Abstract
Abstract only
Objective:
Machine learning (ML) may enhance prediction of outcomes such as mortality or acute kidney injury (AKI) among cardiac patients after coronary artery bypass graft (CABG). In this study, we used meta-analyses of reported ML models to assess what ML has been able to accomplish in this field, by evaluating the model performance in studies with CABG patients.
Methods:
We performed a literature search using Google Scholar and included studies that reported AUC and 95% CI for various models in our analysis. In addition, total participants, year of publication, type of analytical method (gradient boosting, random forest, etc.) and type of outcome (mortality or AKI) were extracted. We combined effect sizes using random effects model, and tested for heterogeneity, and publication bias.
Results:
5 models from 5 studies were included in the analysis (patients= 35,152; with outcome mortality =3,080, AKI=933). Combined mean AUC was 0.796 (95% CI: 0.776, 0.815). Test of heterogeneity showed high variation between studies (I
2
= 66.7%). Egger’s test intercept was -1.03 (95% CI: -7.22, 5.17,
p
> .25) indicating no small study bias. Meta regression showed newer publications had a positive association (
coef =
0.003) and number of variables in the study had a negative association with higher AUC values (
coef
= -0.0002). In subgroup analysis, the pooled AUC values for mortality and AKI groups were 0.795 and 0.805 respectively. The highest individual AUC was from ensemble model predicting AKI with AUC 0.84 and lowest was from gradient boosting model predicting mortality with AUC 0.77.
Conclusion:
Among the presented models for CABG ensemble methods performed well, but surprisingly methods with lesser number of variables tended to have higher predictive power. In near future, ML-based models may form the basis to build intelligent decision support systems for patient selection and risk stratification prior to CABG and could be applied to other cardiac surgeries.