Markers of Exposure to the Colombian Armed Conflict: A Machine Learning Approach.

The Colombian armed conflict has affected in some degree its entire population. Health authorities require markers to determine this exposure and provide proper mental-health interventions. Unsupervised learning techniques allow clustering subjects with similar features. Here, we propose a novel methodology to automatically finds the features that best relate to levels of exposure to the armed conflict and associated risks (drug dependency, alcoholism, etc.) through cluster centers. Unlike previous studies on the armed conflict field, we do not use key predefined labels to cluster the data. We test this methodology with a mixed-response type characterization database of 528 features obtained from 346 volunteers with different estimated levels of exposure to extreme experiences in the frame of the Colombian armed conflict. As a result, using the proposed approach we identified 62 features related to exposure. In order to confirm the selected features as violence exposure markers, we created a model based on artificial neural networks (ANN). The ANN model uses the 62 features as input and it was able to estimate the subjects' level of exposure to conflict with 100 % accuracy in training and over 76% in validation.