The Long March to Identify Patients With Coronary Artery Disease by Pretest Probability Model

Corresponding Author

Coronary computed tomography angiography (CTA) has a high negative predictive value for ruling out obstructive coronary artery disease (CAD), making it an effective gatekeeper to invasive procedures in the catheterization laboratory.¹ It also allows comprehensive evaluation of coronary anatomy, plaque burden, and high-risk plaque features. Moreover, advanced computational assessments derived from coronary CTA, such as computational fractional flow reserve² and inflammatory evaluation of perivascular fat attenuation index³ transcend traditional anatomical assessment. These novel techniques have demonstrated incremental prognostic value in identifying patients who are at elevated risk of adverse cardiovascular events.

However, indiscriminate use of coronary CTA for all patients may strain clinical resources, especially in low-prevalent populations. Current guidelines⁴ recommend initial evaluation of an individual’s pretest probability of obstructive CAD to guide further diagnostic tests. Developing more refined risk stratification models to estimate the clinical likelihood of obstructive CAD helps identify patients most likely to benefit from coronary CTA, thereby reducing unnecessary testing and improving resource efficiency.

Continuous Efforts in Risk Stratification

The journey of risk stratification for CAD has progressed significantly over the decades, evolving from simple clinical scores to sophisticated artificial intelligence (AI)-empowered predictive models. In the European Society of Cardiology Guidelines,⁴ an updated Diamond-Forrester approach based on a pooled analysis⁵ included age, sex, and the nature of symptoms to estimate clinical likelihood of CAD. The RF-CL model refined this approach by weighting risk factors based on their relative contributions to CAD risk.⁶ Other methods, such as the CAD Consortium and the CONFIRM registry scores, incorporate additional contemporary risk factors like smoking, diabetes and hypertension.⁷ Although groundbreaking, these scores are limited by linear statistical models, failing to capture complex risk factor interactions. Machine learning (ML) offers a significant advantage by capturing nonlinear relationships within high-dimensional data.⁸ Thus, recent advancements have leveraged ML to integrate a broader range of risk factors.⁹ However, it should be noted that existing models have predominantly been developed and validated in Western populations, raising concerns about their generalizability to diverse ethnic groups.

In this issue of JACC: Asia, Dou et al¹⁰ developed the C-STRAT Score, an ML-based predictive model based on a large-scale, prospective, multi-center Chinese registry. A total of 27,652 patients were enrolled in the analysis, with 10% identified as obstructive CAD on coronary CTA. The data set was randomly split into training data set (70%) and testing data set (30%). A total of 11 readily available demographic and clinical variables were used to develop an ensemble ML algorithm, ie, extreme Gradient Boosting (XGBoost). The model was specifically designed for Chinese population to overcome the challenge of inconsistent performance of previously developed models among different populations. This paper should be seen as a valuable contribution in the field of risk stratification to streamline patients before referring to coronary CTA, addressing the under-representation of Asian cohorts in prior models.

Validation of a New Model Via Multifaceted Approach

As new models driven by AI thrive in the risk stratification field, it is important to evaluate them thoroughly to fully understand their strengths and limitations. The authors should be congratulated by performing a comprehensive evaluation of the newly proposed model, demonstrating strong performance across discrimination, calibration, reclassification, and clinical utility. The C-STRAT Score achieved a superior discriminative performance with AUC of 0.769 (95% CI: 0.753-0.784) in the internal testing data set. The integrated discrimination improvement (IDI) and net reclassification improvement (NRI) methods were used to evaluate the incremental value of the C-STRAT score. IDI compares the mean predicted probabilities between positive and negative cases, whereas NRI evaluates the correctness of changes in predicted risk categories by the new model. The proposed C-STRAT showed improvement compared with traditional models with the smallest IDI of 0.54% (95% CI: 0.09%-1.18%; P = 0.039) and NRI of +3.46% (95% CI: 0.06%-6.98%; P = 0.045), respectively.

Beyond discrimination, calibration evaluation assesses the concordance between predicted probabilities and observed outcomes. A robust predictive model should excel in both discrimination and calibration. For instance, a model with a high AUC but poor calibration produces unreliable probability estimates despite good class separation. Conversely, a well-calibrated model with low discrimination may lack clinical utility caused by insufficient predictive power. The proposed C-STRAT score showed good calibration via calibration plot (slop of 1.077) and Brier Score (0.081).

The performance of a predictive model in real-world decision making is another crucial aspect of evaluation, ensuring the model can be practically implemented. Decision curve analysis addresses this by weighing the trade-offs between benefits of improved outcomes and potential harm. Decision curve analysis demonstrated the practical value of the C-STRAT score, supporting its potential clinical applications.

Study Limitations and Prospects

Although AUC is commonly used to assess models’ discriminative performance, it cannot fully represent the accuracy of the model in imbalanced data sets because the majority of negative class will be correctly classified. In such scenarios, the precision-recall curve is more informative because it focuses on the rare (positive) class. In this study, the proposed C-STRAT score outperformed other models in AUC; however, given the relatively low positive rate (10%), the precision-recall curve would provide a more clinically relevant assessment. Unfortunately, the granular analysis is not available.

Another limitation of this study is the absence of external validation in independent cohorts, which is essential to confirm the generalizability and robustness of the proposed model. Although population-specific models like C-STRAT are critical for addressing regional disparities, the trend in the era of big data lies in developing universally applicable models to ensure robustness across ethnicities, geographies, and clinical contexts. To date, despite notable progress, current AI-empowered models still lack the generalization and precision required to be a true game-changer. However, emerging AI technologies, such as foundation models, hold great potential by incorporating multimodal data beyond traditional risk factors—such as biomarkers, monitoring data from wearable devices, historical health records, and so on. As high-quality, diverse data sets accumulate and AI algorithms continue to advance, the full potential of AI-driven risk stratification would revolutionize CAD diagnosis and management. As these innovations unfold, rigorous external validation and prospective clinical trials will be critical for translating these technologies into practice—ultimately improving patient outcomes and health care efficiency.

Funding Support and Author Disclosures

Dr Tu is the cofounder of Pulse Medical; and has received research grants and consultancy from Pulse Medical outside of this work. Dr Chu has reported that they have no relationships relevant to the contents of this paper to disclose.