Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
1
Towards Transformative Artificial Intelligence
in Life Science and Health Care
Assoc.Prof. Mag. Dr. Matthias Samwald
Section for Artificial Intelligence and Decision Support
Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Medical University of Vienna
• One of the largest hospitals in Europe
• 25 university clinics
• 8000 students
2
Assoc.Prof. Dr. Matthias Samwald
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Section for Artificial Intelligence
and Decision Support
• History dating back to 1977
• Multidisciplinary work on AI, medicine,
biology, cognitive science
3
„From data to knowledge to action“
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
4
Diagnosing diseases based on lab values or images
Natural language processing
• Identifying key statements and their semantic relations in biomedical texts
• Deep learning for biomedical text summarization
• Creating high-precision biomedical search engines
• Re-purposing deep NLP models for biological sequences (proteins)
Predicting novel links in large biomedical knowledge bases
Decision support systems for enabling personalized medicine
Strategic considerations for harnessing AI for global progress
Raw data
Knowledge
Action
„From data to knowledge to action“
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
5
Diagnosing diseases based on lab values or images
Natural language processing
• Identifying key statements and their semantic relations in biomedical texts
• Deep learning for biomedical text summarization
• Creating high-precision biomedical search engines
• Re-purposing deep NLP models for biological sequences (proteins)
Predicting novel links in large biomedical knowledge bases
Decision support systems for enabling personalized medicine
Strategic considerations for harnessing AI for global progress
Raw data
Knowledge
Action
Predicting novel links in large biomedical
knowledge bases
Section for Artificial Intelligence and Decision Support
Matthias Samwald
6
We are using neural models for predicting yet unknown
links in biomedical knowledge bases
We are using neural models for predicting yet unknown
links in biomedical knowledge bases
Knowledge graph embeddings
• low dimensional
• vector space
• certain properties can be represented
by notions of distance
similar entities are close together
relationships based on geometric regularities
We are using neural models for predicting yet unknown
links in biomedical knowledge bases
Gene
Drug
Disease
Anatomy
Phenotype
GO Term
Pathway
Relation
Multiple Relation Types
Negative Relation
Ontology Relation
We are using neural models for predicting yet unknown
links in biomedical knowledge bases
SIDER
OpenBioLink
Decision support systems for enabling
personalized medicine
Section for Artificial Intelligence and Decision Support
Matthias Samwald
11
The promise of pharmacogenomics (PGx)
 Genetic polymorphisms in small set of
genes impact pharmacokinetics of a wide
variety of medications
 Hope: Genetic patient data could be used
to make pharmacotherapy safer and more
effective
 Reality: PGx is still rarely used in clinical
routine
Section for Artificial Intelligence and Decision Support
Matthias Samwald
12
Section for Artificial Intelligence and Decision Support
Matthias Samwald
13
The frequency of actionable pharmacogenes. Over 95% of the population
carry at least one actionable genotype for one of the genes covered by
the DPWG guidelines.
We found that patients often receive multiple PGx drugs
over time
Section for Artificial Intelligence and Decision Support
Matthias Samwald
14
Age 40 – 64
(n = 26 561 525)
Age ≥ 65
(n = 5 429 266)
Prescribed one or more PGx
drugs
42,2% 50,6%
Prescribed three or more
PGx drugs
7,5% 13,8%
Expected co-occurence of
drug with high-risk genotype
1,9% 4,1%
Matthias Samwald, Hong Xu, Kathrin Blagec, Philip E. Empey, Daniel C. Malone, Seid Mussa Ahmed, Patrick Ryan, Sebastian Hofer and Richard D. Boyce. “Incidence of
Exposure of Patients in the United States to Multiple Drugs for Which Pharmacogenomic Guidelines Are Available.” PLOS ONE 11, no. 10 (October 20, 2016):
e0164972
Wolfgang Kuch, Christoph Rinner, Walter Gall and Matthias Samwald. “How Many Patients Could Benefit From Pre-Emptive Pharmacogenomic Testing and Decision
Support? A Retrospective Study Based on Nationwide Austrian Claims Data.” Studies in Health Technology and Informatics 223 (2016): 253–58.
Patients in the United States with novel prescriptions of one or more PGx drugs within a
four-year time window
How can we create IT-based workflows for facilitating
clinical PGx?
Section for Artificial Intelligence and Decision Support
Matthias Samwald
15
 Genetic patient data needs to be available easily and quickly
 Ideally, a single test would yield data on all important PGx genes
and data could be re-used in all care processes, without requiring
multiple tests (pre-emptive PGx).
 Clinical decision support algorithms based on up-to-date
guidelines to guide medical practitioners
 Ideal IT solutions
need to be available
across a wide
variety of care settings
We developed a barrier-free system for PGx data access
and decision support
Section for Artificial Intelligence and Decision Support
Matthias Samwald
16
 Created open-source inference engine for PGx decision support based on raw
genetic test results
 Developed a decentralized, mobile-based system for making PGx data and decision
support available across healthcare settings
 „Medication Safety Code“ system
