High Profile Call Girls Jaipur Vani 8445551418 Independent Escort Service Jaipur
ITU/WHO Focus Group Introduction
1. FGAI4H-M-002
E-meeting, 28-30 September 2021
Source:
Title:
Purpose:
Contact:
Abstract:
Chairman FG-AI4H
Introduction to ITU/WHO Focus Group on AI for Health (FG-AI4H)
Information
Wiegand, Thomas E-mail: thomas.wiegand@hhi.fraunhofer.de
This PPT contains an overview of the history, goals and activities of the
ITU/WHO Focus Group on AI for Health (FG-AI4H).
2. ITU/WHO Focus Group on
Artificial Intelligence for Health
Thomas Wiegand
Fraunhofer HHI & TU Berlin, Germany
Funding support by:
3. • Standardization of health AI needed to facilitate safe use on a local and global scale
• Health is much more complex and critical than other areas where AI has been successfully
implemented (e.g., advertisement, entertainment, …)
• Focus Group on AI for Health established in 2018 as ITU/WHO collaboration
to address challenges related to AI and health
AI offers large potential for
public & clinical health
4. • Benefits through automation with AI
• Examples for microbiology/hygiene: lab automation, automated microscopy in
high-throughput screenings, classification of genomic and protein data, ...
• Shortage of qualified healthcare professionals esp. in low/middle income
countries – extension of scope of practice
Source: WHO
+
AI offers large potential for
public & clinical health
5. Challenge: Healthcare provision & regulation
is a multidisciplinary & complex undertaking
Medicine
AI/ML
Public health
Government
Regulation
Ethics
Research in
academia/
industry
6. Challenges: trustworthy AI
Models must be trustworthy
• Proven good performance
• Tested with consistent validation criteria, metrics, data
• Robust, explainable, can quantify uncertainty
• Effective + safe in the application
Training + test data
• Must be representative to avoid bias
• Reflect population characteristics
(regional, gender, and age variations)
• Reflect differences in equipment and hospitals
• Of sufficient quantity; bottleneck: annotated data →
• Reproducible test data collection and annotation
Demonstration of web-based annotations tool; Prof. Klauschen (LMU & Charité)
7. Challenges: complex reality
Simple assumption:
• Train once and use (linear deployment)
• Fixed input-output relationship
(stationarity assumption)
Reality of AI:
• Cyclical with updates
• Non-stationarity (data nor input-output)
Data Train Test Apply
New data
Retraining
Retesting
Application
Initial AI System
8. About the ITU/WHO focus group
• Members: Experts from around the globe
• Leadership:
Thomas Wiegand, Fraunhofer HHI & TU Berlin, Germany (chair)
Shan Xu, CAICT, China (vice-chair)
Stephen Ibaraki, ACM, Canada (vice-chair)
Naomi Lee, The Lancet, United Kingdom (vice-chair)
Sameer Pujari, WHO (vice-chair)
Manjula Singh, ICMR, India (vice-chair)
Ramesh Krishnamurthy, WHO (vice-chair)
9. Starting a global dialogue
on AI & health: Meetings & Workshops
5/2018 11/2018 4/2019 9/2019 1/2020 5/2020 1/2021 9/2021
AI for Good
Geneva
Columbia University
NYC
World Expo
Shanghai
UCSAF
Zanzibar
PAHO/WHO
Brasilia
Online Online Online
10/2018 1/2019 5/2019 11/2019 3/2020 9/2020 5/2021
WHOHQ EPFL AI for Good ICMR & NICF AI in Online Online
Geneva Lausanne Geneva New Delhi Singapore
10. Matrix structure in working x topic groups
WG A
WG B
WG C
…
TG 1
ITU/WHO
Focus Group
AI for Health
Standards for Health AI
…
TG 2 TG 3
TG = topic group
WG = working group
11. Topic groups: dedicated to specific use cases
• Big variety of specific health use cases in the context of AI
• Aim at producing evidence and case studies and bringing together experts and data
• Propose procedures to benchmark AI models for a given task within a health topic
12. Working Group - Ethics
• Coordinator: Andreas Reis (WHO; chair WG-Ethics of ITU/WHO FGAI4H)
• WHO ethics experts published 'Ethics and governance of AI for health' in July '21
• Definition of 6 principles for guidance of the design, development, deployment of any AI4 health technology
13. WG - Regulatory Considerations
• Collaboration: members of FDA (USA), EMA (EU), BfArM (Germany), National Medical Products
Administration (China), CDSCO (India), WHO, etc.
