the data and the problem - integrative biomedical imaging...

6
EMR Data mining Nigam Shah [email protected] 1 0 -1 1 0 -1 1 0 -1 1 0 -1 1 0 -1 thousands millions hundreds engineered The Data and the problem ß Individuals (N)à ß Features (p) à Procedures Devices Diseases Drugs Sequence, Expression (gene, protein), Metabolites … Claims Quantified self We do a lot, without knowing what works. 1. Clinical questions 2. Insights from the data 3. Predictive models Understanding the question What proportion of the population has familial hyperlipidemia? What are the subtypes of autism? Is there a relation between gut bacteria and depression? Which patients will be high-cost next year? Is mitral valve repair or replacement better for patients with prolapse? How does androgen deprivation increase dementia Risk? 5 tid cui str Note freq syn Medline freq % noun 2933 C0020255 hydrocephalus 29,634 NNS 19,541 64.61 42612 C0020255 hydrocephaly 113 NN 275 49.81 90773 C0020255 water on the brain 8 ROOT 1 50 Phenotyping Clinical studies

Upload: others

Post on 30-Jan-2021

1 views

Category:

Documents


0 download

TRANSCRIPT

  • EMRDatamining

    [email protected]

    1 0 -11 0 -1

    1 0 -11 0 -1

    1 0 -1

    thousands

    millions

    hundreds

    engineered

    TheDataandtheproblem

    ßIndividu

    als(N)à

    ß Features(p)à

    ProceduresDevicesDiseasesDrugs Sequence,Expression(gene,protein),Metabolites…

    Claims Quantifiedself

    Wedoalot,withoutknowingwhatworks.

    1. Clinicalquestions2. Insightsfromthedata3. Predictivemodels

    Understandingthequestion

    Whatproportionofthepopulationhasfamilialhyperlipidemia?

    Whatarethesubtypesofautism?

    Istherearelationbetweengutbacteriaanddepression?

    Whichpatientswillbehigh-costnextyear?

    Ismitralvalverepairorreplacementbetterforpatientswithprolapse?

    HowdoesandrogendeprivationincreasedementiaRisk?

    5

    90773 water*on*the*brain3107785 hydrocephalus*[disease/finding]42612 hydrocephaly2889104 hydrocephalus*(disorder)1936514 hydrocephalus,*unspecified2001177 hydrocephalus*nos

    2933 hydrocephalus*************************

    4010253C0020255

    hydrocephalus

    tid cui str Note+freq syn Medline+freq %+noun2933 C0020255 hydrocephalus 29,634 NNS 19,541 64.6142612 C0020255 hydrocephaly 113 NN 275 49.8190773 C0020255 water>on>the>brain 8 ROOT 1 50

    Phenotyping

    Clinicalstudies

  • Terms,Concepts

    Counts

    Variables

    Anyallergymed?

    Profilingriskfactorsforchronicuveitis PAST MEDICAL/SURGICAL HISTORY: Positive for atrial fibrillation. The patient had AVR 6 years ago. Peripheral arterial disease with hypertension, peripheral neuropathy, atherosclerosis, hemorrhoids, proctitis, CABG, and cholecystectomy.

    FAMILY HISTORY: Positive for atherosclerosis, hypertension, autoimmune diseases in the family.

    REVIEW OF SYSTEMS: Weight loss of 25 pounds within the last 6 months, shortness of breath, constipation, bleeding from hemorrhoids, increased frequency of urination, muscle aches, dizziness and faintness, focal weakness and numbness in both legs, knees and feet.

    LABORATORY DATA AND RADIOLOGICAL RESULTS: The patient had a chest x-ray, which showed cardiomegaly with atherosclerotic heart disease, pleural thickening and small pleural effusion, a left costophrenic angle which has not changed when compared to prior examination, COPD pattern. The patient also had a head CT, which showed atrophy with old ischemic changes. No acute intracranial findings.

    DISCHARGE DIAGNOSIS: Syncope.

