Mixed-effects Model For Classification And Prediction In Longitudinal Data Analysis

Longitudinal studies use repeated measures to study change over prolonged periods of time. We propose a method for classification of longitudinal data. This work is motivated by a real world dataset obtained following a clinical trial conducted as part of an ICMR funded project titled "Prevention of Pneumonia in HIV-infected Children". We tried to predict the prevalence of a disease if an intervention was not given and compare it with the actual prevalence after an intervention to assess the indirect impact of the intervention. A k-NN based imputation method was used to fill a majority of the missing data. In this paper, we trained two ensemble and two distance-based classifiers on data from previous visits to predict for a subsequent visit. A weighted average of the outputs of each classifier was compared with the actual observations. We found that Gradient Boosting (GB) and Support Vector Machine (SVM) outperformed the other classifiers. We were able to predict that the intervention decreased the prevalence by 12%.