Predicting operative mortality in patients who undergo elective open thoracoabdominal aortic aneurysm repair

Abstract

Background: We have developed a model aimed at identifying preoperative predictors of operative mortality in patients who undergo elective, open thoracoabdominal aortic aneurysm (TAAA) repair. We converted this model into an intuitive nomogram to aid preoperative counseling.

Methods: We retrospectively analyzed data from 2884 elective, open TAAA repairs performed between 1986 and 2023 in a single practice. Using clinical and selected operative variables, we built 4 predictive models: multivariable logistic regression (MLR), random forest, support vector machine, and gradient boosting machine. Each model's predictive effectiveness was evaluated with the C-statistic. Test C-statistics were computed using an 80:20 cross-validation scheme with 1000 iterations.

Results: Operative death occurred in 200 patients (6.9%). Test set C-statistics showed that the MLR model (median, 0.68; interquartile range [IQR], 0.65-0.71) outperformed the machine learning models (0.61 [IQR, 0.59-0.64] for random forest; 0.61 [IQR, 0.58-0.64] for support vector machine; 0.65 [IQR, 0.62-0.67] for gradient boosting machine). The final MLR model was based on 7 characteristics: increasing age (odds ratio [OR], 1.04/y; P < .001), cerebrovascular disease (OR, 1.54; P = .01), chronic kidney disease (OR, 1.53; P = .008), symptomatic aneurysm (OR, 1.42; P = .02), and Crawford extent I (OR, 0.66; P = .08), extent II (OR, 1.61; P = .01), and extent IV (OR, 0.41; P = .002). We converted this model into a nomogram.

Conclusions: Using institutional data, we evaluated several models to predict operative mortality in elective TAAA repair, using information available to surgeons preoperatively. We then converted the best predictive model, the MLR model, into an intuitive nomogram to aid patient counseling.

Keywords: aortic aneurysm; clinical decision rules; health care; nomograms; operative mortality; outcome assessment; patient counseling; prognosis; surgical risk.

Receiver operating characteristic curve for predicting operative mortality in TAAA repair patients.

Figure 1

Nomogram constructed from the predictive multiple logistic regression model. The number of points each predictor contributes can be determined from the corresponding point count on the top line. For binary variables, the “no” value serves as the reference value. Regarding the 4 Crawford extents of repair, extent III serves as the reference value. The sum of these points is then mapped onto the “total points” line, and the corresponding location on the “predicted value” line corresponds to the patient’s predicted risk of operative death. For example, a patient age 60 years (62 points) with chronic kidney disease (CKD; 14 points) undergoes an extent II repair (16 points, plus 14 points for not undergoing an extent I repair and 31 points for not undergoing an extent IV repair). This produces a total score of 137 points, which corresponds to an approximate operative mortality risk of 8%. CVD, Cerebrovascular disease.

Figure 2

Unmarked nomogram for the predictive multiple logistic regression model for use in patient counseling. The number of points that each predictor contributes can be determined from the corresponding point count on the top line. The sum of these points is then mapped onto the “total points” line, and the corresponding location on the “predicted value” line corresponds to the patient’s predicted risk of operative death. CVD, Cerebrovascular disease; CKD, chronic kidney disease.