Learning from heterogeneous temporal data in electronic health records.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Data Science

Background:

• Electronic health records (EHRs) offer valuable longitudinal data for research.
• Machine learning (ML) presents an alternative to manual analysis of complex EHR data.
• Standard ML algorithms struggle with the temporal and uneven nature of clinical event data.

Purpose of the Study:

• To explore novel representations of temporal data in EHRs.
• To develop methods that retain sequential information for ML compatibility.
• To improve the modeling of temporality in clinical data for prediction tasks.

Main Methods:

• Utilized symbolic sequence representations of time series data from EHRs.
• Developed and tested methods for creating these representations.
• Employed a distance measure to random subsequences for comparison against original and clustered sequences.

Main Results:

• The novel representations significantly improved predictive performance on 19 real-world EHR datasets.
• The distance measure to random subsequences outperformed using original or clustered sequences.
• Symbolic sequence representations demonstrated quality comparable to expert-generated sequences.

Conclusions:

• The proposed method effectively captures the temporality of clinical events in EHRs.
• These new representations enhance ML algorithm compatibility and predictive power.
• This approach addresses a key challenge in leveraging longitudinal EHR data for biomedical research.