서브메뉴
검색
Domain-Guided Representations are Superior for Machine Learning in Healthcare
Domain-Guided Representations are Superior for Machine Learning in Healthcare
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152748
- ISBN
- 9798342108980
- DDC
- 616.94
- 서명/저자
- Domain-Guided Representations are Superior for Machine Learning in Healthcare
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 189 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-04, Section: B.
- 주기사항
- Advisor: Brunskill, Emma;Shah, Nigam.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약Machine learning to support decision making in healthcare typically assumes a sufficient representation of the patient's history for learning and inference, such that all information required to perform accurate inference is present in the materialized patient timeline. We propose three principles for designing health data representations that are more expressive and comprehensive compared to traditional representation approaches, in order to enable clinical decision support tasks. Concretely, we propose that patient timeline representations (1) explicitly model clinicians' observation and intervention processes; (2) enable integration of unstructured data and structured data by representing all data as text; and (3) leverage this ability by including clinical notes. We apply these principles to several clinical challenges and provide evidence supporting the superiority of domain-guided representations for machine learning in healthcare. We show that reinforcement learning algorithms built on data models which explicitly include clinician observation and intervention processes yield context-aware policies that perform better in the realistic setting of patient observation costs compared to algorithms which do not model these processes. We show that using text as a unifying representation for both structured and unstructured data can enable large language models to follow electronic health record-based instructions, potentially streamlining common clinical workflows, and that incorporating clinical notes provides a key benefit for propensity score models in causal inference with observational data. By improving downstream model performance, our proposed principles for representing patient data could help clinical machine learning researchers and practitioners increase both the effectiveness and efficiency of healthcare delivery, leading to decreased costs and improved patient outcomes.
- 일반주제명
- Sepsis
- 일반주제명
- Planning
- 일반주제명
- Decision making
- 일반주제명
- Medical research
- 일반주제명
- Taxonomy
- 일반주제명
- Large language models
- 일반주제명
- Filtering systems
- 일반주제명
- Medicine
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-04B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017163745
■00520250211152748
■006m o d
■007cr#unu||||||||
■020 ▼a9798342108980
■035 ▼a(MiAaPQ)AAI31520317
■035 ▼a(MiAaPQ)Stanfordtn305mf5188
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a616.94
■1001 ▼aFleming, Scott Lanyon.
■24510▼aDomain-Guided Representations are Superior for Machine Learning in Healthcare
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a189 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-04, Section: B.
■500 ▼aAdvisor: Brunskill, Emma;Shah, Nigam.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aMachine learning to support decision making in healthcare typically assumes a sufficient representation of the patient's history for learning and inference, such that all information required to perform accurate inference is present in the materialized patient timeline. We propose three principles for designing health data representations that are more expressive and comprehensive compared to traditional representation approaches, in order to enable clinical decision support tasks. Concretely, we propose that patient timeline representations (1) explicitly model clinicians' observation and intervention processes; (2) enable integration of unstructured data and structured data by representing all data as text; and (3) leverage this ability by including clinical notes. We apply these principles to several clinical challenges and provide evidence supporting the superiority of domain-guided representations for machine learning in healthcare. We show that reinforcement learning algorithms built on data models which explicitly include clinician observation and intervention processes yield context-aware policies that perform better in the realistic setting of patient observation costs compared to algorithms which do not model these processes. We show that using text as a unifying representation for both structured and unstructured data can enable large language models to follow electronic health record-based instructions, potentially streamlining common clinical workflows, and that incorporating clinical notes provides a key benefit for propensity score models in causal inference with observational data. By improving downstream model performance, our proposed principles for representing patient data could help clinical machine learning researchers and practitioners increase both the effectiveness and efficiency of healthcare delivery, leading to decreased costs and improved patient outcomes.
■590 ▼aSchool code: 0212.
■650 4▼aSepsis
■650 4▼aPlanning
■650 4▼aDecision making
■650 4▼aMedical research
■650 4▼aTaxonomy
■650 4▼aLarge language models
■650 4▼aFiltering systems
■650 4▼aMedicine
■690 ▼a0800
■690 ▼a0564
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-04B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17163745▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
