OBJECTIVE: The authors sought to train an artificial neural network to predict early outcomes after orthotopic liver transplantation. SUMMARY BACKGROUND DATA: Reliable prediction of outcomes early after liver transplantation would help improve organ use and could have an impact on patient survival, but remains an elusive goal. Traditional multivariate models have failed to attain the sensitivity and specificity required for practical clinical use. Alternate approaches that can help us model clinical phenomena must be explored. One such approach is the use of artificial neural networks, or connectionist models. These are computation systems that process information in parallel, using large numbers of simple units, and excel in tasks involving pattern recognition. They are capable of adaptive learning and self-organization, and exhibit a high degree of fault tolerance. METHODS: Ten feed-forward, back-propagation neural networks were trained to predict graft outcomes, using data from 155 adult liver transplants. The data included information that was available by the second postoperative day. Ten separate training and testing data subsets were prepared, using random sampling, and the ability of the different networks to predict outcomes successfully was evaluated using receiver operating characteristic (ROC) curve analysis. RESULTS: Four of the networks showed perfect discrimination, with an area under the ROC curve (Az) of 1.0. Two other networks also had excellent performance, with an Az of 0.95. The sensitivity and specificity of the combined networks was 60% and 100%, respectively, when using an output neuron activation of 0.6 as the cutoff point to decide class membership. Lowering the cutoff point to 0.14 increased the sensitivity to 77%, and lowered the specificity to 96%. CONCLUSIONS: These results are encouraging, especially when compared to the performance of more traditional multivariate models on the same data set. The robustness of neural networks, when confronted with noisy data generated by nonlinear processes, and their freedom from a priori assumptions regarding the data, make them promising tools with which to develop predictive clinical models.
展开▼
