A preliminary search was conducted on Medical Literature Analysis and Retrieval System Online (MEDLINE) using the MEDLINE (PubMed)i and the International Prospective Register of Systematic Reviews (PROSPERO)ii. No current or in progress similar systematic review or meta-analysis protocols were identified. This study was carried out according to a protocol, registered in PROSPERO (CRD42022374879), and followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA).12,13

We systematically searched all available literature on PubMed, Scopusiii, and Web of Knowledgeiv databases using search queries adapted for each database (described in detail in Supplemental Methods SM1. The search was restricted to articles published from 1 January 1995 (the date of the first CRT publication) through 10 November 2022. No other restrictions were applied. Manual scanning of the reference list of all selected articles was performed to identify additional pertinent articles. Specialists in the field were contacted to confirm if all pertinent information had been retrieved.

Studies that met all the following criteria were considered for inclusion: (a.) included patients with HF and implanted CRT devices; (b.) evaluated response to CRT (symptomatic and/or volumetric); (c.) defined response and stated the criteria used; (d.) predicted response based (isolated or in association) in at least one independently assessable electrocardiographic parameter; (e.) had an analytic study design; (f.) reported objective effect measures (association and/or accuracy tests). Studies that did not meet all the criteria defined above or had any of the following criteria were excluded: (a.) non-medical related research area; (b.) ECG parameter determined only post-implant (i.e. predictor not assessable pre-CRT implantation at baseline conditions; (c.) involving non-standard ECG methods/techniques (e.g. ECG mapping or ECG imaging techniques); (d.) involving animals; (e.) including humans below 18 years of age; (f.) full-text information not available, even after contacting the corresponding author; (g.) articles in languages not fully understood by the review team.

After duplicate removal, each study was screened for inclusion by two out of five reviewers (PDC, DFS, MCP, FMS and JPB). The screening was made blindly and independently, initially by reading the title/abstract and later by reviewing the full text, using Rayyan (Rayyan web application [Computer program]. Rayyan Systems Inc., 2022).14 Results were blinded to reviewers, and any discrepancies were resolved by an independent reviewer (PPR). The level of agreement between the reviewers was measured using overall agreement proportions.

Data from the selected studies was extracted to an Excel spreadsheet (Microsoft 365 Apps for Enterprise [Computer program]. Version 2210, Microsoft Corporation, 2022) by two reviewers (PDC and JPB) using standardized forms. Data extraction forms (Supplemental Methods SM2) included: (a.) study information: such as first author, date and country of publication, number of centers involved, device type(s), study design and period; (b.) population: sample size, age, gender, indication, selection method, and participants lost to follow-up; (c.) exposure: statistical method(s) used, and predictor(s) assessed; and (d.) outcomes: response cut-off, assessment timing, responder proportion, co-variables used, and effect measures.

If appropriate, authors were contacted to request unpublished, missing, or additional data. Any missing study information was reported as ‘Not reported’ (NR) and any information that was not clear or is ill-defined, leading to uncertainty in interpretation, was reported as ‘Unclear’ (U); if an item was not applicable in context, it was labeled as ‘Not applicable’ (NA).

Any disagreements between the reviewers were resolved through discussion.

To assess the risk of bias (RoB), data from the selected studies was extracted to an Excel spreadsheet (Microsoft 365 Apps for Enterprise [Computer program]. Version 2210, Microsoft Corporation, 2022) using a customized form (Supplemental Methods SM3) of the Quality in Prognostic Studies (QUIPS) tool.15 Three of the authors (PDC, DFS and MCP) independently classified all studies for the rating of reporting (‘Yes’ if all relevant signaling items were present, ‘No’ if none of the key items was present, ‘Partial’ if some key items to assess were present, ‘Unsure’ if some key items were ill-defined, unclear, or ambiguous, and ‘Not applicable’ if the key item did not apply to the study), and assigned a rating for RoB (‘High’, ‘Moderate’ and ‘Low’) to each of the six domains assessed (study participation, study attrition, prognostic factor measurement, outcome measurement, study confounding, and statistical analysis and reporting). The overall RoB was categorized into low (if all domains were classified as having low RoB, or up to one moderate RoB), high (if one or more domains were classified as having high RoB, or more than three moderate RoB), or moderate RoB (all other situations).16 The level of agreement (agreement proportions) between the reviewers was measured by overall agreement and determined for each domain/overall RoB. RoB plots were created using the Risk-of-bias VISualization (robvis) tool.17

To analyze and compare findings, categorical data were represented as proportions. Continuous variables were presented as mean (standard deviation) or median [along with interquartile range] when appropriate. Dichotomous data was analyzed by odds ratios (OR) or hazard ratios (HR) and presented as the respective test result [95% confidence interval]; receiver operating characteristic and diagnostic accuracy measures (DAM) – which included sensitivity, specificity, positive predictive value, and negative predictive value – are presented as accuracy or percentages. Data was synthesized and presented through evidence tables and included a descriptive report of eligible studies’ general characteristics, methods, and summarized results (refer to Supplemental Tables ST1–ST4). Since there were not two or more eligible studies assessing the same ECG predictor using similar methods, meta-analysis was not possible.

