This project aims to develop new, easy to use, artificial intelligence (AI) augmented tools that support GPs and pharmacists to find patients with multimorbidity (2+ long-term health conditions) who might be offered a better combination of medicines. We focus on people at risk of rapidly worsening health from multimorbidity and who are taking multiple regular medicines (polypharmacy) who might experience:
1. Hospitalisation (overall and because of medication side effects and adverse drug reactions).
2. Hospitalisation with a fall, fracture, or delirium.
3. Death.
The National Health Service (NHS) introduced Structured Medication Reviews (SMRs) by GPs and pharmacists aiming to reduce the number of people taking potentially harmful drug combinations. However, there is no easy way of predicting who is most likely to benefit from a medication review. The Dynamic Artificial Intelligence for Prescriptions (DynAIRx) project will develop tools to combine information from electronic healthcare records, clinical guidelines and risk-prediction models to ensure that clinicians and patients have the necessary information to prioritise and support SMRs.
The main aim of the study is to find harmful interactions between conditions and medications and investigate potential improvements. We will develop AI augmented tools that combine GP and hospital records to calculate risks of hospital admissions and other adverse outcomes. To ensure this information is easily understandable, we will develop visual summaries of patients’ journeys, showing how health conditions, treatments and risks are changing over time. This will be tested in general practices across northern England and improved based on feedback from stakeholders.
Patients with multimorbidity and polypharmacy should have their medications reviewed regularly. However, these reviews do not always happen, and are conducted suboptimally because it is challenging to assemble the relevant information to make the best (de)prescribing decisions. DynAIRx will address this by predicting where patients would benefit from a medication review, and to support the conduct of that review with supporting models and visualisations. Public health benefit will arise via 1) better targeting of medication reviews to those most likely to benefit, and 2) improving the efficiency of planning/executing a medication review.
We will use CPRD Aurum data to identify adult patients with polypharmacy (>= 5 medications) in a longitudinal retrospective cohort design so that patients have polypharmacy at some point during the study period. Adverse outcomes of interest, emergency hospitalisation and death, will be derived from Aurum, and hospital episode statistics.
We will determine the extent of multimorbidity and polypharmacy using descriptive statistics (e.g. electronic frailty score, number of prescriptions within 84 days), and build statistical prediction models (logistic regression and Cox), augmented with AI techniques for predictor discovery and hypothesis generation, to predict adverse outcomes.
We will use causal analysis to identify which individuals stand to benefit most from a structured medication review, and calculate specific benefits of deprescribing interventions. The causal modelling will use propensity score matching for single-time interventions, and marginal structural models for longitudinal interventions (changing a given medication) – comparing observed effects to clinical trials for robustness, where available.
We will design visualisations of the relevant data, augmented with information from our models, to support GPs in delivering better informed and more efficient structured medication reviews. The decision to modify treatments during medication review will remain with clinicians and patients. We anticipate further CPRD applications to arise for future research, building upon these findings.
Primary Outcomes
all-cause emergency hospital admission; all-cause mortality
Secondary Outcomes
Hospitalization with an admission code for adverse drug reaction; Hospitalization with a fall or fracture; Hospitalization with delirium; GP diagnosis of dementia
Andrew Clegg - Chief Investigator - University of Leeds
Samuel Relton - Corresponding Applicant - University of Leeds
Asra Aslam - Collaborator - University of Leeds
Danushka Bollegala - Collaborator - University of Liverpool
Gary Leeming - Collaborator - University of Liverpool
Harriet Cant - Collaborator - University of Manchester
Iain Buchan - Collaborator - University of Liverpool
Lauren Walker - Collaborator - University of Liverpool
Layik Hama - Collaborator - University of Leeds
Matthew Sperrin - Collaborator - University of Leeds
Maurice O'Connell - Collaborator - University of Leeds
Micheal Abaho - Collaborator - University of Liverpool
Pieta Schofield - Collaborator - University of Liverpool
Roy Ruddle - Collaborator - University of Leeds
Tjeerd van Staa - Collaborator - University of Manchester
Matthew Sperrin - Collaborator - University of Manchester
Maurice O'Connell - Collaborator - University of Manchester
HES Accident and Emergency;HES Admitted Patient Care;HES Outpatient;ONS Death Registration Data;Patient Level Index of Multiple Deprivation;CCG Pseudonyms