We treat each disease a person has separately. This means different medicines are prescribed for each disease, which may not help people with multiple long-term health problems. For example, a drug for one disease can make another disease worse or better. But we don't know the effect of one drug on a second disease and some diseases may have lots of options of medicines that could be prescribed. If we can group such people who have multiple medical problems based on their diseases, then we can study the effects of a particular drug on each disease mix. Artificial intelligence (or AI) is a computer system that can conduct tasks that usually need human intelligence. Newer AI methods can process large amounts of health data in a short time.
We aim to:
· Find the mix of diseases and drug treatments that interact to worsen or improve a patient’s health.
· Predict the next disease people may develop
· Find drugs that help more than one disease
We will do this by linking detailed health records of patients who attend primary care (GP) services and the University Hospital Birmingham Trust. These include all diagnoses, drugs, blood tests, and data taken from scans. Using AI methods, we will model how different mixes of diseases arise over time. The model will tell us what drugs cause or prevent a new disease. These models could guide doctors with their prescribing to reduce the number of drugs patients need or stop multiple drugs from interacting.
At present research on people with multiple long-term health conditions has focussed on clusters of disease with a simple binary classification; not taking into account disease trajectories, the spectrum of disease severity, disease accumulation. Understanding disease trajectories within clusters, taking into account subclinical disease and severity of disease, is important for early identification of populations at high risk of developing further diseases.
Treatment strategies in people with multiple long-term health conditions are based on single diseases leading to multiple medications being prescribed and which makes clinical decision making complex. We do not know the effect of prescribed medications given for one component disease on trajectories of other component diseases. We need a deeper understanding of the interaction between diseases and prescribed medications.
We will link patients’ data from primary (Clinical Practice Research Datalink Aurum) to secondary care (University Hospital Birmingham Trust). This will include all data on prescriptions, blood tests, medical diagnoses and data taken from scans (eg of the eyes and heart) and data extracted from other measurements (eg from electrocardiograms). We will develop a risk prediction algorithm to identify the likely next disease and the next best treatment option in patients with multiple long-term health conditions. We will identify disease clusters and trajectories using biomarkers and physiological measurements of disease severity, and then add in prescription data to model how they interact with trajectories of diseases. Prediction models will be based upon latent embeddings derived using deep neural network approaches to learn suitable representations from irregularly sampled, mixed data type, longitudinal sequence data that have recently been developed for observational electronic medical record data. The models will determine:
i) likely next disease;and
ii) best treatment option where there are multiple choices (e.g. best second line treatment in diabetes in someone with multiple long-term health conditions)
Primary outcomes
Diagnosis of another long-term health condition
Progression of a long-term health condition eg heart failure classified according to New York Heart classification stage I progressing to stage IV
Secondary outcomes
All-cause mortality (ONS death date)
Cause-specific mortality
Acute hospitalisations
Number of medications prescribed
GP consultations with a clinician or nurse
Krishnarajah Nirantharakumar - Chief Investigator - University of Birmingham
Francesca Crowe - Corresponding Applicant - University of Birmingham
Aditya Acharya - Collaborator - University of Birmingham
Alastair Denniston - Collaborator - University of Birmingham
Christopher Yau - Collaborator - University of Oxford
Dominic Danks - Collaborator - University of Birmingham
Georgios Gkoutos - Collaborator - University of Birmingham
Jennifer Cooper - Collaborator - University of Birmingham
Krishna Gokhale - Collaborator - University of Birmingham
Marco Canducci - Collaborator - University of Birmingham
Peter Tino - Collaborator - University of Birmingham
Suzy Gallier - Collaborator - University Hospitals Birmingham
HES Admitted Patient Care;ONS Death Registration Data;Patient Level Index of Multiple Deprivation;Practice Level Index of Multiple Deprivation