Page 20 - Shaping ethics, regulation and standardization in AI for health

Shaping ethics, regulation and standardization in AI for health

(TG-Dermatology, TG-Ophthalmology, TG-Malaria, TG-Symptoms, TG-Psychology, TG-Snakes,
TG-Radiology, TG-Neuro, TG-Outbreaks) during the trial audits, whose main objectives were:

– To facilitate the conduct of trial audit series and to provide a platform for the submission,
discussion and publication of FG-AI4H audit methods and reports under the special
collection titled "Machine Learning for Health: Algorithm Auditing & Quality Control" in
the Journal of Medical Systems (JOMS) [3]
– To provide basic training to FG-AI4H Topic Groups on how to verify and validate the
technical, clinical and regulatory requirements of their ML4H tool by following the audit
workflow over the AI4H Assessment Platform

The collaboration between the OCI platform and the TGs has been mutually beneficial, extending
beyond platform enhancement to include improvements in TGs' Topic Description Documents
(TDDs). Through the process of running trial audits and engaging in detailed discussions,
the OCI experience has enabled TGs to refine their understanding of their data's nature and
characteristics. This, in turn, has empowered them to describe their datasets in a standardized
and comprehensive manner. By gaining clarity on data attributes and standardizing their
documentation, TGs have enhanced the quality and utility of their TDDs. This collaborative
exchange underscores the synergistic relationship between the OCI platform and the TGs,
driving continuous improvement and innovation across the FG-AI4H ecosystem.

In parallel, the OCI team collaborated with various FG-AI4H Working Groups, especially WG-
DAISAM and WG-Clinical Evaluation, to implement a range of features aimed at enhancing the
platform's functionality. These features include:
– Data annotation campaign management: This feature facilitates the streamlined
management of data annotation campaigns, allowing for efficient organization and
tracking of tasks within the platform.
– Health metadata management: Enabling comprehensive management of health metadata
ensures that relevant information is effectively captured and utilized across various aspects
of the platform, enhancing data accuracy and usability.
– FHIR standard implementation for Data acquisition: Integration of the HL7 Fast Healthcare
Interoperability Resources (FHIR) standard for data acquisition ensures compatibility with
existing healthcare systems, promoting seamless data exchange and interoperability.
– Providing an integration layer for data annotation UI tools: This feature establishes
a flexible integration layer, allowing for the seamless incorporation of various data
annotation user interface (UI) tools into the platform, catering to diverse user preferences
and requirements. For example, integration with Visian, a student project from HPI
Potsdam, offers a flexible human-in-the-loop AI solution tailored for effortless utilization
of machine learning (ML) models on medical image data. It also facilitates comfortable
project management for data and model versioning, enhancing overall workflow efficiency
and collaboration.
– Collaboration with industry on privacy-preserving encryption: The OCI team has partnered
with leading industry players such as Inpher.io to implement state-of-the-art privacy-
preserving encryption techniques. By leveraging homomorphic encryption, the platform
ensures robust protection of sensitive data through advanced cryptographic methods,
fostering trust and compliance with privacy regulations.

Furthermore, as part of the development process, comprehensive documentation was created
and hosted on GitHub (https:// github .com/ fg -ai4h). This documentation serves as a valuable
resource for users, providing guidance on platform usage, feature implementation, and
troubleshooting. Feedback from stakeholders was actively solicited and utilized to refine several
FG-AI4H deliverables, ensuring alignment with user needs and preferences.

10