Machine-learning-based classification of ACS-causing coronary plaque morphologies to accelerate precision management in acute coronary syndrome

Elaaha Anwari (Berlin)1, Y. Abdelwahed (Berlin)1, C. Seppelt (Frankfurt am Main)2, D. Meteva (Berlin)1, T. Gerhardt (Berlin)1, J. Musfeld (Berlin)1, L. Sieronski (Berlin)1, N. Kränkel (Berlin)1, L. M. Seegers (Frankfurt am Main)2, U. Landmesser (Berlin)1, D. Leistner (Frankfurt am Main)2

1Charité - Universitätsmedizin Berlin CC 11: Med. Klinik für Kardiologie Berlin, Deutschland; 2Universitätsklinikum Frankfurt Med. Klinik III - Kardiologie, Angiologie Frankfurt am Main, Deutschland


Background: Identification and characterization of acute coronary syndrome (ACS)-causing culprit plaques by high-resolution intracoronary optical coherence tomography (OCT) allows personalized management of patients with ACS. This is especially crucial when distinguishing between culprit lesions (CL) with ruptured - (RFC) and intact-fibrous cap (IFC) as such classifications play a pivotal role in clinical decision-making and in evaluating prognosis. While OCT-evaluation of the ACS-causing culprit plaque morphology is rather effective, its invasivity limits broad clinical implementation.

ObjectiveThe aim of this study was to develop a machine-learning-based model that allows effective differentiation between RFC- and IFC-ACS, utilizing routinely available clinical variables of both routine-invasive and non-invasive diagnostic parameters.

Methods: A representative subset of the OPTICO-ACS study (NCT03129503) was investigated. The dataset was split into a training set (70%) and a test set (30%). The optimum discriminatory cutoff point for age, Body-Mass-Index, LDL, creatinine, aortic pulse pressure (APP), pulse wave velocity (PWV), heart rate-corrected augmentation index, intima-media-thickness (IMT) and the distance to the next coronary branch was determined using the Youden's J statistic and included as a binary variable in the prediction models. We used Least Absolute Shrinkage and Selection Operator (LASSO) regressions with 10-fold cross-validation and subsequently assessed their performance using the Area under the Curve (AUC). 
We established three distinct models: The Basic model, which incorporates regularly available clinical variables; the Advanced model, which adds arterial stiffness metrics and IMT; and the Full model, which further includes angiographic parameters. Derived from the model coefficients, a normalized risk score (0-100) was defined for the classification of RFC and IFC.

Results: In total, 96 patients (81% male) with RFC-ACS (69%) and IFC-ACS (31%), with a median age of 62.0 years were included in this study. The predictive power improved with the addition of arterial stiffness metrics (Advanced model: AUC = 0.74) and improved even further with the subsequent inclusion of angiographic findings (Full model: AUC = 0.82) (Fig. 1).
Reclassification analysis revealed that a model containing a combination of clinical covariates, arterial stiffness parameters and angiographic findings significantly outperforms its isolated components (Full vs. Basic model: Net reclassification index (NRI) 0.87 [95% CI 0.46-1.27], p<0.001), Integrated discrimination improvement (IDI) 0.16 [95% CI 0.05-0.27], p=0.005; Full vs. Advanced model: NRI 1.13 [95% CI 0.73-1.54], p<0.001, IDI 0.11 [95% CI 0.03-0.18], p=0.004) (Fig. 2).

Conclusion: A model that integrates commonly available clinical data, arterial stiffness parameters, and angiographic findings offers potential as an effective tool for differentiating between RFC- and IFC-ACS and therefore allows for the first time for a personalized, pathophysiology-guided management of patients with ACS.

Fig. 1:
 Receiver Operating Characteristic (ROC) curves

Fig. 2: Model development and reclassification metrics

Legend: BMI = Body-Mass-Index; DM = Diabetes mellitus; AHT = Arterial hypertension; LDL-C = Low-density lipoprotein cholesterol; PWV = Puls wave velocity; APP = Aortic pulse pressure; AIx@75 = Heart rate-corrected augmentation index; IMT = Intima-media-thickness; MVD = Multivessel disease.

Diese Seite teilen