#ResBaz2019 Keystory - Applying data science to health care

Can we predict how many emergency admissions would occur in the next hour? Or is it possible to provide better care by automating cumbersome but essential clinical tasks? Discover how data science can help shape the 'health and care', in an industry which has data locked into many disparate systems. In this talk, we will touch upon 'the data and digital skills' required to overcome some of the challenges in realising the real value of healthcare data.

Transcript

1. SO, WHAT IS AI?
4. 4 Systems that exhibit human-like intelligence
5. Source: https://towardsdatascience.com/role-of-data-science-in-artificial-intelligence-950efedd2579
6. APPLYING DATA SCIENCE TO HEALTHCARE MANISH KUKREJA https://datamak.ml Disclaimer: The views and opinions expressed in this talk are mine and do not necessarily reflect the official policy or position of my employer Auckland District Health Board. #ResBaz2019 https://resbaz.auckland.ac.nz/schedule/#session-103
7. IN THIS TALK.. •Motivation •Challenges •Solutions 7
8. WHY UPSKILL IN HEALTH DATA SCIENCE, ML, AI ? 8 In 2015 $B Health care and social assistance LABOUR EFFICIENCIES FROM AI IN NEW ZEALAND IN 2035 Healthcare efficiencies $1.6B - $3.6B Source: https://aiforum.org.nz/wp-content/uploads/2018/07/AI-Report-2018_web-version.pdf
9. "How might this improve the care of the patient and the working environment of the clinician?" - Dr. Keith Grimes, GP 9 Source: https://youtu.be/5YW3zahlfYs
10. WHAT ARE THE NEAR & LONG-TERM IMPACTS ? • Patient Flow • ED Forecast, outpatient DNA, inpatient LOS • Personalisation (or Precision Medicine) • Individualised drug dosage plan • Physician burnout reduction • Automating clerical tasks • AI-assisted screening and diagnosis 10 Source: http://med.stanford.edu/content/dam/sm/school/documents/Health-Trends-Report/Stanford-Medicine-Health-Trends-Report-2018.pdf
11. WHAT ARE THE NEAR & LONG-TERM IMPACTS ? • Patient Flow • ED Forecast, outpatient DNA, inpatient LOS • Personalisation (or Precision Medicine) • Individualised drug dosage plan • Physician burnout reduction • Automating clerical tasks • AI-assisted screening and diagnosis 11 Source: http://med.stanford.edu/content/dam/sm/school/documents/Health-Trends-Report/Stanford-Medicine-Health-Trends-Report-2018.pdf Strategic and Operational Decision Support
12. WHAT ARE THE NEAR & LONG-TERM IMPACTS ? • Patient Flow • ED Forecast, outpatient DNA, inpatient LOS • Personalisation (or Precision Medicine) • Individualised drug dosage plan • Physician burnout reduction • Automating clerical tasks • AI-assisted screening and diagnosis 12 Source: http://med.stanford.edu/content/dam/sm/school/documents/Health-Trends-Report/Stanford-Medicine-Health-Trends-Report-2018.pdf Clinical Decision Support
13. NZ HEALTH AND DISABILITY SECTOR 13 Source: https://www.health.govt.nz/sites/default/files/images/nz-health-system/structure-nz-health-disability-sector-oct16.png Central Govt. Ministry of Health Purchase Agreement Reporting DHB provider arms Health Services Private and NGO providers Health Services NZ population and businesses Private health insurance Reporting Reporting Reporting Some fees/ co-payments 20 District Health Boards (DHBs) Service Agreement Service Agreement
14. CHALLENGES WITH HEALTH DATA • Siloed data, locked into legacy or standalone systems • Often unstructured, messy and/or in proprietary formats • Inflexible or rigid, if in a structed form e.g. in data warehouse • Mostly transactional reporting against funding (not patient-centric) • Lack of Data Management plans and policies • Paper based or not in digital form (e.g. tissue slides) • Not shared within services, region or entities 14
15. CHALLENGES WITH HEALTH DATA • Siloed data, locked into legacy or standalone systems • Often unstructured, messy and/or in proprietary formats • Inflexible or rigid, if in a structed form e.g. in data warehouse • Mostly transactional reporting against funding (not patient-centric) • Lack of Data Management plans and policies • Paper based or not in digital form (e.g. tissue slides) • Not shared within services, region or entities 15 Digital Enablement
