Article In Press : Article / Volume 5, Issue 2

Total Arterial Grafting and Long-Term Survival after Isolated Coronary Artery Bypass Surgery: An Overlap-Weighted Analysis of a 27-Year Single-Centre Registry

Salik Akhtar1Amina Khalil2Nandor Marczin3Sunil K. Bhudia2Shahzad G. Raja2*

1Faculty of Medicine, Imperial College London, London, United Kingdom
2Department of Cardiac Surgery, Harefield Hospital, London, United Kingdom.
3Department of Anaesthesia and Critical Care, Harefield Hospital, London, United Kingdom.

Correspondng Author:

Shahzad G. Raja, Department of Cardiac Surgery, Harefield Hospital, London, United Kingdom.

Copyright:

© 2026 Shahzad G. Raja. This is an open-access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

  • Received Date: 30-10-2026   
  • Accepted Date: 10-11-2026   
  • Published Date: 19-11-2026
Abstract Keywords:

Coronary artery bypass grafting, Internal thoracic artery, Propensity score, Overlap weighting, Survival analysis, Total arterial grafting.

Abstract

Background: Total arterial grafting avoids venous conduits and may improve long-term outcomes after coronary artery bypass grafting, but its survival advantage over mixed arterial–venous grafting remains uncertain because it is preferentially offered to younger, fitter patients. We estimated the association between total arterial grafting and long-term survival among patients compatible with either strategy and examined whether treatment effects varied across clinically relevant subgroups.

Methods: In a single centre retrospective cohort of 13,831 consecutive isolated coronary artery bypass operations between February 1996 and September 2023, 11,864 multigraft operations were analysed, 1,843 total arterial and 10,021 mixed. Propensity score overlap weighting targeted the population in clinical equipoise. Outcomes were mortality by overlap weighted Cox regression, adjusted survival and restricted mean survival time, seven prespecified interactions, ten early outcomes and an exploratory cross fitted personalised model.

Results: There were 1,424 deaths over a median 12.5 years. Weighting removed all imbalance (maximum standardised mean difference 0.001). The hazard ratio for death was 0.885 (95% CI 0.743 to 1.041), the adjusted survival difference at 15 years +1.0 percentage points (-0.7 to +2.9), and the restricted mean survival difference to 15 years +0.07 years. No interaction was significant after multiplicity adjustment, and the personalised model over-stated benefit and did not calibrate.

Conclusions: Total arterial grafting was not associated with a clear population-average long-term survival advantage, and no reproducible beneficiary subgroup was identified. These findings support individualised conduit selection based on patient characteristics, coronary anatomy, conduit suitability, anticipated longevity and surgical expertise rather than routine use of a total arterial strategy.

Introduction

The left internal thoracic artery graft to the left anterior descending artery has been the foundation of coronary artery bypass grafting since the long term survival advantage of that single arterial conduit was demonstrated [1]. Whether extending the arterial principle to every distal target confers a further survival advantage is less certain. Saphenous vein grafts remain in wide use despite their inferior long term patency, and contemporary European and North American guidance recommends multiple arterial grafting for selected patients rather than for all [2-3].
The randomised evidence is not settled. The Arterial Revascularization Trial found no significant difference in mortality between bilateral and single internal thoracic artery grafting at ten years in its intention to treat analysis, although substantial crossover complicates interpretation [4-5]. Pooled randomised data support better patency and fewer adverse events with a radial artery than with a saphenous vein graft as the second conduit [6-7], and network comparisons of candidate second conduits point the same way [8], while observational series consistently report larger survival benefits than trials do, a divergence that has been attributed to selection of younger and fitter patients for arterial strategies [9-10]. The balance of harm is also debated, since bilateral internal thoracic artery harvesting devascularises the sternum and has been linked to deep sternal wound infection, particularly with pedicled harvesting and in patients with diabetes [11-12]. A central and largely unanswered question is therefore not whether total arterial grafting helps on average, but whether a subgroup of patients can be identified in whom it reliably helps [13].
We analysed twenty seven years of consecutive isolated coronary artery bypass operations from a single high volume centre. Our primary objective was to estimate the association between total arterial grafting and long term survival among patients whose characteristics were compatible with either conduit strategy, using propensity score overlap weighting to define that population explicitly. Our secondary objectives were to test seven prespecified clinical effect modifiers using formal interaction tests, to estimate adjusted absolute risk differences for ten early postoperative outcomes, and to examine, as an explicitly exploratory analysis, whether a cross fitted doubly robust model could predict individual treatment benefit reliably enough to support patient level conduit selection.

Methods and Patients

We studied 13,831 consecutive isolated coronary artery bypass operations performed at a single United Kingdom cardiac surgical centre between February 1996 and September 2023 and recorded in the departmental cardiac surgical registry. The unit of analysis is the operation. Operations were eligible if at least two valid distal coronary anastomoses were recorded. Operations performed through a limited access route were excluded on the operative approach field alone, independently of graft burden. Vein only operations and operations whose conduit construct could not be determined were excluded, as were operations with an impossible survival interval, operations in which any conduit was coded as other so that arterial against venous reconstruction could not be validated, and operations whose distal anastomosis count lay outside the plausible range. The resulting analytical cohort comprised 11,864 operations. The cohort algorithm and the number removed at each step are shown in Figure 1.

Grafting Strategies

The exposure was total arterial grafting, defined as an operation in which every recorded conduit was arterial, against mixed arterial and venous grafting. Classification used the validated combined registry grafting variable as the principal integrated source, cross checked against all eight graft level conduit description fields and against the recorded arterial and venous graft counts. We use the term total arterial grafting rather than complete arterial revascularisation throughout, because the registry can establish that all recorded conduits were arterial but cannot establish that every anatomically important target was revascularised. No operation in the analytical cohort was assigned to different exposure groups by different sources. There were 1,843 total arterial and 10,021 mixed operations.

Study Design and Outcomes

This was a retrospective cohort study. The primary outcome was all cause mortality, taken from the registry survival status together with the supplied updated survival interval, which incorporates later mortality tracing. Prespecified early postoperative outcomes were 30 day mortality, in hospital mortality, postoperative stroke, low cardiac output syndrome, reoperation for bleeding, any return to theatre, deep sternal wound infection, new postoperative dialysis, registry recorded multisystem failure, and postoperative hospital stay longer than 14 days. Thirty day and in hospital mortality were both retained because they capture related but distinct endpoints, and neither was substituted for the other.

Data Collection and Variable Definitions

Every field was profiled before any analytical rule was written, and completeness was mapped year by year so that genuine errors could be distinguished from fields never collected in a given period. The original data file was never modified; all derived variables were written to new columns and every eligibility decision was recorded as a reversible flag. The recorded age field was used as the primary age variable. Numeric ejection fraction values of zero were treated as missing and the harmonised registry category was used as the primary left ventricular function variable. A functioning renal transplant was retained as its own descriptive category and was included within any renal dysfunction for the binary renal interaction. Previous myocardial infarction was reconstructed across registry eras using the interval from last infarction as the principal source, supplemented by the number of previous infarctions where the principal field was blank and the secondary field was unambiguous; residual blanks were retained as an explicit missing category and were not read as absence.
Two coding conventions require explicit statement. First, the registry uses positive event coding for postoperative stroke: a recorded entry identifies a stroke and the absence of an entry indicates that no postoperative stroke occurred. Blank entries for this field are therefore not missing data, and postoperative stroke was analysed across the whole eligible cohort. Second, deep sternal wound infection and new postoperative dialysis were not recorded uniformly across the whole registry period. Outcomes that were structurally uncollected were not imputed; instead each early outcome was analysed using its own ascertainment denominator, which is reported alongside every estimate.