A total of 13503 records (13375 from databases plus 128 from citation cross-referencing) were identified and uploaded into the Rayyan platform. After duplicate removal (5742), 7633 records were screened using the title and abstract, and, of those, 182 were assessed for eligibility through full-text review. Following full-text examination for compliance, 120 articles were excluded, resulting in 62 eligible articles included in this review.18–79Figure 1 illustrates in detail the study selection process. The overall agreement between reviewers in the screening and eligibility stages was 98% and 86%, respectively.

Figure 1.

Preferred reporting items for systematic reviews and meta-analyses study selection flowchart.

Adapted from Page et al.12

The most frequent inclusion criteria comprised patients in New York Heart Association (NYHA) class II–IV, with left ventricular ejection fraction (LVEF) ≤35% and QRSd ≥120 ms, analyzing a total of 11774 patients (median of 100 patients, ranging from 30 (79) to 1718 (40)), of which 8535 (71%) were male. All studies were observational, comprising mostly cohort studies (57; 92%) from single centers (42; 68%). Studies were published between 2007 and 2022, 22 of which (35%) within the last five years. Most studies were set in European (27; 44%) and Asian countries (24; 39%). Left ventricle (LV) lead delivery was described to be via the coronary sinus in 49 (79%) studies; the remaining 13 studies (21%) did not report where the LV lead was positioned.

These results are summarized in Table 1. A detailed description of the general characteristics of all included studies is available in Supplemental Table ST1.

QRS axis revealed a marginal ability to predict response to CRT in three of the studies,37,48,67 with an OR of around one in univariate analysis. The presence of left axis deviation (LAD) appears to be associated with higher odds of responding to CRT in the study by Garcia-Seara et al.39 (adjusted OR: 5.04), albeit with a very wide CI. Similarly, Storkas et al.70 suggested that the presence of LAD is associated with lower odds of non-response to CRT (adjusted OR: 0.21).

PR interval prolongation in the studies by Gasparini et al.40 and Sabbag et al.65 showed a decreased likelihood of response, however with several concerns regarding the degree of uncertainty and the true effect size of the estimates. In contrast, according to Gasparini et al.40, a ‘normal’ PR interval is associated with a higher likelihood of response (adjusted OR: 2.51).

Two studies addressed the value of the QT interval, either by assessing super response33 or non-response.77 The work by Yu et al.77 highlights the value of the ratio between the T wave from peak to end interval (TpTe) and QTc (TpTe/QTc) in predicting CRT non-response (adjusted OR=5.2). These results need to be interpreted with caution as this study had a small size sample from a single center and, since TpTe was not statistically significant, TpTe/QTc could have been largely determined by QTc. Moreover, it showed moderate discriminatory power for TpTe/QTc in predicting non-response to CRT (AUC: 0.616; Se: 57.3%; Sp: 63.8%).

A summary of all relevant ECG predictors is presented in Table 2.

Results of association and diagnostic accuracy measures are available, respectively, in the Supplemental Tables ST2 and ST3.

Our review shows that classical predictors, such as baseline QRSd and LBBB morphology are good markers in predicting response to CRT. We found that the benefit of CRT appears to be restricted to patients with QRS >150 ms (but not longer than 178 ms), and that this benefit disappears for QRSd ≤120 ms, which is supported by guidelines changes that recognize the limited use of CRT in these patients.83 Equally, our findings suggest that LBBB is not only a good predictor of response but can also predict super response to CRT. Also, we have found that the ESC 2013 and Strauss criteria have the highest sensitivity while the AHA definition has the highest specificity in predicting response to CRT, suggesting that the highly stringent AHA definition might limit access to therapy.85 These different definitions, however, reflect not only how difficult it is to define LBBB but, more importantly, how it can impact patient selection as potential responders may be denied CRT, while others who may not benefit can receive this therapy; this was shown in the study by Van Stipdonk et al.86 who found that only 14% of patients could be classified as LBBB using the AHA and ESC definitions.