Matthias Samwald, Klaus-Peter Adlassnig. „Pharmacogenomics in the pocket of every patient? A prototype based on Quick Response (QR) codes“ Journal of the American Medical Informatics Association 20, no. 3
(January 5, 2013): 409–412.
Jose Antonio Miñarro-Giménez, Kathrin Blagec, Richard D. Boyce, Klaus-Peter Adlassnig and Matthias Samwald. „An Ontology-Based, Mobile-Optimized System for Pharmacogenomic Decision Support at the Point-of-
Care.” PLoS ONE 9, no. 5 (May 2, 2014): e93769. doi:10.1371/journal.pone.0093769.
We developed a barrier-free, decentralised system for PGx
data access and decision support
Section for Artificial Intelligence and Decision Support
Matthias Samwald
17
The Medication Safety Code system
We developed a barrier-free, decentralised system for PGx
data access and decision support
Section for Artificial Intelligence and Decision Support
Matthias Samwald
18
We developed a barrier-free, decentralised system for PGx
data access and decision support
Section for Artificial Intelligence and Decision Support
Matthias Samwald
19
We conducted a mixed-methods study on system design
Section for Artificial Intelligence and Decision Support
Matthias Samwald
20
• Mixed methods study with PGx experts, clinicians and pharmacists (n
= 114)
• System proved capable of enabling PGx guided decision making
of physicians and pharmacists in fictional patient scenarios
• Majority of participants agreed (57,6) or strongly agreed (9,8%) that
user experience was appealing
• Feedback was used to make further improvements to the system
• System adopted in other research projects (e.g., at Mayo Clinic, US)
Kathrin Blagec, Katrina M. Romagnoli, Richard D. Boyce and Matthias Samwald. “Examining Perceptions of the Usefulness and Usability of a Mobile-Based System for
Pharmacogenomics Clinical Decision Support: A Mixed Methods Study.” PeerJ 4 (February 8, 2016): e1671
The Ubiquitous Pharmacogenomics (U-PGx) H2020
project
Section for Artificial Intelligence and Decision Support
Matthias Samwald
21
Implement and evaluate IT-enabled,
pre-emptive PGx in clinical care
(Implementation study)
Five years duration
€ 15 million budget
10 EU countries
“Multicenter, multi-healthcare
system, multigene, multidrug, multi-
ethnic, multilingual”
Kathrin Blagec, Rudolf Koopmann, Mandy Crommentuijn – van Rhenen, Inge Holsappel, Cathelijne H van der Wouden, Lidija Konta, Hong Xu, Daniela Steinberger, Enrico Just, Jesse J Swen, Henk-Jan Guchelaar
and Matthias Samwald „Implementing Pharmacogenomics Decision Support across Seven European Countries: The Ubiquitous Pharmacogenomics (U-PGx) Project“. Journal of the American Medical
Informatics Association, Feb 9 2018.
The Ubiquitous Pharmacogenomics (U-PGx) H2020
project
Section for Artificial Intelligence and Decision Support
Matthias Samwald
22
Panel of 13 genes
43 drugs of interest
Both inpatients and outpatients
Across all medical specialties
Kathrin Blagec, Rudolf Koopmann, Mandy Crommentuijn – van Rhenen, Inge Holsappel, Cathelijne H van der Wouden, Lidija Konta, Hong Xu, Daniela Steinberger, Enrico Just, Jesse J Swen, Henk-Jan Guchelaar
and Matthias Samwald „Implementing Pharmacogenomics Decision Support across Seven European Countries: The Ubiquitous Pharmacogenomics (U-PGx) Project“. Journal of the American Medical
Informatics Association, Feb 9 2018.
Section for Artificial Intelligence and Decision Support
Matthias Samwald
23
PGx gene Affected drugs
CYP1A2 Clozapine
CYP2B6 Efavirenz
CYP2C19 Warfarin, Citalopram, Escitalopram, Sertraline, Imipramine, Voriconazole
CYP2C9 Phenytoin
CYP2D6 Flecainide, Propafenon, Codeine, Oxycodone, Tramadol, Tamoxifen, Paroxetine, Venlafaxine,
Amitriptyline, Clomipramine, Doxepin, Imipramine, Notriptyline, Metoprolol, Aripiprazole,
Haloperidol, Pimozide, Zuclopenthixol, Atomoxetine
CYP3A5 Tacrolimus
DPD Capecitabine, Fluorouracil, Tegafur
HLA B*5701 Flucloxacillin
SLCO1B1 Atorvastatin, Simvastatin
TPMT Azathioprine, Mercaptopurine, Thioguanine
UGT1A1 Irinotecan
VKORC1 Acenocoumarol, Phenprocoumon, Warfarin
FVL Estrogen containing drugs
Adapted from the guidelines of the Dutch Pharmacogenetics Working group (DPWG). CYP, cytochrome P450; DPD, dihydropyrimidinedehydrogenase;
FVL, factor five Leiden HLA, human leucocyte antigen; SLCO, solute carrier organic anion transporter; TPMT, thiopurine S-methyltransferase; UGT,
UDP-glucuronosyltransferase; VKORC, vitamin K oxide reductase complex.
DPWG classification of clinical effect
Section for Artificial Intelligence and Decision Support
Matthias Samwald
24
We found health informatics capabilities to differ
strongly between institutions and countries
Section for Artificial Intelligence and Decision Support
Matthias Samwald
25
NL GB IT ES AT SI GR
EHR
inpatient
setting
Yes Yes Partially Yes Partially Partially No
EHR
outpatient
setting
Yes Partially Partially Yes No No No
Active CDS
Yes (for
PGx, DDIs,
contraind.)
Yes (for
allergies
and DDI)
No Partially No No No
Passive
CDS
No Yes Yes Yes Yes No No
Structured
laboratory
results
Yes Yes Yes Yes Yes No No
The decision support solutions meet the needs of
clinicians
Section for Artificial Intelligence and Decision Support
Matthias Samwald
26
Strongly
disagree
Disagree Neutral Agree Strongly agree
The CDS tools integrates
well with my work routine
1 2 11 31 6
I feel that I have received
enough training to
confidently use the different
CDS tools.
0 4 13 30 4
Very Dissatisfied Dissatisfied Neutral Satisfied Very Satisfied
Overall satisfaction 1 0 9 34 7
