In this cross-sectional study, we enrolled 346 PLWH who underwent 64-detector row coronary CTA evaluation between December 2019 and June 2023. The inclusion criteria for PLWH included (1) diagnosis of HIV infection, (2) age between 20 and 80 years, (3) undergoing ART. The exclusion criteria included: (1) previous history or symptoms CAD, (2) contraindications of CCTA examination, (3) poor quality of coronary images, and (4) not have complete clinical information (Figure 1). The final dataset contained no missing values. Eventually, 304 patients were enrolled for final analysis. Fifty of these patients have been used for the earlier MRI study, but the current CT study is analyzing more patients with a completely different goal and result (14). This study was approved by the local Ethics Review Board and conducted in accordance with the Declaration of Helsinki.

Prior to the coronary CTA examination, demography (age, sex), clinical information (cardiovascular risk factors, antihypertensive therapy, HIV acquisition mode, hepatitis C, intravenous drug use, the time since HIV diagnosis), hematologic and immunological factors (estimated glomerular filtration rate, HIV viral load, current CD4⁺ T-cell count, CD8⁺ T-cell count, and nadir CD4⁺ T-cell count, CD4⁺/CD8⁺), and treatment-related factors (duration on ART, ART regimen) was prospectively collected for all PLWH participants.

Coronary CTA was performed using a third-generation dual-source CT (SOMATOM Force, Siemens Healthcare, Forchheim, Germany) and 45 to 100 mL of intravenous contrast (Iopamidol Injection, 370 mg/mL, Bracco Si, Shanghai, China) at a flow rate of 4.0–5.0 mL/s. Retrospective electrocardiographic gating was used in all of the participants. The main scan parameters were as follows: tube voltage = 100 kV, tube current = 300–400 mas, slice thickness = 0.75 mm, detector collimation = 2 mm × 64 mm × 0.6 mm, and matrix = 512 × 512. A non-contrast CT examination was performed for coronary calcium scoring.

All the scans were sequentially analyzed and reviewed on a workstation (Syngo.via version VB20, Siemens Healthineer, Forchheim, Germany) by two radiologists (with three and fifteen years of experience, respectively) using axial images and multiplanar reconstructions. The coronary segment was analyzed by using the 16-segment American Heart Association definitions from 1975 (15). The outcome was the presence of obstructive CAD demonstrated by coronary CTA, which has been shown in previous studies to have a high diagnostic accuracy for obstructive CAD (16, 17). Obstructive CAD was defined as at least 1 coronary segment with a lesion of ≥50% luminal stenosis in diameter (18). Coronary artery calcium score (CACS) measurement was performed using the Agatston method (19). Observations between the two readers were analyzed for interobserver agreement.

We developed a total of 11 models to predict obstructive CAD. There are three traditional conventional pre-test probability models of coronary CTA. All our patients were asymptomatic, so clinical symptoms could not be included as a feature in the traditional model. In addition to gender, age, hypertension, diabetes, and smoking status, CAD consortium clinical score requires hyperlipidemia and BMI, CRS requires cardiovascular family history, and GCM requires hyperlipidemia. The logistic regression (LR) model was built using all available features, including traditional cardiovascular and HIV-specific variables. Moreover, we added CACS to above four models to create four additional models: CAD + CACS, CRS + CACS, GCM + CACS, and LR + CACS. A CACS model using only the calcium score was included for comparison.

Furthermore, three ML models were developed based on the random forest algorithm. Model 1 incorporated all available clinical, laboratory, and HIV-specific variables. After adding CACS, Model 2 was reconstructed with renewed feature selection and hyperparameter tuning. To evaluate the incremental value of CACS, Model 3 retained the features and hyperparameters selected in Model 1, with CACS added as an additional variable.

Consistent with the 2019 ESC Guidelines, our machine learning model incorporates multiple cardiovascular risk factors to estimate the likelihood of obstructive CAD (20). To facilitate clinical risk interpretation, we adopted the risk stratification thresholds proposed in the 2013 ESC Guidelines based on pretest probability: low risk (<15%), intermediate risk (15–85%), and high risk (>85%) (21). Patients with a low pretest probability are less likely to benefit from additional testing, while those with an intermediate probability are most likely to benefit from initial noninvasive testing.

