Improving Young Stroke Prediction by Learning with Active Data Augmenter in a Large-Scale Electronic Medical Claims...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Epidemiology

Background:

• Electronic medical claims (EMC) databases are valuable for disease prediction but struggle with rare conditions due to imbalanced data.
• Existing models show inadequate predictive performance for rare disease occurrences, necessitating improved methodologies.

Purpose of the Study:

• To enhance stroke prediction accuracy in the young adult demographic (25-45 years) within a large-scale EMC database.
• To address challenges of insufficient training samples and class imbalance in predicting stroke in younger populations.

Main Methods:

• Development of a deep neural network model for young stroke prediction.
• Implementation of a novel active data augmentation strategy to select informative electronic health record (EHR) data from older stroke patients.

Main Results:

• Achieved a 9.3% improvement in Area Under the ROC Curve (AUC) compared to training solely on young adult data.
• Demonstrated an 8.2% AUC improvement over models trained on all age groups.
• Analyzed AUC performance based on training data size and augmentation characteristics.

Conclusions:

• The proposed active data augmentation method significantly improves stroke prediction for young adults.
• This approach effectively mitigates issues of data scarcity and class imbalance in predictive modeling.
• The findings highlight the potential of targeted data augmentation for enhancing clinical prediction models.