An explainable machine-learning model that incorporated traditional cardiovascular risk factors and inflammatory biomarkers showed reasonable discrimination for angiographic coronary artery disease among patients without acute myocardial infarction who had already been referred for coronary angiography, researchers reported.
The retrospective, single-center study initially enrolled 4,656 patients treated at Beijing Anzhen Hospital in China between August 8, 2021, and November 10, 2023. All patients underwent coronary angiography for suspected coronary artery disease (CAD). After exclusions, 3,482 patients were included in the final analysis. CAD was defined as at least 50% stenosis in at least one major coronary artery.
Patients were excluded for acute myocardial infarction, severe valvular disease, advanced heart failure, significant arrhythmias, major hepatic or renal dysfunction, malignancy, autoimmune disease, active infection, or recent use of anti-inflammatory or immunosuppressive drugs.
The analytic cohort had a high prevalence of angiographic disease. Of the 3,482 patients included, 3,180 had CAD and 302 did not. Because of that imbalance, the study’s balanced accuracy and specificity metrics may be more informative than accuracy alone for interpreting model performance in this data set. The researchers reported that balanced accuracy was included to better characterize model discrimination under class imbalance.
Researchers collected demographic characteristics, clinical variables, routine laboratory measures, and 12 inflammatory cytokines. Least absolute shrinkage and selection operator regression identified 21 candidate features, and 10 machine-learning models were evaluated in a 70% training cohort and a 30% independent testing cohort. Researchers used stratified sampling, repeated cross-validation, and oversampling methods to address class imbalance.
In the independent testing cohort, the generalized linear model selected as the final model achieved 81% accuracy, 82% sensitivity, 71% specificity, 77% balanced accuracy, and an area under the curve of 0.815. The researchers selected this model for interpretability, clinical transparency, and relatively balanced overall performance, not because it had the highest discrimination.
Several more complex models had numerically higher areas under the curve. The C5.0 model had the highest area under the curve at 0.847, while XGBoost had 92% accuracy and 97% sensitivity. However, those models had lower specificity, including 39% for C5.0 and 47% for XGBoost, indicating more false-positive classifications among patients without CAD. Pairwise DeLong tests among the top-performing models did not show statistically significant differences in area under the curve after adjustment, according to the researchers.
SHapley Additive exPlanations analysis identified hypertension, hyperlipidemia, sex, diabetes, and triglyceride levels as the most influential predictors in the final generalized linear model. Smoking history, low-density lipoprotein cholesterol, albumin, lipoprotein(a), and uric acid also contributed to model output.
The role of inflammatory biomarkers was less clear. Interleukin-6, interleukin-8, and interferon-alpha were retained after feature selection, but interleukin-6 ranked 14th among the 17 features in the stepwise SHapley Additive exPlanations analysis. The study did not report a direct model comparison quantifying whether the 12 inflammatory cytokines improved performance beyond traditional cardiovascular risk factors and routine laboratory measures alone.
Baseline findings also complicated interpretation of the inflammatory signal. Patients with CAD had lower interleukin-6, interleukin-8, interferon-alpha, and low-density lipoprotein cholesterol levels compared with patients without CAD. In model interpretation, higher interleukin-6 and low-density lipoprotein cholesterol levels were also associated with lower predicted probabilities of CAD.
The researchers suggested that prior treatment, particularly statin therapy, may have contributed to these inverse interleukin-6 and low-density lipoprotein cholesterol findings. However, medication use was not directly assessed in the model, and the researchers noted that medications could have confounded biomarker interpretation.
The study’s intended use remains uncertain. The model was developed in patients who had already undergone coronary angiography for suspected CAD, and the researchers noted that the model has not been validated in an independent external cohort. The study did not establish performance in lower-prevalence outpatient or primary care populations.
The model should also be interpreted as a classifier of angiographic stenosis, not ischemia, plaque vulnerability, or future cardiovascular events. CAD was defined by stenosis severity alone, without fractional flow reserve, instantaneous wave-free ratio, intravascular ultrasound, optical coherence tomography, or coronary computed tomography angiography.
Other limitations included the retrospective single-center design, limited inflammatory biomarker panel, absence of electrocardiography and echocardiography as candidate predictors, and the relatively small non-CAD group. The researchers reported discrimination and classification metrics but did not report calibration or decision-curve analyses.
The researchers concluded that prospective multicenter validation and additional multimodal data will be needed to determine the model’s clinical utility.
Disclosures: The study was funded by the Noncommunicable Chronic Diseases–National Science and Technology Major Project and the National Natural Science Foundation of China. The researchers reported no competing interests.
Source: Open Heart