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

Aixia Guo, Rahmatollah Beheshti, Yosef M. Khan, James R. Langabeer Ii, Randi E. ForakerBMC Medical Informatics Decis. Mak 2021 · DOI · OpenReview

Introduction

Cardiovascular disease (CVD) remains the leading cause of death in the United States. While the American Heart Association has established clear metrics for Cardiovascular Health (CVH)—including smoking status, BMI, and blood pressure—clinical management is frequently reactive rather than proactive. The primary challenge in shifting toward preventive care lies in the nature of Electronic Health Records (EHR). EHR data is longitudinal, irregular, and categorical, making it difficult for traditional statistical models to accurately forecast a patient's health trajectory.

This study explores whether deep learning, specifically Long Short-Term Memory (LSTM) networks, can transform these complex time-series records into a predictive tool. By mapping 13 years of EHR data, researchers aimed to allow clinicians to intervene before a patient’s health status deteriorates from "intermediate" to "poor."

Methods and Data Processing

The researchers employed LSTM networks, a type of Recurrent Neural Network (RNN) designed to capture long-term dependencies in sequential data. The methodological pipeline involved several key stages of data refinement and feature engineering.

Methodological pipeline for data extraction, processing, and LSTM model training across five CVH submetrics.

Key technical steps included:

  • Data Categorization: Five CVH submetrics (Smoking, BMI, Blood Pressure, A1C, and LDL cholesterol) were mapped into three clinical categories: Poor, Intermediate, and Ideal.
  • Medical Embeddings: To process categorical data, the team used Word2Vec to map health states (e.g., "bmipoor") into 32-dimensional numerical vectors. This approach treats a patient's health history as a sequence of "words" in a medical "sentence."
  • Time Integration: The model incorporated the time interval (in days) between clinical visits as a feature, recognizing that the frequency and timing of check-ups carry diagnostic weight.
  • Benchmarking: The LSTM performance was compared against two "shallow" machine learning baselines: Logistic Regression (LR) and Random Forest (RF).

Dataset Scope

The study utilized The Guideline Advantage (TGA) dataset, a massive ambulatory quality clinical registry covering over 70 outpatient clinics across the US from 2004 to 2016. The final analysis included 216,445 unique patients who had at least two measurements for any given submetric.

Key Insights and Results

The study revealed a phenomenon described as "Silo Prediction." Surprisingly, multi-metric models (using all five health factors to predict one) did not outperform single-metric models. For instance, a patient's BMI history is a powerful predictor of future BMI, but adding smoking status or blood pressure data to the model did not significantly improve accuracy. This is likely due to the low correlation between metrics; the researchers found Pearson correlation coefficients as low as 0.1 between factors like A1C and Blood Pressure.

Model Performance

The LSTM consistently outperformed traditional models across all submetrics. Performance was evaluated using the Area Under the Receiver Operating Characteristic curve (AUROC):

  • Smoking Status: 0.99 AUROC
  • BMI: 0.97 AUROC
  • LDL Cholesterol: 0.93 AUROC
  • Hemoglobin A1C: 0.91 AUROC
  • Blood Pressure: 0.84 AUROC
Receiver Operating Characteristic (ROC) curves for smoking status (SMK) prediction, showing near-perfect accuracy.

Blood pressure proved the most difficult to predict, which the authors attribute to the inherent volatility of BP readings in clinical settings, often influenced by factors like "white coat syndrome."

Discussion and Future Directions

While the study demonstrates the superiority of deep learning in modeling medical trajectories, several limitations remain. The conversion of raw numerical data into categories (Poor/Intermediate/Ideal) results in a loss of granular detail. Future research should focus on predicting raw numerical values to provide more precise clinical insights.

Furthermore, while the model successfully predicts what will happen, it does not yet identify the optimal window for intervention. The next step in this research is to translate these predictive trajectories into actionable clinical alerts that notify physicians exactly when a medical "nudge" is most likely to alter a patient's long-term cardiovascular path.

Summary updated

Abstract

Background Cardiovascular disease (CVD) is the leading cause of death in the United States (US). Better cardiovascular health (CVH) is associated with CVD prevention. Predicting future CVH levels may help providers better manage patients’ CVH. We hypothesized that CVH measures can be predicted based on previous measurements from longitudinal electronic health record (EHR) data. Methods The Guideline Advantage (TGA) dataset was used and contained EHR data from 70 outpatient clinics across the United States (US). We studied predictions of 5 CVH submetrics: smoking status (SMK), body mass index (BMI), blood pressure (BP), hemoglobin A1c (A1C), and low-density lipoprotein (LDL). We applied embedding techniques and long short-term memory (LSTM) networks – to predict future CVH category levels from all the previous CVH measurements of 216,445 unique patients for each CVH submetric. Results The LSTM model performance was evaluated by the area under the receiver operator curve (AUROC): the micro-average AUROC was 0.99 for SMK prediction; 0.97 for BMI; 0.84 for BP; 0.91 for A1C; and 0.93 for LDL prediction. Model performance was not improved by using all 5 submetric measures compared with using single submetric measures. Conclusions We suggest that future CVH levels can be predicted using previous CVH measurements for each submetric, which has implications for population cardiovascular health management. Predicting patients’ future CVH levels might directly increase patient CVH health and thus quality of life, while also indirectly decreasing the burden and cost for clinical health system caused by CVD and cancers.
BibTeX
@article{guo2021predicting,
  author = {Aixia Guo and Rahmatollah Beheshti and Yosef M. Khan and James R. Langabeer II and Randi E. Foraker},
  title = {{Predicting cardiovascular health trajectories in time-series electronic health records with LSTM models}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2021},
  doi = {10.1186/s12911-020-01345-1},
}
Figure