Parkinson's Disease Detected Early with AI...Diagnostic Accuracy Up to 96%

Samsung Medical Center AI Research Center Develops Integrated Gait, Voice, and Brain Imaging Analysis AI
Enables Differential Diagnosis Even in Early Stages with Subtle Symptoms

The AI Research Center at Samsung Medical Center announced on the 23rd that it has confirmed the potential for early diagnosis and prognosis prediction of Parkinson's disease and Parkinson-plus syndromes through a multimodal artificial intelligence (AI) model that jointly analyzes diverse clinical data such as gait, voice, and brain imaging (MRI and PET).

Researchers at the AI Research Center at Samsung Medical Center who developed an AI for early diagnosis and prognosis prediction of Parkinson's disease and Parkinson-plus syndromes. From left: Yang Gwangmo, Director of the AI Research Center; Jo Jinhwan, Professor of Neurology; Jeong Myeongjin, Professor of Radiology. Samsung Medical Center

Parkinson's disease often has no distinct early symptoms, leading to delayed diagnosis in many cases. In some patients, by the time hand tremors or gait abnormalities appear, the disease is already in an advanced stage. Parkinson-plus syndromes such as progressive supranuclear palsy and multiple system atrophy are representative conditions in which early differential diagnosis is difficult because the symptoms are similar.

Over approximately four years, the research team collected data from about 500 patients, including 363 with Parkinson's disease, 67 with progressive supranuclear palsy, and 61 with multiple system atrophy, and built an integrated database. They standardized clinical information such as gait patterns, voice signals, and brain MRI and PET images to establish an analytical foundation.

Based on this, they developed three separate models: a gait data-based fall-risk prediction model, a voice test-based Parkinson's classification AI, and an MRI-based automatic brain structure analysis model. In clinical evaluations, the voice-based severity classification model achieved an AUC of 0.96, and the MRI-based disease differentiation model achieved an AUC of 0.91. The fall prediction model that combined gait and brain imaging also recorded an AUC of 0.84.

The research team designed the AI not only to provide simple prediction results but also to present the grounds for its decisions. It automatically extracts indicators such as gait stability indices, brain structural changes, and voice features so that they can be used as reference metrics in diagnostic decision-making.

The developed models were implemented to run on a network-attached storage (NAS) system for data storage and analysis within the hospital's internal network. This allows analysis without transferring medical data outside the institution, thereby achieving both personal information protection and research efficiency. The research team plans to expand the application to other neurological diseases such as dementia and to develop it into multi-center collaborative research.