Use of a probabilistic neural network to estimate the risk of mortality after cardiac surgery

Objective. To develop a probabilistic neural network (PNN) to estimate mortality risk following cardiac surgery. Design and setting. The PNN model was created using an institutional database obtained as part of routine quality assurance activity. Patient records (from 1991 to 1993) were randomly divided into training (n = 732) and validation (n = 380) sets. The model uses seven variables, each obtainable during routine clinical patient care. After completion of the initial validation phase, newer data (1994) became available and were used as an independent source of validation (n = 365). Patients. 1,477 consecutive cardiac surgery patients operated on in a teaching hospital during a four-year period (1991-94). Results. The overall accuracy of the neural network was 91.5% in the training set; it was 92.3% in the validation set. The model was well calibrated (p = 0.21 for the Hosmer-Lemeshow goodness-of-fit test) and discriminated well (areas under the ROC curves were 0.72 and 0.81 for the training and validation sets). The trained network also performed well on the 1994 data (ROC = 0.74, p = 0.19 for the Hosmer-Lemeshow test), albeit with a slight decrement in overall accuracy (88.2%). Conclusion. A neural network may be implemented to estimate mortality risk following cardiac surgery. Implementation is relatively rapid, and it is an alternative to standard statistical approaches.