    DISCHARGE MEDICATIONS: The patient was discharged on the following medications; Cardizem 90 mg p.o. thrice daily, digoxin 0.125 mg p.o. once daily, allopurinol 100 mg two times daily, Coumadin 4 mg p.o. q.h.s., and Remeron 15 mg p.o. q.h.s.

    past medical surgical history: positive for atrial fibrillation. patient avr years ago. peripheral arterial disease hypertension, peripheral neuropathy, atherosclerosis, hemorrhoids, proctitis, cabg, and cholecystectomy.

    family history: positive for atherosclerosis, hypertension, autoimmune diseases family.

    review of systems: weight loss pounds months, shortness of breath, constipation, bleeding hemorrhoids, increased frequency of urination, muscle aches, dizziness and faintness, focal weakness and numbness both legs, knees and feet.

    laboratory data and results: patient chest x-ray, which cardiomegaly atherosclerotic heart disease, pleural thickening and small pleural effusion, left costophrenic angle which not changed compared prior examination, copd pattern. patient head ct, which atrophy old ischemic . no acute intracranial findings.

    discharge diagnosis: syncope.

    discharge medications: patient following medications; cardizem mg . . daily, digoxin 0.125 mg . . once daily, allopurinol 100 mg two times daily, coumadin mg . . . . ., and remeron mg . . . . .

    Input Text Annotations

    True Internal Representation(with some keys shown for illustration)

    53 844 58539 83 16 (colon)996 31363 13 (atrial fibrillation)1 19 (period)8 75087 12129 6158 61 316091 3 (peripheral arterial disease)254 332 124624 22 21 (comma)6198 2 (atherosclerosis)2 15 (comma)2835 22 1110647 2 (proctitis)2 92026 2 (cabg)2 411 21907 4 (cholecystectomy)1 15 (period)...

    Reconstructed Representation

    PAST MEDICAL/SURGICAL HISTORY: Positive for atrial fibrillation. The patient had AVR 6 years ago. Peripheral arterial disease with hypertension, peripheral neuropathy, atherosclerosis, hemorrhoids, proctitis, CABG, and cholecystectomy.

    FAMILY HISTORY: Positive for atherosclerosis, hypertension, autoimmune diseases in the family.

    REVIEW OF SYSTEMS: Weight loss of 25 pounds within the last 6 months, shortness of breath, constipation, bleeding from hemorrhoids, increased frequency of urination, muscle aches, dizziness and faintness, focal weakness and numbness in both legs, knees and feet.

    LABORATORY DATA AND RADIOLOGICAL RESULTS: The patient had a chest x-ray, which showed cardiomegaly with atherosclerotic heart disease, pleural thickening and small pleural effusion, a left costophrenic angle which has not changed when compared to prior examination, COPD pattern. The patient also had a head CT, which showed atrophy with old ischemic changes. No acute intracranial findings.

    DISCHARGE DIAGNOSIS: Syncope.

    DISCHARGE MEDICATIONS: The patient was discharged on the following medications; Cardizem 90 mg p.o. thrice daily, digoxin 0.125 mg p.o. once daily, allopurinol 100 mg two times daily, Coumadin 4 mg p.o. q.h.s., and Remeron 15 mg p.o. q.h.s.

    past medical surgical history: positive for atrial fibrillation. patient avr years ago. peripheral arterial disease hypertension, peripheral neuropathy, atherosclerosis, hemorrhoids, proctitis, cabg, and cholecystectomy.

    family history: positive for atherosclerosis, hypertension, autoimmune diseases family.

    review of systems: weight loss pounds months, shortness of breath, constipation, bleeding hemorrhoids, increased frequency of urination, muscle aches, dizziness and faintness, focal weakness and numbness both legs, knees and feet.

    laboratory data and results: patient chest x-ray, which cardiomegaly atherosclerotic heart disease, pleural thickening and small pleural effusion, left costophrenic angle which not changed compared prior examination, copd pattern. patient head ct, which atrophy old ischemic . no acute intracranial findings.

    discharge diagnosis: syncope.

    discharge medications: patient following medications; cardizem mg . . daily, digoxin 0.125 mg . . once daily, allopurinol 100 mg two times daily, coumadin mg . . . . ., and remeron mg . . . . .

    Input Text Annotations

    True Internal Representation(with some keys shown for illustration)

    53 844 58539 83 16 (colon)996 31363 13 (atrial fibrillation)1 19 (period)8 75087 12129 6158 61 316091 3 (peripheral arterial disease)254 332 124624 22 21 (comma)6198 2 (atherosclerosis)2 15 (comma)2835 22 1110647 2 (proctitis)2 92026 2 (cabg)2 411 21907 4 (cholecystectomy)1 15 (period)...