To little Just right Too much
Do the CDS tools provide too much,
too little or just the right amount of
information?
3 39 9
Randomization
U-PGx
Standard
Standard
U-PGx
01-01-2017
01-06-2018
01-01-2020
Approx.
8000
patients
in total
Analysis of clinical outcome
Clinical trial measuring outcome still underway
The CDSS systems developed in U-PGx are also
being made available outside of the project
Section for Artificial Intelligence and Decision Support
Matthias Samwald
28
How can we utilize
Artificial Intelligence
for transformative progress?
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
29
The truth about ‚deep learning‘ is that it makes lots of
people sad
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
30
In late 2018, Google DeepMind
had it‘s debut at the CASP13
protein structure prediction
competition and won against
established players
The truth about ‚deep learning‘ is that it makes lots of
people sad
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
31
AI progress is best steered and measured by public
benchmarks. Progress has been significant in last years.
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
32
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
33
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
34
We observe increasing model generality and algorithmic
progress becoming a public good
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
35
A small set of models and algorithms seems to be applicable to a wide variety of data
modalities
• Sequence (Text? SMILES string? Processes? DNA? Protein?)
• Graph (Social network? Citation network? Molecular structure? Molecular interaction
network? Ontologies and knowledge bases?)
• 2D/3D data array (Photo? Microscopic image? MRI?)
We observe increasing model generality and algorithmic
progress becoming a public good
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
36
Global progress bottlenecked by availability of good (semi-)public train and test sets and
solid benchmarks
Rather than facing ever-increasing model complexity, we keep finding model structures
that are simpler yet more powerful and more generally applicable (even thought they
might be scaled up a lot)
AI is both an enabling ‘General Purpose Technology’
& an ‘Invention of a Method of Invention’
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
37
NO YES
NO
YES
‚Enabling General Purpose Technology‘?
„Invention of
a Method
of Invention“?
Autonomous vehicles
Combustion engine
Artificial Intelligence
Any specific drug
MRI machine
CRISPR
WWW
Unsupervised and multi-task pre-training generates models that
have a general understanding of the phenomena in the world
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
38
• Unsupervised pre-training = ‚Predict everything from everything else‘
• Multi-task training = ‚Train a single model to do many different things‘
• Very recent example: BERT (2018) and BioBERT (2019) architectures achieving state of
the art across many NLP tasks
Unsupervised and multi-task pre-training generates models that
have a general understanding of the phenomena in the world
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
39
Type 2 ________ ________ is characterized by insulin __________
Type 2 diabetes mellitus is characterized by insulin resistance
Fill in the blanks!
Unsupervised and multi-task pre-training generates models that
have a general understanding of the phenomena in the world
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
40
• Train/test algorithms not only on raw data of single modality, but data aggregated across
different modalities
• ‚Multi-omics‘
• Increase spatial and temporal resolution
• Better Integrate raw data / machine learning and higher-level knowledge / knowledge-based
systems
Strategic analysis of meta-layers of progress:
Incremental vs. radical improvements
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
41
What should be prioritized to see most progress 20 years from now?
Today ??? 2040
Section for Artificial Intelligence and Decision Support
Matthias Samwald
42
We should not think only
about AI, but about
hybrid intelligent
systems
Section for Artificial Intelligence and Decision Support
Matthias Samwald
43
R&D processes,
organisations,
societies
Human
Knowledge &
Decision Making
Artificial
Intelligence
Section for Artificial Intelligence and Decision Support, Medical University of Vienna
Assoc.Prof. Dr. Matthias Samwald
44
 AI to aid doctor’s decision making
 AI for creating better drugs for known targets
 AI for fundamental, basic biological research
 AI for developing completely novel research tools and apparatuses
 AI for general human intelligence augmentation (high-precision information
retrieval, analysis, visualization, sharing insights)
 Autonomous AI (closing the intervention-inference loop w/o humans)
 … AI for AI (recursive self-improvement) …
Section for Artificial Intelligence and Decision Support
Matthias Samwald
45
Need to find strategies to apply
AI with
* Greatest leverage
* Greatest versatility
* Lowest risk
The AI Strategies project (work in progress)
Section for Artificial Intelligence and Decision Support
Matthias Samwald
46
Thanks!
This work has received funding from the European Union’s Horizon 2020 research and Innovation programme under grant
agreement No 668353 and the Austrian Science Fund (FWF; P 25608-N15)
Section for Artificial Intelligence and Decision Support
Matthias Samwald
Project members:
Kathrin Blagec
Hong Xu
Anna Breit
http://samwald.info/

