Prediction task guided representation learning of medical codes in EHR.

Area of Science:

• Health Informatics
• Machine Learning
• Natural Language Processing

Background:

• Machine learning models are increasingly used for predictive analytics in Electronic Health Records (EHR).
• Effective feature vector representation of medical codes is crucial for model performance.
• Existing unsupervised methods for medical code vectorization are independent of prediction tasks and require large datasets.

Purpose of the Study:

• To develop a novel method, Prediction Task Guided Health Record Aggregation (PTGHRA), for constructing training corpora for medical code representation learning.
• To improve the predictive capability of machine learning models utilizing EHR data, particularly when training samples are limited.

Main Methods:

• Developed Prediction Task Guided Health Record Aggregation (PTGHRA) to aggregate health records guided by specific prediction tasks.
• Integrated PTGHRA with representation learning models for medical code vectorization.
• Evaluated the performance of the integrated models compared to unsupervised approaches.

Main Results:

• Representation learning models integrated with PTGHRA demonstrated significant improvements in predictive capability.
• The proposed method showed particular effectiveness when dealing with limited training samples.
• PTGHRA facilitates the generation of more appropriate feature vectors for specific prediction tasks.

Conclusions:

• PTGHRA offers a task-guided approach to enhance medical code representation learning for EHR predictive analytics.
• The method addresses the limitations of unsupervised approaches, especially in data-scarce scenarios.
• This work contributes to improving the quality of hospital services and healthcare resource utilization through more accurate predictive models.