Item 41 — Implications for practice, including intended use and/or clinical role

Describe the implications for practice, including the intended use and possible clinical role of the AI model. Describe the key impact the work may have on the field, including changes from current practice. Envision the next steps that one might take to build upon the results. Discuss any issues that would impede successful translation of the model into practice.

Explanation and Elaboration: It is essential to clearly define the model’s intended use to ensure a clear understanding of its role. Describe how the model may improve clinical practice, as well as potential unintended consequences, risks, new burdens imposed by working with and supervising the AI and new or additional requirements for successful deployment. If there are unsolved problems that stand as barriers to clinical implementation, identify these as areas where additional research is needed.

Examples
Using model diagnosis and interpretable attention maps improved radiologists’ diagnostic performance compared with the unassisted group. The attention maps aided radiologists in assessing the model’s diagnosis importance and accurately locating lesions, reducing missed diagnoses…. This approach is especially valuable in regions with a specialist shortage. [Discussion (7), CC BY 4.0]
One major barrier to the widespread clinical deployment of machine learning models is the need for improved generalization. Models must maintain high levels of performance accuracy on data from diverse sources encompassing heterogeneous populations and differing technical parameters. To address this, we employed a semi-supervised learning paradigm to detect and segment intracranial hemorrhage on head CT images using a broad complement of both labeled and unlabeled data to improve model generalization capability. [Discussion (62), CC BY 4.0]
In conclusion, replacing one or both radiologists in double-read mammography screening with AI seems feasible, but there were important differences in accuracy and workload depending on the deployment site in the workflow. A triage-based approach was found to have the best outcome across all performance metrics. However, as it is not possible to replace both readers because of current national guidelines, replacing the first reader with AI represents a feasible solution for implementation. Prospective trials are warranted to validate these scenarios, with particular focus on the effect of AI-integrated screening on clinically relevant outcomes, as well as the impact of replacing either the first or second reader because of their differing experience and accuracy. [Discussion (22), CC BY 4.0]