Transitive Sequencing Medical Records for Mining Predictive and Interpretable Temporal Representations.

Area of Science:

• Computational biology
• Health informatics
• Machine learning in healthcare

Background:

• Electronic health records (EHRs) capture valuable longitudinal patient data.
• Extracting meaningful temporal patterns from EHRs for machine learning remains a challenge.
• Existing methods often aggregate EHR data, losing critical sequential information.

Purpose of the Study:

• To propose and evaluate a transitive sequencing approach for constructing temporal representations from EHR observations.
• To enhance machine learning models for disease progression prediction and classification using sequential EHR data.
• To demonstrate the superiority of temporal representations over aggregated, atemporal EHR data.

Main Methods:

• Developed a transitive sequencing method to create temporal representations from EHR medication and diagnosis records.
• Applied the method to a cohort of patients with congestive heart failure.
• Compared the performance of the 'bag-of-sequences' approach against aggregated vector representations using various classifiers.

Main Results:

• Transitive sequential representations significantly outperformed aggregated, atemporal EHR records in classification and prediction tasks.
• The proposed method identified EHRs as better phenotype differentiators.
• Temporal representations revealed novel insights into disease progression not apparent from independent clinical data.

Conclusions:

• Transitive sequencing of EHR data provides a powerful method for building accurate predictive models.
• This approach enhances the understanding of disease trajectories by preserving temporal relationships.
• The findings suggest a paradigm shift towards utilizing sequential, temporal data in EHR-based machine learning.