S66 Towards implementation of live AI-based prognostic risk-prediction scores in a COPD MDT. (11th November 2022)

• Record Type: Journal Article
• Authors: Burns, S; Subasic, G; Morgan, D; Taylor, A; McGinness, P; Lowe, DJ; Carlin, C
• Is Part Of: Thorax. Volume 77(2022)Supplement 1
• Page Start: A41
• Page End: A42
• Publication Date: 2022-11-11
• Subjects: Chest -- Diseases -- Periodicals; Thorax; Chest -- Diseases; Periodicals; Periodicals; 617.54
• DOI: 10.1136/thorax-2022-BTSabstracts.72
• Languages: English
• ISSNs: 0040-6376

Abstract

Abstract : Introduction: COPD is a common, progressive, preventable and treatable respiratory disorder effecting approximately 1.2 million people in the UK. It is estimated to become the third leading cause of death worldwide by 2030. Accurately identifying high-risk patients in a live clinical setting is essential for proactive care re-orientation and prioritisation. Aims and Objectives: Machine Learning (ML) models were developed using clinical data to predict risk in COPD patients with the aim to optimise care and improve patient outcomes. This AI is being operationalised in live patient care as part of a feasibility study. Methods: We used de-identified demographics, hospital admissions, diagnosis, prescribing and labs data from 60, 000 patients to develop risk prediction models. We focused on risk of mortality, respiratory-related hospital readmission, and exacerbation prediction. 15% of all patients were held out for a final test set. All sets were checked to ensure they were drawn from a sample representative of the full population. An 80:20 split of the remaining 85% of patients, and cross validation methods, were used for model training and validation. The 12-month mortality prediction model was tested on the hold-out test dataset. Patients who were recruited to our RECEIVER COPD digital service trial (support.nhscopd.scot) had been omitted from model training and validation. We were therefore able to undertake a further retrospective evaluation on this trial data, running synthetic AI-MDTs with the model applied at patient onboarding and at monthly intervals for the following 6 months, with comparison of model predictions to events over the subsequent 12 months. Results: The model performed well achieving an averaged ROC-AUC of 0.83 (0.76–0.92) and an averaged PR-AUC of 0.57 (0.38–0.76) averaged over the 7 runs. Local explainability for each patient prediction was also calculated using SHAP. The feasibility study using live AI models within our COPD MDT is ongoing. Conclusions: Our 12-month mortality prediction model performed well and is ready for live adoption in the AI insights app and propspective evaluation in the DYNAMIC-AI trial. The other two models are at advanced development stage. The ongoing feasibility study will demonstrate how AI can be used as part of live patient care.