Revisión sistemática de modelos de aprendizaje automático para la estimación del tiempo de desaparición de las personas

Abstract

This study presents a systematic review of the application of machine learning models to the analysis of missing persons cases, focusing on time prediction, characterization of the social phenomenon, and comparison of algorithmic performance. Following the PRISMA methodology, 72 studies were initially identified, of which only 12 met the inclusion criteria related to relevance, methodological quality, and availability of technical information. The articles, published between 2020 and 2025, reveal a sustained increase in the use of machine learning to address complex social problems, with a greater concentration in the Americas and Europe, especially in Colombia, Ecuador, Spain, and Poland. The models analyzed include supervised methods (Random Forest, Decision Trees, KNN, SVM), deep neural networks (CNN, MLP, LSTM, GRU), and econometric models (VAR, VECM), applied according to the nature of the data. Based on a comprehensive analysis of the studies, tree-based models (Random Forest and Decision Trees) are recommended when stability, relative interpretability, and balanced performance are desired; deep neural networks when the data are nonlinear, sequential, or visual; and SVMs when accuracy is paramount and the dataset structure is clear and well-defined. This review shows that machine learning is a promising tool for strengthening the search for and investigation of missing persons, although challenges remain regarding data quality, the standardization of metrics, and the need for more explainable models for institutional adoption.