Long Short-Term Memory-Based Deep Learning Models for Screening Parkinson's Disease Using Sequential Diagnostic...

Area of Science:

• Neurology
• Data Science
• Medical Informatics

Background:

• Early detection of Parkinson's disease (PD) remains a significant challenge.
• Existing methods often rely on cross-sectional data, failing to capture the disease's temporal progression.
• There is a need for advanced techniques that can process time-series information for more accurate PD diagnosis.

Purpose of the Study:

• To develop and evaluate a novel machine learning method for early Parkinson's disease detection.
• To leverage time-series data from patient diagnostic histories for improved predictive accuracy.
• To utilize the long short-term memory (LSTM) algorithm for processing sequential diagnostic information.

Main Methods:

• A dataset of 926 Parkinson's disease patients and 9,260 controls was analyzed using medical-claims data.
• The long short-term memory (LSTM) algorithm was employed to process diagnostic codes and their temporal information.
• Prediction performance was compared across different time windows (1-4 years prior to diagnosis).

Main Results:

• The LSTM model trained on the most recent 12-month diagnostic codes achieved the highest accuracy (94.25%), sensitivity (82.91%), and specificity (95.26%).
• Models using data from 2, 3, and 4 years prior showed comparable performance, with accuracies ranging from 91.81% to 92.27%.
• Area Under the Curve (AUC) values ranged from 0.839 to 0.923, indicating strong predictive power.

Conclusions:

• The proposed machine learning method, utilizing LSTM for time-series analysis, demonstrates significant potential for early Parkinson's disease detection.
• This approach can provide valuable support for Parkinson's disease specialists in identifying the disease at its nascent stages.
• The findings suggest that incorporating temporal diagnostic patterns can enhance the accuracy of automated PD detection systems.