Risk-based Evaluation of ML Classification Methods Used for Medical Devices

Abstract Background: In the future, more and more medical devices will be based on machine learning (ML) methods. For such medical devices, the rating of risks is a crucial aspect and should be considered when evaluating their performance. This means that an integration of risks and their associated costs into the corresponding metrics should be taken into account. This paper addresses three key issues towards a risk-based evaluation of ML-based classification models. Methods: First, it analyzes a selected set of scientific publications for determining how often risk-based metrics are currently utilized in the context of ML-based classification models. Second, it introduces an approach for evaluating such models where expected risks and associated costs are integrated into the corresponding performance metrics. Additionally, it analyzes the impact of different risk ratios on the resulting overall performance. For this purpose, an artificial model was used which allows to easily adapt key parameters. Third, the paper elaborates how such risk-based approaches relate to regulatory requirements in the field of medical devices. A set of use case scenarios were utilized to demonstrate necessities and practical implications, in this regard. Results: With respect to the first research question, it was shown that currently most scientific publications do not include risk-based approaches for measuring performance. For the second topic, it was demonstrated that risk-based considerations have a substantial impact on the outcome. The relative increase of the resulting overall risks can go up 198%, i.e. the risk value almost triples, when the ratio between different types of risks (risk of false negatives in comparison to false positives) goes down/up to 0.1 or 10.0. As discussed within the third research question, this situation typically represents a case where the risk increases one level in the corresponding risk matrix. Based on this, it was demonstrated that differences in parameter settings lead to a substantially different behavior when risk factors are not addressed properly. Conclusion: In summary, the paper demonstrates the necessity of a risk-based approach for the evaluation of ML-based medical devices, develops basic steps towards such an approach, and elaborates consequences which occur, when these steps are neglected.