Identifying markers of exposure to the Colombian armed conflict: a machine-learning approach

Abstract

ABSTRACT : The Colombian armed conflict has affected, to some degree, its entire population. Health authorities require markers to determine the consequences of this exposure and provide appropriate mental health interventions. In this thesis, we propose a novel methodology to automatically find the features that best relate to the level of exposure to the armed conflict and related risks (drug use, college desertion, among others) using the information provided by unsupervised techniques. We use clustering techniques that do not use predefined labels to cluster the data and obtain relevant information from cluster centers. This methodology was tested on two databases with more than 500 mixed response variables (dichotomous, categorical, Likert scale, etc.), the first with 346 subjects with a direct measure of their level of exposure in the context of the Colombian armed conflict, and the second one with 9467 subjects, but without a direct measure of their level of exposure. For the latter, the missing data problem was addressed by finding the appropriate parameters for eliminating and imputing missing values. As a result, 60 features related to exposure were identified as markers, which divided the subjects into three groups, in which characteristics related to high levels of exposure were highlighted. We created an artificial neural network (ANN) based model to confirm the features found as markers of exposure to violence. The model was able to estimate the level of exposure with an accuracy of 99% in training and 76% in validation using the selected features as input.