[DSC Europe 23][DigiHealth] Dimitrios Kalogeropoulos A Sustainable Future for Learning AI in Healthcare

Dr Dimitrios Kalogeropoulos
Chief Executive, Global Health & Digital Innovation Foundation, UK
Health Executive in Residence, UCL Global Business School For Health, UK
A Sustainable Future for
Open Learning AI in Healthcare
2003-2023
Belgrade DSC - DigiHealth, 23 November 2023

Risks With (Open Learning) AI, 2003-2023
• AI may restrict, discriminate and exclude patients from treatment options
e.g., RCT inclusion and exclusion criteria or mammography s/w
• AI may be the perpetrator in privacy and security attacks
• Deep fakes— the manipulation of society
• Adversarial fakes— data poisoning and collateral damage as in exclusion
• Threat to Digital sovereignty— ability to directly control the Digital Twin
of a jurisdiction
• Threat to Sovereign economies— ability to control the digital economy
• Misuse, power concentration
• Climate impact— direct (by AI/LLMs) and indirect (over-diagnosis)
Underdeveloped, unvalidated
or misappropriated data
assets
Trivial, fake, or biased data
relationships may be
recovered or “planted” by
to exacerbate under or over-
diagnosis and treatment
Dr Dimitrios Kalogeropoulos 2
Kalogeropoulos DA, Carson ER, Collinson PO. Towards knowledge-based systems in clinical practice: development of an integrated
clinical information and knowledge management support system.
Comput Methods Programs Biomed. 2003 Sep;72(1):65-80. doi: 10.1016/s0169-2607(02)00118-9.
Kalogeropoulos, D. and Barach, P. (2023). Telehealth’s Role Enabling Sustainable Innovation and Circular Economies in Health.
Telehealth and Medicine Today, 8(1). https://doi.org/10.30953/thmt.v8.409
Interwoven Realms: Data Governance as the Bedrock for AI Governance, By Stefaan G. Verhulst and Friederike Schüür -
https://medium.com/data-policy/interwoven-realms-data-governance-as-the-bedrock-for-ai-governance-ffd56a6a4543

State Of The Art – then back 2 steps
• We made significant progress from 2003 to 2023 - where?
• AI holds important potential for meaningful health care innovation, but suffers a poor adoption ecosystem
• Radiology accounts for 70% of AI approvals by the FDA as Medical Devices (October 2023 update had 155
new premarket notifications)1
• Countless research projects out there2
• Yet, a 2020 systematic review3 of more than 8,000 supervised deep learning studies reported to focus on
comparing with expert clinicians AI performance in medical imaging, found:
• Only 10 RCTs – 8 were ongoing, 9-China, 1-US
• 81 retrospective non-randomised clinical trials identified
• Only 9 prospective; just 6 tested in a real-world clinical setting
• High-quality studies are lacking
1. FDA AI/ML-Enabled Medical Devices, https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices | 2. Artificial Intelligence Applications for
Biomedical Cancer Research- A Review. DOI: 10.7759/cureus.48307 | 3. Artificial intelligence versus clinicians: systematic review of design, reporting standards, and claims of deep learning studies.
htttp://dx.doi.org/10.1136/bmj.m689

Path to Sustainable Adoption— Focus
The gatekeeper to adoption of AI in clinical settings are clinical trials. We never get there; why?
• Clinical value is not necessarily from distal outcomes such an overall survival– It can come from proving
proximal outcomes, e.g., improving patient satisfaction and engagement, compliance, that translate to
improvements in set distal outcomes1
• Proving safety toward proximal outcomes depends on the claim
• a hard to prove claim is expert systems— AI that compares or exceeds the diagnostic expert clinicians
• a feasible claim might be a reduction in recalls of breast cancer screens or missed cancer with AI as double-
reader of mammography
• With focused claims, methods exist for external validation of learning models in downstream tasks with
better access to better data
• Agility is required in the adoption ecosystem to merge Health Technology Assessment (and RCTs) with
external validation and Explainable AI
1. Weiqi Jiao, Xuan Zhang, And Fabian D’ Souza. The Economic Value and Clinical Impact of Artificial Intelligence in
Healthcare: A Scoping Literature Review. 2023. DOI: 10.1109/ACCESS.2023.3327905

