Limits...
Long-term mortality prediction after operations for type A ascending aortic dissection.

Macrina F, Puddu PE, Sciangula A, Totaro M, Trigilia F, Cassese M, Toscano M - J Cardiothorac Surg (2010)

Bottom Line: Training SVM, using a larger number of covariates, showed no false negative or false positive cases among 118 randomly selected patients (error = 0%, AUC 1.0) whereas validation SVM, among 117 patients, provided 5 false negative and 11 false positive cases (error = 22%, AUC 0.821, p < 0.01 versus NN results).Both NN and SVM accurately selected a few operative and immediate post-operative factors and the Marfan habitus as long-term mortality predictors in AAD Type A.Although these factors were not new per se, their combination may be used in practice to index death risk post-operatively with good accuracy.

View Article: PubMed Central - HTML - PubMed

Affiliation: Department of the Heart and Great Vessels Attilio Reale, University of Rome La Sapienza, Rome, Italy.

ABSTRACT

Background: There are few long-term mortality prediction studies after acute aortic dissection (AAD) Type A and none were performed using new models such as neural networks (NN) or support vector machines (SVM) which may show a higher discriminatory potency than standard multivariable models.

Methods: We used 32 risk factors identified by Literature search and previously assessed in short-term outcome investigations. Models were trained (50%) and validated (50%) on 2 random samples from a consecutive 235-patient cohort. NN were run only on patients with complete data for all included variables (N = 211); SVM on the overall group. Discrimination was assessed by receiver operating characteristic area under the curve (AUC) and Gini's coefficients along with classification performance.

Results: There were 84 deaths (36%) occurring at 564 +/- 48 days (95%CI from 470 to 658 days). Patients with complete variables had a slightly lower death rate (60 of 211, 28%). NN classified 44 of 60 (73%) dead patients and 147 of 151 (97%) long-term survivors using 5 covariates: immediate post-operative chronic renal failure, circulatory arrest time, the type of surgery on ascending aorta plus hemi-arch, extracorporeal circulation time and the presence of Marfan habitus. Global accuracies of training and validation NN were excellent with AUC respectively 0.871 and 0.870 but classification errors were high among patients who died. Training SVM, using a larger number of covariates, showed no false negative or false positive cases among 118 randomly selected patients (error = 0%, AUC 1.0) whereas validation SVM, among 117 patients, provided 5 false negative and 11 false positive cases (error = 22%, AUC 0.821, p < 0.01 versus NN results). An html file was produced to adopt and manipulate the selected parameters for practical predictive purposes.

Conclusions: Both NN and SVM accurately selected a few operative and immediate post-operative factors and the Marfan habitus as long-term mortality predictors in AAD Type A. Although these factors were not new per se, their combination may be used in practice to index death risk post-operatively with good accuracy.

Show MeSH

Related in: MedlinePlus

Receiver operating characteristic plots by randomly selected training (50%) and validation (50%) support vector machine (SVM) with optimal C search on overall study patients (N = 235). A semi-quantitative graphic presentation of the covariates relevance is presented for training and validation models. Full names of coded variables are reported in Additional File 2, Table S1. Keep = 1 means that covariate may stay in the model. Using 27 of 32 covariates, Gini's coefficient by training SVM was 1.00 and no false negative or false positive cases were identified among 118 randomly selected AAD Type A patients (error = 0%). However, validation SVM on the remaining 117 patients provided 15 false negative and 11 false positive cases (error = 22%) and the Gini's coefficient was 0.642 (ROC AUC 0.821), which is statistically lower (p < 0.01) than the results obtained by neural network model, shown in Figure 1. Of note that validation and training SVM use different covariates to predict outcome and a relatively different ranked importance. Nevertheless, with both training and validation SVM, apart from extracorporeal circulation time, the other 4 covariates were also selected by neural network models.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC2902218&req=5

Figure 2: Receiver operating characteristic plots by randomly selected training (50%) and validation (50%) support vector machine (SVM) with optimal C search on overall study patients (N = 235). A semi-quantitative graphic presentation of the covariates relevance is presented for training and validation models. Full names of coded variables are reported in Additional File 2, Table S1. Keep = 1 means that covariate may stay in the model. Using 27 of 32 covariates, Gini's coefficient by training SVM was 1.00 and no false negative or false positive cases were identified among 118 randomly selected AAD Type A patients (error = 0%). However, validation SVM on the remaining 117 patients provided 15 false negative and 11 false positive cases (error = 22%) and the Gini's coefficient was 0.642 (ROC AUC 0.821), which is statistically lower (p < 0.01) than the results obtained by neural network model, shown in Figure 1. Of note that validation and training SVM use different covariates to predict outcome and a relatively different ranked importance. Nevertheless, with both training and validation SVM, apart from extracorporeal circulation time, the other 4 covariates were also selected by neural network models.

Mentions: Figure 2 shows the results of the SVM run on the overall study group (N = 235), since by this method there is no limitation to confine the analysis to patients with complete data for all variables, as with neural networks. A somewhat different picture is provided by SVM as compared to neural network. First, SVM make use of a larger number of covariates, some of which provided little if any information, yet globally enabled to obtain a Gini's coefficient of 1.00 (using 27 of 32 covariates) with no false negative or false positive cases identified among 118 randomly selected AAD Type A patients (error = 0%). Second, when validation SVM were run on the remaining 117 patients, the Gini's coefficient was 0.642 (with an ROC AUC = 0.821), a statistically lower (p < 0.01) result as compared to those obtained by neural network model. There were 15 false negative and 11 false positive cases (error = 22%) identified. Third, validation and training SVM used different covariates to predict outcome and there was a relatively different ranked importance.


