Albert Lai - Information Fusion Poster

download Albert Lai - Information Fusion Poster

of 1

Transcript of Albert Lai - Information Fusion Poster

  • 8/7/2019 Albert Lai - Information Fusion Poster

    1/1

    Abstract

    An Information Fusion Pipeline for Longitudinal Health RecordsAlbert M. Lai,1,2 Preethi Raghavan,2 and Chunhua Weng3

    1Departments of Biomedical Informatics and2Computer Science and Engineering, The Ohio State University, Columbus, OH3Department of Biomedical Informatics, Columbia University, New York, NY

    Conclusion

    http://bmi.osu.eduDEPARTMENT OF BIOMEDICAL INFORMATICS

    BMI

    Existing EHRs often store unstructured clinical narrativesand structured data in an uncoordinated manner, which

    poses complex challenges for extracting a longitudinal viewof a patient across clinical encounters. The various clinical

    data sources may contain redundant or inconsistent infor-

    mation, and hence involve a large number of temporally orsemantically overlapping or contradictory events, compli-

    cating automatic generation of a coordinated Longitudinal

    Health Record (LHR). We propose an information fusionpipeline, which leverages both semantic and temporal infor-

    mation in conjunction with a machine learning approach to

    integrate structured and unstructured clinical information, togenerate the foundations of an LHR. This LHR offers the

    potential to improve the efficiency and accuracy of decisionsupport tools.

    In the proposed research, we will characterize the overlapof clinical information across various data sources at two

    large integrated medical centers. We will develop algo-

    rithms for extracting temporal expressions and relationsfrom clinical narratives obtained from the institutional infor-

    mation warehouses. Utilizing the temporal information ex-tracted from these narratives, we will develop information

    fusion algorithms and combine the temporal information

    with structured data from EHR systems, leveraging both se-mantic and temporal information, developing a longitudinal

    health record.

    We will perform temporal relation extraction using a ma-

    chine learning approach, mapping medical events and tem-

    poral expressions.

    Temporal interval overlap and semantic alignment of onto-logically coded medical events that can be used in conjunc-

    tion with heuristics or as features in a machine learning ap-

    proach to resolving intra-narrative event coreferences.

    We have presented an information fusion approach toLHRs. Successfully creating an LHR using an information

    fusion approach would benefit multiple clinical scenarios.

    For instance, it could be used to improve the automaticidentification of cohorts of patients who satisfy specific clini-

    cal trial eligibility criteria and are hence eligible for a particu-lar clinical trial. The ability to utilize an automated, data-

    driven approach to cohort identification would enable the

    pace at which clinical research progresses. However, with-out a coherent LHR created through information fusion, au-

    tomated approaches are likely to be fraught with errors due

    to poor information consistency when extracting clinical in-formation from various clinical data sources.

    Acknowledgements

    This study is partially funded by NLM R01 LM009886 andNCRR UL1-RR025755.

    We will also perform inter-narrative co-reference resolutionby integrating the list of events generated from clinical nar-

    ratives with various other sources of information. This in-

    cludes similar lists from other narratives, such as relevantdata from structured narratives (lab values), tagged con-

    cepts (like UMLS), entities and ontology mappings obtainedfrom MetaMap. We propose to combine all events gener-

    ated from clinical narratives together with the aforemen-

    tioned structured data and perform information fusion.

    This includes performing both inter-narrative and multi-

    source co-reference resolution, leveraging semantic andtemporal information. In addition, we intend to also leverage

    the metadata (e.g. medical record number, admission

    number, encounter number, admission date, encounterdate, date note was written) associated with clinical narra-

    tives to improve the quality of the information fusion.Once the inter-narrative coreferences are resolved resulting

    in a distinct set of chronologically ordered events, we have

    a longitudinal model of the clinical record for a patient.

    Name: John Doe Admit Date: 4/3/2010

    MRN: #AB123 Discharge Date: 4/13/2010Attending Physician: Dr. Mike Smith

    HISTORY OF PRESENT ILLNESS

    Mr. Doe came in with a problem of a leaking G-tube.

    The G-tube had been in place since 2008...HOSPITAL COURSE

    The G-tube was surgically removed at laparotomywith closure of the stomach. After approximately 2

    days of nothing by mouth status, we started bringing

    him around. It took some time for his diet to be appre-ciated and he was returned to an ECF (Extended care

    facility) with a regular diet and wound care.

    Sample Discharge Summary

    Information

    Fusion

    UMLS &Knowledge

    Base

    Longitudinal

    Health

    Record

    StructuredData

    (Labs, ADT)

    Clinical

    Narratives

    Clinical

    Narrative1

    Clinical

    Narrativen

    NLPextracted

    concepts

    Temporal

    Relations

    Clinical

    Events

    NLP

    extracted

    concepts

    Temporal

    Relations

    Clinical

    Events

    Events with TemporalConstraints & Ontologically

    Coded Entities

    Methods