ABSTRACT
We aimed to develop machine learning-based predictive models for identifying inappropriate implantable cardioverter-defibrillator (ICD) therapy. Our study included 182 consecutive cases (average age 62.2 ± 4.5 years, 169 men) and employed 14 non-deep learning models for prediction (hold-out method). These models utilized selected electrocardiogram parameters and clinical features collected after ICD implantation. From the feature importance analysis of the best ML model, we established easily calculable scores. Among the patients, 25 (13.7%) experienced inappropriate therapy, and we identified 16 significant predictors. Using recursive feature elimination with cross-validation, we reduced the features to six with high feature importance: history of atrial arrhythmia (Atr-arrhythm), ischemic cardiomyopathy (ICM), absence of diabetes mellitus (DM), lack of cardiac resynchronization therapy (CRT), V3 ST level at J point (V3 STJ), and V5 R-wave amplitudes (V5R amp). The extra-trees classifier yielded the highest area under receiver operating characteristics curve (AUROC; 0.869 on test data). Thus, the Cardi35 score was defined as [+ 5.5*Atr-arrhythm - 1.5*CRT + 1.0*V3STJ + 1.0*V5R - 1.0*ICM - 0.5*DM], which demonstrated a hazard ratio of 1.62 (P < 0.001). A cut-off value of the score + 5.5 showed high AUROC (0.826). The ML approach can yield a robust prediction model, and the Cardi35 score was a convenient predictor for inappropriate therapy.
Subject(s)
Cardiac Resynchronization Therapy , Defibrillators, Implantable , Diabetes Mellitus , Heart Failure , Male , Humans , Middle Aged , Aged , Defibrillators, Implantable/adverse effects , Electric Countershock/adverse effects , Cardiac Resynchronization Therapy/adverse effects , Cardiac Resynchronization Therapy/methods , Arrhythmias, Cardiac/therapy , Cardiac Resynchronization Therapy Devices , Diabetes Mellitus/therapy , Risk Factors , Treatment Outcome , Heart Failure/therapyABSTRACT
PURPOSE: Histopathological or intracoronary image assessment of healed plaques (HPs) has been reported. However, the lesion characteristics of HPs remains undetermined yet. We assessed the healed plaque components in patients with coronary artery lesions using multiple imaging modalities. METHODS: We enrolled 33 stable angina pectoris (SAP) patients with 36 native coronary culprit lesions with angiography severe stenosis and without severe calcification undergoing pre-intervention optical coherence tomography (OCT) and coronary angioscopy (CAS). HPs were defined as layered phenotype on OCT. Lesion morphologies and plaque characteristics of HPs were assessed using OCT and CAS. RESULTS: HPs were observed in 19 lesions (52.8%). HP lesions had higher frequent B2/C lesions (89.4% vs. 52.9%, p = 0.02), worse pre-PCI coronary flow (corrected thrombolysis in myocardial infarction count 21.6 ± 13.5 vs. 13.8 ± 6.2, p = 0.047) and greater lumen-area stenosis (79.6 ± 10.6% vs. 68.0 ± 21.6%, p = 0.047) than non-HP lesions. HP lesions had higher prevalence of OCT-thin-cap fibroatheroma (TCFA) (31.6% vs. 0.0%, p = 0.02), OCT-macrophage (89.5% vs. 41.2%, p = 0.004), and CAS-red thrombus (89.5% vs. 41.2%, p = 0.004) than non-HP lesions. The combination of 3 features including OCT-TCFA, macrophages, and CAS-red thrombus showed higher predictive valuer for HPs on OCT than each single variable. Post-PCI irregular tissue protrusion was more frequently observed in lesions with HPs than in those without (52.6% vs. 13.3%, p = 0.03). CONCLUSIONS: SAP lesions with HPs might have more frequent vulnerable plaques with intraplaque inflammation and thrombus than those without, suggesting that layered phenotype on OCT might reflect not only healing process but also potential risks for future coronary events.