Path to Sustainable Adoption— Actions
• Academia/Industry: Change your research question, this is critical when applying for pre-market authorization
• Government: Improve the ecosystem– 1) Drive innovation in the medical device clinical trials industry to
safeguard the quality and clinical robustness of AI— 2) Adopt U.S. Continual Learning Regulatory Policies
• Quintuple Helix Collaboration: Deliver agility with RWD that meets several pre-conditions for responsible AI
 Expose unquantified, hidden biases and confounders with Standards and XAI Lifecycles
 Adopt a systems medicine approach to knowledge
 Cascade simple target AIs to avoid bias-variance translation chasms (failure)
 Innovate for patient-centricity, e.g., AI checks for “hidden” comorbidity in EHRs
 Scale with inclusion and equity for population representation
 Expand precision medicine with temporal disease features
 Guarantee ecological validity
• Collaborate for data scale and scope !!
Dr Dimitrios Kalogeropoulos
5
Environment
Academia
Business/
Industry
Government
Civil Society

Weakness: Ecological or Person-level
validity
Specialisation of external validity (reproducibility-generalisability) focusing on whether
behaviours or outcomes observed in a study can occur in real-world settings
DEFINITION
Involves examining whether the experimental
conditions and assumptions in a study mirror
the complexity and dynamics of the natural
environment being investigated.
High ecological validity suggests that the
results of studies are at least reproducible
and likely to be applicable and generalizable
to real-world situations
EXAMPLES
A study that investigates sleep patterns with
participants sleeping in a lab connected to
machines that monitor their sleep has low
ecological validity compared to sleeping in the
comfort of our home
A retrospective study conducted to aggregate
data from multi-country data sources to
phenotype Covid-19 disease profiles and
trajectories and mobilise federated
knowledge

Ground We Have (and Haven’t)
Covered
• Generative AI and Large Language Models point to a sustainable path with a new Productivity
Paradigm— Continual Learning AI
• Deep Learning with LLMs scales to manage complex semantics
• But in healthcare we hit the same old wall regardless of the internal validation potential of the tech
• This is due to a lack of access to high-quality RWD with ecological validity, required for Next-Gen
prospective studies
Belkin, M et al. Reconciling modern machine-learning practice and the classical bias–variance trade-off. PNAS, 116(32), 15849–15854. 2019. https://www.pnas.org/doi/10.1073/pnas.1903070116

Verification
Analytical
Validation
Clinical
Validation
New Culture in Translation Science
In-Silico testing
e.g., breast cancer detection learning
using curated retrospective
study/data
Clinical Generalizability
Translation
bridge
Producing Meaningful Innovation in a
specified population
e.g., reduced HPV-induced HNC adverse
treatment effects, or non-inferior breast
cancer screens
Chasm-1 Chasm-2
New patient-centric
adaptive Clinical Trials
© Dr. Dimitrios Kalogeropoulos, 23 November 2023

Data Science — Information Model
Standards
Wrong
cohort?
Ecological,
Person-level validity
Dimitris
Jelena
Sara
Selim
Norma

The Clinical Trials Industry—
Gatekeeper for Social Innovation
• Current global clinical trials market* is $50 billion
annually
• Roughly 3.5% of the global pharmaceutical market
revenue ($521B)
• 20% of that is data curation
and this is one of the innovation delivery stages
• In the case of AI model training, data curation
accounts for a staggering 80% of budget
• Quality-of-care consensus conferences keep
reiterating the lack of research data
• Yet we generate so much of it, it has become a
climate issue
• Evidently, clinical research and clinical trials are in
desperate need of innovation themselves
* pharmaceutical & biotechnology companies, medical device companies, and academic institutes.