More Related Content

PPTX
Artificial Intelligence and Expediting Drug Development
PPTX
Sara Gerke: "AI in Drug Discovery and Clinical Trials"
PDF
Artificial intelligence in medicine (projeck)
PDF
Paradigmshift in precision medicine
PPTX
Thomas Willkens-El impacto de las ciencias ómicas en la medicina, la nutrició...
PPTX
AI applications in life sciences - drug development
PDF
Ai Application in Life Sciences
PDF
A decision support system for tuberculosis
Artificial Intelligence and Expediting Drug Development
Sara Gerke: "AI in Drug Discovery and Clinical Trials"
Artificial intelligence in medicine (projeck)
Paradigmshift in precision medicine
Thomas Willkens-El impacto de las ciencias ómicas en la medicina, la nutrició...
AI applications in life sciences - drug development
Ai Application in Life Sciences
A decision support system for tuberculosis

What's hot (19)

PPTX
iOMICS Clinical & Omnia
PPTX
Digital platforms could disrupts how pharma companies plan and excecute clini...
PDF
PDF
Precision medicine the future of_healthcare
PDF
Healthcare Conference 2013 : Genes, Clouds and Cancer - dr. Andrew Litt
PDF
Big Medical Data – Challenge or Potential?
PDF
NETWORK OF DISEASES AND ITS ENDOWMENT TOWARDS DISEASE
PPTX
Artificial intelligence and its applications in healthcare and pharmacy
PDF
2021 06-14 EATRIS-Plus summer school, Alain van Gool
PPTX
Artificial intelligence in drug discovery and development
PPTX
Emergency Department Triage and Digital Health
PDF
Machine Learning in Medicine A Primer
PDF
Big data, big knowledge big data for personalized healthcare
PPTX
AI in translational medicine webinar
PDF
Clinical genomics spx
PDF
NEURO-FUZZY APPROACH FOR DIAGNOSING AND CONTROL OF TUBERCULOSIS
PDF
Improving health care outcomes with responsible data science #escience2018
PDF
Open Educational Resources for Big Data Science
PPTX
From Experimental to Applied Predictive Analytics on Big Data - Milan Vukicevic
iOMICS Clinical & Omnia
Digital platforms could disrupts how pharma companies plan and excecute clini...
Precision medicine the future of_healthcare
Healthcare Conference 2013 : Genes, Clouds and Cancer - dr. Andrew Litt
Big Medical Data – Challenge or Potential?
NETWORK OF DISEASES AND ITS ENDOWMENT TOWARDS DISEASE
Artificial intelligence and its applications in healthcare and pharmacy
2021 06-14 EATRIS-Plus summer school, Alain van Gool
Artificial intelligence in drug discovery and development
Emergency Department Triage and Digital Health
Machine Learning in Medicine A Primer
Big data, big knowledge big data for personalized healthcare
AI in translational medicine webinar
Clinical genomics spx
NEURO-FUZZY APPROACH FOR DIAGNOSING AND CONTROL OF TUBERCULOSIS
Improving health care outcomes with responsible data science #escience2018
Open Educational Resources for Big Data Science
From Experimental to Applied Predictive Analytics on Big Data - Milan Vukicevic
Ad

Similar to Towards Transformative Artificial Intelligence in Life Science and Health Care (20)

PDF
Day 2: Innovation to optimise therapeutic options for prevention and treatmen...
 
PPTX
The Medication Safety Code initiative: Towards a global IT system for persona...
PPTX
Very brief overview of AI in drug discovery
PPTX
Big Data, AI, and Pharma
PPT
Jack Tuszynski Accelerating Chemotherapy Drug Discovery with Analytics and Hi...
PPT
Personalized Medicine: Current and Future Perspectives Personalized Medicin...
PPTX
Incidence of exposure to p gx drugs rdb
PPTX
BioVariance - Pediatric Pharmacogenomics in Drug Discovery
PDF
Big Data Analytics in the Health Domain
PPTX
Role of bioinformatics and pharmacogenomics in drug discovery
PDF
5.+keynote o donnell
PDF
Krithara meetup may18_final (1)
PPT
Indications discovery and drug repurposing
PPTX
Machine learning in health data analytics and pharmacovigilance
PDF
Henning Müller et Michael Schumacher pour la journée e-health 2013
PPTX
Bridging theory and practice: Clinical decision support systems for personali...
PDF
Elsevier Medical Graph – mit Machine Learning zu Precision Medicine
PPTX
[DSC Europe 23][DigiHealth] Branka Zukic Big Data Personalized medicine
PDF
eHealth unit HES-SO in Sierre
PDF
AI at AZ Festival of Genomics 2025 final.pdf
Day 2: Innovation to optimise therapeutic options for prevention and treatmen...
 
The Medication Safety Code initiative: Towards a global IT system for persona...
Very brief overview of AI in drug discovery
Big Data, AI, and Pharma
Jack Tuszynski Accelerating Chemotherapy Drug Discovery with Analytics and Hi...
Personalized Medicine: Current and Future Perspectives Personalized Medicin...
Incidence of exposure to p gx drugs rdb
BioVariance - Pediatric Pharmacogenomics in Drug Discovery
Big Data Analytics in the Health Domain
Role of bioinformatics and pharmacogenomics in drug discovery
5.+keynote o donnell
Krithara meetup may18_final (1)
Indications discovery and drug repurposing
Machine learning in health data analytics and pharmacovigilance
Henning Müller et Michael Schumacher pour la journée e-health 2013
Bridging theory and practice: Clinical decision support systems for personali...
Elsevier Medical Graph – mit Machine Learning zu Precision Medicine
[DSC Europe 23][DigiHealth] Branka Zukic Big Data Personalized medicine
eHealth unit HES-SO in Sierre
AI at AZ Festival of Genomics 2025 final.pdf
Ad

More from Matthias Samwald (10)

PPTX
CRISPR as a potential tool for malaria eradication
PPTX
One man's *1 is another man's *13? Trouble with nomenclatures in personalized...
PPT
VO Taxonomie und Ontologie (SS 2016)
PPTX
6 Worlds Collide
PPTX
The FindMeEvidence project: An open-source, mobile-friendly search engine for...
PPTX
Samwald ore 2014
PPTX
Hcls call september 2013 clinical pharmacogenomics
PDF
Medinfo2013 - An RDF/OWL Knowledge Base for Query Answering and Decision Supp...
PDF
Genomic CDS: an example of a complex ontology for pharmacogenetics and clinic...
PDF
Samwald cshals2013
CRISPR as a potential tool for malaria eradication
One man's *1 is another man's *13? Trouble with nomenclatures in personalized...
VO Taxonomie und Ontologie (SS 2016)
6 Worlds Collide
The FindMeEvidence project: An open-source, mobile-friendly search engine for...
Samwald ore 2014
Hcls call september 2013 clinical pharmacogenomics
Medinfo2013 - An RDF/OWL Knowledge Base for Query Answering and Decision Supp...
Genomic CDS: an example of a complex ontology for pharmacogenetics and clinic...
Samwald cshals2013

Recently uploaded (20)

