Big Data Meets Healthcare: The case for comparability and consistency

Transcription

1 Big Data Meets Healthcare: The case for comparability and consistency Christopher G. Chute, MD DrPH Professor Biomedical Informatics Vice-Chair, Data Governance Mayo Clinic Chair, International Classification of Disease Revision, WHO Member, HIT Standards Committee, HHS/ONC (USA) Chair, ISO Technical Committee 215 on Health Informatics HISA Big Data: 2013; Melbourne 18 April, 2013

2 Declarations No real or apparent financial conflicts of interest All products are open-source Comments represent beliefs of the author He has an odd sense of humor

3 Health Care Is An Information Intensive Industry Control of Health Care Costs... Improved Quality of Care... Improved Health Outcomes... Appropriate Use of Health Technology... Compassionate Resource Management depend upon information Ultimately Patient Data 3

4 Biocomputing Unimaginable potential, next 50 years Moore s law still reigns fold increase in computing power / 50 yrs World-class Supercomputers of 1990 Game platforms and cell phones of 2012 (GPUs) Extensions to networks, memory, & storage 110 baud to 100Gb backbone (10 9 / 50 yrs) Vacuum tubes to 100Gb RAM (10 11 / 50 yrs) Paper-tape to Pb drives (10 15 / 50 yrs) Cloudy computing (10 3 / 3 yrs) Synergistic Summary / 50 yrs 4

5 Big Science Collaboration and Semantics Communication of findings and results Human publication Sharing of data resources as building blocks Foundation for incremental, big-science Harnessing computing requires formalization Data format and structures Discrete terms, vocabulary, ontology Non-ambiguous concepts, non-overlapping terms Information models and problem architectures Standards, conventions, and shared context 5

6 Some Fashionable Biology Databases This Week 6

7 Some Fashionable Clinical Databases This Week 7

8 The Continuum Of Biomedical Informatics Bioinformatics meets Medical Informatics Biology Medicine Chasm of Semantic Despair 8

9 Mayo Clinic College of Medicine

10 Origins of Big Science Astronomy 10

11 Sloan Digital Sky Survey III DR9 Images Spectra Object catalog Metadata Total area of imaging 31,637 square degrees Image field size 1361x2048 pixels Number fields 938,046 (excluding supernovae runs) Catalog objects 1,231,051,050 Number of unique, primary sources Total 469,053,874 Stars 260,562,744 Galaxies 208,478,448 Unknown 12,682 11

12 Rutherford Table-top Experiment Mayo Clinic College of Medicine

13 That Higgs Boson 600 institutions 10,000 scientists 800 trillion collisions 200PB of data = bytes of data Boarding on an astronomical number in its own right! $13.25B USD Mayo Clinic College of Medicine

14 Dimensionality of Higgs Big Data Mass Direction Energy Mayo Clinic College of Medicine

15 Dimensionality of Big Data Broad Small amounts of data; Huge number observations National Claims data Deep Large amounts of data; Few observation NGS Complete Genome Rich Broad and Deep Clinical Phenotyping data (EMRs) Labs, Vitals, Exam, Waveform, Images, Omics, Social, environmental, diet, Mayo Clinic College of Medicine

16 Comparable and Consistent Data Inferencing from data to information requires sorting information into categories Statistical bins Machine learning features Accurate and reproducible categorization depends upon semantic consistency Semantic consistency is the vocabulary problem Almost always manifest as the value set problem 16

17 From Practice-based Evidence to Evidence-based Practice Data Patient Encounters Clinical Databases Registries et al. Shared Semantics Standards Terminologies & Data Models Inference Medical Knowledge Decision support Expert Systems Clinical Guidelines Knowledge Management Foundations for Learning Health System 17

18 World Health Organization; ICD-11 18

19 World Health Organization; ICD-11 19

20 The nomenclature is of as much importance in this department of inquiry, as weights and measures in the physical sciences, and should be settled without delay. First Annual Report of the Registrar-General of Births, Deaths, and Marriages in England. London: 1839 p. 99. Weights and Measures William Farr 20

21 Carolus Linnaeus Carl von Linné Genera Morborum (1763) Flawed Information Model Underscored Content Difficulty Pathophysiology vs Manifestation e.g. Rabies as psychiatric disease Lacked the Germ Theory of Disease Was not incorporated into an information model 21

22 The Genomic Era The genomic transformation of medicine far exceeds the introduction of antibiotics and aseptic surgery The binding of genomic biology and clinical medicine will accelerate The implications for shared semantics across the basic science and clinical communities are unprecedented 22

23 The Challenge Most clinical data in the United States (world) is heterogeneous non-standard Within Institutions Between Institutions Meaningful Use is mitigating, but has not yet solved the problem Achieving standardization in Meaningful Use is sometimes minimized