Statistical Analysis

The statistical analysis plan was written and the analytical dataset frozen before any model was fitted, and no analytical result or subgroup definition was altered after the estimates were seen. A fixed random seed was used throughout and the complete executable code is provided in the supplement.

Estimand 

The target estimand was the association between grafting strategy and survival among operations whose characteristics were compatible with either strategy. This is the overlap weighted estimand, which targets the population in clinical equipoise rather than an average over the whole registry [14-15].

Propensity Model and Weighting 

The propensity score for total arterial grafting was estimated by logistic regression on the covariates listed in the locked analysis plan: age and operative year, each with a four knot restricted cubic spline; body mass index; logistic EuroSCORE; sex; diabetes at three levels; left ventricular function; renal status; operative urgency; extent of vessel disease; left main disease; previous myocardial infarction with an explicit missing category; smoking; hypertension; hypercholesterolaemia; pulmonary disease; extracardiac arteriopathy; cerebrovascular disease; previous percutaneous intervention; previous cardiac surgery; preoperative rhythm; cardiogenic shock; pump strategy; and distal anastomosis count. Continuous covariates with missing values were median filled with an accompanying missing indicator, and categorical covariates retained an explicit missing category. Overlap weights were applied, equal to one minus the propensity score for total arterial operations and to the propensity score for mixed operations [15-16]. Balance was assessed by absolute standardised mean differences before and after weighting, with 0.10 taken as the threshold of concern [17], and effective sample sizes were calculated as the squared sum of the weights divided by the sum of their squares.

Outcome Models 

The treatment association was estimated by an overlap weighted Cox proportional hazards model with robust standard errors. Absolute measures were reported alongside the hazard ratio: overlap weighted adjusted survival at 5, 10 and 15 years, the absolute survival differences at those times, and restricted mean survival time to 10 and 15 years [18]. Confidence intervals came from 1,000 stratified bootstrap replicates in which the propensity score was re-estimated within every replicate. Twenty valid day zero deaths were retained and entered at half a day, because a proportional hazards model cannot use a survival time of exactly zero; a sensitivity analysis entering them at one day is reported.

Early Outcomes 

For each early outcome we report the number of operations with valid ascertainment, the number and unweighted percentage of events in each group, overlap weighted risks in each group, the adjusted absolute risk difference, its bootstrap confidence interval, the nominal P value and the Benjamini and Hochberg adjusted P value across the family of ten outcomes [19]. Absolute risk differences are the principal effect measures.

Interaction Analyses 

Seven prespecified effect modifiers were tested by likelihood ratio tests comparing overlap weighted models with and without treatment by modifier interaction terms: age modelled continuously with a restricted cubic spline, diabetes at three levels, left ventricular function at three levels, renal dysfunction as a binary variable, sex, operative urgency, and a registry based coronary complexity index restricted to 2001 onwards. Missing categories were retained in the propensity model but were not treated as clinical levels of an effect modifier; operations with a missing modifier were excluded from the corresponding interaction test. Multiplicity across the seven tests was addressed by the Benjamini and Hochberg procedure.

Exploratory personalised treatment effect model. 

An explicitly exploratory cross fitted doubly robust analysis was performed [20-21]. Operations were eligible if their 10 year status was observable, meaning either death within 10 years or known survival to 10 years. Five fold cross fitting was used, all nuisance models were fitted within training folds and every prediction was evaluated only in held out observations. Propensity scores were truncated to the range 0.05 to 0.95 rather than used to exclude operations, and the proportion lying outside that range before truncation is reported. Benefit was defined as the 10 year survival probability under total arterial grafting minus that under mixed grafting, so that positive values favour total arterial grafting. Performance was judged by outcome model discrimination, the Brier score, the treatment effect calibration intercept and slope, the correlation between predicted effects and doubly robust pseudo outcomes, and doubly robust observed estimates within quintiles of predicted benefit.

Governance and reporting. 

This project was registered with the institutional Clinical Information and Research Service as a service evaluation (CIRIS 8272). It evaluated routinely collected, anonymised clinical data and did not alter patient care. In accordance with UK Health Research Authority guidance and institutional policy, Research Ethics Committee review and individual informed consent were not required. Data were processed within institutional information-governance controls. The study is reported in accordance with the STROBE and RECORD statements [22-23], and a completed checklist is supplied.

Results

Cohort Construction and Grafting Strategy

Of 13,831 consecutive isolated coronary artery bypass operations, 11,864 met the prespecified criteria and formed the analytical cohort, of which 1,843 (15.5 per cent) were total arterial and 10,021 (84.5 per cent) were mixed arterial and venous. Figure 1 shows the number removed at each step. Median follow up was 12.5 years and the longest follow up was 27.6 years, during which 1,424 deaths occurred.

Figure1: Cohort flow. Operations removed at each step of the prespecified cohort algorithm, from the full registry extract to the final analytical cohort. CABG denotes coronary artery bypass grafting.

