook subtle signs, according to the researchers.

Those gaps create a technical problem. Conventional machine-learning systems often estimate missing responses, a process called imputation. Yan says that can introduce errors that obscure useful patterns. Models may also learn where a question appears rather than what it means, making them vulnerable when a questionnaire’s wording or order changes.

Quest2Dx uses a large language model to interpret the meaning and context of questions and answers. Its framework was designed to distinguish respondents with Parkinson’s symptoms from healthy controls while addressing three practical obstacles:

• Missing responses: work with unanswered questions rather than depend on a fully completed form.
• Different questionnaires: work across differently worded forms even when trained on only one.
• Individual relevance: identify which question-and-answer pairs are most informative for each patient.

187 questionsPPMI

The PPMI questionnaire examined in the study contained 187 questions.

More than 300 itemsFox Insight

The Fox Insight questionnaire contained more than 300 items.

30 questionsTargeted survey

Researchers assembled a shorter survey from the questions Quest2Dx ranked as most pertinent.

Patterns, not a single telltale answer

The questions Quest2Dx ranked highly covered more than familiar movement symptoms. They included difficulty getting out of bed or eating, along with exposure to industrial solvents and degreasers. Other potential predictors included anxiety, elevated tension or stress, and REM sleep behavior disorder.

The authors caution that these conditions overlap with other illnesses: no single answer establishes Parkinson’s. Solvent exposure, for example, becomes relevant alongside other early signs. The intended signal comes from evaluating several factors together, not treating one reported symptom or exposure as a diagnosis.

A shorter route to further testing

In the paper’s analysis, Quest2Dx outperformed earlier models, including an approach built on the pretrained language model Llama 2. RPI’s summary does not provide numerical accuracy rates or the size of that advantage, leaving the improvement unquantified in the university’s announcement.

The separate shortening exercise tested whether all those questions were necessary. Researchers assembled a 30-question survey using the items Quest2Dx ranked as most pertinent. They reported nearly the same predictive accuracy as with the full questionnaires. The authors see that result as a basis for designing simpler screening forms for future studies and eventual primary care use.

Quest2Dx still needs refinement. It can produce both false positives, flagging people incorrectly, and false negatives, missing potential cases. The researchers frame its role as helping identify patients who might need testing earlier, rather than replacing that testing. Easier access to early screening remains the intended benefit, not a demonstrated improvement in treatment.