DETECT: Feature extraction method for disease trajectory modeling in electronic health records

Area of Science:

• Computational biology
• Health informatics
• Biostatistics

Background:

• Electronic health record (EHR) data presents challenges in modeling due to high dimensionality, redundancy, and noise.
• Identifying meaningful patient disease trajectories is crucial for advancing precision medicine and predicting disease risk.

Purpose of the Study:

• To develop a novel algorithm, Disease Trajectory feature extraction (DETECT), for feature extraction and trajectory generation in high-throughput temporal EHR data.
• To improve the prediction of disease risk and enhance precision medicine strategies by analyzing longitudinal EHR data.

Main Methods:

• The DETECT algorithm simulates longitudinal EHR data with user-defined parameters for scale, complexity, and noise.
• It employs a convergent relative risk framework to identify predictive intermediate codes between index and outcome codes.
• The temporal range for predictor analysis can be specified to focus on specific time windows before outcome onset.

Main Results:

• DETECT was benchmarked on simulated data to validate its performance.
• The algorithm successfully generated real-world disease trajectories from a large cohort (145,575 individuals) with hypertension.
• Predictive features for severe cardiometabolic outcomes were identified within the hypertension cohort.

Conclusions:

• DETECT provides a robust method for feature extraction and trajectory generation from complex temporal EHR data.
• The algorithm facilitates the identification of predictive features for disease outcomes, supporting precision medicine.
• Application to a hypertension cohort demonstrates DETECT's utility in uncovering disease trajectories and risk predictors for cardiometabolic outcomes.