Characteristic Total arterial (n=1,843) Mixed 
(n=10,021)
SMD Total 
arterial, weighted
Mixed, 
weighted
SMD
Age at operation, years, mean (SD) 61.6 (10.1) 65.4 (9.7) 0.385 62.8 (10.1) 62.8 (10.1) 0.000
Sex, n (%)
    Male 1,572 (85.3) 8,241 (82.2) 0.083 84.4 84.4 0.000
    Female 271 (14.7) 1,775 (17.7) 0.082 15.6 15.6 0.000
    Not recorded 0 (0.0) 5 (0.0) 0.032 0.0 0.0 0.000
Body mass index, kg/m2, mean (SD) 28.1 (4.4) 28.0 (4.7) 0.006 28.0 (4.4) 28.0 (4.8) 0.000
Logistic EuroSCORE, mean (SD) 3.1 (4.3) 4.3 (6.0) 0.219 3.4 (4.6) 3.4 (4.6) 0.000
Diabetes, n (%)
    No diabetes 1,401 (76.0) 6,765 (67.5) 0.190 73.7 73.7 0.000
    Non-insulin-treated diabetes 315 (17.1) 2,290 (22.9) 0.144 18.6 18.6 0.000
    Insulin-treated diabetes 126 (6.8) 942 (9.4) 0.094 7.6 7.6 0.000
    Not recorded 1 (0.1) 24 (0.2) 0.048 0.1 0.1 0.000
Left ventricular function, n (%)
    Preserved 1,455 (78.9) 7,517 (75.0) 0.094 77.9 77.9 0.000
    Moderate 325 (17.6) 1,927 (19.2) 0.041 18.3 18.3 0.000
    Severe 41 (2.2) 434 (4.3) 0.119 2.7 2.7 0.000
    Not recorded 22 (1.2) 143 (1.4) 0.021 1.1 1.1 0.000
Renal status, n (%)
    No recorded renal dysfunction 1,569 (85.1) 7,670 (76.5) 0.220 82.5 82.4 0.000
    Any renal dysfunction 267 (14.5) 2,253 (22.5) 0.207 17.1 17.1 0.000
    Not recorded 7 (0.4) 98 (1.0) 0.073 0.4 0.4 0.000
Operative urgency, n (%)
    Elective 1,365 (74.1) 6,513 (65.0) 0.198 72.2 72.2 0.000
    Urgent 433 (23.5) 3,179 (31.7) 0.185 25.3 25.3 0.000
    Emergency/Salvage 39 (2.1) 312 (3.1) 0.063 2.4 2.4 0.000
    Not recorded 6 (0.3) 17 (0.2) 0.031 0.2 0.2 0.000
Extent of vessel disease, n (%)
    3VD 1,043 (56.6) 7,778 (77.6) 0.459 61.0 61.0 0.000
    2VD 714 (38.7) 1,902 (19.0) 0.447 34.6 34.7 0.000
    1VD 51 (2.8) 137 (1.4) 0.099 2.5 2.5 0.000
    Not recorded 22 (1.2) 156 (1.6) 0.031 1.2 1.2 0.000
    0VD 13 (0.7) 48 (0.5) 0.030 0.7 0.7 0.000
Left main disease, n (%)
    No 1,015 (55.1) 5,834 (58.2) 0.063 57.1 57.1 0.000
    Yes 473 (25.7) 2,387 (23.8) 0.043 24.6 24.6 0.000
    Not recorded 355 (19.3) 1,800 (18.0) 0.033 18.3 18.3 0.000
Previous myocardial infarction, n (%)
    Yes 779 (42.3) 4,881 (48.7) 0.130 44.1 44.1 0.000
    No 668 (36.2) 3,896 (38.9) 0.054 38.1 38.2 0.000
    Not recorded 396 (21.5) 1,244 (12.4) 0.244 17.7 17.7 0.001
Current smoker, n (%)
    No 1,626 (88.2) 8,895 (88.8) 0.017 88.6 88.6 0.000
    Yes 208 (11.3) 1,052 (10.5) 0.025 10.9 10.9 0.000
    Not recorded 9 (0.5) 74 (0.7) 0.032 0.5 0.5 0.000
Hypertension, n (%)
    Yes 1,278 (69.3) 7,497 (74.8) 0.122 71.1 71.1 0.000
    No 556 (30.2) 2,463 (24.6) 0.126 28.4 28.4 0.000
    Not recorded 9 (0.5) 61 (0.6) 0.016 0.5 0.5 0.000
Hypercholesterolaemia, n (%)
    Yes 1,429 (77.5) 7,915 (79.0) 0.035 78.3 78.3 0.000
    No 385 (20.9) 1,864 (18.6) 0.058 20.1 20.1 0.000
    Not recorded 29 (1.6) 242 (2.4) 0.060 1.6 1.6 0.000
Pulmonary disease, n (%)
    No 1,065 (57.8) 5,013 (50.0) 0.156 54.2 54.2 0.000
    Yes 775 (42.1) 4,977 (49.7) 0.153 45.6 45.6 0.000
    Not recorded 3 (0.2) 31 (0.3) 0.030 0.2 0.2 0.000
Extracardiac arteriopathy, n (%)            
    No 1,689 (91.6) 8,979 (89.6) 0.070 90.9 90.9 0.000
    Yes 150 (8.1) 984 (9.8) 0.059 8.8 8.8 0.000
    Not recorded 4 (0.2) 58 (0.6) 0.057 0.3 0.3 0.000
Cerebrovascular disease, n (%)
    No 1,758 (95.4) 9,288 (92.7) 0.114 94.8 94.9 0.000
    Yes 73 (4.0) 649 (6.5) 0.113 4.5 4.5 0.000
    Not recorded 12 (0.7) 84 (0.8) 0.022 0.7 0.7 0.000
Previous percutaneous intervention, n (%)
    No 1,580 (85.7) 8,527 (85.1) 0.018 84.8 84.8 0.000
    Yes 251 (13.6) 1,391 (13.9) 0.008 14.5 14.5 0.000
    Not recorded 12 (0.7) 103 (1.0) 0.041 0.7 0.7 0.000
Previous cardiac surgery, n (%)
    No 1,779 (96.5) 9,666 (96.5) 0.004 96.5 96.5 0.000
    Yes 35 (1.9) 175 (1.7) 0.011 2.0 2.0 0.000
    Not recorded 29 (1.6) 180 (1.8) 0.017 1.5 1.5 0.000
Preoperative rhythm, n (%)
    Sinus 1,133 (61.5) 7,263 (72.5) 0.236 66.1 66.1 0.000
    Not recorded 661 (35.9) 2,399 (23.9) 0.263 30.7 30.7 0.000
    Non-sinus 49 (2.7) 359 (3.6) 0.053 3.2 3.2 0.000
Cardiogenic shock, n (%)
    No 1,836 (99.6) 9,921 (99.0) 0.075 99.6 99.6 0.000
    Yes 7 (0.4) 66 (0.7) 0.039 0.4 0.4 0.000
    Not recorded 0 (0.0) 34 (0.3) 0.083 0.0 0.0 0.001
Year of operation, mean (SD) 2008.2 (7.8) 2009.6 (7.9) 0.174 2008.9 (7.8) 2008.9 (7.9) 0.000
Pump strategy, n (%)
    OPCAB 1,174 (63.7) 5,077 (50.7) 0.266 61.4 61.4 0.001
    ONCAB 665 (36.1) 4,909 (49.0) 0.263 38.3 38.3 0.001
    Not recorded 4 (0.2) 35 (0.3) 0.025 0.3 0.3 0.000
Distal anastomoses, mean (SD) 2.38 (0.54) 2.92 (0.68) 0.887 2.47 (0.58) 2.47 (0.55) 0.001

Values before weighting are n (%) for categorical variables and mean (SD) for continuous variables. Weighted columns are overlap weighted percentages or means. SMD, absolute standardised mean difference; a value below 0.10 is conventionally taken as acceptable balance. The maximum absolute SMD after weighting was 0.001. Effective sample sizes after weighting were 1,686 for total arterial and 4,684 for mixed. Not recorded denotes an explicit missing category retained in the propensity model. SD, standard deviation; EuroSCORE, European System for Cardiac Operative Risk Evaluation.

Table1:  Baseline and operative characteristics before and after overlap weighting

Baseline Characteristics and Treatment Overlap

Baseline and operative characteristics before and after overlap weighting are shown in Table 1. Before weighting, 20 of 55 characteristics differed by a standardised mean difference of 0.10 or more, with total arterial operations performed in younger patients, at lower recorded operative risk and with fewer distal anastomoses. After overlap weighting no characteristic exceeded that threshold and the maximum absolute standardised mean difference was 0.001. The effective sample size fell from 10,021 to 4,684 in the mixed group and from 1,843 to 1,686 in the total arterial group, which is the expected cost of weighting towards the region of equipoise. Propensity score distributions and the Love plot are shown in the supplement.

Intraoperative Data

Pump strategy and distal anastomosis burden differed substantially between the groups before weighting and both were included in the propensity model, as prespecified. Their distributions before and after weighting are given in Table 1. Because each may lie on the causal pathway between conduit strategy and outcome, a prespecified baseline only propensity model omitting both was also fitted and is reported in the supplement; it gave a hazard ratio of 0.876 (0.736 to 1.015), close to the primary estimate.

Long Term Survival

Total arterial grafting was associated with an overlap weighted hazard ratio for death of 0.885 (95 per cent bootstrap confidence interval 0.743 to 1.041). Overlap weighted adjusted survival curves are shown in Figure 2. Absolute adjusted survival differences were +0.7 percentage points at 5 years, +0.1 at 10 years and +1.0 at 15 years. The difference in restricted mean survival time was +0.04 years to 10 years and +0.07 years to 15 years, corresponding to a few weeks. Full estimates with confidence intervals are given in Table 2.

Measure Estimate 95% bootstrap 
confidence interval
Hazard ratio for death, overlap weighted Cox 0.885 0.743 to 1.041
Adjusted survival at 5 years, total arterial 95.9%  
Adjusted survival at 5 years, mixed 95.2%  
Absolute difference at 5 years, percentage points +0.74 -0.37 to +1.76
Adjusted survival at 10 years, total arterial 92.6%  
Adjusted survival at 10 years, mixed 92.5%  
Absolute difference at 10 years, percentage points +0.10 -1.39 to +1.54
Adjusted survival at 15 years, total arterial 90.4%  
Adjusted survival at 15 years, mixed 89.4%  
Absolute difference at 15 years, percentage points +1.00 -0.74 to +2.87
RMST to 10 years, total arterial, years 9.568  
RMST to 10 years, mixed, years 9.523  
RMST difference to 10 years, years +0.044 -0.050 to +0.142
RMST to 15 years, total arterial, years 14.135  
RMST to 15 years, mixed, years 14.070  
RMST difference to 15 years, years +0.065 -0.099 to +0.231

All estimates are overlap weighted and are based on 11,864 operations with 1,424 deaths. Confidence intervals come from 1,000 stratified bootstrap replicates with the propensity score re-estimated within each replicate. A positive difference favours total arterial grafting. RMST, restricted mean survival time. Estimates describe associations and are not causal effects.

