AI reads everyday health questionnaires to flag Parkinson’s disease

Researchers have built an artificial intelligence model that can screen for Parkinson’s disease using nothing more than the answers people give to ordinary health questionnaires. The model, called Quest2Dx, was developed by Diego Machado Reyes, Juergen Hahn, Li Shen and Pingkun Yan, and is described in a study published in Nature Aging. Instead of relying on expensive scans or invasive procedures, the system reads self-reported responses and turns them into a probability that a person has the disease.

The motivation comes from a persistent gap in how Parkinson’s disease is detected. Because populations around the world are aging, Parkinson’s is becoming a growing public health challenge, and the researchers argue that a cheap, non-invasive screening tool could help catch it earlier in primary care settings.

Quest2Dx is built on language modeling, a technique more commonly associated with chatbots and text analysis. Rather than treating each questionnaire item as a bare number, the model encodes the meaning of both the question and the answer as numerical representations. Conventional machine learning models, by contrast, are typically built for one specific questionnaire and cannot easily transfer to another.

That flexibility matters because real-world health data is messy. People skip questions, and different clinics and studies use different forms. The researchers note that missing responses degrade the performance of conventional models, while Quest2Dx’s language-based encoding lets it cope with gaps and move between questionnaires. The model can transfer across questionnaires while also enhancing interpretability, since the meaning encoded for each question can be traced back to identifiable predictors.

The team evaluated the model on two large datasets of patient-reported information: the Parkinson’s Progression Markers Initiative Online cohort and the Fox Insight dataset. On these datasets, Quest2Dx achieved areas under the receiver operating characteristic curve, or AUROCs, of 0.977 and 0.974 respectively, outperforming existing methods.

The more demanding test was whether the model could generalize to a questionnaire it had never seen. In cross-questionnaire validation, Quest2Dx achieved an AUROC of 0.920 when trained on PPMI Online and applied to Fox Insight, and 0.952 in the reverse direction.

Beyond classification, the model also identified the key predictors among the questions, offering insight into which self-reported items carry the most diagnostic weight. A figure in the study lists the top ten questions identified by Quest2Dx, and another compares classification performance when using all questions versus three subsets of questions. This interpretability is presented as a feature of the approach: because each question and answer carries semantic meaning, the model’s decisions can be examined rather than treated as a black box.

The authors frame these results as a promising path toward Parkinson’s disease screening in primary care. The study also includes a graphical overview of the Quest2Dx pipeline and a breakdown of participant distribution across levels of data completeness in the two cohorts, indicating attention to how much missing data the approach can tolerate.

As with any screening model trained on research cohorts, real-world performance will depend on how well the findings hold up in broader, more diverse populations and in clinical workflows outside specialist centers.

Subject of Research: An AI language model that screens for Parkinson's disease from self-reported health questionnaire answers

Article Title: Language modeling screens Parkinson’s disease with self-reported questionnaires

Article References: Machado Reyes, D., Hahn, J., Shen, L., & Yan, P. (2026). Language modeling screens Parkinson’s disease with self-reported questionnaires. Nature Aging. https://doi.org/10.1038/s43587-026-01260-6

Image Credits: AI Generated

DOI: 10.1038/s43587-026-01260-6

Keywords: Parkinson's disease, Quest2Dx, language modeling, health questionnaires, machine learning, screening, AUROC, PPMI, Fox Insight, Nature Aging, interpretability, primary care