Publicação
Explainable Artificial Intelligence in Time Series Applied to Physiological Data
| Resumo: | Over the past few decades, we have seen Artificial Intelligence (AI) agents move from controlled laboratory experiments to becoming an integral part of our daily lives. In our evolving society, algorithms govern many aspects of our daily routines, shaping our choices, decisions, perceptions, interactions, and aspirations. To support the impactful, responsible, and legal implementation of these systems, we must understand how they make decisions and why a given decision was made. Explainable Artificial Intelligence concerns designing and developing methods that explain how machine learning models arrive at their decisions. The sequential nature of time series data, coupled with its high dimensionality, poses distinct challenges compared to other data types like images, tables, and text. This underscores the necessity for research in developing methods that enhance explainability and interpretability for time series classification. As we navigate the complexities of time series data, the importance of this data type in the realm of biosignals analysis becomes evident. Biosignals, such as heart rate, human motion, breathing, and muscular activity, stand out as classic examples of time series datasets in electrophysiology. They are essential indicators of physiological health and play a pivotal role in diagnostic and prognostic decisions in healthcare. This thesis introduces innovative methods to enhance the explainability of time se- ries classification applied to biosignals. Our advancements are structured into three complementary research lines: (a) method-oriented, (b) application-oriented, and (c) human-oriented. In the method-oriented line, we introduce new techniques for explain- ing models tailored for electrocardiogram interpretation and processing multimodal biosignal datasets. Within the application-oriented approach, we used feature-based explanations to study human motion, demonstrating their value as an instrument for vali- dating scientific discoveries. We centered this analysis around a use case that examines the relationship between physiological tremors during slow goal-directed movements and the level of movement expertise. Lastly, in the human-oriented research, we propose a protocol for evaluating human-AI collaboration in clinical electrocardiogram diagnosis, highlighted by a user study with cardiologists. |
|---|---|
| Autores principais: | Folgado, Duarte Miguel dos Santos Nunes |
| Assunto: | Time Series Explainability Biosignals Electrocardiogram Human Motion |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Over the past few decades, we have seen Artificial Intelligence (AI) agents move from controlled laboratory experiments to becoming an integral part of our daily lives. In our evolving society, algorithms govern many aspects of our daily routines, shaping our choices, decisions, perceptions, interactions, and aspirations. To support the impactful, responsible, and legal implementation of these systems, we must understand how they make decisions and why a given decision was made. Explainable Artificial Intelligence concerns designing and developing methods that explain how machine learning models arrive at their decisions. The sequential nature of time series data, coupled with its high dimensionality, poses distinct challenges compared to other data types like images, tables, and text. This underscores the necessity for research in developing methods that enhance explainability and interpretability for time series classification. As we navigate the complexities of time series data, the importance of this data type in the realm of biosignals analysis becomes evident. Biosignals, such as heart rate, human motion, breathing, and muscular activity, stand out as classic examples of time series datasets in electrophysiology. They are essential indicators of physiological health and play a pivotal role in diagnostic and prognostic decisions in healthcare. This thesis introduces innovative methods to enhance the explainability of time se- ries classification applied to biosignals. Our advancements are structured into three complementary research lines: (a) method-oriented, (b) application-oriented, and (c) human-oriented. In the method-oriented line, we introduce new techniques for explain- ing models tailored for electrocardiogram interpretation and processing multimodal biosignal datasets. Within the application-oriented approach, we used feature-based explanations to study human motion, demonstrating their value as an instrument for vali- dating scientific discoveries. We centered this analysis around a use case that examines the relationship between physiological tremors during slow goal-directed movements and the level of movement expertise. Lastly, in the human-oriented research, we propose a protocol for evaluating human-AI collaboration in clinical electrocardiogram diagnosis, highlighted by a user study with cardiologists. |
|---|