Clinician Developed AI and CDS - Built in real time, case by case A - - PowerPoint PPT Presentation

Proprietary and confidential — do not distribute ADD-00065154

Clinician Developed AI and CDS

• Built in real time, case by case

September | 2018

A SHANE BROWN PHD MEDICAL AND SCIENTIFIC LIAISON ABBOTT DIAGNOSTICS DIVISION

Proprietary and confidential — do not distribute

Value in Laboratory Data

Laboratory data is unquestionably of immense value:

• 1984 – ICU data included 41% of total patient record1
• 2000 - Mayo Clinic 94% of data in enquiry system pre digital

radiology2

• 2011 – Aurora Health 82% of stored data generated in pathology3
• 70% often quoted but more accurately “is integral to many clinical

decisions providing . . [HCP] . . With often pivotal information”4

1) Bradshaw, KE et al. Int J Clin Monit Comput 1984;1:81-91 2) Forsman R Clin Leadership Manag Rev 2000: 14:292-5 3) Feist. K http://www.executivewarcollege.com/2010/PDFs/Feist.pdf 4) The Value of Laboratory Medicine to Health Care. Chapter 1. In: The Lewin Group: Laboratory Medicine – A National Status Report. May 2008:19–65. See http://www.ascls-sd.org/sitebuildercontent/sitebuilderfiles/laboratory_medicine_-_a_national_status_reportmay08.pd

Proprietary and confidential — do not distribute

Contribution of AI in Financial Services

• Transactions involve only a single

item – currency or derivatives representing currency

• International standardisation of

currency values

• Global regulation of activities
• Consumers predictable - mostly

Proprietary and confidential — do not distribute

Contribution of AI in Business Intelligence

• Transactions involve multiple items –

however centred around individual entities e.g. retail, production, logistics

• Standard and controlled operations

within each entity

• Local control of activities
• Transactions / activity controlled

Proprietary and confidential — do not distribute

Contribution of AI in Health Management

• Transactions involve multiple items

in multiple entities across multiple sectors

• Practices vary widely entity to entity

and sector to sector

• Limited control of activities
• Imprecise predictability of

individual health needs

Proprietary and confidential — do not distribute

AI In Health

Case Study: IBM Watson/MD Anderson

• Project commenced in June 2012 – budget $2.4M
• Project terminated in September 2016 – expense $39.2M
• Goal was to “help community oncologists provide MD Anderson-quality

cancer care to patients who cannot seek treatment directly from MD Anderson physicians”

• Outcome: it “never guided the treatment of any community-based patients”

https://www.medscape.com/viewarticle/876070

Proprietary and confidential — do not distribute

AI in Health (2)

Case Study: IBM Watson/MD Anderson

• IBM claimed that Watson would “continually ingest patient and research data,

medical literature, and treatment options, to offer care advice”

• However, its recommendations are not based on computed insights from this

data.

• Instead recommendations rely exclusively on supervised training from clinical

experts

• At the end of the MD Anderson trail with work from computer engineers and

doctors, Watson was able to deal with only seven types of cancer

https://www.medscape.com/viewarticle/876070 https://www.statnews/2017/09/05/Watson-ibm-cancer

Proprietary and confidential — do not distribute

What is different about the questions we ask in Health and Medicine

Q1 :

When was the Battle of Hastings? 1066

Q2:

Who discovered insulin? Banting and Best (Sharpey-Schafer)

Q3:

Why is the sky blue? Rayleigh scattering

Q4:

Why do I have a headache? ?????

Proprietary and confidential — do not distribute

Why do I have a headache?

• Context
• Multiple data sources
• Query ability
• Assimilation capability
• Experience in the specific domain to interpret the

information

i.e. An Expert is required

Proprietary and confidential — do not distribute

Expert Systems

• Developed extensively in the 1990’s, were the first truly successful form of AI.
• Solve complex problems by application of a corpora of knowledge through rules

rather than attempting to classify via coded algorithms.

• Sophisticated expert systems expose their user interface such that non-

programmers create the rules.

• Have been adopted by many application suite vendors as integral components of

their AI technology.

Proprietary and confidential — do not distribute

Expert Systems

Advantages

• Rules development provided by a

domain expert so delivery of system outputs is independent of an IT specialist.

• Easy to maintain since no

conventional code.

• Systems are built incrementally.
• Training data sets are provided

incrementally.

Disadvantages

• System development relies heavily
• n domain experts.
• Supervised training sets with

defined end points are needed.

• Needs on-line system integration.

Proprietary and confidential — do not distribute

Ripple Down Rules (RDR)

• Developed by Prof Paul Compton and colleagues at Garvan Institute of

Medical Research in 1988.

• Uses “case based reasoning” to incrementally acquire knowledge with each

creation of a new rule.

• Built on the premise that at any point in time more WILL be known about

any given topic – poorly classified cases are used to improve the quality of the knowledge.

• Assumes that

a) knowledge is only correct in context and b) over time that knowledge will be modified – eg the atom.

Proprietary and confidential — do not distribute

RDR Expert technology

Proprietary and confidential — do not distribute

Expert Controlled Rules Engine

7 rules added 8 rules added

Proprietary and confidential — do not distribute

Rapid, Consistent Rule Building – In production Environments

Write a rule

AlinIQ-CDS Knowledge Base

Potential Conflicting case?

No Conflict or No more cases - Finish rule session Conflict ? No Yes

Add conditions

Yes Update rules

Proprietary and confidential — do not distribute

Rapid, Consistent Rule Building – Case Studies OhioHealth USA

Dr Eugenio Zabaleta

Proprietary and confidential — do not distribute

Proprietary and confidential — do not distribute

5 10 15 20 25 30 35 2013 2014 2015 2016 2017 2018 2019

Cumulative Number of Kbase

18

Dr Peter Cole Dr Emma Wypkema

Rapid, Consistent Rule Building – Case Studies Lancet Laboratories South Africa

Proprietary and confidential — do not distribute

Organisational use of Expert System in Lancet Pathologists

• Specimen reception
• Data entry
• Accounts
• Appropriate testing (MTF)
• Customer service

(Referrers and Funders)

• Marketing
• Workflow
• Laboratory alerts
• Appropriate testing (intelligent

reflex)

• Patient centred reporting
• Patient / Referrer tailored

reports

Proprietary and confidential — do not distribute

Clinical Decision Making is Complex

1. Graber ML, Franklin N, GordonR. Arch Intern Med 2005; 165: 1493-9. 2. Singh H et al, JAMA Intern Med 2013; 173(6); 418-25.

• 3. Kachalia A, et al. Ann Emerg Med 2007; 49: 196-205.

1 2 3