MACHINE LEARNING MODEL FOR PREDICTION OF POSTPARTUM DEPRESSION, A CASE OF MOMBASA COUNTY

Abstract

Postpartum depression is a medical condition which affects many mothers. The condition exposes the mother and the newborn baby to illnesses that can lead to death. Management of the condition requires heavy expenditure incurred by the family, government, and stakeholders. The condition is also a source of many social problems. Manual systems are currently used to predict the condition which is slow and inconsistent. Machine learning technology which has reliably been used in prediction modelling in other domains can also be employed to build a model to predict mothers at risk of postpartum depression during pregnancy for primary prevention. In this study, perinatal records were collected from 324 mothers attending postnatal healthcare clinics at the Coast General Teaching and Referral Hospital and Tudor Sub-County Hospital in Mombasa County. The filter feature in WEKA was used to split the data into 70% and 30% for model training and testing respectively. Models were built on WEKA machine learning platform using logistic regression, support vector machine, extremely randomized trees, random forest and adaptive boosting algorithms which were identified from literature. A positive case of postpartum depression was defined as diagnosis or treatment of major depression within one year after delivery. Random forest model produced the best performance with a receiver operating characteristic (ROC) curve area of 0.863867 which is comparable within the bracket of high performing models. With this level of performance, the model can be used by healthcare staff to make quick and consistent prediction for early mitigation measures. Further research could be done with more data collected from other counties in Kenya