PPT
Geologic Time for studying geology for geologist
PPTX
2018-HIPAA-Renewal-Training for executives
PDF
A review of recent deep learning applications in wood surface defect identifi...
PDF
A contest of sentiment analysis: k-nearest neighbor versus neural network
PPT
Galois Field Theory of Risk: A Perspective, Protocol, and Mathematical Backgr...
PDF
sustainability-14-14877-v2.pddhzftheheeeee
PDF
STKI Israel Market Study 2025 version august
PDF
From MVP to Full-Scale Product A Startup’s Software Journey.pdf
PPTX
Chapter 5: Probability Theory and Statistics
PDF
Architecture types and enterprise applications.pdf
PPTX
Microsoft Excel 365/2024 Beginner's training
PDF
Enhancing emotion recognition model for a student engagement use case through...
PDF
Taming the Chaos: How to Turn Unstructured Data into Decisions
PDF
sbt 2.0: go big (Scala Days 2025 edition)
PDF
How ambidextrous entrepreneurial leaders react to the artificial intelligence...
PPT
Module 1.ppt Iot fundamentals and Architecture
PDF
A comparative study of natural language inference in Swahili using monolingua...
PPTX
Benefits of Physical activity for teenagers.pptx
PPTX
The various Industrial Revolutions .pptx
PDF
TrustArc Webinar - Click, Consent, Trust: Winning the Privacy Game
Geologic Time for studying geology for geologist
2018-HIPAA-Renewal-Training for executives
A review of recent deep learning applications in wood surface defect identifi...
A contest of sentiment analysis: k-nearest neighbor versus neural network
Galois Field Theory of Risk: A Perspective, Protocol, and Mathematical Backgr...
sustainability-14-14877-v2.pddhzftheheeeee
STKI Israel Market Study 2025 version august
From MVP to Full-Scale Product A Startup’s Software Journey.pdf
Chapter 5: Probability Theory and Statistics
Architecture types and enterprise applications.pdf
Microsoft Excel 365/2024 Beginner's training
Enhancing emotion recognition model for a student engagement use case through...
Taming the Chaos: How to Turn Unstructured Data into Decisions
sbt 2.0: go big (Scala Days 2025 edition)
How ambidextrous entrepreneurial leaders react to the artificial intelligence...
Module 1.ppt Iot fundamentals and Architecture
A comparative study of natural language inference in Swahili using monolingua...
Benefits of Physical activity for teenagers.pptx
The various Industrial Revolutions .pptx
TrustArc Webinar - Click, Consent, Trust: Winning the Privacy Game

