Cargando...
Fecha
2025
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editorial
Citas
0 citas en 
0 citas en 
Resumen
Aplicar técnicas de inteligencia artificial en el ámbito sanitario supone un desafío debido a la escasez de datos etiquetados, el tamaño reducido de los conjuntos disponibles y la necesidad de interpretabilidad en los modelos.
Detectar el deterioro cognitivo de forma precoz es una prioridad en el ámbito clínico, especialmente en fases iniciales donde las intervenciones tempranas pueden mitigar significativamente su progresión. Este trabajo explora la aplicabilidad de técnicas de aprendizaje contrastivo, incluyendo redes siamesas (offline y online), SimCLR, SupCon, Joint Contrastive Classifier y Triplet Loss, para generar representaciones discriminativas a partir de un conjunto real de datos clínicos con solo 314 muestras. Para ello, se ha desarrollado una metodología experimental, que incluye la comparación de funciones de distancia, estrategias de generación de pares y técnicas de balanceo, tanto clásicas como aplicadas durante el entrenamiento.
Los modelos contrastivos se han evaluado sobre tres subconjuntos binarios y uno multiclase, y sus resultados se han comparado con modelos clásicos de aprendizaje automático. Los experimentos muestran que, aunque las técnicas contrastivas resultan competitivas en tareas binarias más simples, los modelos clásicos mantienen una ventaja general, especialmente en escenarios multiclase más complejos. Los resultados obtenidos mediante aprendizaje contrastivo refuerzan el valor de este enfoque como herramienta complementaria para el diagnóstico temprano en contextos reales con escasez de datos, donde el tamaño limitado de las muestras representa una barrera habitual para la aplicación de métodos avanzados de inteligencia artificial.
Applying artificial intelligence techniques in the healthcare domain poses a challenge due to the scarcity of labeled data, the limited size of available datasets, and the need for model interpretability. Early detection of cognitive impairment is a clinical priority, especially in its initial stages, where timely interventions can significantly slow down its progression. This work explores the applicability of contrastive learning techniques, including Siamese networks (offline and online), SimCLR, SupCon, Joint Contrastive Classifier, and Triplet Loss, to generate discriminative representations from a real clinical dataset consisting of only 314 samples. Accordingly, a experimental methodology has been developed, including the comparison of distance functions, pair generation strategies, and balancing techniques, both classical and integrated into training. Contrastive models have been evaluated on three binary subsets and one multiclass setting, and their results were compared with traditional machine learning models. The experiments show that, although contrastive techniques are competitive in simpler binary tasks, traditional models still hold an overall advantage, particularly in more complex multiclass scenarios. The results obtained through contrastive learning reinforce the value of this approach as a complementary tool for early diagnosis in real-world settings with limited data availability, where small sample sizes represent a common barrier to the application of advanced artificial intelligence methods.
Applying artificial intelligence techniques in the healthcare domain poses a challenge due to the scarcity of labeled data, the limited size of available datasets, and the need for model interpretability. Early detection of cognitive impairment is a clinical priority, especially in its initial stages, where timely interventions can significantly slow down its progression. This work explores the applicability of contrastive learning techniques, including Siamese networks (offline and online), SimCLR, SupCon, Joint Contrastive Classifier, and Triplet Loss, to generate discriminative representations from a real clinical dataset consisting of only 314 samples. Accordingly, a experimental methodology has been developed, including the comparison of distance functions, pair generation strategies, and balancing techniques, both classical and integrated into training. Contrastive models have been evaluated on three binary subsets and one multiclass setting, and their results were compared with traditional machine learning models. The experiments show that, although contrastive techniques are competitive in simpler binary tasks, traditional models still hold an overall advantage, particularly in more complex multiclass scenarios. The results obtained through contrastive learning reinforce the value of this approach as a complementary tool for early diagnosis in real-world settings with limited data availability, where small sample sizes represent a common barrier to the application of advanced artificial intelligence methods.
Descripción
Categorías UNESCO
Palabras clave
deterioro cognitivo, aprendizaje contrastivo, redes siamesas, datos tabulares, clasificación supervisada, diagnóstico clínico, cognitive impairment, contrastive learning, Siamese networks, tabular data, supervised classification, clinical diagnosis
Citación
López Pombero, Berta. Trabajo Fin de Máster: "Aplicación de redes siamesas y técnicas contrastivas sobre datos tabulares para el diagnóstico del deterioro cognitivo". Universidad Nacional de Educación a Distancia (UNED), 2025
Centro
E.T.S. de Ingeniería Informática