24 ACO Operations and Success Learning Health System Commitment to discovering best practice Critical examination of practice ACO rationale Align incentives for reimbursement Practice integration, communication, exchnage Quality measurement, ultimate improvement Best Practice Discovery, monitor adherence Improve health, wellness, disease management Reduce overall costs 24

25 What Does It Take To Become Learning Health System Culture, Culture, and Culture Data issues Comparable and consistent data representation Comprehensive and longitudinal EMR/EHR Data quality, reliability, and completeness Curation, monitoring, management Near real-time warehouse Analytic capacity First order Decision Support implementation 25

26 What Does It Take To Become Learning Health System Culture, Culture, and Culture Data issues Comparable and consistent data representation Comprehensive and longitudinal EMR/EHR Data quality, reliability, and completeness Curation, monitoring, management Near real-time warehouse Analytic capacity 26

27 Meaningful Use 13 January, 2010 Reporting Requirements 169pp Interoperability Standards 35pp 27

28 Meaningful Use (in the USA) and Comparable Data Phase I an into Phase II dawn of effective standards Phase III deployment of useful comparability Biannual follow-ons ratcheting consistency Learning Health System cannot be supported by Meaningful Use standards alone Not yet sufficient granularity/specificity 28

29 Counterfactual: Mining Gold from Dirty Big Data Mar 2013 Mayo Clinic College of Medicine

30 Does Big Data Change Criteria for Scientific Evidence? that society will need to shed some of its obsession for causality in exchange for simple correlations: not knowing why by only what. This overturns centuries of established practices and challenges our most basic understanding of how to make decisions and comprehend reality. Mayer-Schönberger, V. and K. Cukier, Big data : a revolution that will transform how we live, work, and think. 2013, Boston: Houghton Mifflin Harcourt. 242 p. 30

31 Actionable Knowledge: More than just Google Search Well known spurious association problem Reproducibility Power Issues [Have I seen a patient like that ] Single drug vs. single side effect Stratify across cells by: Age and Sex Co-morbidity Lab values (nomal vs non-normal) Image findings Waveform findings Genetic profile (and other omics ) 31

32 Clinical Terminology and Classification SNOMED-CT evolving into utility Not a human entry terminology Increasingly comprehensive and rational Platform for data comparability Meaningful Use Phase II requirement (problem list) Consolidated CDA Vast improvement over CCD Specifies Value Sets NLM National Value-set Center 32

33 How Well do Clinical Classifications Work? Study of Coding efficacy to measure content capture Chute et al., JAMIA 1996;3(3): Text reports from four academic medical centers Data parsed into 3,061 clinical observations Coded twice by national panel of experts Updated re-coding using most current versions of ICD-10-CM and ICD-9-CM Coding over-read by CDC/NCHS National Center for Health Statistics

34 Scoring Example ICD-9-CM 2 <Primary Diagnosis>:melanoma 0 <Extent>:Clark s level 2 0 <Quantitative>:0.84 mm [depth of invasion] 2 <Histology>:superficial spreading 1 <Anatomy>:thigh 0 <Topology>:left Code: Malignant melanoma of skin, lower limb Code: M8743/3 Superficial spreading melanoma

35 Average value of coverage scores Biomedical Informatics Comparison of Clinical Content Coverage between ICD-9-CM and ICD-10-CM ICD-9-CM ICD-10CM Diagnosis Findings Modifier Other Overall

36 Some Observations on American Clinical Modifications of the ICD ICD-9-CM Diagnoses codes Roughly 13,000 codes Injury codes: 15% Disease codes: 8,500 ICD-10-CM Diagnoses codes Roughly 68,000 codes Injury codes: 60% Disease codes: 18,000 Permuted by: Right, Left, Both, Unspecified First Episode, subsequent Adds up to 6 codes per disease Done for a bit more than half disease codes 36

37 International Classification of Disease ICD11 Use Cases Scientific consensus of clinical phenotype Public Health Surveillance Mortality Public Health Morbidity Clinical data aggregation Metrics of clinical activity Quality management Patient Safety Financial administration Case mix Resource allocation 37

38 Traditional Hierarchical System ICD-10 and family 38

39 Addition of structured attributes to concepts Concept name Definition Language translations Preferred string Language translations Synonyms Language translations Index Terms 39

40 THE CONTENT MODEL Any Category in ICD is represented by: 1. ICD Concept Title 1.1. Fully Specified Name 1.2 Preferred Name 1.3 Synonyms 2. Classification Properties 2.1. Parents 2.2 Type 2.3. Use and Linearization(s) 3. Textual Definition(s) 4. Terms 4.1. Base Index Terms 4.2. Inclusion Terms 4.3. Exclusions 5. Body System/Structure 5.1. Body System(s) 5.2. Body Part(s) [Anatomical Site(s)] 5.3. Morphological Properties 6. Manifestation Properties 6.1. Signs & Symptoms 6.2. Investigation findings 7. Causal Properties 7.1. Etiology Type 7.2. Causal Properties - Agents 7.3. Causal Properties - Causal Mechanisms 7.4. Genomic Linkages 7.5. Risk Factors 8. Temporal Properties 8.1. Age of Occurrence & Occurrence Frequency 8.2. Development Course/Stage 9. Severity of Subtypes Properties 10. Functioning Properties Impact on Activities and Participation Contextual factors Body functions 11. Specific Condition Properties 11.1 Biological Sex Life-Cycle Properties 12. Treatment Properties 13. Diagnostic Criteria

