ABSTRACT
Logistic regression models are used to forecast outcomes from binary response events. Examples of such events are patient survival or a blood transfusion. Logistic regression models are calculated using input parameters correlated to the event of interest. The purpose of this research was to compare logistic regression and neural networks forecast performance on simulated data. Methods: Two-615 record data sets were created. Each record included 24 independent and one dependent (outcome) binary variables. In the first data set (probability set), outcome was computed using a probability function. In the second set (rule set), outcome resulted from the application of logic rules. The first 400 records of each set were used for neural network training and logistic regression estimation. The remaining records were used for testing. Results: Both models estimated correctly 66% of outcomes in the probability set. In the rule set, the neural network predicted outcome significantly better (p<0.0001) than the logistic model (neural network 96% vs. logistic regression 78% correct). Both models estimated outcome significantly better in the rule set than in to the probability set (p<0.0001). Conclusion: This work compared neural network and logistic regression forecasting performance on two binary data sets. These data sets were selected as extreme examples of cause-effect problems observed in medical research. Our results indicate that neural networks may yield better results that logistic regression. This finding may encourage the use neural networks in problems approached traditionally with logistic regression models.