Is Gen AI Social Innovation?— If not,
What’s missing?
• Resists the ephemeral, reductionist, siloed view to innovation, in
favour of scalable, transformational, sustainable, health innovation
that drives systemic change with equity and inclusion
• Innovation usually provides novel solutions to “wicked” problems –
e.g., Gen AI
• Solutions for which the value created accrues primarily to society
rather than individuals
• Democratic institutional voids are exploited to allow new forms of
participation by a range of actors with complementary objectives—
leading to collaborative innovation ecosystems
van Niekerk, L., Manderson, L. & Balabanova, D. The application of social innovation in healthcare: a scoping review. Infect Dis
Poverty 10, 26 (2021). https://doi.org/10.1186/s40249-021-00794-8

13
Evidence-based medicine
“Studies of antibiotics for
bronchitis fail to show a benefit
for most patients”
RCTs
Peer-reviewed journals
Gold standards
Guidelines
Old School Expert/Heuristic
Systems (AI)
Learning (Digital Residency)
“Since bronchitis involves infection of
the airways in the lungs, antibiotics
should help”
Deep Learning
Knowledge Graphs
Large Language Models
Generative AI
Personalised Medicine
“Studies of antibiotics for
bronchitis fail to show a benefit
for most patients apart from this
cohort/ group/ profile provided
treatment plan-X is followed
RWE
Discovery with AI
Access and Inclusion
with Telehealth
Coherence
Covid-19 - 2030
Correspondence
1990-2020
Precision
2030-
What’s missing?
An AI Knowledge
Ecosystem
Systems medicine
© Dr. Dimitrios Kalogeropoulos,
Patient-Centred
Collaborative Learning

II. Going forward, two Claims We Must
Support

Knowledge Mobilisation Reciprocity-
Recycling Claim
• By generating ecological knowledge during clinical
practice, “AI companions” (e.g. comorbidity scanner or
precision treatment outcomes predictor) enable clinical
applications of patient-centric Systems Medicine, or P4
Medicine— Predictive, Preventive, Personalised,
Participatory
• A patient-centric ecology for knowledge “Opens the
floor” to participatory care with multi-stakeholder
collaboration – Integrated, value-based care?
• Cooperating participants reciprocate by contributing
evidence for inclusive knowledge models, making AI
accepted into practice to generate more knowledge and
stronger collaboration
• Standards enable the necessary Governance
Multi-Stakeholder
Collaboration
Advancing Standards
Cross-
Validation
Knowledge
Integration
e.g., with KGs
Applications of
Systems/P4
Medicine
AI research & applications
for patient-centric care
Digital applications that work alongside patients and doctors to improve patient engagement and adherence to treatment, and adherence to clinical
guidelines

Telehealth Value Capture-Recycling
Claim
• To participate in participatory, patient-centric and ecological
knowledge generation, AI services are combined with Telehealth
services, reaching out to underrepresented populations and
enabling continuous and fused streams of data
• This means wider and deeper understanding of population needs,
and better AI that addresses the differential circumstances of
patients
• Better understanding of needs means more targeted inclusion
strategies and expanded access to care, i.e., more inclusion and
better, collaborative, cross-evaluated AI
Reach-out
Differentiate
Include
incl. patients,
providers, more and
better AI
Advancing AI
with
telehealth

Digital Resident & Attending Physician
Technology (DRAPHT)
Digital Resident Oncologist
(DRO): AI may systematically
observe and review cancer
patient outcomes (DL) to
optimise lines of multimodal
therapy
Digital Resident Radiologist
(DRR): AI may read radiology
reports to match annotations
for self-supervision
Digital Attending Radiologist (DAR): AI
participates in double reading of
mammography to reduce variability between
radiologists in diagnostic accuracy
(unnecessary recalls and missed cancer)

Collaborating DRAPHTs
• Collaborating DRAPHTs are key for scaling a patient-centric, positive innovation
ecosystem
• DRAPHTs aimed at simple targets hold the potential to safely improve workflow
efficiencies and address workforce shortages
• Reciprocally, DRAPHTs accelerate the discovery of clinical insights in ecosystem data
spaces that improve patient-centricity
ACTION-1: Focus on narrow fields and simple targets with good clinical trial results – e.g.,
Digital Radiologist in non-inferiority study in breast cancer screening double reading
• If we succeed in validating narrow downstream applications, we can create ecosystems
of Reciprocal and Collaborative uses of AI with DRAPHTs
Collaborative
AI
Multiple
connected
targets
Knowledge
mobilization
acceleration

