Objectives: Our objective was to compare artificial neural networks (ANNs) with logistic regression for prediction of in-hospital death after percutaneous transluminal coronary angioplasty and to assess the impact of guiding initial ANN variable selection with univariate analysis.
Background: ANNs can detect complex patterns within data. Criticisms include the unpredictability of variable selection. They have not previously been applied to outcomes modeling for percutaneous coronary interventions.
Methods: A database of consecutive (n = 3019) percutaneous transluminal coronary angioplasty procedures from an academic tertiary referral center between July 1994 and July 1997 was used. An ANN was developed for 38 variables (unguided model) (n = 1554). A second model was developed with predictors from an univariate analysis (guided model). Both were compared with a logistic regression model developed from the same database. Model validation was performed on independent data (n = 1465). Model predictive accuracy was assessed by the area under receiver operating characteristic curves. Goodness of fit was assessed with the Hosmer-Lemeshow statistic.
Results: Sixty unguided and guided ANNs were developed. Predictive accuracy and model calibration for all models were similar for training data but were significantly better for logistic regression for independent validation data. Overestimation of event rate in higher risk patients accounted for the majority of discrepancy in model calibration for the ANNs. This difference was partially amended by guiding variable selection.
Conclusion: ANNs were able to model in-hospital death after percutaneous transluminal coronary angioplasty when guiding variable selection. However, performance was not better than traditional modeling techniques. Further investigations are needed to understand the impact of this methodology on outcomes analysis.