Abstract & Commentary
Source: Das A, et al. Prediction of outcome in acute lower-gastrointestinal haemorrhage based on an artificial neural network: Internal and external validation of a predictive model. Lancet 2003;362:1261-1266.
Evaluation of emergency department (ED) patients with acute lower gastrointestinal bleeding (LGIB) would be much improved if there was a reliable decision tool that accurately could predict which patients are at risk for recurrent bleeding, death, and need for therapeutic interventions to control hemorrhage. The aim of this study was to develop and validate an artificial neural network (ANN) model to predict these outcomes in LGIB patients and to compare this with the predictive performance of a conventional multiple logistic regression (MLR) model and the BLEED (bleeding, low systolic blood pressure, elevated prothrombin time, erratic me tal status, unstable comorbid disease) classification system, a previously validated model.
In the first part of the study, the ANN was trained with 26 prospectively collected clinical variables from 120 patients with LGIB admitted to an academic, urban, U.S. medical center in a 12-month period. For the next six months, data from 70 patients were used for internal validation of the ANN, during which time 19 variables were used to predict outcome. Finally, the ANN model was validated externally and direct comparison was made with an MLR model and the BLEED classification system in 142 patients admitted to an independent institution in another state.
The predictive accuracy (sum of correct predictions divided by total predictions) of the ANN in the internal validation group was significantly better than that of BLEED (predictive accuracy for death, 87% vs 21%; for recurrent bleeding, 89% vs 41%; for intervention, 96% vs 46%) and was similar to that of MLR. During the external validation, the ANN performed well in predicting death (97%; 95%, CI 95-99%), recurrent bleeding (93%; 95% CI 89-97%), and need for intervention (94%; 95% CI 90-98%) and it was superior to MLR (70%, 73%, and 70%). Although the positive predictive value of the ANN was not high in either validation group, the negative predictive value was high for the three outcome variables in both the internal and external validation groups (98-100%).
Commentary by Stephanie B. Abbuhl, MD, FACEP
The ANN’s excellent negative predictive value suggests that this tool may have a role in determining a low-risk population of LGIB patients who safely could be discharged for further outpatient evaluation. The next step would be for the ANN to be tested prospectively against unassisted clinical judgment. If found to improve clinical outcome, this decision tool could be of great significance.
One of the most impressive findings in this study was the superior performance of the ANN in the external cohort of patients, who differed significantly from the internal group, both in the clinical characteristics of the patients and in the second institution’s management approach to patients with acute LGIB. This supports the hypothesis that, unlike conventional predictive models, ANN models are more universally applicable.
ANN models use non-linear analysis to reveal previously unrecognized relations between given input variables and an output variable. This method has tremendous appeal due to its potential power to analyze complex interactions, and yet, there is inherent resistance to ANNs in that the network logic cannot be broken down into simple elements of clinical reasoning.
There are other potential barriers to the widespread clinical use of ANN predictive instruments. They typically use more variables than MLR models, and entering 26 clinical variables into a computer is not practical in most EDs at the present time. However, with the increasing use of fully computerized charting systems, one could imagine that most of the variables already would be routinely entered by various members of the ED team and the outcome score would be determined as easily as clicking the mouse on the correct icon.
Dr. Abbuhl, Medical Director, Department of Emergency Medicine, The Hospital of the University of Pennsylvania; Associate Professor of Emergency Medicine, University of Pennsylvania School of Medicine, Philadelphia, is on the Editorial Board of Emergency Medicine Alert.