Usefulness of computer-assisted ECG analysis in the pre-operative evaluation of noncardiac surgery

Editor, High levels of pre-operative N terminal B-type natriuretic propeptide (NT-proBNP) have been recently identified as a promising biomarker of postoperative cardiovascular complications.1 As elevated values for these markers indicate a cardiac origin, the question of whether traces of these biomarkers could be tracked down in alterations of the ECGs certainly deserves to be addressed. This letter explores the possibility of identifying ECG predictors of postoperative major cardiovascular complications in major noncardiac surgery by using electrocardiographic biomarkers extracted automatically from pre-operative ECGs of adult patients with intermediate and high surgical risk. The original study on which this one is based1 was approved on 22 September 2011 by an independent ethics committee (Comite Etico de Investigacion Clinica del Hospital Universitario de La Princesa, Chairperson: Francisco Abad Santos, Madrid, Spain. Protocol number EC1835). A retrospective cohort study was conducted from data taken at the Hospital Universitario de La Princesa (Madrid, Spain) between 1 October 2011 and 1 July 2013. Written consent was obtained from all patients. We analysed 242 out of 304 ECGs from the originally recruited patients. The remaining ECGs were not analysed due to: missing ECGs (n=39), artefactual ECG (n=19), and patients with pacemakers (n=5). All patients’ ECGs were recorded 1 month or less before surgery. The inclusion criteria were patients older than 45 years, scheduled for major noncardiac surgery with an estimated surgical risk of either intermediate or high. Patients’ characteristics and specific cardiovascular risk factors are detailed in the original article.1 Routine 10-s 12-lead ECGs were recorded at baseline and then processed by the Philips 12-Lead algorithm,2 which automatically provides a quantified analysis of heart rate, rhythms, abnormal Q waves and amplitudes for peaks and means amplitudes durations, and morphologies.2 Thirty-seven ECGs morphological measures were quantified for each lead, such as the duration of the positive part of P (pdur according to Philips nomenclature), duration of positive + negative parts of P (ppppdur), area of QRS (qrsarea) and offset of ST-segment (stend), among many others. The primary composite endpoint of this study was cardiovascular complications and all-cause mortality, comprising: death from all causes (except those clearly associated with the surgery procedure), resuscitated cardiac arrest, stroke, exacerbation of congestive heart failure (CHF) and acute coronary syndrome. A follow-up period of 30 days after surgery was established. Univariate logistic regression with dichotomic output was used to study the association between the ECG measures and the outcome. Because the assumption of linearity between most of the continuous measures/variables and the logit of the outcome was not satisfied, the ECG measures were dichotomised by the 85th percentile or, in the case of those measures related to the ST-segment, by zero. A multivariate logistic model was thereafter constructed by using as initial independent variables the significant (with P < 0.01 to reduce overfitting and overdispersion) measures obtained from the prior univariate analysis. Independence of errors and multicollinearity were checked and the ECG measure stmid (midpoint of ST-segment) in lead II was excluded due to signs of collinearity. To determine the final model, the Akaike information criterion was used with an automated and a stepwise model selection with similar results on both procedures. The confidence interval for all estimates was 95% (95% CI). Model values were considered statistically significant when P less than 0.05. Missing values − 802 in 107 448 data points – were discarded. The incidence of major cardiovascular complications and all-cause mortality during 30 days after surgery was 5.3% (n=13). The all-cause mortality rate was 3.4% (n=8), the frequency of resuscitated cardiac arrest was 0.4% (n=1), acute myocardial infarction 0.4% (n=1) and acute episodes of CHF 2% (n=5). Twenty-seven (out of 444) ECG measures were selected (P < 0.01) by the univariate analysis, which were then used as input for the multivariate regression, obtaining the model composed of four variables: pdur in aVL (marginally significant), ppppdur in V1, qrsarea in V5 and stend in V4, independently related to the outcome. Values of pdur in aVL, ppppdur in V1 and qrsarea in V5 higher than the 85th percentile (102, 110 ms and 90.7 Ashman units, respectively) were risk factors for cardiovascular complications (OR 3.62; 95% CI: 1.06 to 11.96; P = 0.076), (OR 4.96; 95% CI: 1.51 to 16.36; P = 0.024) and (OR 4.3; 95% CI: 1.27 to 14.7; P = 0.043), respectively. Also, values of stend in V4 lower than zero convey a risk factor (OR 9.4; 95% CI: 1.8 to 68.1; P = 0.037). Figure 1 illustrates the documented independent ECG marker. Two or three-factor significant interactions were not found. When proBNP was included in the multivariate model, similar model parameters were obtained. Moreover, pre-operative ProBNP showed a significant correlation – Spearman test – with the aforementioned measures: 0.20 (P = 0.003) with pdur (aVL), 0.14 (P = 0.04) with ppppdur (V1), and −0.32 (P < 0.001) with stend (V4), with no correlation for the case of QRS area in V5 −0.07 (P = 0.31). A further assessment by bootstrapping with 1000 replicas confirmed that all correlation values were within a 95% CI.Fig. 1: ECG measurements by Philips algorithm. The figure illustrates how the Philips algorithm recognises relevant portions of ECG. The QRS area is calculated by the sum of small trapezoids enclosed by the QRS complex. Area above the baseline is positive while area below the baseline is negative. Philips nomenclature in bold parentheses.In this letter, we have explored the possibility of using computer-assisted ECGs predictors of postsurgical cardiac complications, as an alternative to high levels of pre-operative NT-proBNP levels. We found that duration of the P wave in V1, QRS area in V5 and ST offset depression in V4 are independent predictors of complications. These results are consistent with previous findings in other clinical scenarios different from the surgery risk stratification: P wave prolongation has been related to several complications3,4 even with no sudden cardiac death,5 ST-segment deviation is associated with hypertrophic cardiomyopathy6 and R amplitude in V5 is part of the Sokolow–Lyon criteria. Remarkably, the Cornell product – a surrogate of the QRS area – showed to be the ECG criteria with the best performance for detecting hypertrophy among obese patients.7 This predictor highlights the importance of computer-assisted evaluation of ECG, since QRS area is not routinely measured in the clinical practice. An extensive research on a wider sample could shed further light on the potential clinical utility of these preliminary results, fundamentally considering that most commercial implementations, for example General Electric Marquette 12L, are based on similar algorithms as the one used by Philips.