서브메뉴
검색
Harnessing Data and Deep Learning for Stroke Rehabilitation
Harnessing Data and Deep Learning for Stroke Rehabilitation
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260331111230
- ISBN
- 9798381730739
- DDC
- 004
- 저자명
- Kaku, Aakash.
- 서명/저자
- Harnessing Data and Deep Learning for Stroke Rehabilitation
- 발행사항
- [Sl] : New York University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 178 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 85-08, Section: B.
- 주기사항
- Advisor: Fernandez-Granda, Carlos;Razavian, Narges.
- 학위논문주기
- Thesis (Ph.D.)--New York University, 2024.
- 초록/해제
- 요약Stroke is a leading cause of long-term disability, necessitating rehabilitation for improved function and quality of life. However, quantifying rehabilitation doses traditionally lacks objectivity, often proving time-consuming and costly. Thus, there's a pressing need for more efficient methods to automatically quantify rehabilitation doses and make therapy objective.In this study, we explore innovative methods to automatically quantify rehabilitation doses and enhance therapy objectivity. We introduce a novel approach to identify functional primitives, basic activity components, from inertial measurement unit (IMU) data. Our model, equipped with a sensor embedding module and adaptive normalization techniques, outperforms traditional methods, demonstrating its superior ability to quantify rehabilitation training doses. These adaptive normalization techniques also enhance model robustness to distributional shifts, as observed in the CIFAR-10 corrupted dataset.Conventional action recognition primarily focuses on recognizing broad actions, limiting applications requiring high temporal precision for elemental movements. To address this limitation, we present the StrokeRehab dataset, a substantial benchmark containing video and kinematic data from stroke-affected individuals and healthy subjects engaged in everyday activities. This dataset's inclusion of data from both healthy and impaired subjects poses realistic distributional challenges.While state-of-the-art action segmentation models yield noisy predictions on the StrokeRehab dataset, we propose a novel high-resolution action identification approach inspired by speech recognition techniques. This method employs a sequence-to-sequence model to directly predict action sequences, improving accuracy within the StrokeRehab dataset.Furthermore, we leverage data from healthy subjects to develop a data-driven tool for quantifying impairment in stroke patients. By training the model on healthy subject data and using model confidence as an indicator of movement abnormalities in stroke patients, we establish a strong correlation between reduced model confidence and stroke patient impairment.These findings highlight the potential of deep learning and data-driven methods to enhance stroke rehabilitation outcomes. Future research should focus on developing and integrating deep learning-based tools into clinical practice to optimize stroke rehabilitation further.
- 일반주제명
- Computer science
- 일반주제명
- Statistics
- 키워드
- Deep learning
- 키워드
- Motor impairment
- 키워드
- Rehabilitation
- 키워드
- Stroke
- 기타저자
- New York University Center for Data Science
- 기본자료저록
- Dissertations Abstracts International. 85-08B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260324s2024 us c eng d■001000017410831
■00520260331111230
■006m o d
■007cr#unu||||||||
■020 ▼a9798381730739
■035 ▼a(MiAaPQ)AAI30811832
■040 ▼aMiAaPQ▼cMiAaPQ
■082 ▼a004
■1001 ▼aKaku, Aakash.
■24510▼aHarnessing Data and Deep Learning for Stroke Rehabilitation
■260 ▼a[Sl]▼bNew York University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a178 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 85-08, Section: B.
■500 ▼aAdvisor: Fernandez-Granda, Carlos;Razavian, Narges.
■5021 ▼aThesis (Ph.D.)--New York University, 2024.
■520 ▼aStroke is a leading cause of long-term disability, necessitating rehabilitation for improved function and quality of life. However, quantifying rehabilitation doses traditionally lacks objectivity, often proving time-consuming and costly. Thus, there's a pressing need for more efficient methods to automatically quantify rehabilitation doses and make therapy objective.In this study, we explore innovative methods to automatically quantify rehabilitation doses and enhance therapy objectivity. We introduce a novel approach to identify functional primitives, basic activity components, from inertial measurement unit (IMU) data. Our model, equipped with a sensor embedding module and adaptive normalization techniques, outperforms traditional methods, demonstrating its superior ability to quantify rehabilitation training doses. These adaptive normalization techniques also enhance model robustness to distributional shifts, as observed in the CIFAR-10 corrupted dataset.Conventional action recognition primarily focuses on recognizing broad actions, limiting applications requiring high temporal precision for elemental movements. To address this limitation, we present the StrokeRehab dataset, a substantial benchmark containing video and kinematic data from stroke-affected individuals and healthy subjects engaged in everyday activities. This dataset's inclusion of data from both healthy and impaired subjects poses realistic distributional challenges.While state-of-the-art action segmentation models yield noisy predictions on the StrokeRehab dataset, we propose a novel high-resolution action identification approach inspired by speech recognition techniques. This method employs a sequence-to-sequence model to directly predict action sequences, improving accuracy within the StrokeRehab dataset.Furthermore, we leverage data from healthy subjects to develop a data-driven tool for quantifying impairment in stroke patients. By training the model on healthy subject data and using model confidence as an indicator of movement abnormalities in stroke patients, we establish a strong correlation between reduced model confidence and stroke patient impairment.These findings highlight the potential of deep learning and data-driven methods to enhance stroke rehabilitation outcomes. Future research should focus on developing and integrating deep learning-based tools into clinical practice to optimize stroke rehabilitation further.
■590 ▼aSchool code: 0146.
■650 4▼aComputer science
■650 4▼aStatistics
■653 ▼aDeep learning
■653 ▼aDeep neural networks
■653 ▼aInertial measurement unit
■653 ▼aMotor impairment
■653 ▼aRehabilitation
■653 ▼aStroke
■690 ▼a0984
■690 ▼a0800
■690 ▼a0463
■71020▼aNew York University▼bCenter for Data Science.
■7730 ▼tDissertations Abstracts International▼g85-08B.
■790 ▼a0146
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17410831▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


