AI screens for Parkinson’s on questionnaires it has never seen
AI models trained to detect Parkinson’s from health questionnaires typically fail when presented with a different questionnaire. A new system takes a different approach, and screens accurately even on questionnaires it was never trained on.
Parkinson’s disease is the world’s fastest-growing neurological disorder. Many patients experience symptoms for years before receiving a diagnosis. Questionnaires covering sleep, smell, mood, and motor function can carry early warning signals, but the way questions are phrased varies considerably across questionnaires. That variability has made it difficult for AI models to transfer knowledge between them.
Encoding meaning, not wording
The new system uses a language model to capture the meaning of each question-and-answer pair as a number, rather than memorising exact wording. The researchers published their findings in Nature Aging. By capturing meaning rather than wording, the system was able to screen for Parkinson's on questionnaires it had never seen during training.
Detection without expensive equipment
Many early-detection approaches for Parkinson’s require brain scanners or specialised hardware. Questionnaires are cheap and scalable. If an AI system can reliably transfer between validated questionnaires, existing health databases could be used for large-scale screening without re-examining patients. The source describes this as a promising direction; detailed performance figures for specific clinical settings are not provided in the available text.
Search terms to explore further: language model medical questionnaire generalisation | Parkinson’s early detection AI screening | semantic representation symptom data