Dynamic evaluation and updating of Covid-19 risk prediction models over time: Methods and pipeline

Since the first cases of COVID-19 were confirmed in the UK in early 2020, the pandemic has impacted all of our daily lives. It became clear that some people were at greater risk of hospitalisation or death if they became infected. Researchers developed a prediction model called QCOVID to help identify those most at risk of getting infected and then being hospitalised or dying due to COVID based on their individual characteristics such as age, sex, ethnicity and longstanding illnesses. The QCOVID model was used to prioritise people for vaccination and to inform the government’s shielding list. The model was developed using data from a GP records database called QResearch and it was extended in 2021 to account for people having one or two doses of covid-19 vaccination. The QResearch data is anonymous; there is no information that would allow an individual person to be identified.

The QCOVID models were developed in a rapidly changing environment in terms of availability of vaccines, vaccine uptake, infection prevalence, levels of immunity in the population, new variants and availability of treatments. As the pandemic continues to evolve, it is important that the QCOVID models are kept up to date both to provide accurate predictions of risk and to enable their use, if needed, to prioritise groups in the population for new treatments or vaccine boosters.

Our objective is to find out how best to update a clinical prediction model such as QCOVID to account for changes over time. The updating method must be flexible because changes related to COVID-19 can happen quickly and are not always easy to foresee. An example is the speed with which the Omicron variant became dominant in the UK at the end of 2021. The updating method must also be able to include new treatments or vaccines (e.g. the booster dose) , and for their effects to possibly wane over time.

Our study will compare a variety of statistical techniques for updating prediction models over time, taking into account the specific factors that influence risk in the COVID-19 context. To do this, we will use computer-generated data to represent hypothetical future scenarios and also actual data from QResearch, the database of health information used for the development of the original QCOVID models. Based on the performance of different model updating methods, we will make a recommendation for how to update QCOVID going forward.

(These are the same data points as were used in the final QCOVID3 model)
Data regarding:

Demographic variables:
Age (continuous variable)
Sex
Townsend deprivation score
Ethnicity (9 categories)
Lives in a care home (nursing home or residential care)
Homelessness

+++
Lifestyle factors :
Body mass index (continuous variable)

+++
Vaccination: first, second, booster dose
Record of a SARS-CoV-2 positive test result

+++
Co-morbidities (includes variables in the QCovid models)
Chronic kidney disease
Chemotherapy in previous 12 months
Type 1 or type 2 diabetes (with glycated haemoglobin (HbA1c) levels <59 or ≥59 mmol/mol)
Blood cancer
Bone marrow transplantation in past six months
Respiratory cancer
Radiotherapy in past six months
Solid organ transplantation
Chronic obstructive pulmonary disease
Asthma
Rare lung diseases (cystic fibrosis, bronchiectasis, or alveolitis)
Pulmonary hypertension or pulmonary fibrosis
Coronary heart disease
Stroke
Atrial fibrillation
Heart failure
Venous thromboembolism
Peripheral vascular disease
Congenital heart disease
Dementia
Parkinson’s disease
Epilepsy
Rare neurological conditions (motor neurone disease, multiple sclerosis, myasthenia gravis, or Huntington’s chorea)
Cerebral palsy
Osteoporotic fracture
Rheumatoid arthritis or systemic lupus erythematosus
Liver cirrhosis
Bipolar disorder or schizophrenia
Inflammatory bowel disease
Sickle cell disease
HIV/AIDS
Severe combined immunodeficiency

Data regarding admission and treatment for the above conditions.
For example, OPCS and ICD10 codes for transplantation, chemotherapy, radiotherapy, operative treatment for congenital heart disease.

Linkage to mortality records to enable ascertainment of deaths from COVID-19 during the study period.
There will also be linkage to SGSS data regarding the results of COVID-19 testing.

systemic anticancer and radiotherapy dataset will be requested to identify patients on these treatments.

Professor Carol Coupland - University of Nottingham

Professor Ruth Keogh - London School of Hygiene & Tropical Medicine

Professor Karla Diaz-Ordaz - London School of Hygiene & Tropical Medicine

Dr Kamryn Tanner - London School of Hygiene & Tropical Medicine