Prediction of obstructive coronary artery disease in people living with HIV: value of machine learning incorporating HIV-specific factors

Hui Liu, Haibo Ding, Yue Zheng, Yue Li, Yu Yang, Zhaodi Geng, Jie Zhou, Huaibi Huo, Han Li, Xin Peng, Zhaoxin Tian, Xiaolin Li, Hong Shang, Ting LiuFrontiers in Medicine 2025 · DOI

This study, published in Frontiers in Medicine, demonstrates that a machine learning (ML) approach—specifically a Random Forest model—significantly outperforms traditional clinical scores in predicting obstructive coronary artery disease (CAD) in asymptomatic people living with HIV (PLWH). By integrating HIV-specific biomarkers like CD4+ T-cell counts and duration of antiretroviral therapy (ART) with traditional cardiovascular risk factors, the researchers developed a tool that could prevent unnecessary diagnostic imaging while identifying high-risk patients who might otherwise be missed.

TL;DR

Researchers developed a Random Forest machine learning model that predicts heart disease in asymptomatic HIV patients with far greater accuracy (AUC 0.946) than standard clinical tools. The model highlights that HIV-specific factors, such as ART duration and nadir CD4 counts, are critical non-linear predictors of cardiovascular risk that traditional linear models fail to capture.

What Problem Does This Paper Address?

People living with HIV (PLWH) face a significantly higher risk of cardiovascular disease than the general population, driven by chronic inflammation, immune activation, and the metabolic side effects of long-term antiretroviral therapy (ART).

The current clinical challenge is "pre-test probability." Doctors need to decide which asymptomatic patients should undergo Coronary CT Angiography (CTA)—a test that involves radiation and contrast agents. Existing prediction models (like the CAD consortium or CONFIRM scores) were designed for the general population and typically for patients already exhibiting chest pain. These models perform poorly for PLWH because they ignore the unique biological footprint of HIV. This paper seeks to fill that gap by building an ML model tailored to the specific risk profile of the HIV population.

Methods

The researchers conducted a prospective cross-sectional study using a pipeline that moved beyond simple linear statistics:

  • Algorithm: They employed a Random Forest (RF) model, chosen for its ability to handle non-linear relationships and interactions between complex variables.
  • Feature Selection: The Boruta algorithm was used to rank and select the most relevant features from a pool of clinical, laboratory, and HIV-specific data.
  • Comparison: The ML model was pitted against three traditional clinical scores (CAD consortium, CONFIRM, and Genders clinical model) and a standard Logistic Regression (LR) model.
  • Validation: The dataset was split 8:2 into training and held-out test sets, with 5-fold cross-validation used for hyperparameter tuning.
  • Incremental Value: They also tested whether adding the Coronary Artery Calcium Score (CACS)—a common imaging-based risk marker—improved the models.

Datasets

The study analyzed a dataset of 304 PLWH (mean age 48, 91% male) recruited from the First Hospital of China Medical University.

  • Inclusion Criteria: Confirmed HIV diagnosis, aged 20–80, currently on ART, and asymptomatic (no chest pain).
  • Ground Truth: Obstructive CAD was defined as ≥50% luminal stenosis confirmed via third-generation dual-source CT angiography.
  • Features: Included demographics, traditional lipids, eGFR, and HIV-specific metrics (nadir CD4+ count, current CD4+ count, CD4+/CD8+ ratio, and years on ART).

Key Insights and Ideas

  • The Non-Linearity of HIV Risk: A key "aha moment" was that the Logistic Regression model—even when fed HIV-specific data—did not significantly outperform traditional scores. Only the Random Forest model saw a massive jump in performance. This suggests that the relationship between HIV markers (like CD4 counts) and heart disease is complex and interactive, rather than a simple straight-line correlation.
  • HIV History Matters: "Nadir CD4+ count" (the lowest CD4 level a patient ever had) and "years of ART use" emerged as top predictors. This confirms that a patient's historical immune health is a lasting "legacy" risk factor for their heart, even if their current viral load is suppressed.
  • Redundancy of Calcium Scoring: Surprisingly, adding the imaging-based Calcium Score (CACS) did not significantly improve the best ML model (Model 1). This implies that the clinical and HIV-specific data already capture the risk information that CACS provides, potentially allowing clinicians to skip the calcium scan in favor of a purely data-driven pre-test assessment.

Key Results

  • Superior Discrimination: The ML model (Model 1) achieved an AUC of 0.946, vastly outperforming the CAD consortium (0.734), CONFIRM (0.736), and Genders (0.737) models ($p < 0.05$).
  • High Accuracy: In the independent test set, the ML model showed a sensitivity of 84.6% and a specificity of 95.8%, with an overall accuracy of 93.4%.
  • Clinical Utility: Decision Curve Analysis (DCA) showed that the ML model provided a higher "net benefit" across almost all threshold probabilities (0–70%) compared to traditional methods, meaning it makes better clinical decisions about who to test.
  • Calibration: The model showed excellent calibration (Brier score 0.076), meaning its predicted probabilities closely matched the actual observed frequency of disease.

Limitations and Future Work

The authors acknowledge several hurdles before this can be used in a clinic:

  • Single-Center Study: The data comes from one hospital in China; the model needs to be validated on more diverse, multi-ethnic cohorts.
  • Lack of External Validation: While the "held-out" test set is a good start, true external validation on a completely different patient population is required.
  • Ground Truth Limits: The study used CT angiography as the gold standard rather than invasive coronary angiography (the absolute gold standard), though CTA is widely accepted for this type of research.
  • Future Direction: The researchers suggest integrating this model into a Clinical Decision Support System (CDSS) to help doctors triage HIV patients for cardiac care in real-time.

Summary updated

Abstract

The ML model incorporating cardiovascular risk factors and HIV-specific factors can more accurately estimate the pretest likelihood of obstructive CAD in PLWH than traditional models. ML improves risk stratification in HIV populations and may help guide management.
BibTeX
@inproceedings{liu2025prediction,
  author = {Hui Liu and Haibo Ding and Yue Zheng and Yue Li and Yu Yang and Zhaodi Geng and Jie Zhou and Huaibi Huo and Han Li and Xin Peng and Zhaoxin Tian and Xiaolin Li and Hong Shang and Ting Liu},
  title = {{Prediction of obstructive coronary artery disease in people living with HIV: value of machine learning incorporating HIV-specific factors}},
  booktitle = {Frontiers in Medicine},
  year = {2025},
  doi = {10.3389/fmed.2025.1661990},
}
Figure