Table2:  Primary long term survival estimates

Figure2: Overlap-weighted adjusted survival. Adjusted survival curves for total arterial grafting and mixed arterial–venous grafting are accompanied by the overlap-weighted Cox hazard ratio, 95% bootstrap confidence interval and two-sided P value. The unweighted numbers at risk are shown at 0, 5, 10, 15 and 20 years. The lower panel presents adjusted absolute survival differences with 95% bootstrap confidence intervals at the three prespecified time points. The estimates describe adjusted associations and not causal effects.

Treatment Effect Modification

None of the seven prespecified interactions reached conventional significance, either nominally or after Benjamini and Hochberg adjustment (Table 3). The smallest nominal P value was 0.182. The multipanel heterogeneity figure (Figure 3) shows the prespecified subgroup estimates alongside the continuous age effect, neither of which indicates reliable effect modification. A sensitivity analysis restricting the coronary complexity interaction to 2006 onwards, when ascertainment is approximately complete, did not change this conclusion.

Figure3: Treatment-effect heterogeneity. Prespecified subgroup estimates are shown as hazard ratios with 95% confidence intervals on a logarithmic scale. The dashed line marks unity and the dotted line marks the overall estimate. Interaction P values are from the prespecified formal interaction tests; subgroup-specific confidence intervals are descriptive and do not themselves establish effect modification.

Effect modifier Operations Deaths df Chi-square Nominal P BH-adjusted P
Age, continuous with restricted cubic spline 11,864 1,424 3 4.25 0.236 0.724
Diabetes, three level 11,839 1,421 2 3.41 0.182 0.724
Left ventricular function, three level 11,699 1,411 2 0.95 0.621 0.724
Renal dysfunction, binary 11,759 1,411 1 0.10 0.749 0.749
Sex 11,859 1,422 1 0.27 0.607 0.724
Operative urgency, three level 11,841 1,423 2 1.29 0.524 0.724
Coronary complexity, 2001 onwards, complete classification 9,468 1,290 2 1.81 0.405 0.724

Likelihood ratio tests for treatment by modifier interaction in the overlap weighted Cox model. Operations with a missing value of the modifier were excluded from the corresponding test, so degrees of freedom reflect clinical levels only; missing categories were retained in the propensity model. Multiplicity across the seven tests was addressed by the Benjamini and Hochberg procedure. Prespecified sensitivity analyses: coronary complexity restricted to 2006 onwards, 7,427 operations, 2 df, nominal P 0.888; diabetes collapsed to present or absent, 11,839 operations, 1 df, nominal P 0.078. df, degrees of freedom; BH, Benjamini and Hochberg.

Table3: Prespecified treatment effect interaction analyses

Early Postoperative Outcomes

Adjusted absolute risk differences for the ten early outcomes are shown in Table 4, each with its own ascertainment denominator. Deep sternal wound infection was recorded in 23 of 1,171 total arterial operations and 100 of 7,563 mixed operations with valid ascertainment, giving overlap weighted risks of 2.10 and 1.25 per cent and an adjusted absolute risk difference of +0.86 percentage points (+0.03 to +1.81), with a nominal P value of 0.042 and a Benjamini and Hochberg adjusted P value of 0.420. Postoperative stroke, analysed across the full eligible cohort under the registry positive event coding convention, showed an adjusted absolute risk difference of -0.26 percentage points (-0.76 to +0.25). No early outcome remained statistically supported after Benjamini and Hochberg adjustment across the family of ten endpoints; the deep sternal wound infection difference is therefore a nominal association only and requires cautious interpretation.

Early 
outcome
Valid n Events/n total arterial (%) Events/n 
mixed (%)
Weighted risk, 
TA (%)
Weighted risk, 
mixed (%)
ARD 
(pp)
95% 
bootstrap CI
Nominal P BH P
30-day mortality 11,836 19/1,843 (1.03) 215/9,993 (2.15) 1.15 1.46 -0.30 -0.87 to +0.33 0.348 0.580
In-hospital mortality 11,864 14/1,843 (0.76) 151/10,021 (1.51) 0.90 0.95 -0.05 -0.57 to +0.49 0.844 0.844
Postoperative stroke 11,864 14/1,843 (0.76) 128/10,021 (1.28) 0.81 1.07 -0.26 -0.76 to +0.25 0.302 0.580
Low cardiac output syndrome 11,761 302/1,836 (16.45) 2,274/9,925 (22.91) 17.80 19.27 -1.47 -3.75 to +0.58 0.184 0.580
Reoperation for bleeding 11,527 46/1,785 (2.58) 368/9,742 (3.78) 2.53 2.89 -0.35 -1.13 to +0.42 0.414 0.591
Any return to theatre 11,713 90/1,829 (4.92) 513/9,884 (5.19) 4.75 4.61 +0.15 -0.97 to +1.26 0.754 0.844
Deep sternal wound infection 8,734 23/1,171 (1.96) 100/7,563 (1.32) 2.10 1.25 +0.86 +0.03 to +1.81 0.042 0.420
New postoperative dialysis 8,989 90/1,214 (7.41) 712/7,775 (9.16) 7.16 8.09 -0.94 -2.58 to +0.72 0.244 0.580
Multisystem failure 11,689 15/1,815 (0.83) 174/9,874 (1.76) 0.91 1.36 -0.46 -0.98 to +0.12 0.124 0.580
Postoperative stay over14 days 11,842 144/1,841 (7.82) 1,023/10,001 (10.23) 8.34 8.12 +0.22 -1.26 to +1.77 0.774 0.844

Each outcome is analysed using its own ascertainment denominator, shown as Valid n. Deep sternal wound infection and new postoperative dialysis were not recorded uniformly across the registry period and structurally uncollected outcomes were not imputed. The registry uses positive event coding for postoperative stroke, so a blank entry denotes absence of stroke rather than missing data. Thirty day mortality excludes 28 operations censored before 30 days. Any return to theatre excludes records whose multi-select entry was internally contradictory. Absolute risk differences are the principal effect measure; a positive value indicates a higher risk after total arterial grafting. P values are two sided and derived from the bootstrap distribution; BH-adjusted P values apply the Benjamini and Hochberg procedure across the family of ten outcomes, of which 0 remained significant after adjustment. All estimates describe adjusted associations and not causal effects. TA, total arterial; ARD, absolute risk difference; pp, percentage points; CI, confidence interval; BH, Benjamini and Hochberg.

Table4: Adjusted early postoperative outcomes

Sensitivity Analyses

The complete sensitivity table is in the supplement. The estimate was materially unchanged when day zero deaths were entered at one day rather than half a day (0.886), when the analysis was restricted to operations with complete source corroboration of exposure (0.895), and across the cohort definition variants. Administrative censoring at the census date moved the estimate towards the null (0.931), as expected given that it discards later mortality tracing. The prespecified baseline only propensity model, which omits pump strategy and distal anastomosis burden because each may lie on the causal pathway, gave 0.876 (0.736 to 1.015), so the estimate does not depend on conditioning on those two intraoperative variables. The most consequential specification choice was the propensity model itself: omitting logistic EuroSCORE and preoperative rhythm, both of which were listed in the locked analysis plan, moved the hazard ratio to 0.786 (0.668 to 0.907). The prespecified model is reported as primary and this difference is discussed below. Differences between the on pump, off pump, two distal and three distal estimates are not interpreted as effect modification, since formal interaction testing does not support a reliable subgroup effect.