Towards Transformative Artificial Intelligence in Life Science and Health Care

  • 1. Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 1 Towards Transformative Artificial Intelligence in Life Science and Health Care Assoc.Prof. Mag. Dr. Matthias Samwald Section for Artificial Intelligence and Decision Support Medical University of Vienna
  • 2. Assoc.Prof. Dr. Matthias Samwald Section for Artificial Intelligence and Decision Support, Medical University of Vienna Medical University of Vienna • One of the largest hospitals in Europe • 25 university clinics • 8000 students 2
  • 3. Assoc.Prof. Dr. Matthias Samwald Section for Artificial Intelligence and Decision Support, Medical University of Vienna Section for Artificial Intelligence and Decision Support • History dating back to 1977 • Multidisciplinary work on AI, medicine, biology, cognitive science 3
  • 4. „From data to knowledge to action“ Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 4 Diagnosing diseases based on lab values or images Natural language processing • Identifying key statements and their semantic relations in biomedical texts • Deep learning for biomedical text summarization • Creating high-precision biomedical search engines • Re-purposing deep NLP models for biological sequences (proteins) Predicting novel links in large biomedical knowledge bases Decision support systems for enabling personalized medicine Strategic considerations for harnessing AI for global progress Raw data Knowledge Action
  • 5. „From data to knowledge to action“ Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 5 Diagnosing diseases based on lab values or images Natural language processing • Identifying key statements and their semantic relations in biomedical texts • Deep learning for biomedical text summarization • Creating high-precision biomedical search engines • Re-purposing deep NLP models for biological sequences (proteins) Predicting novel links in large biomedical knowledge bases Decision support systems for enabling personalized medicine Strategic considerations for harnessing AI for global progress Raw data Knowledge Action
  • 6. Predicting novel links in large biomedical knowledge bases Section for Artificial Intelligence and Decision Support Matthias Samwald 6
  • 7. We are using neural models for predicting yet unknown links in biomedical knowledge bases
  • 8. We are using neural models for predicting yet unknown links in biomedical knowledge bases Knowledge graph embeddings • low dimensional • vector space • certain properties can be represented by notions of distance similar entities are close together relationships based on geometric regularities
  • 9. We are using neural models for predicting yet unknown links in biomedical knowledge bases
  • 10. Gene Drug Disease Anatomy Phenotype GO Term Pathway Relation Multiple Relation Types Negative Relation Ontology Relation We are using neural models for predicting yet unknown links in biomedical knowledge bases SIDER OpenBioLink
  • 11. Decision support systems for enabling personalized medicine Section for Artificial Intelligence and Decision Support Matthias Samwald 11
  • 12. The promise of pharmacogenomics (PGx)  Genetic polymorphisms in small set of genes impact pharmacokinetics of a wide variety of medications  Hope: Genetic patient data could be used to make pharmacotherapy safer and more effective  Reality: PGx is still rarely used in clinical routine Section for Artificial Intelligence and Decision Support Matthias Samwald 12
  • 13. Section for Artificial Intelligence and Decision Support Matthias Samwald 13 The frequency of actionable pharmacogenes. Over 95% of the population carry at least one actionable genotype for one of the genes covered by the DPWG guidelines.
  • 14. We found that patients often receive multiple PGx drugs over time Section for Artificial Intelligence and Decision Support Matthias Samwald 14 Age 40 – 64 (n = 26 561 525) Age ≥ 65 (n = 5 429 266) Prescribed one or more PGx drugs 42,2% 50,6% Prescribed three or more PGx drugs 7,5% 13,8% Expected co-occurence of drug with high-risk genotype 1,9% 4,1% Matthias Samwald, Hong Xu, Kathrin Blagec, Philip E. Empey, Daniel C. Malone, Seid Mussa Ahmed, Patrick Ryan, Sebastian Hofer and Richard D. Boyce. “Incidence of Exposure of Patients in the United States to Multiple Drugs for Which Pharmacogenomic Guidelines Are Available.” PLOS ONE 11, no. 10 (October 20, 2016): e0164972 Wolfgang Kuch, Christoph Rinner, Walter Gall and Matthias Samwald. “How Many Patients Could Benefit From Pre-Emptive Pharmacogenomic Testing and Decision Support? A Retrospective Study Based on Nationwide Austrian Claims Data.” Studies in Health Technology and Informatics 223 (2016): 253–58. Patients in the United States with novel prescriptions of one or more PGx drugs within a four-year time window
  • 15. How can we create IT-based workflows for facilitating clinical PGx? Section for Artificial Intelligence and Decision Support Matthias Samwald 15  Genetic patient data needs to be available easily and quickly  Ideally, a single test would yield data on all important PGx genes and data could be re-used in all care processes, without requiring multiple tests (pre-emptive PGx).  Clinical decision support algorithms based on up-to-date guidelines to guide medical practitioners  Ideal IT solutions need to be available across a wide variety of care settings
  • 16. We developed a barrier-free system for PGx data access and decision support Section for Artificial Intelligence and Decision Support Matthias Samwald 16  Created open-source inference engine for PGx decision support based on raw genetic test results  Developed a decentralized, mobile-based system for making PGx data and decision support available across healthcare settings  „Medication Safety Code“ system Matthias Samwald, Klaus-Peter Adlassnig. „Pharmacogenomics in the pocket of every patient? A prototype based on Quick Response (QR) codes“ Journal of the American Medical Informatics Association 20, no. 3 (January 5, 2013): 409–412. Jose Antonio Miñarro-Giménez, Kathrin Blagec, Richard D. Boyce, Klaus-Peter Adlassnig and Matthias Samwald. „An Ontology-Based, Mobile-Optimized System for Pharmacogenomic Decision Support at the Point-of- Care.” PLoS ONE 9, no. 5 (May 2, 2014): e93769. doi:10.1371/journal.pone.0093769.
  • 17. We developed a barrier-free, decentralised system for PGx data access and decision support Section for Artificial Intelligence and Decision Support Matthias Samwald 17 The Medication Safety Code system
  • 18. We developed a barrier-free, decentralised system for PGx data access and decision support Section for Artificial Intelligence and Decision Support Matthias Samwald 18
  • 19. We developed a barrier-free, decentralised system for PGx data access and decision support Section for Artificial Intelligence and Decision Support Matthias Samwald 19
  • 20. We conducted a mixed-methods study on system design Section for Artificial Intelligence and Decision Support Matthias Samwald 20 • Mixed methods study with PGx experts, clinicians and pharmacists (n = 114) • System proved capable of enabling PGx guided decision making of physicians and pharmacists in fictional patient scenarios • Majority of participants agreed (57,6) or strongly agreed (9,8%) that user experience was appealing • Feedback was used to make further improvements to the system • System adopted in other research projects (e.g., at Mayo Clinic, US) Kathrin Blagec, Katrina M. Romagnoli, Richard D. Boyce and Matthias Samwald. “Examining Perceptions of the Usefulness and Usability of a Mobile-Based System for Pharmacogenomics Clinical Decision Support: A Mixed Methods Study.” PeerJ 4 (February 8, 2016): e1671
  • 21. The Ubiquitous Pharmacogenomics (U-PGx) H2020 project Section for Artificial Intelligence and Decision Support Matthias Samwald 21 Implement and evaluate IT-enabled, pre-emptive PGx in clinical care (Implementation study) Five years duration € 15 million budget 10 EU countries “Multicenter, multi-healthcare system, multigene, multidrug, multi- ethnic, multilingual” Kathrin Blagec, Rudolf Koopmann, Mandy Crommentuijn – van Rhenen, Inge Holsappel, Cathelijne H van der Wouden, Lidija Konta, Hong Xu, Daniela Steinberger, Enrico Just, Jesse J Swen, Henk-Jan Guchelaar and Matthias Samwald „Implementing Pharmacogenomics Decision Support across Seven European Countries: The Ubiquitous Pharmacogenomics (U-PGx) Project“. Journal of the American Medical Informatics Association, Feb 9 2018.
