Use of Attention Maps to Enrich Discriminability in Deep Learning Prediction Models Using Longitudinal Data from Electronic Health Records
Carrasco-Ribelles, Lucía Amalia (Institut Universitari d'Investigació en Atenció Primària Jordi Gol)
Cabrera-Bean, Margarita (Universitat Politècnica de Catalunya. Departament de Teoria del Senyal i Comunicacions)
Llanes-Jurado, Jose (Universitat Politècnica de València. Instituto Universitario de Investigación en Tecnología Centrada en el Ser Humano.)
Violán, Concepció (Universitat Autònoma de Barcelona. Departament de Medicina)
Universitat Autònoma de Barcelona. Departament de Medicina

Data:	2025
Resum:	Featured Application: A better discrimination in a prediction model does not imply a better interpretability. In healthcare, where transparency is crucial, both discriminability and interpretability should be checked before stating that a new model is better. Background: In predictive modelling, particularly in fields such as healthcare, the importance of understanding the model's behaviour rivals, if not surpasses, that of discriminability. To this end, attention mechanisms have been included in deep learning models for years. However, when comparing different models, the one with the best discriminability is usually chosen without considering the clinical plausibility of their predictions. Objective: In this work several attention-based deep learning architectures with increasing degrees of complexity were designed and compared aiming to study the balance between discriminability and plausibility with architecture complexity when working with longitudinal data from Electronic Health Records (EHRs). Methods: We developed four deep learning-based architectures with attention mechanisms that were progressively more complex to handle longitudinal data from EHRs. We evaluated their discriminability and resulting attention maps and compared them amongst architectures and different input processing approaches. We trained them on 10 years of data from EHRs from Catalonia (Spain) and evaluated them using a 5-fold cross-validation to predict 1-year all-cause mortality in a subsample of 500,000 people over 65 years of age. Results: Generally, the simplest architectures led to the best overall discriminability, slightly decreasing with complexity by up to 8. 7%. However, the attention maps resulting from the simpler architectures were less informative and less clinically plausible compared to those from more complex architectures. Moreover, the latter could give attention weights both in the time and feature domains. Conclusions: Our results suggest that discriminability and more informative and clinically plausible attention maps do not always go together. Given the preferences within the healthcare field for enhanced explainability, establishing a balance with discriminability is imperative.
Ajuts:	Instituto de Salud Carlos III PI19/00535 Agencia Estatal de Investigación PID2022-138648OB-I00
Drets:	Aquest document està subjecte a una llicència d'ús Creative Commons. Es permet la reproducció total o parcial, la distribució, la comunicació pública de l'obra i la creació d'obres derivades, fins i tot amb finalitats comercials, sempre i quan es reconegui l'autoria de l'obra original.
Llengua:	Anglès
Document:	Article ; recerca ; Versió publicada
Matèria:	Attention mechanism ; Clinical plausibility ; Discriminability ; Electronic health record ; Recurrent neural network ; Longitudinal data
Publicat a:	Applied sciences (Basel), Vol. 15 Núm. 1 (january 2025) , p. 146, ISSN 2076-3417

DOI: 10.3390/app15010146

14 p, 796.1 KB

El registre apareix a les col·leccions:
Documents de recerca > Documents dels grups de recerca de la UAB > Centres i grups de recerca (producció científica) > Ciències de la salut i biociències > Institut d'Investigació en Ciencies de la Salut Germans Trias i Pujol (IGTP)
Articles > Articles de recerca
Articles > Articles publicats