Evaluación y aplicación de algoritmos de inteligencia artificial explicada para apoyar la toma de decisiones médicas en la salud fetal.

Abstract

This work answers the need for understand the decisions of AI models in the classification of fetal health. In order to do this, a comparison of machine learning algorithms and the implementation of a demonstrator model of Explained Artificial Intelligence through LIME has been proposed to support decision making on medical diagnoses. The project elaboration process has been supported by the CRISP-DM methodology, taking advantage of its approach oriented to data analysis and intelligent models’ implementation. The results of its execution indicate the lack of literature regarding XAI models in obstetrics; moreover, after the comparison and contextualized evaluation of the AI models: SVM, ANN and Random Forest, it is the latter that grants the best results. Regarding the XAI algorithm, the decision-influencing characteristics determine that for an instance classified as "Normal", low values in accelerations and prolonged decelerations support this decision. For an instance classified as "Suspicious", the percentage of prolonged time with abnormal long-term variability supports this classification. Finally, for an instance classified as "Pathological", prolonged and severe decelerations at minimum values contradict this result, therefore, higher values could warn of a pathology in the fetal health. These explanations have been validated by an expert in the obstetric field through the Technology Acceptance Model (TAM) in terms of ease of use and perceived usefulness. In addition, an implementation strategy has been proposed through a non-linear process with activities that facilitate to use the results of the models and seek their continuous improvement. This can mark a route from the reception of data to the delivery of value in terms of explanations.