Rule Extraction for Health Care Decision Making from Artificial Neural Networks Models

摘要

Artificial neural networks (ANN) are now widely used for modeling non-linearrnrelationships in many complex systems where analytical equations are not available.rnHistorical data is used to train a model, which can then be used for prediction,rnknowledge extraction and decision making. Health care is one of the areas in whichrnANN models can be useful. However, the acceptance of such models depends on thernconfidence of the user in the ANN recommendation. This confidence will bernenhanced if understandable rules can be developed from the ANN models.rnIn this paper several rule extraction methods will be shown, using two medicalrndatabases available in data repositories. The first method is the Causal Index (CI)rncalculation from the trained ANN. The CI is a relative numerical causal relationshiprnbetween each input and each output, from which qualitative rules may be formed. Inrnsome cases better rules can be extracted if the numerical inputs are fuzzified.rnA second method is based on clustering the data using the outputs of the hiddenrnneurons of a trained ANN model. In many cases “binary” patterns can be found, eachrnpattern defining a cluster. Calculation the mean values of the inputs of the examplesrnin each cluster relative to the mean input values of the rest of the examples reveals therncontribution of each input on the clustering result.rnThe third method is the extension of the previous two methods to auto-associativernANNs, in which the AA-ANN models the relationships in the data, the outputs of thernmodels are the predicted inputs. Both CI and hidden-neuron-output clustering rulernextraction are possible.rnThe health-care examples used are the post-operative decision making and the cardiacrncondition diagnosis. In the first example the patient is sent to intensive care unit,rnregular hospital department or home based on several attributes of his condition. Inrnthe second example the existence of heart artery blocking is diagnosed from thernpatient test results and other attributes. Efficient ANN training algorithms whichrnemploys non-random initial connection weights and local minima escaping were used.rnIn both cases a high percentage of prediction is achieved, and understandable rulesrnand relationships are extracted.