key: cord-0841848-v2s89wpw authors: van Royen, F. S.; Joosten, L. P.; van Smeden, M.; Rutten, F. H.; Geersing, G.-J.; van Doorn, S. title: Cardiovascular vulnerability predicts hospitalisation in primary care clinically suspected and confirmed COVID-19 patients: a model development and validation study date: 2021-05-14 journal: nan DOI: 10.1101/2021.05.12.21257075 sha: 55235f3deabbe653d4e7c654a0aabdc3dd98b20e doc_id: 841848 cord_uid: v2s89wpw Introduction Cardiovascular disease and diabetes have shown to be predictive of clinical deterioration towards critical illness or death in the hospitalised COVID-19 patient population. The aim of this study was to determine the incremental value of cardiovascular vulnerability - defined by the number of cardiovascular diseases and/or diabetes - in predicting the risk of escalation of care towards hospital referral in primary care patients with clinically suspected or confirmed COVID-19. Methods Data were retrospectively collected from three large Dutch primary care registries with routine care data of approximately 850,000 people. A prognostic prediction model was developed in two databases to assess the incremental value of cardiovascular vulnerability. Data from the first wave of COVID-19 infections in the Netherlands (March 1, 2020 to June 1, 2020) was used for derivation. A multivariable logistic regression model was fitted to predict hospital referral within 90 days follow-up after first consultation in consecutive adult patients seen in primary care for COVID-19 symptoms. Age, sex, the interaction between age and sex, and the number of underlying cardiovascular diseases and/or diabetes (0, 1, or <1) were pre-specified as predictors prior to the analyses. The model was (i) compared to a simpler model without the predictor number of cardiovascular diseases and/or diabetes and (ii) externally validated in COVID-19 confirmed patients during the second wave (June 1, 2020 to April 15, 2021) in all three databases. Results There were 5,475 patients included for model development and 6.8% had the primary outcome hospital referral. The model with number of cardiovascular diseases included as a predictor performed better than a model without this predictor (likelihood ratio test p<0.001). Older male patients with multiple cardiovascular diseases and/or diabetes had the highest predicted risk of hospital referral, reaching risks above 15-20% in these patients. The model was externally validated in a population of 16,693 COVID-19 patients. The observed risk was lower in this temporal validation cohort (4.7% versus 6.8%). The temporally validated c-statistic was 0.747 (95%CI 0.729-0.764) and the model showed good calibration. Conclusion In this general population study, risk of clinical deterioration after suspected or confirmed COVID-19 was on average 5.1% in the development and validation cohorts combined. This risk increased with age and was higher in males compared to females. Importantly, patients with concurrent cardiovascular disease and/or diabetes had higher predicted risks. Identifying those at risk for hospital referral could have clinical implications for COVID-19 early disease management in primary care. in the Netherlands (March 1 2020 to June 1 2020) that was temporally validated in a cohort of 86 patients from the 'second wave' of infections in the Netherlands (June 1 2020 to April 15 2021). 87 Where appropriate for this study, we adhered to the TRIPOD guideline for reporting prediction 88 For identification of the study population and data collection, the same methods were applied 111 in all three databases. Dutch primary care physicians link diagnoses and clinical symptoms to 112 the electronic medical records as diagnostic codes using the International Classification of 113 Primary Care (ICPC) coding system. The primary care physicians supplying clinical data to the 114 JGPN, AHA and ANH databases are trained in and experienced with using ICPC codes. For 115 the development cohort, COVID-19 suspected patients were identified using the ICPC codes 116 At the time of reporting, primary care physicians were recommended to use R81 and R83 for 118 indicating COVID-19 suspected and COVID-19 confirmed cases respectively. Records of 119 patients labelled with ICPC R74 (unspecified acute upper respiratory infection) were manually 120 screened for COVID-19 suspicion in the consultation text by three clinical scientists (FSvR, 121 LPTJ, and SvD) and cases of doubt were discussed until agreement was reached. Patients 122 with ICPC R74 and not having a synonym of or reference to COVID-19 suspicion or related 123 symptoms in this text were excluded. Of all included patients, baseline characteristics (i.e. age; 124 sex; comorbidities; and where available Body Mass Index (BMI), oxygen saturation and C-125 Reactive Protein (CRP)) were collected. Comorbidities (i.e. cardiovascular disease, pulmonary 126 disease, cancer), and history of relevant diseases were identified using ICPC. The model development cohort in JGPN and ANH yielded 5,475 eligible patients with an event 164 fraction of 0.068 (6.8%, n=373) for the primary outcome referral to the hospital. Prior to 165 prediction analysis, the number of allowed candidate predictors was determined. Based on the 166 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted May 14, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 number