More importantly, our review shows that the role of standard ECG can go beyond the classical predictors described above and can provide additional predictive value. QRS notching and fragmentation results from the heterogeneous ventricular activation and dyssynchronous contraction,87 and was initially described as a predictor of cardiac events and HF hospitalizations in patients with cardiac disease.88 Our results show that notched/fragmented morphologies and fragmentation duration can be good predictors of non-response, in line with the findings of Balci et al.89 that found that a notched QRS duration >67.5 ms was an independent predictor of non-response to CRT; however, the study by Rickard et al.63 did not corroborate these findings. In contrast, our findings also suggest that the presence of nQRS in lateral leads or mid-QRS notching (in at least one lead) could predict response to CRT due to initial high septum activation resulting in conduction through the whole ventricular tissue and leading to more mechanical dyssynchrony, coinciding with a functional line of block toward the apex of the LV, and therefore the target of CRT.90,91

It is speculated that the presence of S-waves in right precordial leads might indicate conduction disturbance in the LV and that it is associated with smaller left atrium diameters, probably representing less left atrial remodeling and therefore within scope for CRT.92,93 Likewise, while the mechanism behind the presence of an S wave in V6 in patients with complete LBBB is not yet fully understood, Leonelli et al.94 speculate that an S wave in V6 may represent a more posterosuperior late LV vector due to biventricular enlargement while based on Upadhyay et al.95 it could be related to widespread diffuse LBBB lesion indicating intramyocardial disease, leading to poor outcome. Even though our results found a significant association of this parameter in predicting (non)response, Poposka et al.11 found no such significance in multivariate models.

Finally, literature suggests that LAD in the presence of LBBB is an independent predictor of poor prognosis96 and that patients without LAD seem to benefit more from CRT,97 with our review showing a marginal value of LAD in the prediction of response to CRT, despite one study showing higher odds of response.39 Similarly, we found that the PR interval had a limited role in response prediction, although it was suggested that patients with PR intervals between 150 and 170 ms were more likely to respond. Current literature seems to corroborate these findings, with patients with prolonged PR intervals showing worse outcomes when compared to those with a ‘normal’ PR interval,98,99 although Kutyifa et al.100 reported that a PR interval ≥230 ms was able to identify responders with non-LBBB patterns. Figure 2 summarizes the key electrocardiographic parameters, which are useful in assessing response to CRT.

Figure 2.

Key electrocardiographic parameters useful in assessing response to cardiac resynchronization therapy.

Our results are in line with those reported in the literature, showing similar response rates101 and a significant disagreement around the definition of response.102

Our study shows higher response rates in studies using symptomatic criteria when compared to those using volumetric criteria, although the agreement between these parameters is good.103 Volumetric criteria are considered not to be reliable predictors of outcome and symptomatic criteria are patient-dependent, therefore making it difficult to obtain a consistent measure.99 However, our study showed compounding symptomatic and volumetric criteria yields the highest response proportions, suggesting that this could be a more sensible approach when trying to assess response.

In this systematic review, we found a significant heterogeneity in inclusion criteria, with most studies including patients with mild to severe HF (NYHA II–IV) mostly with LVEF ≤35% and varying QRSd cut-offs. A potential explanation for this heterogeneity might be that most studies have an inclusion period that spans over several years (and, sometimes, using populations with devices implanted almost 10 years before the study was published), are based on retrospective cohorts, and use a convenience sampling often from single centers, possibly to make the study sample as large as possible.

Heterogeneity is reflected in the ill-defined inclusion criteria found in some studies.22,23,25,27,32,34,35,38,45,51,57,60,65,67,69,73,74,76 This, combined with the relatively small samples, could explain the overall high RoB of the studies included in our review. On that note, it is important to say that only a small number of studies had a low RoB regarding study participation and attrition, with some studies failing to describe the target population in detail, not giving information regarding exclusion criteria or not describing how missing data was handled, thus contributing to the high risk of selection and attrition bias. Interestingly, the moderate or high RoB shown in all the studies analyzed stresses the potential RoB for the relationship between predictors and outcome, ultimately impacting the generalization of results.

Clinically, the possibility of using simple ECG criteria that can accurately predict response is of immense value in selecting patients who are more likely to benefit from CRT. The ubiquitous, inexpensive, and non-invasive nature of the test enhances its practicality and utility, and by basing shared decision-making on robust evidence healthcare providers can more effectively optimize patient outcomes. For patients and healthcare systems, expediting the selection process while reducing the potential for iatrogeny and costs offers substantial benefits. The ability to quickly and accurately determine which patients will benefit most from CRT helps avoid unnecessary procedures and associated complications, leading to a more efficient resource allocation. Nevertheless, ECG criteria have to be reproducible and easy to use in order to be useful in clinical practice. As such, it is expected that computer-aided automation and artificial intelligence will play a vital role in CRT response prediction,104 especially if based in simple, easily accessible tests.