Exploratory Personalised Treatment Effects

The complete horizon cohort comprised 7,974 operations, 1,264 total arterial, with 993 deaths within 10 years; the eligibility rule and a flow table are given in the supplement. 2,602 operations (32.6 per cent) had a propensity score outside 0.05 to 0.95 before truncation; scores were truncated to that range and no operation was excluded on that basis. The outcome model achieved a discrimination of 0.836 with a Brier score of 0.0785. Predicted benefit had a median of 0.83 percentage points and a 5th to 95th percentile range of -11.03 to 10.06. Across out of fold quintiles of predicted benefit the observed doubly robust estimates ran from -5.22 to +3.90 percentage points, so the broad ordering was preserved, but the two lowest quintiles were not separated from one another and neither were the three highest. The treatment effect calibration slope was 0.706 against a target of one, with an intercept of 0.0005, indicating that the model over-states the spread of benefit, and the correlation between predicted effects and doubly robust pseudo outcomes was 0.045. Individual predictions therefore carry very little information even where group averages align, and no patient level calculator was produced.

Discussion

In 11,864 multigraft isolated coronary artery bypass operations performed over twenty seven years at a single centre, total arterial grafting was associated with a hazard ratio for death of 0.885 within the population of patients whose characteristics were compatible with either conduit strategy. The absolute adjusted survival differences at 5, 10 and 15 years were all of the order of one percentage point or less, and the restricted mean survival differences amounted to a few weeks. None of the seven prespecified clinical interactions showed convincing treatment effect heterogeneity, and an exploratory cross fitted doubly robust model did not predict individual benefit with sufficient calibration to support clinical use.These findings sit between the randomised and the observational literature. The Arterial Revascularization Trial reported no significant mortality difference between bilateral and single internal thoracic artery grafting at ten years [4], whereas large observational series and their meta-analyses have repeatedly reported substantial survival advantages for multiple arterial grafting [9]. That divergence has been attributed directly to selection, and a meta-analytic comparison of randomised against observational evidence has shown it to be systematic rather than incidental [10]. Adding a radial artery to one or two internal thoracic grafts has also been examined within the randomised setting [24]. Our estimate is closer to the randomised result than to the conventional observational one, which is what would be expected when the estimand is deliberately restricted to the region of equipoise rather than extrapolated across the whole case mix. The pooled randomised evidence favouring the radial artery over the saphenous vein as a second conduit [6,7] concerns graft patency and composite events rather than all-cause mortality in an unselected surgical population, and is therefore not in conflict with what we observe.
Several explanations deserve consideration. Treatment selection is the most important: total arterial grafting was offered to younger patients with lower recorded operative risk, and although overlap weighting removed measured imbalance completely, it cannot address unmeasured selection such as conduit quality, target vessel calibre or surgeon preference. Conduit durability may require longer follow up to express itself in mortality, particularly where competing non cardiac mortality is substantial in an ageing cohort. Completeness and complexity of revascularisation are only partially captured by a registry, and the number of distal anastomoses is an imperfect proxy. Total arterial grafting is also a bundled strategy rather than a single intervention, combining conduit choice with harvesting technique, target selection and often an off pump approach, so an estimate of its association cannot be attributed to the arterial conduit alone. Finally, the sensitivity of the estimate to the propensity specification deserves emphasis: omitting logistic EuroSCORE and preoperative rhythm moved the hazard ratio from 0.885 to 0.786 and changed whether the interval excluded one. We report the prespecified model as primary because both variables were listed in the locked analysis plan written before any model was fitted, but the dependence of the conclusion on that single choice is itself a material finding and is reported in full rather than resolved in favour of either answer.
This study has several strengths: a large consecutive single centre series with long follow up, a cohort and exposure definition reconciled at record level before any outcome model was fitted, a prespecified analysis plan with a frozen dataset, and an estimand chosen to match the clinical question. Its limitations are substantial and should temper interpretation. Residual confounding by indication cannot be excluded in any observational conduit comparison. The absence of a linkage key means the unit of analysis is the operation and not the patient. Structural missingness is era related: left main status, deep sternal wound infection and new dialysis were not collected uniformly across the registry period, and we have used outcome specific denominators rather than imputing fields that were never collected. A further 160 otherwise eligible operations carried a confirmed death with no usable time to death; these are interval censored observations whose event status is known but whose timing is not, and they could not enter conventional time to event models. They were excluded from the survival analysis and this is reported as a limitation rather than addressed by assigning an arbitrary event time. That exclusion was not differential by grafting strategy: 19 of 1,862 total arterial and 141 of 10,162 mixed operations were lost on this ground, 1.02 against 1.39 per cent, P 0.227, so the loss is unlikely to have moved the estimate in either direction. Positivity is limited at the extremes of graft burden, in dialysis dependence and at extreme age, and these regions are described as not reliably estimable rather than as evidence against the strategy. Outcome ascertainment for early endpoints depends on registry coding conventions, which we have stated explicitly.
For practice, these data do not support a general expectation of a long term survival advantage from extending an arterial strategy to every target in patients of the kind who currently receive either approach at this centre, nor do they identify a subgroup in whom such an advantage is reproducible using the variables a registry records. The adjusted association between total arterial grafting and deep sternal wound infection is consistent with the known effect of internal thoracic artery harvesting on sternal perfusion [12], and with the higher risk reported in patients with diabetes and obesity [11]; because bilateral harvesting is an integral component of many total arterial procedures, it offers a plausible explanation for the association rather than acting as an independent confounder, and ascertainment for this outcome is incomplete across registry eras. Generalisability beyond a single high volume centre with a particular case mix and surgical culture is uncertain. The questions these data cannot answer are precisely those that the ongoing randomised comparison of single against multiple arterial grafting is designed to address [25], and prospective work linking conduit strategy to patient level outcomes with complete death timing and independent ascertainment of postoperative complications remains the priority.

Conclusion

In this 27-year single-centre registry, total arterial grafting was not shown to confer a clear population-average long-term survival advantage within the overlap population, and no reproducible subgroup with differential benefit was identified. The findings support individualised conduit selection integrating patient characteristics, coronary anatomy, conduit suitability, anticipated longevity and surgical expertise, while the exploratory model was insufficiently calibrated to provide an algorithmic treatment rule.

Funding

No external funding was received for this service evaluation.

Conflict of Interest

The authors declare no conflicts of interest.

Data Availability

The underlying operation-level data are not publicly available because they are governed by institutional information-governance requirements. Access requests are subject to approval by the data custodian and the institution.

Online Supplement

Total Arterial Grafting and Long Term Survival After Isolated Coronary Artery Bypass Surgery: An Overlap Weighted Analysis of a Twenty Seven Year Single Centre Registry
This supplement contains the full balance diagnostics, the complete sensitivity analyses, the missingness summary, the exposure reconciliation, the personalised model performance and calibration results, and the supplementary figures.