of candidate predictors that can be modelled was 30 with a R 2 Cox-Snell (R 2 cs) of 168 0.0495. As this R 2 cs was estimated in absence of a known value, varying R 2 cs from 0.0395 to 169 0.0595 yielded a minimum of 24 and a maximum of 37 candidate predictors, including 170 interaction terms. By using the candidate predictors age, sex, the interaction between age and 171 sex, and the number of cardiovascular diseases with three categories, the amount is 172 considerably lower than the maximum of allowed candidate predictors, and therefore the 173 Baseline characteristics were summarised using descriptive statistics with categorical 183 variables as numbers with percentages and continuous variables as means with standard 184 deviations or medians with interquartile ranges (IQR). A multivariable logistic regression 185 modelling approach was used. All included patients were entered in a fixed model (full model) 186 with the predictors age, sex, the interaction between age and sex, and the number of 187 cardiovascular diseases with three categories. To determine the incremental value of the 188 predictor number of cardiovascular diseases, a second model (simple model) was fitted using 189 only the predictors age, sex, and the interaction between age and sex. Age was considered as 190 a continuous variable and was studied using a restricted cubic spline function to account for (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted May 14, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 test. To determine incremental value, an alpha of 0.05 was used for the likelihood ratio test. 196 The full model was internally validated using Harrell's bootstrapping with 100 repetitions to 197 obtain optimism corrected estimates of the c-statistic and R 2 , and slope were calculated. For years and 44.3% were male. In ANH, the median age was 47 (IQR 34-59) years and 42.5% 232 were male and in the AHA dataset, the median age was 49 (IQR 36-60) years and 39.2% were 233 male. The differences between these three datasets in the validation cohort were also minor. 234 Around 15-20% suffered from one or more cardiovascular disease, again most often type 2 235 diabetes and coronary artery disease. 236 237 All 5,475 patients were used for model development. 373 patients (6.8%) had the outcome 239 hospital referral. All predefined model regression coefficients of the two models (i.e. age, sex, as a function of (increasing) age, stratified by sex and by the number of underlying 249 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Predicted risks were overall slightly higher than the observed risk (6.2% versus 4.6%) and the 255 calibration slope was 1.36. Overall discrimination showed an AUC of 0.747 (95%CI 0.729-256 0.764). All combined and individual database validation performance measures are shown in 257 table 5. Figure 2 shows the overall calibration plot and the calibration plots per database 258 separately are given in the supplementary materials. The outcome prevalence was lower in 259 the validation datasets than in the development datasets (4.7% versus 6.8%), probably 260 explaining the overestimation in predicted risks by the model. Although most COVID-19 patients experience a favourable prognosis without the need for 275 referral for hospital care, studies on COVID-19 mainly focussed on those seen in the hospital 276 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. identify specific patients at far greater risk of hospitalisation. In fact, for female patients without 281 cardiovascular comorbidity, the risk of clinical deterioration for which hospital referral was 282 deemed necessary remains well below 10% even in the eldest elderly (aged 80+), while in the 283 presence of cardiovascular diseases and/or diabetes, patients experience higher risks already 284 at younger ages, notably males. For instance, a male patient with two or more underlying 285 cardiovascular diseases will reach a predicted risk of 15% already at the age of around 57 286 years and this predicted risk will even further increase to above 20% from the age of 80 287 onwards. This indicates the incremental effect of cardiovascular diseases and/or diabetes in 288 addition to age and sex in predicting the risk for complicated COVID-19 disease trajectories in 289 primary care patients. patients, it has been demonstrated that there is an association between cardiovascular disease 298 and COVID-19 complicated disease trajectories, with higher prevalence of cardiovascular 299 disease and diabetes described in those with critical illness. (3-7,25) Our study shows that this 300 prognostically unfavourable effect is already present much earlier on in the COVID-19 disease 301 course, at the start of symptoms in primary care. 302 303 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. This research contributes to the evidence-based prognostication of community COVID-19. We 330 were able to use a large and representative database capturing both the 'first' and 'second' 331 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Furthermore, the outcome hospital referral was based upon a rigorous manual extraction of 344 medical records by pairs of researchers, yet not based upon actual linkage to hospital records. 