Predicting mortality following gastrointestinal surgery

Compared to survival data the accurate prediction of mortality following gastrointestinal surgery is a complex task. This is because mortality is rare and this makes the data unbalanced for classification. Consequently, the prediction accuracy tends to be biased towards the more abundant class i.e. survival. This essentially requires pre-processing of the data as well as an intelligent model to classify these cases. Our study aimed to identify the risk factors and develop an improved prediction model using such imbalanced data. We examined retrospective data on 3500 patients admitted to our surgical gastroenterology unit and applied a linear logistic model to identify the risk factors for mortality as well as an artificial neural network(ANN) technique for predicting the mortality with greater accuracy. Logistic regression indicated that patients requiring inotropic support or having gastrointestinal haemorrhage were at an almost four times greater and patients admitted as emergencies were at almost two and a half times greater risk of dying. Among the eight ANN models, we identified two based on ten predictors one specifically for predicting survival with a high accuracy (93%) & sensitivity (98%); and the other for predicting mortality with a high accuracy (85%) & sensitivity (83%) using the synthetically modified oversampling technique(SMOTE) along with under-sampling of the majority class. The ANN models with the SMOTE applied to the mortality class along with under-sampling of the survival class data provided a high prediction accuracy and sensitivity for mortality. However, the developed models need further testing on unseen cases.