Co-developing and operationalising machine learning models for COPD care in a clinical setting: a workflow framework

<bold>Background:</bold> A wide range of machine learning applications to support earlier diagnosis, triage, and treatment selection in COPD have been described. There has however been limited translation of the models presented in the literature into tools used in live clinical practice. There is a need for a shift towards coordinated and collaborative working process involving clinicians and technical teams, to allow developed tools to be made available to patients and clinicians at the point of care. <bold>Methods:</bold> Using participatory core design and following best practice guidance for machine learning formulation and software as a medical device, we have created a workflow framework covering the complex multi-stage process of developing and operationalising machine learning models trained on routinely collected electronic health record data. <bold>Results:</bold> Figure 1 shows the stages for model co-development, adoption, supervision and adaptation. We have successfully used this framework in the DYNAMIC-AI clinical investigation (<ext-link>NCT05914220</ext-link>), delivering explainable 12-month mortality and 3-month readmission risk scores for patients with COPD within the Lenus ‘AI insights’ app. <bold>Conclusions:</bold> This framework provides an exemplar for development and adoption of machine learning-based risk prediction models to optimise proactive preventative care for patients with COPD and other long-term conditions. <fig><object-id>erj;64/suppl_68/OA2770/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig> Figure 1.