Clinical Decision Support of the Future

Transcription

1 Clinical Decision Support of the Future Session #4, March 5, 2018 Scott Weingarten, MD Senior Vice President Chief Clinical Transformation Officer Professor of Medicine Cedars-Sinai Health System 1

2 Conflict of Interest (30pt Verdana) Scott Weingarten, MD Senior Vice President Cedars-Sinai Health System Receives salary from Cedars-Sinai Health System Chairman of the Board Acting CEO Stanson Health 2

3 Conflict of Interest (30pt Verdana) Scott Weingarten, MD Salary: No Royalty: No Receipt of Intellectual Property Rights/Patent Holder: No Consulting Fees (e.g., advisory boards): No Fees for Non-CME Services Received Directly from a Commercial Interest or their Agents (e.g., speakers bureau): No Contracted Research: No Ownership Interest (stocks, stock options or other ownership interest excluding diversified mutual funds): Stock from Stanson Health 3

4 Agenda Clinical Decision Support of the Future 4

5 Learning Objectives Evaluate what the future of CDS will look like for healthcare organizations 5

6 Why CDS? Provider consolidation Risk shifting to providers Many health systems have more than 1 inpatient EHR and many different provider ambulatory EHRs Will seek a consistent approach to patient care Potentially scalable and sustainable approach to value 6

7 Why CDS? Predictors of Success Adjusted OR Automatic provision of decision support as part of workflow Provision of decision support at the time and location of decision making Provision of recommendation rather than just an assessment Computer-based generation of decision support 6 Source: Kawamoto K, Houlihan CA, Balas EA, Lobach DF. Improving clinical practice using clinical decision support systems: a systematic review of trials to identify features critical to success. BMJ Apr 2;330(7494):765. PMID:

8 What About Alert Fatigue? Remove junk CDS High false-positive rate High number needed to change Low impact on quality Low impact on cost 8

9 What To Replace the Junk With? 9

10 What Do You Replace the Junk With? 10

11 200 Tuesdays or Thursdays CDS - Means to an end Impact on Clinical and business priorities Quality, safety Cost Value Value-based contracts 11

12 Reducing Health Care Costs McWilliams JM, Schwartz AL. N Engl J Med 2017;376:

13 CDS Cycle Wet lab testing CDS Authoring create test optimize publish CDS Analytics monitor CDS Delivery or in the Cloud Embedded 13

14 Cloud CDS Direct integration into EMR workflows Alert Engine Rule Runtime Automatically provides recommendations at key workflow points; supports continuous, noninterruptive suggestions (nudges) NLP and ML NLP & ML (Cloud CDS) EMR or Web Browser NLP, ML enable the use of unstructured data Consume additional data sources Ex: Claims, Genomics, Immunization registries Interactive App EHR-integrated and browser-based 14

15 Limitations of Indications or Pick Lists 25.4% of imaging requests could be rated by structured indications alone Additional Challenges Lists can become lengthy First fit picking can lead to degradation of CDS recommendations Elements can often be obtained from medical records 88.7% could be rated when free text information was considered J Am Coll Radiol 2017;14: Ask not what you can do for your EHR, ask what your EHR can do for you 15

16 Changing Clinical Care 16

17 Strategies for CDS Implementation 17

18 Hardwiring Care - Before Ordering Vitamin D intervention, Kaiser Permanente Making it harder to do the wrong thing Education Removing items from preference list Hard stop alerts Results Screening declined 3-fold, from 74 tests/1,000 to 24 tests/1,000 (p<0.001), 67% decrease Appropriate care increased 56% to 70% Inappropriate care decreased 44% to 30% Followed rates decreased 10% to 4%, best practice internalized JAMIA 2017;24:

19 Hardwiring Care Order Set Side Effects 19

20 Non-interruptive physician starts order in EMR likely unnecessary likely appropriate order placed inappropriate order cancelled 20

21 Using Free Text Information NLP & ML Score Calculation Computer-Aided Documentation Auto-Auth 21

22 Measuring Impact Cancelled orders If you trigger the same alert 10 times, do you order and cancel or anticipate? Blood tests, colonoscopies, PAP smears Reduced rate of ordering/inappropriate orders avoided Harder to measure Interrupted time series design Inappropriate orders avoided design Adjusted ordering rates 22

23 After Ordering Peer-Comparison Feedback Randomized controlled trial Low value care Antibiotics for URIs 248 providers 14,753 patient visits Control (24.1% to 13.1%) Peer-comparison feedback (19.9% to 3.7%,p<0.001) Accountable justification (23.2% to 5.2%, p<0.001) 1 year later Drift Peer-comparison feedback still has some impact 23

24 inappropriate ordering of Lyme disease tests 24

25 Peer-comparison feedback Affiliation Physician 1 32% Physician 2 7% Physician 3 7% % of potentially low-value care (n=176) 25

