Assessing Infant Mortality in Nigeria Using Artificial Neural Network and Logistic Regression Models

Abstract

Aim: To examine the suitability of Artificial Neural Network (ANN) in predicting infant mortality and compare its performance with Logistic Regression (LR) model. Study Design: A cross-sectional population based study was conducted. The 2013 Nigeria Demographic Health Survey (NDHS) data were used. Place and Duration of Study: The study was conducted in Nigeria and the fieldwork was carried out from February 15, 2013, to May 31, 2013. Methodology: Data were partitioned into training and testing sets with ratio 7:3. Logistic and ANN models were fitted on the training set and were validated using the testing sample. Akaike Information Criterion (AIC) and Area under curve (AUC) were used as criteria for comparing the two models. The discriminative ability was measured using sensitivity and specificity. Variable importance analysis was also conducted to determine the magnitude of contribution of each predictor to the outcome. Results: The sensitivity of the classification model was 67% and 76% for the LR and the ANN models respectively. Specificity of the prediction was 94% for the two models. Overall accuracy was approximately 81% and 83% for LR and ANN respectively. The AIC values were 9462 and 9614 for ANN model and LR model respectively. Area under curve was 0.621 and 0.637 for the LR model and the ANN model respectively. The variable importance analysis showed that preceding birth interval less than 24 months and not receiving tetanus toxoid injection during pregnancy had the highest positive contribution to infant mortality. Conclusion: The artificial neural network model had a higher sensitivity than the logistic regression model. Preceding birth interval of less than 24 months and non-reception of tetanus toxoid injection by mothers’ during pregnancy were important predictors of infant mortality in Nigeria.