Long-term mortality prediction after operations for type A ascending aortic dissection.

Macrina F, Puddu PE, Sciangula A, Totaro M, Trigilia F, Cassese M, Toscano M - J Cardiothorac Surg (2010)

Receiver operating characteristic plots by randomly selected training (50%) and validation (50%) support vector machine (SVM) with optimal C search on overall study patients (N = 235). A semi-quantitative graphic presentation of the covariates relevance is presented for training and validation models. Full names of coded variables are reported in Additional File 2, Table S1. Keep = 1 means that covariate may stay in the model. Using 27 of 32 covariates, Gini's coefficient by training SVM was 1.00 and no false negative or false positive cases were identified among 118 randomly selected AAD Type A patients (error = 0%). However, validation SVM on the remaining 117 patients provided 15 false negative and 11 false positive cases (error = 22%) and the Gini's coefficient was 0.642 (ROC AUC 0.821), which is statistically lower (p < 0.01) than the results obtained by neural network model, shown in Figure 1. Of note that validation and training SVM use different covariates to predict outcome and a relatively different ranked importance. Nevertheless, with both training and validation SVM, apart from extracorporeal circulation time, the other 4 covariates were also selected by neural network models.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC2902218&req=5

Figure 2: Receiver operating characteristic plots by randomly selected training (50%) and validation (50%) support vector machine (SVM) with optimal C search on overall study patients (N = 235). A semi-quantitative graphic presentation of the covariates relevance is presented for training and validation models. Full names of coded variables are reported in Additional File 2, Table S1. Keep = 1 means that covariate may stay in the model. Using 27 of 32 covariates, Gini's coefficient by training SVM was 1.00 and no false negative or false positive cases were identified among 118 randomly selected AAD Type A patients (error = 0%). However, validation SVM on the remaining 117 patients provided 15 false negative and 11 false positive cases (error = 22%) and the Gini's coefficient was 0.642 (ROC AUC 0.821), which is statistically lower (p < 0.01) than the results obtained by neural network model, shown in Figure 1. Of note that validation and training SVM use different covariates to predict outcome and a relatively different ranked importance. Nevertheless, with both training and validation SVM, apart from extracorporeal circulation time, the other 4 covariates were also selected by neural network models.
Mentions: Figure 2 shows the results of the SVM run on the overall study group (N = 235), since by this method there is no limitation to confine the analysis to patients with complete data for all variables, as with neural networks. A somewhat different picture is provided by SVM as compared to neural network. First, SVM make use of a larger number of covariates, some of which provided little if any information, yet globally enabled to obtain a Gini's coefficient of 1.00 (using 27 of 32 covariates) with no false negative or false positive cases identified among 118 randomly selected AAD Type A patients (error = 0%). Second, when validation SVM were run on the remaining 117 patients, the Gini's coefficient was 0.642 (with an ROC AUC = 0.821), a statistically lower (p < 0.01) result as compared to those obtained by neural network model. There were 15 false negative and 11 false positive cases (error = 22%) identified. Third, validation and training SVM used different covariates to predict outcome and there was a relatively different ranked importance.

Bottom Line: Training SVM, using a larger number of covariates, showed no false negative or false positive cases among 118 randomly selected patients (error = 0%, AUC 1.0) whereas validation SVM, among 117 patients, provided 5 false negative and 11 false positive cases (error = 22%, AUC 0.821, p < 0.01 versus NN results).Both NN and SVM accurately selected a few operative and immediate post-operative factors and the Marfan habitus as long-term mortality predictors in AAD Type A.Although these factors were not new per se, their combination may be used in practice to index death risk post-operatively with good accuracy.

View Article: PubMed Central - HTML - PubMed

Affiliation: Department of the Heart and Great Vessels Attilio Reale, University of Rome La Sapienza, Rome, Italy.

ABSTRACT

Background: There are few long-term mortality prediction studies after acute aortic dissection (AAD) Type A and none were performed using new models such as neural networks (NN) or support vector machines (SVM) which may show a higher discriminatory potency than standard multivariable models.

Methods: We used 32 risk factors identified by Literature search and previously assessed in short-term outcome investigations. Models were trained (50%) and validated (50%) on 2 random samples from a consecutive 235-patient cohort. NN were run only on patients with complete data for all included variables (N = 211); SVM on the overall group. Discrimination was assessed by receiver operating characteristic area under the curve (AUC) and Gini's coefficients along with classification performance.

Results: There were 84 deaths (36%) occurring at 564 +/- 48 days (95%CI from 470 to 658 days). Patients with complete variables had a slightly lower death rate (60 of 211, 28%). NN classified 44 of 60 (73%) dead patients and 147 of 151 (97%) long-term survivors using 5 covariates: immediate post-operative chronic renal failure, circulatory arrest time, the type of surgery on ascending aorta plus hemi-arch, extracorporeal circulation time and the presence of Marfan habitus. Global accuracies of training and validation NN were excellent with AUC respectively 0.871 and 0.870 but classification errors were high among patients who died. Training SVM, using a larger number of covariates, showed no false negative or false positive cases among 118 randomly selected patients (error = 0%, AUC 1.0) whereas validation SVM, among 117 patients, provided 5 false negative and 11 false positive cases (error = 22%, AUC 0.821, p < 0.01 versus NN results). An html file was produced to adopt and manipulate the selected parameters for practical predictive purposes.

Conclusions: Both NN and SVM accurately selected a few operative and immediate post-operative factors and the Marfan habitus as long-term mortality predictors in AAD Type A. Although these factors were not new per se, their combination may be used in practice to index death risk post-operatively with good accuracy.

Show MeSH
Related in: MedlinePlus