Comparison of neural networks, decision trees, discriminant analysis and logistic regression for predicting unwanted pregnancy of multiparous women in Khorramabad

Background and Objective: Unwanted pregnancy is a pregnancy that is considered to be unwanted by at least one member of the couple, and has adverse consequences for the family and community. Using four classification models, this study predicted unwanted pregnancy in the urban population of Khorramabad and compared these classification models.

Materials and methods: In this cross-sectional study, 467 multiparous pregnant women attending the urban health care centers of Khorramabad in 2011 were selected using stratified and cluster sampling and risk factors were collected. The logistic regression model, discriminant analysis, decision trees, andCART artificial neural networks, along with the SPSS and MATLAB software were applied in data modeling. The indices of sensitivity, specificity, area under the ROC curve, and accuracy rate were applied to compare the models.

Results: The prevalence of unwanted pregnancy was 32.3%. Based on the index of area under the ROC curve,ROC the best models were artificial neural networks (0.741), decision tree (0.731), logistic regression (0.712) and discriminant analysis (0.711). The highest. sensitivity was found for decision tree model (73.5%), and the highest specificity was for artificial neural network (62.3%).

Conclusion: Given the relatively high prevalence of unwanted pregnancy in Khorramabad, the revision of the family planning programs seems to be inevitable. Moreover, in selecting the best classification method, decision tree and logistic regression are recommended when the researcher is interested in better interpretability of the results, and the model of neural networks is recommended when a higher prediction power is intended.