A machine learning algorithm to predict a culprit lesion after out of hospital cardiac arrest

Abstract

Abstract Background Between 50-90% of patients suffering from out-of-hospital cardiac arrest (OHCA) have a primary cardiac aetiology of arrest with the presence of a culprit coronary lesion. However, recent trials have failed to show a benefit from an early invasive angiographic approach following OHCA. Identification of patients with high likelihood of a culprit lesion may enable selection of patients likely to benefit from direct conveyance to cardiac arrest centres and for an early invasive approach. However, electrocardiogram (ECG) following return of spontaneous circulation, the current gold-standard, is a poor predictor of presence of a culprit lesion. Purpose We aimed to develop a machine learning algorithm to predict the presence of a culprit lesion in patients with OHCA. Methods We used a retrospective cohort of 398 patients admitted with primary cardiac aetiology OHCA to a tertiary cardiology centre in the United Kingdom, between May 2012 and December 2017. The primary outcome was the presence of a culprit coronary artery lesion, for which a Gradient Boosting model was optimised to predict. The algorithm was then validated in two independent European cohorts comprising 568 patients (346 and 222 patients respectively). Results The development cohort consisted of 398 patients (median age 64.3 years [IQR 53.3-75.4], 74.6% male), and the two validation cohorts consisted of 346 patients (median age 63.0 years [IQR 55.0-73.0], 81.5% male) and 222 patients (median age 63.0 years [IQR 52.0-74.0], 77.0% male). Of all patients, 85.6% had a witnessed arrest, 69.6% received bystander cardiopulmonary resuscitation, and 77.3% had an initial shockable rhythm. A culprit lesion was observed in 67.4% patients receiving early coronary angiography in the development cohort, and 67.9% and 61.1% in the two validation cohorts respectively. A Gradient Boosting model yielded an algorithm, named KOCAR Culprit Predictor, incorporating nine variables including age, a localising feature on ECG (≥2mm of ST change in contiguous leads), regional wall motion abnormality, history of vascular disease and initial shockable rhythm. This model had an AUC of 0.89 in the development and 0.83/0.81 in the validation cohorts with good calibration (Figure 1), and outperformed the current gold-standard ECG alone (AUC 0.69/0.67/0/67). Conclusions A novel simple machine learning derived algorithm (Figure 2) combining nine variables which are readily available to clinicians can be applied to patients with OHCA to predict a culprit coronary artery lesion with high accuracy. The algorithm, KOCAR Culprit Predictor, outperforms the current gold-standard ECG alone.Figure 1Figure 2