Predicting Parkinson's Disease Using a Deep-Learning Algorithm to Analyze Prodromal Medical and Prescription Data.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Parkinson's disease (PD) presents with prodromal symptoms, but less specific ones complicate early identification.
• Retrospective analysis of prodromal symptoms is common, yet predicting PD risk remains challenging.
• Deep learning offers potential to enhance prediction accuracy by analyzing complex medical data.

Purpose of the Study:

• To improve deep learning-based screening for prodromal Parkinson's disease.
• To leverage medical claims data, including prescription information, for enhanced PD detection.
• To assess the impact of diagnostic and medication codes on prediction accuracy.

Main Methods:

• A deep learning algorithm was developed using Korean National Health Insurance cohort data.
• 820 PD patients and 8,200 matched controls were sampled.
• The algorithm utilized combinations of diagnostic codes, medication codes, and varying prodromal periods.

Main Results:

• Predicting PD using diagnostic codes from year -3 to 0 achieved 0.937 accuracy.
• Adding medication codes in the same period (year -3 to 0) did not significantly improve accuracy (0.931-0.935).
• For the earlier period (year -6 to -3), diagnostic codes alone yielded 0.890 accuracy, which increased to 0.922 with medication codes.

Conclusions:

• Deep learning models integrating prodromal diagnostic and medication codes are effective for PD screening.
• Automated surveillance systems using medical claims data could offer cost-effective early detection of PD risk.
• This approach can facilitate the development of disease-modifying drugs by identifying suitable candidates for clinical trials.