Contents

Supplementary Table S1.  Full covariate balance before and after overlap weighting
Supplementary Table S2.  Missingness in the analytical cohort
Supplementary Table S3.  Exposure reconciliation and source corroboration
Supplementary Table S4.  Complete sensitivity analyses
Supplementary Table S5.  Prespecified subgroup estimates
Supplementary Table S6.  Complete horizon cohort for the personalised analysis
Supplementary Table S7.  Personalised model performance and calibration
Supplementary Table S8.  Observed doubly robust estimates by predicted benefit quintile
Supplementary Figure S1.  Propensity score distributions
Supplementary Figure S2.  Covariate balance
Supplementary Figure S3.  Sensitivity analyses
Supplementary Figure S4.  Personalised model calibration
Supplementary Figure S5.  Use of total arterial grafting over time and by graft burden

Characteristic as entered in the propensity model SMD before 
weighting
SMD after overlap 
weighting
Distal anastomoses 0.887 0.001
Vessel disease: 2VD 0.447 0.000
Age, years 0.385 0.000
Logistic EuroSCORE: not recorded 0.339 0.000
Age, years spline 1 0.337 0.000
Age, years spline 2 0.269 0.000
Pump strategy: ONCAB 0.263 0.001
Preoperative rhythm: Not recorded 0.263 0.000
Previous MI: Not recorded 0.244 0.001
Logistic EuroSCORE 0.212 0.000
Renal status: Any renal dysfunction 0.207 0.000
Urgency: Urgent 0.185 0.000
Operative year 0.174 0.000
Pulmonary disease: Yes 0.153 0.000
Body mass index: not recorded 0.146 0.000
Diabetes: Non-insulin-treated diabetes 0.144 0.000
Operative year spline 1 0.133 0.000
Hypertension: No 0.126 0.000
LV function: Severe 0.118 0.000
Cerebrovascular disease: Yes 0.113 0.000
Vessel disease: 1VD 0.099 0.000
Diabetes: Insulin-treated diabetes 0.094 0.000
Operative year spline 2 0.093 0.000
Cardiogenic shock: Not recorded 0.083 0.001
Sex: Female 0.082 0.000
Renal status: Not recorded 0.073 0.000
Urgency: Emergency/Salvage 0.063 0.000
Hypercholesterolaemia: Not recorded 0.060 0.000
Extracardiac arteriopathy: Yes 0.059 0.000
Hypercholesterolaemia: No 0.058 0.000
Extracardiac arteriopathy: Not recorded 0.057 0.000
Previous MI: No 0.054 0.000
Preoperative rhythm: Non-sinus 0.053 0.000
Diabetes: Not recorded 0.048 0.000
Left main disease: Yes 0.043 0.000
LV function: Moderate 0.041 0.000
Previous PCI: Not recorded 0.041 0.000
Cardiogenic shock: Yes 0.039 0.000
Left main disease: Not recorded 0.033 0.000
Sex: Not recorded 0.032 0.000
Smoking: Not recorded 0.032 0.000
Urgency: Not recorded 0.031 0.000
Vessel disease: Not recorded 0.031 0.000
Vessel disease: 0VD 0.030 0.000
Pulmonary disease: Not recorded 0.030 0.000
Smoking: Current 0.025 0.000
Pump strategy: Not recorded 0.025 0.000
Cerebrovascular disease: Not recorded 0.022 0.000
LV function: Not recorded 0.021 0.000
Previous cardiac surgery: Not recorded 0.017 0.000
Hypertension: Not recorded 0.016 0.000
Previous cardiac surgery: Yes 0.011 0.000
Previous PCI: Yes 0.008 0.000
Smoking: Never 0.006 0.000
Body mass index 0.003 0.000

All 55 terms entering the propensity model are shown, including spline basis terms and explicit missing indicators. 20 terms had an absolute standardised mean difference of 0.10 or more before weighting and 0 afterwards. The maximum absolute standardised mean difference after weighting was 0.001. Effective sample sizes after weighting were 1,686.1 of 1,843 for total arterial and 4,683.5 of 10,021 for mixed. Propensity scores ranged from 0.000 to 0.957; 3,848 operations (32.43 per cent) lay outside 0.05 to 0.95.

Supplementary Table S1:  Full covariate balance

Variable Column Recorded Missing Missing (%) Handling
Age at operation d_age_years 11,864 0 0.0 complete
Sex d_sex 11,859 5 0.0 explicit missing category
Body mass index d_bmi 11,674 190 1.6 explicit missing category
Logistic EuroSCORE d_logistic_euroscore 7,475 4,389 37.0 explicit missing category
Diabetes d_diabetes3 11,839 25 0.2 explicit missing category
Left ventricular function d_lv_function 11,699 165 1.4 explicit missing category
Renal status d_renal5 11,759 105 0.9 explicit missing category
Operative urgency d_urgency3 11,841 23 0.2 explicit missing category
Extent of vessel disease d_vessel_disease 11,686 178 1.5 explicit missing category
Left main disease d_lms 9,709 2,155 18.2 explicit missing category
Previous myocardial infarction d_previous_mi 10,224 1,640 13.8 explicit missing category
Smoking d_smoking3 11,781 83 0.7 explicit missing category
Hypertension d_hypertension 11,794 70 0.6 explicit missing category
Hypercholesterolaemia d_hypercholesterolaemia 11,593 271 2.3 explicit missing category
Pulmonary disease d_pulmonary_disease 11,830 34 0.3 explicit missing category
Extracardiac arteriopathy d_arteriopathy 11,802 62 0.5 explicit missing category
Cerebrovascular disease d_cerebrovascular 11,768 96 0.8 explicit missing category
Previous percutaneous intervention d_previous_pci 11,749 115 1.0 explicit missing category
Previous cardiac surgery d_redo_cardiac 11,655 209 1.8 explicit missing category
Preoperative rhythm d_preop_rhythm 8,804 3,060 25.8 explicit missing category
Cardiogenic shock d_shock 11,830 34 0.3 explicit missing category
Pump strategy d_pump 11,825 39 0.3 explicit missing category
Distal anastomoses d_total_distal 11,864 0 0.0 complete
Coronary complexity index d_complexity 9,653 2,211 18.6 explicit missing category

Missing values of continuous covariates were median filled with an accompanying missing indicator; categorical covariates retained an explicit missing category. Missing categories were used in the propensity model only and were never treated as clinical levels of an effect modifier. Structurally uncollected outcomes were not imputed.

Supplementary Table S2:  Missingness in the analytical cohort

Check Result
Operations in the analytical cohort 11,864
Assigned to different exposure groups by different sources 0
Combined registry classification agrees with graft level reconstruction 11,864 of 11,864
Every validation field present and corroborating exactly 11,814 of 11,864
At least one validation field not corroborating exactly 50
of which differ by one graft only 38
of which the registry arterial count field reads zero despite arterial conduits recorded at graft level 3
Classified total arterial by either source among these 0, all are classified mixed by both sources

Material exposure concordance, meaning that no available source assigns an operation to a different exposure group, is complete across the analytical cohort. Complete source corroboration, meaning that every validation field is present and agrees exactly, holds for 11,814 operations. The 50 operations that fall short differ only in graft count fields, most often by a single graft, and all are classified as mixed by both the combined registry classification and the graft level reconstruction. The corresponding sensitivity analysis is therefore described as a complete source corroboration analysis and should not be read as implying materially discordant exposure classification.

Supplementary Table S3:  Exposure reconciliation and source corroboration

Analysis Operations Deaths Hazard ratio 95% CI
Primary analysis, overlap weighted 11,864 1,424 0.885 0.743 to 1.041
Reduced propensity model omitting logistic EuroSCORE and preoperative rhythm 11,864 1,424 0.786 0.668 to 0.907
Baseline-only propensity model omitting pump strategy and distal anastomosis burden 11,864 1,424 0.876 0.736 to 1.015
Day-zero deaths entered at one day instead of half a day 11,864 1,424 0.886 0.746 to 1.029
Administrative censoring at 1 October 2023 11,864 1,139 0.931 0.774 to 1.121
Complete source-corroboration cohort 11,814 1,409 0.895 0.752 to 1.039
Excluding the operation whose recorded age was corrected 11,863 1,424 0.885 0.746 to 1.029
Excluding operations with discordant ejection fraction 11,555 1,382 0.885 0.740 to 1.029
Excluding left main disease with fewer than two vessel disease 11,764 1,411 0.873 0.732 to 1.023
First-time cardiac surgery only 11,445 1,324 0.870 0.725 to 1.028
Excluding operations converted to full sternotomy 11,859 1,424 0.885 0.744 to 1.029
Off-pump operations only 6,251 815 0.802 0.624 to 0.983
On-pump operations only 5,574 606 1.076 0.792 to 1.393
Two distal anastomoses only 3,783 483 0.945 0.774 to 1.130
Three distal anastomoses only 6,358 760 0.790 0.561 to 1.050