• Objectives: Assist FG-AI4H to navigate the regulatory landscape
• Outline key regulatory considerations with relevance for regulatory agencies for AI development
14. WG - Clinical Evaluation
• Collaboration: Clinicians, research, academia, NGO's, commissioning, WHO
• Objectives: Produce guidance for current best-practice evaluation of AI for health
intended for use by researchers, clinicians, patients, developers, policy-makers
• Facilitate adoption of AI health technologies that are safe, effective, cost-effective
• Special attention to needs of LMICs
15. WG - DASH
• Data & AI Solution Handling (DASH)
• Objectives:
• Consider operational aspects of data processing throughout the data
lifecycle (e.g., dataset submission,- transfer, maintenance, algorithm
benchmarking)
• Specify “how” FG-AI4H should perform operations involving data
• Co-initiators of FG-AI4H Open Code Initiative
16. WG - DAISAM
• Data & AI solution assessment methods (DAISAM)
• Study/standardization of analytical test procedures along AI life cycle
(preparation, training, validation, deployment)
• Leads FG-AI4H's "AI auditing" initiative
17. Output: documentation & software
No. Deliverable title
0 Overview of the FG-AI4H deliverables
1 AI4H ethics considerations
2 AI4H regulatory considerations
3 AI4H requirements specifications
4 AI4H software life cycle specification
5 Data specification
6 AI training best practices specification
7 AI4H evaluation considerations
8 AI4H scale-up and adoption
9 AI4H applications and platforms
10 AI4H use cases: 20+ topic description documents
Software being developed by
Open Code Initiative:
• Benchmarking platform
• Annotation tools
• AI4H auditing
Documentation:
18. Open Code Initiative (by FGAI4H's WG-DASH)
• Develop benchmarking platform with standardized test procedures and metrics
on high-quality, representative, and undisclosed test data
• Develop software tools (e.g., data acquisition, data storage, annotation,
prediction, evaluation, and reporting packages)
• Involve developers, regulators, and medical professionals
• Usable by multiple stakeholders such as notified bodies and doctors
19. • Open-source reference implementation of a benchmarking platform for health AI
• Will support entire assessment process incl. ground truth annotation, (meta) data
management, reporting
• Targeted towards a universal tool applicable across borders
• https://dev.azure.com/mllabai/fg-ai4h & https://github.com/fg-ai4h
Test data
acquisition
Test data
storage
Test data
annotation
Prediction
by AI
Assessment of
AI prediction
Report of
results
Open Code Initiative (by FGAI4H's WG-DASH)
20. Data sourcing (by FGAI4H's WG-DASH)
• AI developers/evaluators face a lack of high-quality health data pools
• FG-AI4H is coordinating compilation of data in regional centers (e.g., for region-
wise generalizability testing of AI, but also federated learning)
21. WG-DAISAM: ML4H auditing
Auditing framework proposed by FG-AI4H:
• Best practices for AI auditing & quality control along entire AI life cycle
• Identify & define methods for data & AI assessment
• Verification & validation of technical/clinical/regulatory/ethical
requirements for AI following a structured audit process
WG-DAISAM tested validation protocol on 3 AI models (teams from
USA/IN/CH/DE): (1) Diagnostic prediction of diabetic retinopathy, (2) of
Alzheimer’s disease, (3) cytomorphologic classification for leukemia
diagnostics
Results of trial audits: presented at ML4H@NeurIPS 2020
& SAIL-conference in October 2021
Editorial under review in Journal of Medical Systems: intro to special
collection "Machine learning for health: algorithm auditing & quality control"
22. Ad-Hoc Group: Digital Technologies
for COVID Health Emergency
• Objective: Studying how AI can be leveraged throughout the cycle of an epidemic emergency
• Prevention & preparedness, outbreak early detection, surveillance & response, recovery, mitigation
• WHO Hub for Pandemic and Epidemic Intelligence (Berlin)
Sources: ITU & WHO
23. AI4H-Webinars & Challenges
Webinars (within ITU AI for Good Global Summit)
• 15 September: Effy Vayena (ETH Zurich)
• 22 September: David Shaywitz (Astounding HT)
• 7 October: Ziad Obermeyer (UC Berkeley)
• 21 October: Eric Horvitz (Microsoft)
• 8 November: Maarten van Smeden (Maastricht
U) & Laure Wynants (U Utrecht)
• 6 December: Enzo Ferrante (Argentina's National
Research Council)
• 17 December: Nigam Shah (Stanford University)
Forthcoming "AI4H-challenges":
• Public Health (Berlin/Oxford/Harvard)
• Platform of Open Code Initiative
24. Publications
• WHO and ITU Establish Benchmarking Process for AI in Health. The Lancet, 2019.
https://doi.org/10.1016/S0140-6736(19)30762-7
• Whitepaper for the ITU/WHO Focus Group on AI for Health, FG-AI4H, 2020.
https://itu.int/go/fgai4h/whitepaper
• Toward Global Validation Standards for Health AI. IEEE CSM, 2020.
https://doi.org/10.1109/MCOMSTD.001.2000006
• ML4H Auditing: From Paper to Practice. PMLR, 2020. In Machine Learning for Health. [Link]
• Full documentation on http://itu.int/go/fgai4h/collab
25. Get involved!
• Visit https://www.itu.int/go/fgai4h/
• Participate in meeting 28-30 September
• Join mailing list, topic/working groups