AI: New testing framework to ensure the quality of training data

A testing framework has been developed at Germany’s national metrology institute, the Physikalisch-Technische Bundesanstalt (PTB), to ensure the quality of training data for medical AI applications. The framework, as well as a tool for the automated testing of algorithms, will be available to all developers, reports project manager Dr. Daniel Schwabe. In an interview with MEDICA-tradefair.com, the mathematician explains the project.

Dr. Schwabe, what problems are being addressed by the project?

Dr. Daniel Schwabe: More and more medical technology products that use AI algorithms are coming onto the market, and these products need to be certified. There is currently the Medical Device Regulation (MDR), and now the AI Act is being incorporated into the regulatory system. Some manufacturers are already overwhelmed, and it should never be the case that regulation slows down innovation.

The EU project TEF-Health (Testing and Experimentation Facility), which is funded with 60 million euros, is designed to certify AI medical devices in the future. The Metric Framework is our first step towards supporting these processes by providing a systematic way of checking training and test data. To increase the trustworthiness and safety of AI medical products, such validation will be necessary in the future.

What is the status quo regarding data quality in medical AI applications?

Schwabe: There is a broad spectrum. We are not trying to evaluate general data for all applications, but want to know: How can we systematically analyze whether this data set that I have available fits the application that I am currently developing? We try to take a very practical approach to the question, always with a view to the manufacturers, especially SMEs and start-ups. They develop innovations, but often do not yet have the necessary infrastructure and knowledge in terms of quality assurance. We want to provide them with a system that they can use to evaluate their data.

How can developers of algorithms use your tools?

Schwabe: We are working on two projects. On the theoretical side, my team has broken down the concept of data quality into individual properties. This system is presented in the “Wheel of Data Quality”. It lays the foundation for systematically analyzing the quality of medical training data. We are now adding metrics to this system that can be used to measure these properties.

On the practical side, we are working on automated tools that put this theory into practice. On the one hand, this is a tool for the simple evaluation of the individual properties of the data quality of medical training data. On the other hand, we want to quantify the extent to which these properties actually influence an AI system. For example, should manufacturers use their limited time to improve certain properties? Such tools will later help us to develop a quality stamp for training data.

In the TEF-Health project, we are working very closely with the TÜV AI Lab and the Fraunhofer Heinrich Hertz Institute to optimize the approval process for AI medical devices. A test stamp from the PTB for the data used represents a safety gain for the approval authorities. This can, for example, result in a smaller scope of clinical studies, which saves companies time and money and speeds up the path to market maturity.

In addition, the software testing system TraCIM (Traceability for Computational-Intensive Metrology) for AI is being further developed at PTB. Here, developers receive test data from us, run their AI model over it and report the results to us. We then issue a test report in which the quality of this prediction is assessed.

How and for whom are the tools accessible?

Schwabe: In principle, it is open to everyone. So far, we have only laid the theoretical foundation with the Metric Framework. The automated tools for evaluating data quality and quantifying its influence are being developed as open-source software. The already established TraCIM system is being updated in parallel with these new findings so that we can also offer validation as a service for AI algorithms in the future. At the same time, our metric framework can form the basis for audits so that training data can be audited using this system.