26 Linking to Incentives Physician Choosing Wisely performance Average 0.74% ignored Choosing Wisely alerts/1,000 orders Range 0% to 8.77% ignored/1,000 orders Example:0.53% Choosing Wisely Performance Rate 26

27 CDS of the Future Today Tomorrow data guidance! delivery structured EMR data rules-based approach episodic: delivered in response to specific provider actions structured + unstructured data (via NLP/ML) rules-based + AI-based (ML) approach episodic + surveillance: delivered whenever and wherever clinical circumstances change 27

28 CDS of the Future Lowered accident claims Mercedes 16% Acura 15% 28 Disease

29 CDS of the Future 29

30 CDS of the Future Past State Closed EHRs Future State EHRs as data platforms (today) Navigator and form-based data entry Traditional input / output (monitors and keyboards) Apps as virtual assistants Voice: Amazon Alexa 30

31 CDS of the Future Mrs. Smith is 64 years old and has diabetes, hypertension. Based upon Mrs. Smith s genetic profile, microbiome information, symptoms, signs, laboratory values, imaging studies, personal preferences, the clinical decision support recommends 31

32 Patient Clinical Decision Support CDS of the Future Mrs. Smith is a 64 year-old woman presents with diabetes and hypertension for follow-up After review of What is the optimal treatment and monitoring of her diabetes and hypertension? 32

33 7 The Next 100 Years of Medicine "Complex but empirically validated algorithms will be embedded in EHR systems as decision support tools to assist in everyday patient care. Those management algorithms will evolve and be modified continuously in accordance with inputs from ongoing clinical observations and from new research. Clinical decision support algorithms will be derived entirely from data, not expert opinion, market incentives, or committee consensus. New England Journal of Medicine December 27,

34 Conclusion During the hour that we are spending together today There may be be approximately 28 deaths in the United States because of medical errors There may be $22 million spent on medical over-treatment Hos will CDS help? 34

35 "Of course it's hard. It's supposed to be hard. If it were easy, everybody would do it. Hard is what makes it great." 35 35

36 36

37 Conflict of Interest Jonathan Teich, MD, PhD Receives salary from InterSystems and Brigham & Women s Hospital 37

38 Learning Objectives Evaluate what the future of CDS will look like for healthcare organizations Identify the impact of CDS from big data and machine learning Discuss how CDS of the future will benefit physicians and patients Discuss the authority of the CDS material used 38

39 Agenda Hot CDS clinical areas Proactive CDS /Care plans Surveillance and analytics CDS New CDS types, new data sources Patient CDS 3rd-party apps and 2-way APIs Sharing content and operations ML/AI is the answer, or isn t 39

40 Hot CDS clinical areas Acute: sepsis, imaging choice, KPIs/QMs (stroke, infection), OR/PACU/ICU summaries, ERAS Transition: care continuity, med rec, pre-/post-procedure, care plans Health targets: opioids, suicide risk, readmission prevention, end of life, find target patients (intelligent chart review) Global: disasters, maternal, immunization, nutrition, epidemics, cancer 40

41 Proactive CDS / care plans Promote timely / effective care, prevent omissions Population health = identify patients for a common broad approach, and execute it (with some personalization) Identify: intelligent chart review Approach: care plan / clinical pathway Execute: communication, tasks, actions, events, common summary view, WDIDN view, measure all flexible modules 41

42 42

43 Surveillance CDS Surveillance: Scan for events occurred: ED visit, metric change Scan for events missed: Rx pickup, f/u appointments, tests while on Harvoni Requires multiple data sources Examples: Chi: ED frequent flyer intervention NY: Alert hospital and community caregivers on patient movement Infection surveillance 43

44 44

45 Analytics CDS Population-level or practice-level surveillance In bulk, CDS is analytics + notification Performance trends (QM) (Flu trends, opioid) Real-time new rules (e.g., antibiotic susceptibility) Learning health system Identify the sickest/most movable 10% Analysis of the CDS performance itself Detect when codes change Detect best/worst alert-to-value 45 CDS

46 New CDS types, new data sources NLP Multi- SDoH Genomics 46

47 Social Media Checklists GIS 47

48 Patient CDS 48

49 Data views for patients Shared decision-making Self-assessments Visualizations Smartphone apps IoT?? 49

50 3rd-party apps and 2-way APIs 50

51 3rd-party apps and 2-way APIs Today: make some orders happen Make some assertions Semaphores Alert on other device Two-way APIs need to happen! 51

52 Sharing content and operations 52

53 What is ML good for? + Diabetes segmentation (cluster) + Radiology CDS (detection)? Predict trajectory o Add to problem list o Broad assessment and risk finding based on set of parameters (score) - Drug regimen optimization - Sequence of tx, pathway (complex)? Fuzzy data sources -- Explaining itself 53

54 The future is Whatever is practical and usable, rather than just fancy Simple > complex We will be helped by future tech after a few bumps 54

55 Questions? or 55

56 Questions? 56