Publicação
Domain Adaptation for Human Activity Recognition: A Case Study in ALS
| Resumo: | Amyotrophic Lateral Sclerosis (ALS) is an incurable neurodegenerative disease that leads to severe motor impairment. This disease presents a heterogeneous progression among the patients, hindering the prognostic prediction. Thus, automatically recognizing the activities performed by the patients could provide relevant information for the prognosis. Nonetheless, collecting enough labeled data from the patients to train the activity classifiers reliably is difficult. In order to surpass such data constraints, this research focused on leveraging motion data from healthy subjects to enhance the performance of activity classifiers for ALS patients using Domain Adaptation (DA) methods. These methods were first developed and tested in public datasets and then used to adapt models from healthy subjects to ALS patients. The importance of domain knowledge in human activity recognition and DA was also addressed by using handcrafted features as an auxiliary input to deep learning models. Regarding the adaptation between public datasets, domain knowledge proved to be more useful than DA methods when few samples from the target domain were available, improving the deep learning baseline results, on average, by 10.56%. In opposition, DA methods were more effective in adapting models trained on healthy subjects to ALS patients. Therefore, this study points towards the usefulness of leveraging data from healthy subjects to enhance the performance of activity classifiers for ALS patients through the use of DA strategies. These findings bring activity-related insights for the remote monitoring of ALS patients, which may be valuable for prognostic prediction and clinical decision-making. |
|---|---|
| Autores principais: | Rebelo, Joana Sofia Gonçalves |
| Assunto: | Human Activity Recognition Amyotrophic Lateral Sclerosis Domain Adaptation Transfer Learning |
| Ano: | 2022 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Amyotrophic Lateral Sclerosis (ALS) is an incurable neurodegenerative disease that leads to severe motor impairment. This disease presents a heterogeneous progression among the patients, hindering the prognostic prediction. Thus, automatically recognizing the activities performed by the patients could provide relevant information for the prognosis. Nonetheless, collecting enough labeled data from the patients to train the activity classifiers reliably is difficult. In order to surpass such data constraints, this research focused on leveraging motion data from healthy subjects to enhance the performance of activity classifiers for ALS patients using Domain Adaptation (DA) methods. These methods were first developed and tested in public datasets and then used to adapt models from healthy subjects to ALS patients. The importance of domain knowledge in human activity recognition and DA was also addressed by using handcrafted features as an auxiliary input to deep learning models. Regarding the adaptation between public datasets, domain knowledge proved to be more useful than DA methods when few samples from the target domain were available, improving the deep learning baseline results, on average, by 10.56%. In opposition, DA methods were more effective in adapting models trained on healthy subjects to ALS patients. Therefore, this study points towards the usefulness of leveraging data from healthy subjects to enhance the performance of activity classifiers for ALS patients through the use of DA strategies. These findings bring activity-related insights for the remote monitoring of ALS patients, which may be valuable for prognostic prediction and clinical decision-making. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.