Random forest algorithm was performed to model our data. The original dataset was randomly divided into training and held-out test sets in an 8:2 ratio to maintain the ratio of obstructive to non-obstructive CAD in both subsets. In addition, model hyperparameters are optimized by five-fold cross-validation, which repeats five iterations by randomly splitting the training set into five equal-sized subsets, four were used for inner training and one was used for inner test. The employ of cross-validation in the training set can be beneficial as its primary use is to empirically determine the optimal model hyperparameters without resorting to the test set (22). Subsequently, the optimal hyperparameters are applied to evaluate the model in the test set (maximum tree depth: 4; number of trees: 5; random seed: 2175). Finally, the area under the curve (AUC) and the associated 95% confidence interval were used to measure the classification performance of the ML model. Feature selection is performed using the Boruta algorithm. The feature selection and modeling process were performed in the training set, and the independent test set was only used to evaluate the performance of the model and was not involved in the training process.

Continuous variables were expressed as median with 25th–75th percentile interquartile range (IQR), depending on the normality of distribution assessed by the Kolmogorov–Smirnov test. Categorical variables were presented as frequencies and percentages. Comparisons between groups were carried out using an independent t-test or the Mann–Whitney U test for normally and non-normally distributed continuous variables and the chi-square test or Fisher’s exact test for categorical variables, as appropriate.

The discriminative power of Model 1 was compared with the 4 traditional models and LR model, respectively, using receiver-operating characteristic analysis and pairwise comparisons according to Delong et al. (23). p < 0.05 indicated statistical significance between the performance of the models. Calibration curves were drawn to explore the agreement between the observed outcome frequencies and predicted probabilities of the models in the testing group. Decision curve analysis was used to assess the clinical utility of these models by quantifying the net benefits at different risk threshold probabilities in testing set.

All statistical analyses were carried out using IBM SPSS software (V26.0; IBM Corporation, Armonk, New York, USA) and R software (version 4.3.1).

Figure 3 presents the feature importance rankings generated by the Boruta algorithm. As shown, age, total cholesterol, and current CD4⁺ T-cell count were the most predictive features in the Model 1. This was followed by years of ART use, triglycerides, nadir CD4⁺ T-cell count, and diastolic blood pressure. Interestingly, the importance of traditional cardiovascular risk factors increased with the addition of the CACS into the model, the most predictive features (after the CACS itself) were triglycerides and age followed by estimated glomerular filtration rate and CD4⁺/CD8⁺. The risk factors for obtaining models after screening are shown in Table 2.

The machine learning model (Model 1, AUC of 0.946), which incorporated both traditional cardiovascular risk factors and HIV-specific metrics, showed the best performance in predicting obstructive CAD compared with CAD Consortium clinical score, CRS, and GCM (AUC of 0.734, 0.736, and 0.737, respectively; p < 0.05 for all comparisons). In order to control for potential confounding due to algorithm differences, we additionally constructed a logistic regression (LR) model using the full feature set (LR model, AUC of 0.782; p < 0.05), confirming the superior predictive capability of the random forest–based Model 1 (p < 0.05). Furthermore, the LR model did not differed significantly from traditional models. The sensitivity, specificity, positive predictive value, negative predictive value, and accuracy for the prediction of obstructive CAD were 0.846, 0.958, 0.846, 0.958, and 0.934 for Model 1, respectively. Interestingly, Model 2, developed by adding CACS and reapplying feature selection, yielded a slightly lower AUC of 0.926 but still outperformed the CAD + CACS model, CRS + CACS model, GCM + CACS model, and LR + CACS model (AUC of 0.740, 0.742, 0.750, and 0.798; p < 0.05 for all comparisons). After adding CACS to the Model 1 feature set without re-optimization, Model 3 demonstrated a reduced AUC of 0.890, with no statistically significant difference compared to Model 1 or Model 2 (p > 0.05). Meanwhile, the performance of Model 1 and Model 2 significantly outperformed the CACS models (AUC of 0.776; p < 0.05), suggesting that clinical and HIV-specific features contributed more substantially to predictive performance than CACS alone. The discriminative performance of all models is shown in Tables 3, 4 and Figure 4.

In both groups, the calibration curves of Model 1 and Model 2 showed the best calibration of the predicted probability with the true probability (Brier scores of 0.076 and 0.102, respectively) (Figure 5). The clinical usefulness is shown as decision curve in Figure 6. The ML Models demonstrated superior net benefit in decision curve analysis when the threshold probability ranged from 0 to 70%.