Purpose Mortality prediction models for patients with\nperforated peptic ulcer (PPU) have not yielded consistent\nor highly accurate results. Given the complex nature of this\ndisease, which has many non-linear associations with outcomes,\nwe explored artificial neural networks (ANNs) to\npredict the complex interactions between the risk factors of\nPPU and death among patients with this condition.\nMethods ANN modelling using a standard feed-forward,\nback-propagation neural network with three layers (i.e., an\ninput layer, a hidden layer and an output layer) was used to\npredict the 30-day mortality of consecutive patients from a\npopulation-based cohort undergoing surgery for PPU. A\nreceiver-operating characteristic (ROC) analysis was used\nto assess model accuracy.\nResults Of the 172 patients, 168 had their data included\nin the model; the data of 117 (70 %) were used for the\ntraining set, and the data of 51 (39 %) were used for the\ntest set. The accuracy, as evaluated by area under the ROC\ncurve (AUC), was best for an inclusive, multifactorial\nANN model (AUC 0.90, 95 % CIs 0.85ââ?¬â??0.95; p\\0.001).\nThis model outperformed standard predictive scores,\nincluding Boey and PULP. The importance of each variable\ndecreased as the number of factors included in the\nANN model increased.\nConclusions The prediction of death was most accurate\nwhen using an ANN model with several univariate\ninfluences on the outcome. This finding demonstrates that\nPPU is a highly complex disease for which clinical prognoses\nare likely difficult. The incorporation of computerised\nlearning systems might enhance clinical judgments to\nimprove decision making and outcome prediction.
Loading....