Page 33 - Shaping ethics, regulation and standardization in AI for health

Shaping ethics, regulation and standardization in AI for health

approach using the LGBM algorithm performed best. The XGB algorithm also showed strong
predictive capabilities. The benchmarking provided insights into effective predictive models for
neonatal health. Post-benchmarking, data retention for future research is crucial. Establishing
benchmarks in maternal and child health is challenging due to diverse socio-economic contexts,
requiring collaboration for accurate and reliable data.

A�4�7 DEL 10�8: FG-AI4H Topic Description Document for the Topic group
on Neurological disorders (TG-Neuro)

Summary: This TDD specifies a standardized benchmarking for AI in neurological diseases. It
covers scientific, technical, and administrative aspects relevant for setting up this benchmarking.

The document explores the role of AI in diagnosing neuro-cognitive disorders, particularly
Alzheimer's and Parkinson's, using real-world brain imaging and genetic data. With an
aging population and rising dementia cases, AI-driven diagnostics offer more accuracy
and reproducibility compared to traditional visual assessments. Methods like supervised
classification and multivariate pattern recognition have shown promise in diagnosing dementia
and predicting healthy aging, benefiting both patients and researchers.

A key focus is benchmarking AI performance in neuro-cognitive disorders. While no universal
benchmarking system exists, various studies and internal frameworks have assessed AI's clinical
utility. The document reviews these benchmarking efforts, emphasizing the need for improved
assessment methods, metrics, and integration with electronic health records. Solutions like
LORIS, CBrain, Cerner, and Epic Systems help address data privacy and compatibility challenges.

The latest TG-Neuro benchmarking iteration evaluates AI tools using MRI, PET scans, and memory
scores, assessing their technical, operational, scientific, and clinical potential. It sets criteria for
selecting AI use cases, details benchmarking scenarios, and incorporates clinician feedback.
Early results are promising but limited by small, labelled datasets, making expansion to broader
datasets necessary. Future improvements include testing additional imaging techniques like
brain iron density studies.

Overall, TG-Neuro benchmarking provides a strong foundation for further research. Next steps
involve analysing larger datasets, incorporating more imaging modalities, and increasing
clinician collaboration to enhance AI's role in early dementia detection and treatment.

A�4�8 DEL 10�9: FG-AI4H Topic Description Document for the Topic group
on Ophthalmology (TG-Ophthalmo)

Summary: This TDD specifies a standardized benchmarking for AI in ophthalmology. It covers
scientific, technical, and administrative aspects relevant for setting up this benchmarking.

The topic description for benchmarking AI in Ophthalmology provides detailed background
information on how AI can address real-world problems in this field. The TG-Ophthalmo
included subtopics such as Diabetic Retinopathy (DR), Age-related Macular Degeneration
(AMD), Glaucoma (GC), Pathological Myopia (PM), and Red Eye (RE). The document outlines
the AI tasks, current gold standard, relevance and impact of AI solutions, currently available
datasets, systems and benchmarks for the subtopics above.

23