Each row is produced by the same code as the primary analysis, with the propensity score re-estimated within the altered cohort or model. The interval on the primary row is the one reported throughout, from 1,000 stratified bootstrap replicates; the remaining rows use 300 replicates, which is enough to show whether an alteration moves the estimate materially. The primary estimate is 0.885. The reduced propensity model omitting logistic EuroSCORE and preoperative rhythm gives 0.786; both variables were named in the locked statistical analysis plan written before any model was fitted, so the model retaining them is reported as primary and this row quantifies the dependence of the estimate on that specification. The baseline-only model omits pump strategy and distal anastomosis burden, the two intraoperative variables that may lie on the causal pathway between conduit strategy and outcome, and gives 0.876, so conditioning on them does not drive the primary estimate. Differences between the on pump, off pump, two distal and three distal rows are not interpreted as effect modification; formal interaction testing does not support a reliable subgroup effect.

Supplementary Table S4:  Complete sensitivity analyses

Subgroup Operations Deaths Hazard ratio 95% CI
Age under 65 years 5,510 383 0.783 0.567 to 1.082
Age 65 years or over 6,354 1,041 0.954 0.762 to 1.196
No diabetes 8,166 876 0.754 0.595 to 0.955
Non-insulin-treated diabetes 2,605 340 1.328 0.923 to 1.911
Insulin-treated diabetes 1,068 205 1.064 0.631 to 1.794
LV function preserved 8,972 967 0.923 0.743 to 1.147
LV function moderate 2,252 332 0.687 0.452 to 1.045
LV function severe 475 112 1.066 0.514 to 2.210
No recorded renal dysfunction 9,239 1,084 0.877 0.712 to 1.080
Any renal dysfunction 2,520 327 1.021 0.681 to 1.530
Male 9,813 1,132 0.857 0.696 to 1.055
Female 2,046 290 1.014 0.673 to 1.527
Elective 7,878 884 0.786 0.622 to 0.993
Urgent 3,612 469 1.068 0.770 to 1.480
Emergency or salvage 351 70 1.201 0.522 to 2.766
Complexity low 1,684 197 0.901 0.605 to 1.342
Complexity intermediate 5,108 683 0.770 0.564 to 1.052
Complexity high 2,676 410 1.156 0.847 to 1.576

Subgroup estimates are shown for completeness and are displayed in Figure 3A. They are not evidence of effect modification: formal interaction tests are reported in Table 3, and none is significant nominally or after multiplicity adjustment. Intervals use robust standard errors from the overlap weighted model within each subgroup. Each set of subgroups covers the operations with that characteristic recorded, so the counts within a characteristic sum to the number reported for it in Table 3 rather than to the whole cohort. The coronary complexity subgroups are restricted to 2001 onwards with a complete classification, as the corresponding interaction test is.

Supplementary Table S5:  Prespecified subgroup estimates

Step Operations Total 
arterial
Deaths 
within 10 years
Frozen primary cohort 11,864 1,843 1,424
Died within 10 years of operation 993 101 993
Known alive at 10 years 6,981 1,163 0
Censored before 10 years with no event, excluded 3,890 579 0
Complete 10-year horizon cohort 7,974 1,264 993
Reconciliation: operations performed at least 10 years before the census date 7,707 1,237 742

An operation contributes an observable 10 year outcome if death occurred within 10 years or if survival to 10 years is known. Operations that died inside 10 years are included even where total recorded follow up is shorter. The resulting cohort is 7,974 operations, 1,264 total arterial, with 993 deaths by 10 years. The final row is a reconciliation against an alternative rule, restricting to operations performed at least 10 years before the census date so that 10 years of potential follow up exists by calendar time: that rule gives 7,707 operations, 1,237 total arterial and 742 deaths within 10 years. The difference between the two rules is the 267 operations performed more recently in which death occurred inside 10 years, which the stated definition includes and the calendar rule does not. The analysis uses the stated definition.

Supplementary Table S6:  Complete horizon cohort for the personalised analysis

Metric Value
Outcome and horizon Death within 10 years, analysed on the survival scale
Estimator Cross fitted doubly robust, five folds
Nuisance models Fitted within training folds; predictions evaluated only in held out folds
Operations analysed 7,974
Propensity scores outside 0.05 to 0.95 before truncation 2,602 (32.63%)
Handling of extreme scores propensity scores truncated to 0.05 to 0.95; no operation excluded
Outcome model discrimination (AUC) 0.836
Brier score 0.0785
Treatment effect calibration slope 0.706
Treatment effect calibration intercept 0.0005
Correlation, predicted effect against doubly robust pseudo-outcome 0.045
Predicted benefit, median (percentage points) 0.83
Predicted benefit, 5th to 95th percentile -11.03 to 10.06
Predicted benefit, full range -31.77 to 51.61
Observed estimates follow the predicted ordering across all quintiles Yes
Rank correlation between two independent cross fits (supplementary) 0.877
Agreement on top benefit quintile membership (supplementary) 79.2%
Patient level calculator proposed No

Positive values of predicted benefit favour total arterial grafting, because benefit is defined as the 10 year survival probability under total arterial grafting minus that under mixed grafting. A calibration slope of one with an intercept of zero would indicate perfect treatment effect calibration. Rank correlation between cross fits and agreement on extreme quintile membership are stability measures and do not substitute for calibration.

Supplementary Table S7:  Personalised model performance and calibration

Quintile of predicted benefit Operations Total arterial Deaths 
within 
10 years
Mean 
predicted 
benefit (pp)
Observed 
doubly 
robust estimate (pp)
95% CI
Q1 (least predicted benefit) 1,595 204 306 -8.82 -5.22 -11.90 to +1.46
Q2 1,595 251 186 -2.82 -5.13 -9.76 to -0.50
Q3 1,598 254 106 +0.83 +3.57 -0.03 to +7.18
Q4 1,591 261 187 +3.29 +3.81 -0.74 to +8.37
Q5 (most predicted benefit) 1,595 294 208 +8.48 +3.90 -0.43 to +8.22

Quintiles are formed on the cross fitted predicted benefit. Observed values are means of the doubly robust pseudo-outcome within each quintile with normal approximation intervals. Predicted spread across quintiles was 17.30 percentage points against an observed spread of 9.12. The ordering is not monotonic, and on the prespecified criteria the model does not achieve adequate treatment effect calibration. No patient level calculator was produced.

Supplementary Table S8:  Observed doubly robust estimates by predicted benefit quintile

Supplementary Figure S1: Propensity score distributions by grafting strategy.

Supplementary Figure S2: Absolute standardised mean differences before and after overlap weighting, showing the 28 largest imbalances.

Supplementary Figure S3: Sensitivity analyses with 95 per cent bootstrap intervals.

Supplementary Figure S4: Calibration of the exploratory personalised model by quintile of predicted benefit.

Supplementary Figure S5: Use of total arterial grafting over time and by graft burden.

