Novel algorithm predicts deaths in COVID-19 patients
The population-based risk prediction algorithm QCOVID demonstrates very high levels of discrimination for deaths and hospital admissions due to the novel coronavirus disease (COVID-19), according to the results of a national derivation and validation cohort study.
However, the absolute risks presented may change over time in keeping with the prevailing SARS-CoV-2 infection rate and the extent of social distancing measures in place. Therefore, caution is warranted when interpreting such risks.
“This study presents robust risk prediction models that could be used to stratify risk in populations for public health purposes in the event of a ‘second wave’ of the pandemic and support shared management of risk,” the researchers said. “We anticipate that the algorithms will be updated regularly as understanding of COVID-19 increases, as more data become available, as behaviour in the population changes, or in response to new policy interventions.”
This study utilized the QResearch database, comprising 1,205 general practices in England with linkage to COVID-19 test results, Hospital Episode Statistics, and death registry data. Overall, 6.08 and 2.17 million adults, aged 19–100 years, were included in the derivation and validation datasets, respectively. The derivation and first validation cohort period covered 24 January to 30 April 2020, and the second temporal validation cohort was from 1 May to 30 June 2020.
The researchers fitted models in the derivation cohort to derive risk equations using a range of predictor variables. Then, they assessed the performance, including discrimination and calibration, in each validation time period.
During follow-up, 4,384 deaths from COVID-19 occurred in the derivation cohort. Moreover, 1,722 died in the first validation cohort period and 621 in the second validation cohort period. The final risk algorithms included the following: age, ethnicity, body mass index, deprivation, and a range of comorbidities. [BMJ 2020;371:m3731]
QCOVID showed good calibration in the first validation cohort, explaining 73.1 percent (95 percent confidence interval [CI], 71.9–74.3 percent) of the variation in time to death (R2) for men who died from COVID-19. The D statistic was 3.37 (95 percent CI, 3.27–3.47), and Harrell’s C was 0.928 (0.919–0.938). For deaths from COVID-19 in women, results were similar for both outcomes and in both time periods.
The algorithm also had a sensitivity of 75.7 percent for identifying deaths within 97 days in the top 5 percent of patients with the highest predicted mortality risks. People in the top 20 percent of predicted mortality risk accounted for 94 percent of all deaths from COVID-19.
“This study represents a substantial improvement on previously developed risk algorithms in terms of the size and representativeness of the study population, the richness of data linkages enabling accurate ascertainment of cases (including both in-hospital and out of hospital deaths) across the health network, and the breadth of candidate predictor variables considered,” the researchers claimed.
In a linked editorial, Matthew Sperrin and Brian MacMillan, both lecturers at the University of Manchester in the UK, explained that QCOVID predicts the risk of catching and dying from COVID-19 in the general population. They also stated that the researchers correctly highlighted the fact that separately predicting either the probability of catching COVID-19 or the probability of dying from it is not possible, owing to incomplete knowledge of who actually carries the disease. [BMJ 2020;371:m3777; BMJ 2020;371:m3731]
“However, this conflation causes limitations in the model’s application. The risk of catching COVID-19 depends on an individual’s behaviour and the local dynamics of the disease, which are not modelled by QCOVID. These dynamics, such as local disease prevalence, change rapidly. Therefore, calibration of the model is likely to deteriorate rapidly,” they said.
With QCOVID described as a “living” model, Sperrin and MacMillan agreed that these problems could be mitigated with regular updating. [Diagn Progn Res 2018;2:23]