Predicting cardiovascular health trajectories in time-series electronic health records with LSTM models

Area of Science:

• Biomedical Informatics
• Cardiovascular Health Research
• Predictive Analytics

Background:

• Cardiovascular disease (CVD) is a leading cause of mortality in the US.
• Improved cardiovascular health (CVH) is crucial for CVD prevention.
• Predicting future CVH can aid in proactive patient management.

Purpose of the Study:

• To investigate the predictability of future cardiovascular health (CVH) levels.
• To determine if longitudinal electronic health record (EHR) data can be used for CVH prediction.
• To assess the efficacy of predictive models for various CVH submetrics.

Main Methods:

• Utilized the Guideline Advantage (TGA) dataset comprising EHR data from 70 US outpatient clinics.
• Focused on predicting five CVH submetrics: smoking status (SMK), body mass index (BMI), blood pressure (BP), hemoglobin A1c (A1C), and low-density lipoprotein (LDL).
• Applied embedding techniques and long short-term memory (LSTM) networks to predict future CVH categories from historical measurements for 216,445 patients.

Main Results:

• The LSTM model demonstrated high predictive accuracy across all CVH submetrics.
• Micro-average area under the receiver operator curve (AUROC) values ranged from 0.84 (BP) to 0.99 (SMK).
• Performance did not significantly improve when incorporating all five submetrics compared to single-measure predictions.

Conclusions:

• Future CVH levels can be reliably predicted using historical CVH measurements for individual submetrics.
• This predictive capability has significant implications for population cardiovascular health management.
• Predicting CVH may enhance patient quality of life and reduce the economic burden of CVD.