345 Additionally, there are differences between our development and validation population: the 346 patients from the 'first wave' are all symptomatic patients that visited their primary care 347 physician for symptoms suggestive of COVID-19, while the patients from the 'second wave' -348 due to government recommendation for individuals to get tested even in the circumstance of 349 only mild symptoms -also include more healthy people that just informed their primary care 350 physician of their positive COVID-19 PCR status. This could also explain the lower event 351 fraction in the validation set (4.7% versus 6.8% in the development population). Finally, the 352 incremental value of cardiovascular disease on prognosticating COVID-19 was assessed in 353 different ways; although we did observe a highly significant change in the likelihood ratio test, 354 the delta in c-statistic and R 2 cs was only small to modest. Possible reasons for this include the 355 overall low risk of hospitalisation in most patients in our cohort, as well as that most patients 356 (80.2%) in fact in our cohort did not suffer from concurrent cardiovascular diseases and/or 357 diabetes, and it has been widely acknowledged that, notably in such scenario's, a change in 358 e.g. the c-statistic is difficult to achieve. 359 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted May 14, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 advantages for risk profiling patient with suspected or confirmed COVID-19 in the primary care 363 and community setting. This can have several important clinical implications. First, it may be 364 possible to identify patients that will benefit from closer monitoring and frequent follow-up at 365 home by predicting the risk of clinical deterioration early on in the COVID-19 disease course. 366 By intensified monitoring of higher risk patients, critical illness may be detected earlier, 367 potentially improving prognosis. Second, risk prediction could also support advanced care 368 planning. Informing both patients and physicians on the risk of severe illness, may help in 369 anticipating a more stringent or more lenient management. Last, risk profiling may be used for 370 targeting treatment. Vaccination strategies to prevent COVID-19, for instance, may focus on 371 those with cardiovascular disease and/or diabetes first. Additionally, experimental regiments 372 to treat COVID-19 may be addressed to high-risk patients that may benefit most. Examples 373 include for instance treatment with budesonide or colchicine; both treatment options likely 374 benefit patients most at higher prior probability of having an adverse prognosis.(29,30) 375 Nevertheless, in the end, risk prediction in primary care has to prove its value in daily practice 376 at the background of changing characteristics of this challenging COVID-19 pandemic and 377 influences of virus mutations. We however do hope that prognostic studies, like ours, may aid 378 physician by making informed, evidence-based decisions and thereby improve patient 379 outcomes. 380 381 In this general population study, risk of clinical deterioration after suspected or confirmed 383 COVID-19 was on average 5.1%. This risk increased with age and was higher in males 384 compared to females. Importantly, patients with concurrent cardiovascular disease and/or 385 diabetes had higher predicted risks. Identifying those at risk for hospital referral could have 386 clinical implications for COVID-19 early disease management in primary care. 387 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. CRP in mg/L (IQR) 6 (2-23) (n=480) 3 (1-9) (n=646) Hospital referrals 185 (6.5%) 188 (7.1%) All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted May 14, 2021. ; https://doi.org/10.1101/2021.05.12.21257075 doi: medRxiv preprint Temporal validation performance measures with 95% confidence intervals (CI). R 2 cs = R 2 535 Cox-Snell. 536 537 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted May 14, 2021. ; https://doi.org/10.1101/2021.05.12.21257075 doi: medRxiv preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted May 14, 2021. ; https://doi.org/10.1101/2021.05.12.21257075 doi: medRxiv preprint Calibration plot in the total validation cohort with hospitalisation as the outcome 549 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted May 14, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 Prediction 389 models for diagnosis and prognosis of covid-19: systematic review and critical 390 appraisal Mild or Moderate Covid-19 Cardiac complications in patients hospitalised with COVID-19 Prevalence and impact of 397 cardiovascular metabolic diseases on COVID-19 in China Predicting Mortality Due to SARS-CoV-2: A 401 Mechanistic Score Relating Obesity and Diabetes to COVID-19 Outcomes in Mexico Features of 404 20 133 UK patients in hospital with covid-19 using the ISARIC WHO Clinical 405 Characterisation Protocol: prospective observational cohort study Prevalence of 408 10 Calculating the 423 sample size required for developing a clinical prediction model Regression modeling strategies: with applications to linear models, 425 logistic and ordinal regression, and survival analysis R: A language and environment for statistical computing. R Foundation 429 for Statistical Computing Regression Modeling Strategies pROC: an 434 open-source package for R and S+ to analyze and compare ROC curves AHA (n=6,284) (IQR) 28 (24-32) (n=1,685) 27 (24-32) (n=1,823) 29 (25-33) (n=2,178) Median oxygen saturation in % (IQR) CRP in mg/L (IQR)