Human-centred machine learning for health and cognitive modelling

Vista sencilla de ítem
dc.contributor.advisorAlmeida, Aitor
dc.contributor.advisorZulaika Zurimendi, Unai
dc.contributor.authorPikatza Huerga, Amaia
dc.date.accessioned2026-08-31T11:56:13Z
dc.date.available2026-08-31T11:56:13Z
dc.date.issued2026-06-19
dc.description.abstractMachine learning is increasingly transforming research in healthcare, mental health, and cognitive science by enabling predictive and adaptive systems that support human decision-making. Yet the most accurate models often remain opaque, limiting their interpretability and trustworthiness in domains where transparency is essential. This doctoral research proposes a unified methodological framework for explainable and multimodal machine learning, validated across four representative contexts: prediction of hospital readmission in patients with heart failure (ReIC), prediction of readmission in patients with multiple chronic conditions (RePluris), prediction of eating disorder risk and recovery, and multimodal assessment of creativity from drawings and textual titles. The framework combines data preprocessing, imbalance management, multimodal feature integration, and embedded explainability through SHapley Additive exPlanations (SHAP) and attention-based analyses. Each case study adapts this structure to its domain while maintaining methodological coherence. In cardiovascular and multimorbidity settings, ensemble models integrating clinical, functional, and psychosocial indicators achieved up to 30\% improvement in discrimination compared with traditional Cox and logistic regression baselines. In addition, they highlighted the prognostic value of frailty, anxiety, and depression. In the eating-disorder study, resilience and quality-of-life measures emerged as strong determinants of one-year outcomes, confirming that self-perception and adaptability are central to recovery prediction. The creativity assessment demonstrated that combining image and text embeddings provides complementary information: textual inputs enhanced sensitivity to originality, whereas visual features better captured elaboration and complexity. Across all four domains, the integration of explainability within the modelling process produced systems that balance predictive precision with interpretability, enabling domain experts to trace and validate decision mechanisms. This thesis advances the methodological foundation of human-centred artificial intelligence by demonstrating that explicit incorporation of explainability enhances not only trust but also model performance. The resulting approach contributes to the development of transparent, adaptive, and ethically grounded predictive systems applicable to complex clinical, psychological, and cognitive contexts.en
dc.identifier.urihttps://hdl.handle.net/20.500.14454/6530
dc.language.isoeng
dc.publisherUniversidad de Deusto
dc.subjectMatemáticas
dc.subjectCiencia de los ordenadores
dc.subjectInteligencia artificial
dc.subjectCiencias Tecnológicas
dc.subjectTecnología de los ordenadores
dc.subjectDispositivos de transmisión de datos
dc.titleHuman-centred machine learning for health and cognitive modellingen
dc.typedoctoral thesis
Archivos
Bloque original
Mostrando 1 - 1 de 1
Miniatura
Nombre:
Pikatza_human_2026.pdf
Tamaño:
2.78 MB
Formato:
Adobe Portable Document Format
Descargar
Colecciones
Tesis doctorales