16. CHALLENGES WITH HEALTH DATA • Siloed data, locked into legacy or standalone systems • Often unstructured, messy and/or in proprietary formats • Inflexible or rigid, if in a structed form e.g. in data warehouse • Mostly transactional reporting against funding (not patient-centric) • Lack of Data Management plans and policies • Paper based or not in digital form (e.g. tissue slides) • Not shared within services, region or entities 16 Data Management & Governance
17. ”.. data scientist: those who use both data and science to create something new.” - Dr. DJ Patil 17 Source: http://radar.oreilly.com/2011/09/building-data-science-teams.html
18. WHICH ARE THOSE ‘DATA AND DIGITAL’ SKILLS? • Translating business needs • Conduct experiments • Summarize, interpret and visualize Data Analysis • Supervised learning • Unsupervised learning • Custom algorithm development Data Modelling • Data pipeline construction • Production, automation, integration • Software engineering Data Engineering • Data formatting, fixing importing errors • Handling missing values • Querying, joining, slicing Data Mechanics • Test and monitor model fairness • Gather ‘informed’ consent • Plans to protect and secure user data Data Ethics 18 Source: https://brohrer.github.io/data_science_archetypes.html https://www.amazon.com/dp/B07GTC8ZN7/
19. HOW CAN WE APPLY THEM? – USE CASES • ED presentations predictions • Diabetic Retinopathy • Digital pathology • Sleep Apnea auto-annotation • Chest X-ray screening 19
20. ED arrivals follow set patterns, but react to external variables 1. Daily: less patients in the am 2. Weekly: Monday spikes, midweek lulls 3. Monthly: Flu season, school schedules 4. Holidays: sees drop and day after sharp increase 20 HOW CAN WE APPLY THEM? – ED PRESENTATIONS Source: https://www.kensci.com/solutions/ops/
21. HOW CAN WE APPLY THEM? – DIABETIC RETINOPATHY 21 Source: https://www.eyediagnosis.co/idx-dr-eu-1
22. HOW CAN WE APPLY THEM? – DIGITAL PATHOLOGY • Teach-by-example on Whole Slide Imaging (WSI) • No-programming needed • Tumor segmentation and cell identification on both H&E and IHC stained tissue sections 22 Source: https://www.visiopharm.com/images/grafik/AI/AI_animation_final.gif
23. HOW CAN WE APPLY THEM? – SLEEP APNEA 23 Source: http://www.easmed.com/main/wp-content/uploads/Natus-Remlogic-Brochure.pdf • Learn from manual annotations of Central and Mixed Apnoea and Hypopnea • Automatically annotate the signals • Sleep physician can validate and generate report
24. HOW CAN WE APPLY THEM? – RADIOLOGY • Chest X-ray screening/ segmentation in the browser • https://public.md.ai /hub/model/x9N20 BZa 24
25. FINAL THOUGHTS AI enabled health systems must: •help transition from volume to value i.e. patient-centric outcomes •promote equitable outcomes for all patients i.e. free from bias •protect patient data i.e. adhere to patient’s privacy rights •use the data which patient has consented for and for that purpose only 25
26. THANK YOU See you tomorrow in the workshop: Accessing APIs using R (Rm:G30, 11-12pm)
27. BACKUP SLIDES 27
28. HOW READY ARE WE ? • Govt is building AI strategy/plan • Ministry of Social Development (MSD) and Accident Compensation Corporation New Zealand (ACC) have started using predictive models • Ministry of Health (MoH) is drafting guidelines for safely developing and using algorithms in healthcare 28 Source: https://www.oxfordinsights.com/government-ai-readiness-index/ https://www.cs.otago.ac.nz/research/ai/AI-Law/NZLF%20report.pdf
29. DATA SCIENTIST ARCHETYPES These archetypes are all proficient at data mechanics and ethics, additionally, • The Generalist is proficient at everything • The Detective is a master of data analysis • The Oracle is a master of data modelling • The Maker is a master of data engineering Source: https://brohrer.github.io/data_science_archetypes.html 29