References

  1. Loop, Floyd D., Bruce W. Lytle, Delos M. Cosgrove, Robert W. Stewart, Marlene Goormastic, George W. Williams, Leonard AR Golding et al. "Influence of the internal-mammary-artery graft on 10-year survival and other cardiac events." New England Journal of Medicine 314, no. 1 (1986): 1-6.
  2. Sousa-Uva, Miguel, Franz-Josef Neumann, Anders Ahlsson, Fernando Alfonso, Adrian P. Banning, Umberto Benedetto, Robert A. Byrne et al. "2018 ESC/EACTS Guidelines on myocardial revascularization." European Journal of Cardio-Thoracic Surgery 55, no. 1 (2019): 4-90.
  3. Lawton, Jennifer S., Jacqueline E. Tamis-Holland, Sripal Bangalore, Eric R. Bates, Theresa M. Beckie, James M. Bischoff, John A. Bittl et al. "2021 ACC/AHA/SCAI guideline for coronary artery revascularization: executive summary: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines." Circulation 145, no. 3 (2022): e4-e17.
  4. Taggart, David P., Umberto Benedetto, Stephen Gerry, Douglas G. Altman, Alastair M. Gray, Belinda Lees, Mario Gaudino et al. "Bilateral versus single internal-thoracic-artery grafts at 10 years." New England journal of medicine 380, no. 5 (2019): 437-446.
  5. Taggart, David P., Douglas G. Altman, Alastair M. Gray, Belinda Lees, Stephen Gerry, Umberto Benedetto, and Marcus Flather. "Randomized trial of bilateral versus single internal-thoracic-artery grafts." New England Journal of Medicine 375, no. 26 (2016): 2540-2549.
  6. Gaudino, Mario, Umberto Benedetto, Stephen Fremes, Giuseppe Biondi-Zoccai, Art Sedrakyan, John D. Puskas, Gianni D. Angelini et al. "Radial-artery or saphenous-vein grafts in coronary-artery bypass surgery." New England Journal of Medicine 378, no. 22 (2018): 2069-2077.
  7. Gaudino, Mario, Umberto Benedetto, Stephen Fremes, Karla Ballman, Giuseppe Biondi-Zoccai, Art Sedrakyan, Giuseppe Nasso et al. "Association of radial artery graft vs saphenous vein graft with long-term cardiovascular outcomes among patients undergoing coronary artery bypass grafting: a systematic review and meta-analysis." Jama 324, no. 2 (2020): 179-187.
  8. Gaudino, Mario, Roberto Lorusso, Mohamed Rahouma, Ahmed Abouarab, Derrick Y. Tam, Cristiano Spadaccio, Gaëlle Saint-Hilary et al. "Radial artery versus right internal thoracic artery versus saphenous vein as the second conduit for coronary artery bypass surgery: a network meta-analysis of clinical outcomes." Journal of the American Heart Association 8, no. 2 (2019): e010839.
  9. Yi, Gijong, Brian Shine, Syed M. Rehman, Douglas G. Altman, and David P. Taggart. "Effect of bilateral internal mammary artery grafts on long-term survival: a meta-analysis approach." Circulation 130, no. 7 (2014): 539-545.
  10. Gaudino, Mario, Mohamed Rahouma, Irbaz Hameed, Faiza M. Khan, David P. Taggart, Marcus Flather, Giuseppe Biondi-Zoccai, and Stephen E. Fremes. "Disagreement between randomized and observational evidence on the use of bilateral internal thoracic artery grafting: a meta-analytic approach." Journal of the American Heart Association 8, no. 23 (2019): e014638.
  11. Stefil, Maria, Matthew Dixon, Umberto Benedetto, Mario Gaudino, Belinda Lees, Alastair Gray, Stephen Gerry, David Taggart, and Marcus Flather. "Coronary artery bypass grafting using bilateral internal thoracic arteries in patients with diabetes and obesity: a systematic review and meta-analysis." IJC Heart & Vasculature 47 (2023): 101235.
  12. De Paulis, MD Ruggero, MD Stefano de Notaris, MD Raffaele Scaffa, MD Saverio Nardella, MD Jacob Zeitani, MD Costantino Del Giudice, MD Alfonso Penta De Peppo, MD Fabrizio Tomai, and MD Luigi Chiariello. "The effect of bilateral internal thoracic artery harvesting on superficial and deep sternal infection: the role of skeletonization." The Journal of thoracic and cardiovascular surgery 129, no. 3 (2005): 536-543.
  13. Taggart, David P., Mario F. Gaudino, Stephen Gerry, Alastair Gray, Belinda Lees, Arnaldo Dimagli, John D. Puskas et al. "Effect of total arterial grafting in the Arterial Revascularization Trial." The Journal of thoracic and cardiovascular surgery 163, no. 3 (2022): 1002-1009.
  14. Thomas, Laine E., Fan Li, and Michael J. Pencina. "Overlap weighting: a propensity score method that mimics attributes of a randomized clinical trial." Jama 323, no. 23 (2020): 2417-2418.
  15. Li, Fan, Kari Lock Morgan, and Alan M. Zaslavsky. "Balancing covariates via propensity score weighting." Journal of the American Statistical Association 113, no. 521 (2018): 390-400.
  16. Li, Fan, Laine E. Thomas, and Fan Li. "Addressing extreme propensity scores via the overlap weights." American journal of epidemiology 188, no. 1 (2019): 250-257.
  17. Austin, Peter C. "Balance diagnostics for comparing the distribution of baseline covariates between treatment groups in propensityscore matched samples." Statistics in medicine 28, no. 25 (2009): 3083-3107.
  18. Royston, Patrick, and Mahesh KB Parmar. "Restricted mean survival time: an alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome." BMC medical research methodology 13, no. 1 (2013): 152.
  19. Benjamini, Yoav, and Yosef Hochberg. "Controlling the false discovery rate: a practical and powerful approach to multiple testing." Journal of the Royal statistical society: series B (Methodological) 57, no. 1 (1995): 289-300.
  20. Kennedy, Edward H. "Towards optimal doubly robust estimation of heterogeneous causal effects." Electronic Journal of Statistics 17, no. 2 (2023): 3008-3049.
  21. Chernozhukov, Victor, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins. "Double/debiased machine learning for treatment and structural parameters." (2018): C1-C68.
  22. Von Elm, Erik, Douglas G. Altman, Matthias Egger, Stuart J. Pocock, Peter C. Gøtzsche, and Jan P. Vandenbroucke. "The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies." The lancet 370, no. 9596 (2007): 1453-1457.
  23. Benchimol, Eric I., Liam Smeeth, Astrid Guttmann, Katie Harron, David Moher, Irene Petersen, Henrik T. Sørensen, Erik von Elm, Sinéad M. Langan, and RECORD Working Committee. "The REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement." PLoS medicine 12, no. 10 (2015): e1001885.
  24. Ren, Justin, Christopher M. Reid, Julian A. Smith, Colin Royse, Dion Stub, Wiliam Chan, David M. Kaye et al. "Long-term survival advantage of total arterial revascularization in elderly patients following coronary artery bypass grafting." JACC: Advances 4, no. 12_Part_1 (2025): 102226.
  25. Gaudino, Mario, John H. Alexander, Faisal G. Bakaeen, Karla Ballman, Fabio Barili, Antonio Maria Calafiore, Piroze Davierwala et al. "Randomized comparison of the clinical outcome of single versus multiple arterial grafts: the ROMA trial—rationale and study protocol." European Journal of Cardio-Thoracic Surgery 52, no. 6 (2017): 1031-1040.

Become an Editorial Board Member

Become a Reviewer

What our clients say

MEDIRES PUBLISHING

At our organization, we prioritize excellence in supporting the endeavors of researchers and practitioners alike. With a commitment to inclusivity and diversity, our journals eagerly accept various article types, including but not limited to Research Papers, Review Articles, Short Communications, Case Reports, Mini-Reviews, Opinions, and Letters to the Editor.

This approach ensures a rich tapestry of scholarly contributions, fostering an environment ripe for intellectual exchange and advancement."

Contact Info

MEDIRES PUBLISHING LLC,
447 Broadway, 2nd Floor, Suite #1734,
New York, 10013, United States.
Phone: +1 (786) 490-6788
WhatsApp us: WhatsApp - Medires Online
Email: info@mediresonline.org