O slideshow foi denunciado.
Utilizamos seu perfil e dados de atividades no LinkedIn para personalizar e exibir anúncios mais relevantes. Altere suas preferências de anúncios quando desejar.
Dynamic Talks
Bumped HQ
12/4/2019
Peter Graven, PhD
Challenges for AI in Healthcare
Background
 Lead Data Scientist at OHSU
 Assistant Professor (Affiliate) at OHSU-PSU
School of Public Health
 PhD in He...
Why is AI in Healthcare Different?
 Artificial intelligence is changing world in many
sectors
 Predictive typing, intern...
AI in Healthcare Current Status
 Robotic surgery (simple routine steps)
 Image analysis (x-rays, retina scans)
 Genetic...
Basic Challenges in Healthcare
 Decisions need to be right at very high level of
accuracy
 Risk of lawsuits (though none...
A Story about Predictive Modeling
Peter Graven, PhD 6
Let’s predict
risk!
Let’s predict
risk of high
costs!
Let’s predict
...
Implications for AI in Healthcare
 Flip the script!
 It’s not about the cool modeling
 It’s about finding interventions...
Focus on the Decision-making
Peter Graven, PhD 8
(AI)
Artificial Intelligence
(IA)
Intelligent Applications
Black box
Uncl...
Advanced approaches
 If specific interventions exist, build models to feed them
patients. Otherwise,
 Follow workflows a...
More Implications
 As models are deployed within delivery
system, the upkeep and maintenance issues
grow
 Cost of a good...
Some realities
 Electronic Medical Record (EMR) systems are
not easy to integrate with
 FHIR and other interoperability ...
Discussion
 Peter Graven, PhD
graven@ohsu.edu
Peter Graven, PhD 12
Terminou este documento.
Transfira e leia offline.
Próximos SlideShares
What to Upload to SlideShare
Avançar
Próximos SlideShares
What to Upload to SlideShare
Avançar
Transfira para ler offline e ver em ecrã inteiro.

0

Compartilhar

"Challenges for AI in Healthcare" - Peter Graven Ph.D

Baixar para ler offline

Dynamic Talks Portland: The use of AI in many industries has revolutionized operations and efficiency. In healthcare, the progress is just beginning. Despite the promise of AI, why has the development lagged other industries? What issues are unique to healthcare that create challenges for common approaches? How can data scientists overcome these challenges and deliver on the promise of using data to reach multiple goals of improved quality, decreased cost, and greater patient satisfaction?

  • Seja a primeira pessoa a gostar disto

"Challenges for AI in Healthcare" - Peter Graven Ph.D

  1. 1. Dynamic Talks Bumped HQ 12/4/2019 Peter Graven, PhD Challenges for AI in Healthcare
  2. 2. Background  Lead Data Scientist at OHSU  Assistant Professor (Affiliate) at OHSU-PSU School of Public Health  PhD in Health Economics from U of MN  Previous Experience  Academic research (health policy, methodology, program evaluation)  Economic consulting Market research  Any views expressed are not (necessarily) the views of OHSU 2
  3. 3. Why is AI in Healthcare Different?  Artificial intelligence is changing world in many sectors  Predictive typing, internet searching  Speech recognition  Visual perception  Marketing  Marketing examples  Product recommendations  Image recognition (for searches)  Sentiment analysis (social media)  Demand based pricing  Identify customers that might leave  Chatbots 3
  4. 4. AI in Healthcare Current Status  Robotic surgery (simple routine steps)  Image analysis (x-rays, retina scans)  Genetic analysis (review large amount of data)  Pathology (analyze biopsy, not approved)  Clinical-decision (sepsis, deterioration, risk of ED/hospital admit, no-show visits,  Virtual nursing (collect basic info for visit)  Administration (billing and claims)  Mental health (use mobile phone for monitoring depression) Source: Strickland, E. “How IBM Watson Overpromised and Underdelivered on AI Health Care”, IEEE Spectrum, Apr 2, 2019. Peter Graven, PhD 4
  5. 5. Basic Challenges in Healthcare  Decisions need to be right at very high level of accuracy  Risk of lawsuits (though none currently known)  Clinicians are ultimately responsible for decisions  Not outsourced to algorithms  Clinician understanding of algorithms are mostly in infancy (in terms broad-based adoption)  Input factors must be transparent  Otherwise, predictive risk cannot be acted upon 5
  6. 6. A Story about Predictive Modeling Peter Graven, PhD 6 Let’s predict risk! Let’s predict risk of high costs! Let’s predict risk of hospital admissions! Let’s predict who needs Care Management Let’s predict who will respond to Care Management How do we predict who will respond to Care Management? BASIC SCIENCE
  7. 7. Implications for AI in Healthcare  Flip the script!  It’s not about the cool modeling  It’s about finding interventions that work  Old fashioned approach of trials and experiments and science  Then create models to match interventions to people  Tailor the model to the intervention  “There’s a model for that! Peter Graven, PhD 7
  8. 8. Focus on the Decision-making Peter Graven, PhD 8 (AI) Artificial Intelligence (IA) Intelligent Applications Black box Unclear interventions Minimizes need for humans Transparent input factors Oriented around decisions Tailored to existing workflows
  9. 9. Advanced approaches  If specific interventions exist, build models to feed them patients. Otherwise,  Follow workflows and assess places for models to be inserted  The workflow is the intervention. Use the model to make it better  Embedded improvement process with model simply as new technology  Focus on making the decision faster, easier, or more certain  Give the user the right information so they feel confident  Will improve clinician satisfaction  Organic distribution  let users get used to the information before workflow is cemented Peter Graven, PhD 9
  10. 10. More Implications  As models are deployed within delivery system, the upkeep and maintenance issues grow  Cost of a good model embedded is not trivial.  Model itself is just one line of code but easy to underestimate cost of  organizing data to estimate model,  Making model appear in proper location  Training individuals in what it means Peter Graven, PhD 10
  11. 11. Some realities  Electronic Medical Record (EMR) systems are not easy to integrate with  FHIR and other interoperability tools may help but will not likely provide the seamless experience  Very little incentive for EMR companies to really make integration smooth  Cloud based options are growing for more complex (real-time) modeling without being an on premise solution  Many lawsuits about improper sharing of data  Difficult to arrange data for algorithms  1000’s of tables that are linked but not designed for analytic purposes Peter Graven, PhD 11
  12. 12. Discussion  Peter Graven, PhD graven@ohsu.edu Peter Graven, PhD 12

Dynamic Talks Portland: The use of AI in many industries has revolutionized operations and efficiency. In healthcare, the progress is just beginning. Despite the promise of AI, why has the development lagged other industries? What issues are unique to healthcare that create challenges for common approaches? How can data scientists overcome these challenges and deliver on the promise of using data to reach multiple goals of improved quality, decreased cost, and greater patient satisfaction?

Vistos

Vistos totais

403

No Slideshare

0

De incorporações

0

Número de incorporações

3

Ações

Baixados

4

Compartilhados

0

Comentários

0

Curtir

0

×