Remote DRAPHTs
Population,
Cohort, Focus
Population,
Cohort II,
Subgroup Focus?
Positive ecosystems require population coverage
ACTION-2: Lead digital inclusion with Telehealth, e.g. for Hospital @ Home
ACTION-3: Support designs for access to care and improved patient experience pathways

My Key Points
• AI is a 1) third-science research field and 2) RPA niche
• LLMs help focus DL on validating downstream knowledge and proximal tasks with simple targets
• LLMs set a new paradigm for AI, which has been very convincingly demonstrated by industry leaders –
that knowledge development iterations must be (responsibly) accelerated
• New paradigm delivers acceleration with continual, participatory, adaptive, collaborative, Open AI Learning
• Accelerates stakeholder buy-in to deliver responsible and meaningful uses that drive innovation
• The clinical trials industry must become innovative to follow the pace
• Must support the RWD industry with Policy & Standards
• Need a New Governance Culture to Lead Social Change

Collaboration
Enable broad collaboration
with telehealth to
modernize and reinforce
evidence-based practices in
related fields (e.g., RCT
industry)
Focus
Accelerate AI adoption by
disposing developers to a
better understanding of the
needs and challenges of
underserved populations
Participation
Enable early design
participation to improve
the patient & carer
experience
Affordance
Develop evidence-based
reimbursement policies,
e.g. for Hospital at Home
services
Regulate
Enable new regulatory
pathways for Continual,
Adaptive Global Learning
with AI (FDA
Predetermined Change
Control Plans)
Standardise
Introduce evidence
translation Standards to
embed AI into clinician
workflows and data
systems; prioritize needs of
patients and workforce
Finance
Change funding policies
underlying VC,
international development,
government programs and
research funds to support a
systems capital approach
Governance
Adopt a new Governance
Culture of Digital
Collaboration incl.
International data
transfers
Policy Ground We Need to Cover

• Advance Data Standards
Data-coupled Collaboration and Recycling
• Advance Privacy Protection Regulation
Multi-Ownership (GDPR), Unlearning AI, International Data
Transfers (UK initiative)
• Embed-Expand Evidence Standardisation Practices
Minimal Common Oncology Data Elements (mCODE)
• Personalise Disease Models
Systems Medicine, Phenotypes, Ontologies, Interactions
• Consider Use-Case Gold Standards
EU MDR, SaMD
• Expand access with New Care Models
Hospital @ Home
• Modernise HTA, evidence practices
RCTs, RWE/RWD, Micro-Randomization Trials
Evolving
Standards
Patient
experience
Value for
patient
Precision
Personalis
ation
Digital
interface:
Personalised
Experience
Digital
interface:
Value from
Participation
& Precision
Patient-Population
Facing
Provider-Physician
Facing






© Dr. Dimitrios Kalogeropoulos, 11 November 2023
Standards for Open Learning AI in
Healthcare
Requested by the EC to implement AI Act

Cooperate along the AI Continuum
Collaborate For Data— Recycle
New
populations
New
populations &
impacts
Innovation
silos
New impacts
Cooperate for scale
(Telehealth, Reg enablement)
Data Recycling Cooperate for scope
(cross-valuation)
• Data recycling means
interoperable insights -
shared use of health and
clinical data along the
AI/care continuum,
emphasising disease
progression to build up
and integrate clinical and
policy insights into a
common, reused dataset

Please reach out to drive innovative uses of technology for humanity – dimitris@ghdif.org
Thank you !

[DSC Europe 23][DigiHealth] Dimitrios Kalogeropoulos A Sustainable Future for Learning AI in Healthcare

More Related Content

Similar to [DSC Europe 23][DigiHealth] Dimitrios Kalogeropoulos A Sustainable Future for Learning AI in Healthcare

More from DataScienceConferenc1

Recently uploaded

[DSC Europe 23][DigiHealth] Dimitrios Kalogeropoulos A Sustainable Future for Learning AI in Healthcare