Aplicación de técnicas de aprendizaje de máquina para la predicción de riesgos en la salud ginecológica

Abstract

Abstract: Cervical cancer is one of the leading causes of death in women, in 2020, 604,127 cases of cervical cancer were diagnosed, with a mortality of 341,831 cases worldwide.The objective of this work is to analyze, explore, compare and identify the best cervical cancer prediction method using machine learning techniques. To achieve this objective, 17 investigations obtained from six scientific databases of the last 5 years were analyzed, and experiments were carried out with 7 algorithms: 1. Decision Tree (DT), 2. Random Forest (RF), 3. Logistic Regression (LR), 4. XGBoost (XG), 5. Naive Bayes (NB), 6. Multi-layer Perceptron(MLP), 7. K-Nearest Neighbors (KNN). Three data balancing techniques were applied: 1. SMOTETomek, 2. SMOTE, 3. ROS, for the four target variables: Hinselmann, Schiller, Citology and Biopsy. In addition, we worked with accuracy, precision, recall, f-score, AUC and the ROC curve to evaluate the values obtained from the prediction algorithms.The algorithm that stood out the best is Random Forest with an accuracy of 94.57% for selection of variables and reduction of PCA variables.