Health Informa.cs. Lecture 9. Samantha Kleinberg

Transcription

1 Health Informa.cs Lecture 9 Samantha Kleinberg samantha.kleinberg@stevens.edu

2 Next week: journal club For all papers: read, and prepare to comment on For your paper: Read the ar.cle (+ other references if needed for context) Prepare a brief summary Prepare ques.ons to lead discussion

3 Format Brief overview of topic Summary of paper What s the main argument/hypothesis? How is this supported? Overview experiments, figures Discussion Prepare ques.ons Note: Everyone must par.cipate Total of 30 min/paper

4 Phenotyping Data-> cohort with par.cular traits

5 Uses Retrospec.ve research Clinical trials Epidemiology/popula.on health Discovering subgroups

6 Using EHRs for research First challenge is iden.fying cohort Who has disease? Who doesn t? Accuracy affects everything that follows Usually based on manually defined rules Itera.ve process Extremely.me consuming Clinical trials Usually manual iden.fica.on of eligibility

7

8 Usual process: -Define criteria -Evaluate results -Iterate Need some labeled data or ability to determine whether results are correct

9 S. Kleinberg and N. Elhadad. Lessons learned in replica.ng data-driven experiments in mul.ple medical systems and pa.ent popula.ons. In AMIA Annual Symposium, 2013

10 Challenges Representa.on How to describe/define a phenotype Complexity Scale of data, complexity of defini.on Data quality/completeness/bias Are the criteria we need there? Can we trust the data? Narra.ve text Extrac.ng criteria Portability Does phenotype in one hospital work at another?

11 EXAMPLE: CHF case defini.on Geisinger: 2 medica.on orders for CHF or 2 outpa.ent visits with CHF diagnosis or 1 medica.on order and 1 outpa.ent CHF diagnosis or CHF on problem list Columbia: 2 ICD-9 codes for CHF or 1 ICD-9 code for CHF and 1 men.on of CHF in note on the same date or 1 ICD-9 code for CHF and one medica.on typically indicated for CHF

12 OperaAonal criteria 1 outpa.ent visit or 1 medica.on order 2+ medica.on orders only 2+ outpa.ent visits only 1 outpa.ent visit and 1 medica.on order only Problem list only Meet 2 of 3 criteria Meet all 3 criteria none 1 minor, 0 major Number of Framingham criteria 2+ minor, 0 major 0 minor, 1 major 1 minor, 1 major 1 major, 2+ minor 2+ major Total Meets opera.onal criteria and Framingham criteria, N=2294 (35%) Meets opera.onal criteria but not Framingham criteria, N=2900 (45%) Meets Framingham criteria but not opera.onal criteria, N=424 (7%) Does not meet opera.onal criteria or Framingham criteria, N=879 (14%) Total Wu et al. (2010). Prediction Modeling Using EHR Data: Challenges, Strategies, and a Comparison of Machine Learning Approaches. Med Care 48(6):S

13 Phenotyping algorithms

14 Main approaches Logical rules Where do rules come from? Guidelines, expert knowledge Text-based Keywords, deeper seman.c processing Sta.s.cal/ML(classifica.on)

15 High throughput phenotyping Previously: laboriously define criteria for one phenotype at a.me Limita.ons Large-scale GWAS Discovering new phenotypes Stra.fica.on Only as good as prior knowledge

16 Challenges? Missing data Bias Error Standardiza.on

17 Open problems Phenotype detec.on Time Disease state changing over.me, pa.ent popula.on changing over.me, medical process changing over.me Interpreta.on Making automated phenotypes understandable Evalua.on What s the ground truth? Fragmented data Nonclinical data HIE

18 Pharmacovigilance Drug trials involve small popula.ons What about Rare side effects? Interac.ons with other medica.ons? Interac.ons with foods? Repurposing drugs? Postmarket analysis

19 How are ADEs reported? Clinical trials Every event reported FDA AERS Selected events submited by pa.ent, doctor But also indirectly Searching for informa.on Evidence of side effect in EHR

20 Biomedical informa.cs goals Drug side effects Drug-drug interacaons Dosing Ideal treatment Subgroup Finding new indica.ons

21 ADD DETAIL

22 Many pa.ents have comorbidi.es Two common illnesses will lead to many drugpairs in common One effect: raised glucose T2DM increasing, major public health problem DDI may not be reported as such, need to look at co-occurrences

23 Methods 12,627 averse event reports 37 drugs 4 pairs 3 were infrequent combo 1 combina.on leu: Paroxe.ne, Pravasta.n Took glucose measurements before/during treatment for pa.ents at Stanford 374 on paroxe.ne, 449 on pravasta.n, 8 on both Replicated at Vanderbilt, and Partners

24 Tatones, N.P., Denny, J.C., Murphy, S.N., Fernald, G.H., Krishnan, G., Castro, V., Yue, P., Tsau, P.S., Kohane, I., Roden, D.M. & Altman, R., 2011, Detec.ng drug interac.ons from adverse-event reports: interac.on between paroxe.ne and pravasta.n increases blood glucose levels, Clinical Pharmacology & Therapeu4cs, 90(1), pp

25 Today s paper

26 Google flu htp://sta.c.googleusercontent.com/media/ research.google.com/en/us/archive/papers/ detec.ng-influenza-epidemics.pdf

27 htp://

28

29

30

31 Twiter + flu

32

33

34 Main points Data is changing over.me Publicizing results can change searcher behavior Correla.on vs causa.on Robustness Big data not necessarily good data

35 How to discover new uses for drugs? important?) (why How to find out when drugs being used offlabel?

36 a

37 Lots of $$ and.me to develop new drugs Can we repurpose old ones for new indica.ons? 100 diseases, 164 drugs pairings -> 2664 stat. sig Assump.on: if drug leads to opposite expression from disease, could be useful treatment

38

39

40

41 Journal club! Next week