Artificial neural network assessment of substitutive pharmacological treatments in hospitalised intravenous drug users.

Artificial neural networks (ANNs) provide better solutions than linear discriminant analysis (LDA) to problems of classification and estimation involving a large number of non-homogeneous (categorical and metric) variables. In this study, we compared the ability of traditional LDA and a feed-forward back-propagation (FF-BP) ANN with self-momentum to predict pharmacological treatments received by intravenous drug users (IDUs) hospitalised for coexisting medical illness. When medical staff considered detoxification appropriate they usually suggested methadone (MET) and (or) benzodiazepines (BDZ). Given four different treatment options (MET, BDZ, MET+BDZ, no treatment) as dependent variables and 38 independent variables, the FF-BP ANN provided the best prediction of the consultant's decision (overall accuracy: 62.7%). It achieved the highest level of predictive accuracy for the BDZ option (90.5%), the lowest for no treatment (29.6), often misclassifying no treatment as BDZ. The LDA yielded a lower mean accuracy (50.3%). When the untreated group was excluded, ANN improved its absolute recognition rate by only 1.2% and the BDZ group remained the best predicted. In contrast, LDA improved its absolute recognition rate from 50.3 to 58.9%, maximum 65.7% for the BDZ group. In conclusion, the FF-BP ANN was more accurate than the statistical model (discriminant analysis) in predicting the pharmacological treatment of IDUs.