  • 22. The Ubiquitous Pharmacogenomics (U-PGx) H2020 project Section for Artificial Intelligence and Decision Support Matthias Samwald 22 Panel of 13 genes 43 drugs of interest Both inpatients and outpatients Across all medical specialties Kathrin Blagec, Rudolf Koopmann, Mandy Crommentuijn – van Rhenen, Inge Holsappel, Cathelijne H van der Wouden, Lidija Konta, Hong Xu, Daniela Steinberger, Enrico Just, Jesse J Swen, Henk-Jan Guchelaar and Matthias Samwald „Implementing Pharmacogenomics Decision Support across Seven European Countries: The Ubiquitous Pharmacogenomics (U-PGx) Project“. Journal of the American Medical Informatics Association, Feb 9 2018.
  • 23. Section for Artificial Intelligence and Decision Support Matthias Samwald 23 PGx gene Affected drugs CYP1A2 Clozapine CYP2B6 Efavirenz CYP2C19 Warfarin, Citalopram, Escitalopram, Sertraline, Imipramine, Voriconazole CYP2C9 Phenytoin CYP2D6 Flecainide, Propafenon, Codeine, Oxycodone, Tramadol, Tamoxifen, Paroxetine, Venlafaxine, Amitriptyline, Clomipramine, Doxepin, Imipramine, Notriptyline, Metoprolol, Aripiprazole, Haloperidol, Pimozide, Zuclopenthixol, Atomoxetine CYP3A5 Tacrolimus DPD Capecitabine, Fluorouracil, Tegafur HLA B*5701 Flucloxacillin SLCO1B1 Atorvastatin, Simvastatin TPMT Azathioprine, Mercaptopurine, Thioguanine UGT1A1 Irinotecan VKORC1 Acenocoumarol, Phenprocoumon, Warfarin FVL Estrogen containing drugs Adapted from the guidelines of the Dutch Pharmacogenetics Working group (DPWG). CYP, cytochrome P450; DPD, dihydropyrimidinedehydrogenase; FVL, factor five Leiden HLA, human leucocyte antigen; SLCO, solute carrier organic anion transporter; TPMT, thiopurine S-methyltransferase; UGT, UDP-glucuronosyltransferase; VKORC, vitamin K oxide reductase complex.
  • 24. DPWG classification of clinical effect Section for Artificial Intelligence and Decision Support Matthias Samwald 24
  • 25. We found health informatics capabilities to differ strongly between institutions and countries Section for Artificial Intelligence and Decision Support Matthias Samwald 25 NL GB IT ES AT SI GR EHR inpatient setting Yes Yes Partially Yes Partially Partially No EHR outpatient setting Yes Partially Partially Yes No No No Active CDS Yes (for PGx, DDIs, contraind.) Yes (for allergies and DDI) No Partially No No No Passive CDS No Yes Yes Yes Yes No No Structured laboratory results Yes Yes Yes Yes Yes No No
  • 26. The decision support solutions meet the needs of clinicians Section for Artificial Intelligence and Decision Support Matthias Samwald 26 Strongly disagree Disagree Neutral Agree Strongly agree The CDS tools integrates well with my work routine 1 2 11 31 6 I feel that I have received enough training to confidently use the different CDS tools. 0 4 13 30 4 Very Dissatisfied Dissatisfied Neutral Satisfied Very Satisfied Overall satisfaction 1 0 9 34 7 To little Just right Too much Do the CDS tools provide too much, too little or just the right amount of information? 3 39 9
  • 28. The CDSS systems developed in U-PGx are also being made available outside of the project Section for Artificial Intelligence and Decision Support Matthias Samwald 28
  • 29. How can we utilize Artificial Intelligence for transformative progress? Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 29
  • 30. The truth about ‚deep learning‘ is that it makes lots of people sad Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 30 In late 2018, Google DeepMind had it‘s debut at the CASP13 protein structure prediction competition and won against established players
  • 31. The truth about ‚deep learning‘ is that it makes lots of people sad Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 31
  • 32. AI progress is best steered and measured by public benchmarks. Progress has been significant in last years. Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 32
  • 33. Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 33
  • 34. Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 34
  • 35. We observe increasing model generality and algorithmic progress becoming a public good Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 35 A small set of models and algorithms seems to be applicable to a wide variety of data modalities • Sequence (Text? SMILES string? Processes? DNA? Protein?) • Graph (Social network? Citation network? Molecular structure? Molecular interaction network? Ontologies and knowledge bases?) • 2D/3D data array (Photo? Microscopic image? MRI?)
  • 36. We observe increasing model generality and algorithmic progress becoming a public good Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 36 Global progress bottlenecked by availability of good (semi-)public train and test sets and solid benchmarks Rather than facing ever-increasing model complexity, we keep finding model structures that are simpler yet more powerful and more generally applicable (even thought they might be scaled up a lot)
  • 37. AI is both an enabling ‘General Purpose Technology’ & an ‘Invention of a Method of Invention’ Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 37 NO YES NO YES ‚Enabling General Purpose Technology‘? „Invention of a Method of Invention“? Autonomous vehicles Combustion engine Artificial Intelligence Any specific drug MRI machine CRISPR WWW
  • 38. Unsupervised and multi-task pre-training generates models that have a general understanding of the phenomena in the world Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 38 • Unsupervised pre-training = ‚Predict everything from everything else‘ • Multi-task training = ‚Train a single model to do many different things‘ • Very recent example: BERT (2018) and BioBERT (2019) architectures achieving state of the art across many NLP tasks
  • 39. Unsupervised and multi-task pre-training generates models that have a general understanding of the phenomena in the world Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 39 Type 2 ________ ________ is characterized by insulin __________ Type 2 diabetes mellitus is characterized by insulin resistance Fill in the blanks!
  • 40. Unsupervised and multi-task pre-training generates models that have a general understanding of the phenomena in the world Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 40 • Train/test algorithms not only on raw data of single modality, but data aggregated across different modalities • ‚Multi-omics‘ • Increase spatial and temporal resolution • Better Integrate raw data / machine learning and higher-level knowledge / knowledge-based systems
  • 41. Strategic analysis of meta-layers of progress: Incremental vs. radical improvements Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 41 What should be prioritized to see most progress 20 years from now? Today ??? 2040
  • 42. Section for Artificial Intelligence and Decision Support Matthias Samwald 42 We should not think only about AI, but about hybrid intelligent systems
  • 43. Section for Artificial Intelligence and Decision Support Matthias Samwald 43 R&D processes, organisations, societies Human Knowledge & Decision Making Artificial Intelligence
  • 44. Section for Artificial Intelligence and Decision Support, Medical University of Vienna Assoc.Prof. Dr. Matthias Samwald 44  AI to aid doctor’s decision making  AI for creating better drugs for known targets  AI for fundamental, basic biological research  AI for developing completely novel research tools and apparatuses  AI for general human intelligence augmentation (high-precision information retrieval, analysis, visualization, sharing insights)  Autonomous AI (closing the intervention-inference loop w/o humans)  … AI for AI (recursive self-improvement) …
  • 45. Section for Artificial Intelligence and Decision Support Matthias Samwald 45 Need to find strategies to apply AI with * Greatest leverage * Greatest versatility * Lowest risk
  • 46. The AI Strategies project (work in progress) Section for Artificial Intelligence and Decision Support Matthias Samwald 46
  • 47. Thanks! This work has received funding from the European Union’s Horizon 2020 research and Innovation programme under grant agreement No 668353 and the Austrian Science Fund (FWF; P 25608-N15) Section for Artificial Intelligence and Decision Support Matthias Samwald Project members: Kathrin Blagec Hong Xu Anna Breit http://samwald.info/