    Reconstructed Representation

    1

    2

    Countpresent,positivementions,aboutthepatient3

    T-1 T-2 .. .. .. .. .. .. .. .. .. T-iP-1 1 0 -1P-2::::

    P-n

    1.8million

    ~3million

    Deferbindingtoconcepts

    Unitex

    NegEx andConTextPatterns

    Knowledgegraph

    tf df NN JJ … VP T-1 T-2 T-3

    ID Term-1 150,879 90,000 0.90 0.05 … 0.03

    ID :

    ID Term-nSyntactictypesFrequency Semantictypes

    termmentions

    Event/Outcomeconcepts

    JuvenileIdiopathicArthritis

    :

    Uveitits

    Iridocylitis

    OralAntihistamines

    JuvenileIdiopathicArthritisICD 9 codes696.0, 714.0, 714.2, 714.3, 714.9, 720.2, 720.9

    Terms:Juvenile idiopathic arthritis, JIAJuvenile rheumatoid arthritis, JRAPsoriatic arthritisJuvenile spondyloarthropathy,spondyloarthritis,enthesitis related arthritis,sacroiliitis,reactive arthritis

    Terms,Concepts

    Counts

    Variables

    Androgendeprivation&Alzheimer’srisk

    0 1 2 3 4 5 6 7

    Anaphylaxis

    Rhabdomyolysis

    Tuberculosis

    Allergic> rhinitis>

    Abdominal>aortic>aneurysm>

    Age

    Caucasian

    Ever> smoker

    AntiDplatelet>medication>use

    AntiDcoagulant>medication>use

    AntiDhypertensive>medication>use

    Statin>use

    History>of>cardiovascular>disease

    History>of>diabetes

    History>of>malignancy

    ADT>use

    1.06 (1.04-1.08)

  • http://tinyurl.com/inf-consult

    antihypertensives

    Intervention

    Diastolicpressure<90mmHg

    Outcome

    A55yearoldfemaleofVietnameseheritagewithknownasthmapresentstoherphysicianwithnewonsetmoderatehypertension

    MyPatient

    100

    DiastolicBPwithDrugA:245DiastolicBPwithDrugB:989

    1 2 3 4 5 6

    90

    80

    70

    60

    MmHg

    months

    VariablesassociatedwithOutcome

    DrugA

    Asthma

    Ethnicity

    HDL

    3 40 1 2HbA1c>10%

    Example– 1:Choosingdiabetesdrugs

    Scenario:WhichsecondlinedrugtousefortreatingdiabeticswhohavehighHbA1conetotwomonthsafterfirstlinetreatment?

    250millionpatients

    Metf->Metf (n)=377Glip->Glip (n)=30

    Efficacy of monotherapy (Stanford data only)

    Effectivetreatmentpathways

    8.0Metforminà Glipizidevs.Pioglitazone

    MetforminàSitagliptin vs.Pioglitazone

    Metforminà Sitagliptin vs.Glipizide

    Example– 2:Choosingchemotherapy

    Scenario:For55-60yearoldwhitemalepatientwithnewlydiagnosedplasmacellleukemia(PCL),whatisthedifferenceinoverallsurvivalbetweenpatientstreatedwithintensiveversuslessintensivechemotherapy?

    Example– 2:Personalizedestimate

  • Outlineofaninformaticsconsult

    Descriptivesummary• Whathappenedaftertreatment?

    Makingrecommendations• Whattreatmentchoicesaretypicallymade,givenpriormedicalhistory?Whataretypicaloutcomes?

    • Estimation:WhatistheeffectoftreatmentchoiceXonoutcomeY?

    XPRESS- EXtraction of PhenotypesfromclinicalRecordsusingSilver Standards

    !!Time!"!

    ICD-9 RX Terms Labs

    3. Patient history is found after first mention of keyword

    1. Build Keyword list based on terms related to “Myocardial

    infarction”

    Pid Notes

    524

    1234

    765

    834

    ….

    ….

    Annotations

    “Myocardial infarction” (Y/N) ?

