Application of machine learning to discover new multimorbidity phenotypes associated with poorer outcomes

Multi-morbidity is a poorly defined concept in which people suffer from more than one ongoing condition at the same time. The true extend of multi-morbidity is difficult to assess as there is no agreed definition for reporting. However, analysis of prescribing for chronic conditions and simple counts of different illnesses show that multimorbidity is becoming more common and is associated with poorer outcomes, such as how long people stay in hospital or premature mortality. It would be helpful to identify factors that predate the development of different morbidities to help understand how morbidities develop, which ones are commonly associated with others, to better understand the effectiveness of health services and individual treatments and to identify opportunities to prevent or delay the onset of these conditions.

Because we know so little about the development of these conditions we propose to use new analytical approaches from computer science, known as machine learning, to identify previously hidden or unknown relationships between different conditions. We will use detailed information from the medical records of the 3 million people of Wales held in the Secure Anonymised Information Linkage (SAIL) system. SAIL is a privacy protecting system in which records that have been stripped of all personal identifiers can be used to understand the development of diseases.

We will use the availability of new data on the results of laboratory investigations, such as changes in blood chemistry, to see if these predict the onset of conditions. If we do find useful patterns we will provide this knowledge back to NHS organisations to allow them to improve their services and intervene earlier to protect people's health.

By bringing together routinely collected and epidemiologic data at scale, this proposal exploits the potential of the fast-developing UK health informatics environment. Our team includes a mixture of health service researchers, computer scientists, clinical doctors and members of the public who have helped develop this proposal and will continue to be involved in the research and its dissemination.

We will exploit the most deeply phenotyped population e-cohort in the UK, created by HDRUK investment, containing detailed multi-sourced data on a 2.5M population with GP records in Wales from 2000-2020, augmented with demographic, multiple disease registry, hospital inpatient, outpatient data and laboratory results from 2007. No other part of the UK has this depth of records in a stable population with low levels of migration and loss to follow up.
Useful algorithms will be adopted by the NHS with tracking of intervention and subsequent impact.
Objectives
A. complete the most deeply phenotyped population e-cohort in the UK using existing data from the Secure Anonymised Information Linkage (SAIL) system augmented with Census data
B. apply innovative machine learning approaches to validate and refine clusters of conditions detected across the adult life course
C. use cohort data to identify mechanistic pathways underlying disease combinations
D. report on prevalence, social patterning and health inequalities using small area, census, taxation, and household composition data
E. identify potential biomarkers predicting individual and multiple morbidities through longitudinal trajectories of values in routinely collected laboratory data
F. undertake a comprehensive analysis of variables used in established morbidity/comorbidity indices with multiple correspondence analysis and factor analysis of mixed data to identify clusters
G. provide new variables for linkage to the 20,000 Welsh participants in UK Biobank, 7,000 in Airwave and 15,000+ in Healthwise Wales and use algorithms for 40+ cohorts in DPUK for further studies into the genetics of shared mechanistic pathways
H. contribute data on incidence, prevalence and burden to the Global Burden of Diseases
I. contribute validated algorithms into NHS systems to allow for early NHS adoption, supporting precision medicine and impact measurement

Impact summary

The beneficiaries of this research are:

1. People with multiple medical conditions that are being sub-optimally treated by partially effective therapies, or those that cause significant side effects
2. General population
3. Clinicians
4. NHS management
5. Research community
6. Government and NHS policy community
7. UK industry and competitiveness

Specific benefits to these groups are listed below:

Better insight into clustering of morbidities, their causes and the development of better targeted therapies will provide significant benefits to a wide range of beneficiaries through:
A. Fewer people whose lives are blighted by the impact of their morbidities and side effects of sub-optimal therapies will improve their quality of life and creativity, reduce demands on the NHS and for long term social care, increase the proportion of the population able to continue as active contributors to wealth generation (through income tax, reduced call on early pensions, enhanced economic productivity)

B. As better understanding of the causes of multimorbidity are identified and preventive strategies are put in place fewer people should develop these conditions at earlier ages leading to improved population health, wellbeing and productivity. Fewer members of the public will be required to undertake premature carer roles.

C. Improved sustainability and productivity of the NHS through reduced demand, freeing resources to be used for anticipatory and elective care with proven benefits (e.g. earlier identification of cancer and joint replacements that prolong independence and economic activity) and the ability of the NHS to cope better with surges in demand (e.g. influenza). The creation of a total population platform for understanding the development of multimorbidity will also provide the basis for testing novel service redesign and innovative policy approaches to improving NHS efficiency and effectiveness.

D. The development of deeply phenotyped cohorts of patients available to the wider research community through trusted privacy protecting environments will stimulate further research into the aetiology, biological and social determinants of illnesses and support and enhance UK Life Sciences global competitiveness in developing new therapies and approaches to management using artificial intelligence and machine learning. This project will lead to significant advances in multidisciplinary approaches in methodology and application to challenging 'Big data' research questions.

E. The sharing of data and methodological approaches across the two principle sites and NHS organisations (and subsequently through HDR UK and other initiatives) will increase the provision of skill sets and skilled people to the UK workforce.