Editor's Notes

  • #15: We conducted an analysis of large, detailed prescription datasets to see how frequent PGx medications were co-prescribed Total US patient records: 73 M Total Austrian patients: 6,8 M
  • #28: This diagram shows the outline of the trial as well as the time-frame Centers each have their own therapeutic focus  oncology, renal tx
  • #36: David ’s (1990) foundational study of the electric motor showed that this invention brought about enormous technological and organizational change across sectors as diverse as manufacturing, agriculture, retail, and residential construction. Such “GPTs” are usual ly understood to meet three criteria that distinguish them from other innovations: they have pervasive application across many sectors; they spawn further innovation in application sectors, and they themselves are rapidly improving. Some of these advances appear to have great potential across a broad set of domains , beyond their initial application : a s highlighted by Griliches (1957) in his classic studies of hybrid corn, some new research tools are inventions that do not just create or improve a specific product — instead they constitute a new way of creating new products, w ith much broader application. In Griliches ’ famous construction, the discovery of double - cross hybridization “ was the invention of a method of inventin g .” (Hereinafter, “IMI”.) Rather than being a means of creating a single a new corn variety , hybrid corn represented a widely applicable me thod for breeding many different new varieties . When a pplied to the challenge of creating new varieties optimized for m any different localities (and even more broadly, to other crops) the invention of double - cross hybridization had a huge impact on agricultural productivity. One the one hand, AI based learning may be able to substantially “automate discovery” across many domains where classification and prediction tasks play an important role. On the other , they m ay also “ expand the playbook” is the sense of opening up the set of problem s that can be feasib ly addressed, and radically alter ing scientific and technical communities’ conceptual approaches and framing of problems. The invention of optical lenses in the 17 th century had important direct economic impact in applications such as spectacles. But optical lenses in the form of microscopes and telescopes also had enormous and long - lasting indirect effects on the progress of science, technological change, growth, and welfare: b y making very small or very dista nt objects visible for the first time, lenses opened up entirely new domains of inquiry and technological opportunity.
  • #37: David ’s (1990) foundational study of the electric motor showed that this invention brought about enormous technological and organizational change across sectors as diverse as manufacturing, agriculture, retail, and residential construction. Such “GPTs” are usual ly understood to meet three criteria that distinguish them from other innovations: they have pervasive application across many sectors; they spawn further innovation in application sectors, and they themselves are rapidly improving. Some of these advances appear to have great potential across a broad set of domains , beyond their initial application : a s highlighted by Griliches (1957) in his classic studies of hybrid corn, some new research tools are inventions that do not just create or improve a specific product — instead they constitute a new way of creating new products, w ith much broader application. In Griliches ’ famous construction, the discovery of double - cross hybridization “ was the invention of a method of inventin g .” (Hereinafter, “IMI”.) Rather than being a means of creating a single a new corn variety , hybrid corn represented a widely applicable me thod for breeding many different new varieties . When a pplied to the challenge of creating new varieties optimized for m any different localities (and even more broadly, to other crops) the invention of double - cross hybridization had a huge impact on agricultural productivity. One the one hand, AI based learning may be able to substantially “automate discovery” across many domains where classification and prediction tasks play an important role. On the other , they m ay also “ expand the playbook” is the sense of opening up the set of problem s that can be feasib ly addressed, and radically alter ing scientific and technical communities’ conceptual approaches and framing of problems. The invention of optical lenses in the 17 th century had important direct economic impact in applications such as spectacles. But optical lenses in the form of microscopes and telescopes also had enormous and long - lasting indirect effects on the progress of science, technological change, growth, and welfare: b y making very small or very dista nt objects visible for the first time, lenses opened up entirely new domains of inquiry and technological opportunity.
  • #38: David ’s (1990) foundational study of the electric motor showed that this invention brought about enormous technological and organizational change across sectors as diverse as manufacturing, agriculture, retail, and residential construction. Such “GPTs” are usual ly understood to meet three criteria that distinguish them from other innovations: they have pervasive application across many sectors; they spawn further innovation in application sectors, and they themselves are rapidly improving. Some of these advances appear to have great potential across a broad set of domains , beyond their initial application : a s highlighted by Griliches (1957) in his classic studies of hybrid corn, some new research tools are inventions that do not just create or improve a specific product — instead they constitute a new way of creating new products, w ith much broader application. In Griliches ’ famous construction, the discovery of double - cross hybridization “ was the invention of a method of inventin g .” (Hereinafter, “IMI”.) Rather than being a means of creating a single a new corn variety , hybrid corn represented a widely applicable me thod for breeding many different new varieties . When a pplied to the challenge of creating new varieties optimized for m any different localities (and even more broadly, to other crops) the invention of double - cross hybridization had a huge impact on agricultural productivity.