41 Relationships Logical Definitions Etiology Genomic Location Laterality Histology Severity Acuity Addition of semantic arcs - Ontology 41

42 Serialization of the cloud Algorithmic Derivation 42

43 Linear views may serve multiple use-cases Morbidity, Mortality, Quality, 43

44 Relationship with IHTSDO SNOMED content IHT (SNOMED) will require high-level nodes that aggregate more granular data Use-cases include mutually exclusive, exhaustive, Sounds a lot like ICD ICD-11 will require lower level terminology for value sets which populate content model Detailed terminological underpinning Sounds a lot like SNOMED Memorandum of Agreement July 2010! WHO right to use for authoring and interpretation 44

45 ICD-11 Potential Future States SNOMED Ghost ICD Ghost SNOMED 45

46 ICD-11 Joint ICD-IHTSDO Effort Alternate Future SNOMED 46

47 SHARP Area 4: Secondary Use of EHR Data Agilex Technologies Harvard Univ. CDISC (Clinical Data Interchange Intermountain Healthcare Standards Consortium) Centerphase Solutions Deloitte Group Health, Seattle IBM Watson Research Labs University of Utah University of Pittsburgh Mayo Clinic Mirth Corporation, Inc. MIT MITRE Corp. Regenstrief Institute, Inc. SUNY University of Colorado

48 Themes & Projects

49 Mission To enable the use of EHR data for secondary purposes, such as clinical research and public health. Leverage health informatics to: generate new knowledge improve care address population needs To support the community of EHR data consumers by developing: open-source tools services scalable software

50 Modes of Normalization Generally true for both structured and unstructured data Syntactic transformation Clean up message formats HL7 V2, CCD/CDA, tabular data, etc Emulate Regenstrief HOSS pipeline Semantic normalization Typically vocabulary mapping

51 Clinical Data Normalization Data Normalization Comparable and consistent data is foundational to secondary use Clinical Data Models Clinical Element Models (CEMS) Basis for retaining computable meaning when data is exchanged between heterogeneous computer systems. Basis for shared computable meaning when clinical data is referenced in decision support logic.

52

53 Data Element Harmonization Stan Huff CIMI Clinical Information Model Initiative NHS Clinical Statement CEN TC251/OpenEHR Archetypes HL7 Templates ISO TC215 Detailed Clinical Models CDISC Common Clinical Elements Intermountain/GE CEMs

54 Clinical Data Normalization Dr. Huff on Data Normalization Stanley M. Huff, M.D.; SHARPn Co-Principal Investigator; Professor (Clinical) - Biomedical Informatics at University of Utah - College of Medicine and Chief Medical Informatics Officer Intermountain Healthcare. Dr. Huff discusses the need to provide patient care at the lowest cost with advanced decision support requires structured and coded data.

55 That Semantic Bit Canonical semantics reduce to Value-set Binding to CEM objects Value-sets should obviously be drawn from standard vocabularies SNOMED-CT and ICD LOINC RxNorm But others required: HUGO, GO, HL7

56 Normalization Pipelines Input heterogeneous clinical data HL7, CDA/CCD, structured feeds Output Normalized CEMs Create logical structures within UIMA CAS Serialize to a persistence layer SQL, RDF, PCAST like, XML Robust Prototypes now posted Early version production Q3 2013

57 NLP Deliverables and Tools ctakes Releases Smoking Status Classifier Medication Annotator Integrated ctakes(ictakes) an effort to improve the usability of ctakes for end users NLP evaluation workbench ctakes Side Effects module Modules for relation extraction the dissemination of an NLP algorithm requires performance benchmarking. The evaluation workbench allows NLP investigators and developers to compare and evaluate various NLP algorithms. SHARPn NLP Common Type SHARPn NLP Common Type System is an effort for defining common NLP types used in SHARPn; UIMA framework.

58 High-Throughput Phenotyping Phenotype - a set of patient characteristics : Diagnoses, Procedures Demographics Lab Values, Medications Phenotyping overload of terms Originally for research cohorts from EMRs Obvious extension to clinical trial eligibility Quality metric Numerators and denominators Clinical decision support - Trigger criteria

59 Where is This Going? Biomedical practice and research are data, information, and knowledge intensive Comparable and consistent data representation are pre-requisite for efficient clinical analytics Clinical Data is Broad, Deep, and Complex Inferencing clinical big data is not Google Data standards are needed to support comparable and consistent information 59