Abstract
Abstract only
Objective:
Identifying patients at high risk of AE after TAVR is essential to prolong their survival. Current prediction models for AE after TAVR suffer from a lack of accuracy and external validation. Modern ML approaches can account for higher-dimensional relationships among variables, potentially improving the prediction of outcomes. We performed a systematic review and meta-analysis to estimate the discriminative ability of recently developed ML-based models, which predict various AE after TAVR.
Methods:
We searched Pubmed, Google Scholar, and Web of Science for studies (Jan 2019 to Jan 2022) that used ML approaches to predict AE after TAVR. Inputs in the meta-analysis were study-reported c-index values and 95% CI. Subgroup analyses separated models by outcome (mortality or clinical AE). Combined effect sizes using a random-effects model, test for heterogeneity, and Egger’s test to assess publication bias were considered.
Results:
Eight studies were included in the systematic review (patients = 26,023; outcomes = 1,014), of which five models had sufficient data for the meta-analysis. The number of features included in each model ranged from 6 to 107. The two most common models were random forest (n=2) and logistic regression (n=2). The most common outcome was mortality (n=5). The meta-analysis showed that models predicting mortality performed better (0.90; 95% CI: 0.81, 1.01) than models predicting clinical AE (0.80; 95% CI: 0.79, 0.95). The combined mean c-index was 0.87 (95% CI: 0.79-0.95). Test of heterogeneity showed high variation among studies (
I
2
=98.5%). Egger’s test did not indicate publication bias (
β
= 1.48; 95% CI: –18.14, 21.09,
p
= 0.848).
Conclusion:
Although relatively few studies have applied ML for predicting AE after TAVR, the results are very promising. The time of complex sophisticated models has arrived with improved predictive accuracy through advanced ML methods able to help identify patients who are at risk for clinical AE early in their care.