Cardiovascular risk prediction in the acute care setting: a mixed methods evaluation using machine learning, real world evidence and qualitative methods

Student thesis: Phd

Abstract

Background Emergency Department (ED) clinicians use clinical prediction models (CPMs) to assist with the diagnosis of acute myocardial infarction (AMI). CPMs are also used in primary care to predict future cardiovascular disease (CVD). Patient care in the ED could be improved by updating existing CPMs for AMI diagnosis or developing new care pathways to predict future CVD. Aims To update a CPM for AMI diagnosis (the Troponin-only Manchester Acute Coronary Syndromes CPM) and co-develop a novel care pathway for the prediction of CVD in acute care. Methods I used real-world data from three centres, linked to a national dataset, to evaluate TMACS. I used three methods to update the CPM: recalibration, model extension and dynamic Bayesian updating. The novel care pathway for CVD prediction was co-developed with a mixed methods approach. This was informed by a systematic review and meta-analysis of potential prognostic factors. I then studied the prognostic value of factors of interest using a retrospective cohort linked to a national dataset. Finally I conducted 41 semi-structured interviews using a co-production framework to construct a novel care pathway. Results TMACS demonstrated good discrimination with an area under the curve (AUC) of 0.88 (95% CI 0.86 to 0.89), but calibration had deteriorated with a calibration in the large (CITL) of -3.93 (95% CI -4.12 to -3.74). Dynamic updating demonstrated favourable model characteristics with an AUC of 0.87 (95% CI 0.85 to 0.89) and CITL of -1.45 (95%CI -1.63 to -1.27). Several routinely collected variables were found to be predictive of future CVD. Framingham CPM demonstrated favourable prognostic model characteristics in external validation. The qualitative research led to a co-produced care pathway based on five themes including loci of clinical responsibility, poor communication, avoidance of pandemic hospitals, focused EM care, and automation of preventative EM care. Conclusion Calibration drift affected TMACS, and this should be a cautionary tale for other deployed CPMs. However dynamic updating did successfully counter this and could provide a sustainable solution if the digital infrastructure is available. Routinely collected data in the ED is predictive of long term CVD, and a care pathway appears possible. A feasibility randomised control trial should be conducted to further assess this intervention.
Date of Award31 Dec 2023
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorBrian McMillan (Supervisor), Anthony Heagerty (Supervisor), Rick Body (Supervisor), Evangelos (Evan) Kontopantelis (Supervisor) & Glen Martin (Supervisor)

Cite this

'