    2. Find Patients with keyword mentions +" #"

    T" TP# FN#F" FP# TN#

    train

    test

    5. Classifier is built using 5-fold cross validation

    4. Feature Vectors constructed after collapsing patient timeline into normalized frequency counts for all terms, ICD-9s,

    prescriptions and labs

    ICD9 RX Terms Labs

    4. Feature Vectors constructed after collapsing patient timeline into normalized frequency counts for all terms, ICD-9s,

    prescriptions and labs

    Input:config.R – withtermsearchsettingsOutput:keywords.tsv andignore.tsv

    Input:getPatients.R -- config.R,keywords.tsv,ignore.tsvOutput:feature_vectors.Rda

    Input:buildModel.R -- config.R,feature_vectors.RdaOutput:model.Rda

    Phenotype AUC Sens. Spec. PPV

    DM 0.95 91% 83 % 83%

    MI 0.91 89% 91% 91%

    FH 0.90 76.5% 93.6% ~20%

    Celiac 0.75 40 % 90% ~4 %

    Phenotyping- effortprecisiontradeoff

    Acc PPV Time

    0.98 0.96 1900

    Acc PPV Time

    0.90 0.91 2hrInsights

    Learningtruerangesoflabtests DETECT:DataminingEMRsToEvaluateCoincidentTesting

  • Qualitymetrics

    • PostsurgeryDVTrates

    • UrinaryIncontinenceafterprostatecancersurgery

    Time between surgery and DVT/PE event (first year after surgery)

    Time in 10 day intervalsFr

    eque

    ncy

    0 20 50 80 110 140 170 200 230 260 290 320 350

    025

    5075

    100

    PredictiveModeling

    ClassificationorPrediction?

    Patient’sMedicalRecord

    ###.##

    Timeof“X”

    ###.#####.#####.##

    Riskprediction Classification

    SourceoffeaturesChoiceoffeatures

    Predictingdelayedhealingwounds

    59,958 patients

    • Age• Gender• Insurance• Zip code

    180,716 wounds• Wound type• Wound location

    • Dimensions• Edema• Erythema• Rubor• Other wound qualities

    Each wound assessment:68 Healogics Wound Care Centers in 26states• Center Code

    Clinicallyusableperformance

    0.00

    0.25

    0.50

    0.75

    1.00

    0.00 0.25 0.50 0.75 1.00Predicted probability of delayed wound healing

    Obs

    erve

    d fra

    ctio

    n of

    del

    ayed

    wou

    nd h

    ealin

    g

    PredictingCostBlooms

    Panel2013-14TotalMembers:149,038

    Panel2014-15TotalMembers:274,525

    Learnmodel

    Applymodel

  • PredictingCostBlooms

    N=274,525Expensivein

    2014=$18,028

    Expensivein2015=$20,489

    60%- Bloomers

    40%- Persistent

    Predicted,butdon’tbloom

    Bloomerswemiss

    CostBloommodel

    WholePopulation– HighCostmodel

    Correctlypredictedbloomers

    0.810.430.57

    AUCPPVCostCap.

    0.860.510.68

    Fruitfulareasofactivity

    • Riskstratification:cost,latentdisease,decompensation

    • Personalizingevidence:riskforme,whattreatmentwillwork

    • Insightsintodiseaseprogression• Usingpassivelycollecteddata:sensors,featureengineering

    • Practicemanagement:predictingmissedappointments,medicationadherence...

    Openresearchproblems

    • Datanonstationarity• Localvs.Globalmodels• Handlingunstructureddata(text,images,timetraces)• Outcomeascertainment(andcensoring)• Evaluation:Beyonddiscrimination• calibration,net-reclassification

    • Bridgingthe“lastmile”

    AcknowledgementsGroupMembers:• Scientists: SuzanneTamang,KenJung,Alison

    Callahan,JuanBanda• Fellows:RainerWinnenburg,Rohit Vashisht• Engineer: VladimirPolony• BMIStudents:SarahPoole,AlejandroSchuler,

    AlbeeLing,Vibhu Agarwal,DavidStark• MedStudents:GregGaskin,Jassi Pannu

    Alums:AnnaBauer-Mehren (Roche),Srini Iyer(Facebook),Amogh Vasekar (Citrix),SandyHuang(Berkeley),Paea LePendu (Lexigram),RaveHarpaz(Oracle),TylerCole(BarrowInst.),SamFinlayson(Harvard),WillChen(Yale),YenLow(Netflix),ElsieGyang (FellowshipinSurgery)

    Funding:• NIH – NLM,NIGMS,NHGRI,NINDS,NCI,FDA• StanfordInternal– Dept.ofMedicine,

    PopulationHealthSciences,ClinicalExcellenceResearchCenter,Dean’soffice

    • Fellowships – MedScholars,SiebelScholarsFoundation,StanfordGraduateFellowship

    • Industry – Apixio,CollabRx,Healogics,JanssenR&D,Oracle,Baidu USA,Amgen

    IT: AlexSkrenchuk,SCCIteam