November 4, 2026
Scientists at Rensselaer Polytechnic Institute (RPI) have developed an artificial intelligence model to streamline and improve screening for Parkinson’s disease (PD) so doctors can refer patients for early-stage testing.
In a paper published in Nature Aging, a team led by Pingkun Yan, Ph.D., Head of the Department of Biomedical Engineering at RPI described how their Quest2Dx AI system was able to analyze lengthy health questionnaires more accurately than other models for early signs of Parkinson’s.
Diagnosis of Parkinson’s disease relies on observation and measurement of motor symptoms as well as invasive procedures including spinal taps, but the disease may be present for years before physical symptoms appear.
“People with Parkinson's disease get referred to specialists pretty late,” Yan said. “We built this AI model to screen for potential cases earlier.”
Visual observation has actually been part of the diagnostic process since 1817, when British physician James Parkinson published a seminal essay on the condition that now bears his name, based mainly on watching people walking through a London park. Medical technology has advanced, but the visual aspect has not changed much since then.
Yan and colleagues — including Director of the Center for Biotechnology and Interdisciplinary Studies Juergen Hahn, Ph.D., and former RPI doctoral student Diego Machado Reyes — are not trying to diagnose PD, just screen for patients who might need testing years before the physical symptoms Parkinson observed more than 200 years ago are apparent.
Symptoms that indicate early-stage Parkinson’s include loss of smell, sleep disorders, handwriting difficulties, and mood changes. According to the authors, patients and caregivers alike might brush these conditions off as minor, delaying referrals to neurologists for years.
Early detection and treatment would be more feasible if AI could scan health questionnaires looking for signs of Parkinson’s, according to the authors, who also include University of Pennsylvania bioinformatics professor Li Shen, Ph.D.
So many questions
Primary care clinics frequently offer health questionnaires for screening for many diseases, but questionnaires are often so thorough and complicated that healthcare providers might dismiss subtle signs related to Parkinson’s and other conditions.
Data for training the Quest2Dx model came from the Parkinson’s Precision Medicine Initiative (PPMI) database, a public-private effort supported by the Michael J. Fox Foundation for Parkinson’s Research and others. The Fox Insight Study provided a second dataset that the team used to validate the model.
The PPMI questionnaire the RPI-led team studied has 187 questions. Fox Insight has a questionnaire with more than 300 items. Both take a long time to complete, and patients frequently skip questions.
The researchers said that health questionnaires are not particularly friendly to AI analysis because typical machine learning algorithms try to fill in the blanks from skipped responses through a process called imputation. This can lead to errors that make it more difficult for AI to find hidden patterns, according to Yan, who focuses on applying AI and machine learning to medical imaging informatics and image-guided clinician interventions.
Conventional models can also fail when questions are reworded or reordered, because they learn the position of each question rather than its meaning.
The Quest2Dx deep learning framework was designed to differentiate between those with Parkinson’s symptoms and a control group made up of healthy questionnaire respondents.
It also was built to address three “key technical challenges,” the researchers wrote. It can work around missing responses, it can work with multiple questionnaires with different wording even if it’s only trained on one, and it is able to calculate the most relevant pairs of questions and answers for each individual patient.
“We used the large language model to understand the meaning and context of each question and answer,” Yan explained. “We used the AI model to read those questions, to read the answers, and then classify whether the patient may or may not have Parkinson's disease,” Yan said.
Results and next steps
Analysis outlined in the paper showed that Quest2Dx outperforms earlier models, including an approach built on the pre-trained large language model Llama 2.
The questions Quest2Dx ranked as most informative included difficulty with common tasks like getting out of bed and eating, as well as risk factors for Parkinson’s such as exposure to industrial solvents and degreasers.
Other potential predictors of Parkinson’s include REM sleep behavior disorder, anxiety, and elevated feelings of tension or stress. These conditions also overlap with many other illnesses, so no single answer is diagnostic on its own. Solvent exposure, for example, “becomes relevant when considered alongside other early signs,” the authors wrote, which “underscores the importance of evaluating multiple factors together.”
As part of their analysis, the researchers created a more targeted survey with the 30 most pertinent questions, as identified by Quest2Dx, and the AI model was nearly as accurate at predicting Parkinson’s as it was with the full questionnaires.
“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,” according to the authors.
Yan and colleagues framed this as a public health issue as the population ages, since age is a key risk factor for Parkinson’s. “As a result, there is a critical need for developing non-invasive and easily accessible early diagnostic methods that can be seamlessly adapted to current practices in primary care settings to identify PD before extensive motor impairment occurs,” they wrote.
Quest2Dx still needs refinement, as it can produce both false positives and false negatives, though. But as they build it, Yan and colleagues believe that their model could be adapted to screen for many other conditions, including cardiovascular diseases.
“The promising results of Quest2Dx lay a strong foundation for the development of AI-driven screening models based on self-reported data, with the potential to significantly enhance early diagnosis and accessibility across a range of diseases in the future,” the research team concluded.