Since the beginning of the pandemic, hospitals have been constantly overcrowded, with several observed waves of infected cases and hospitalisations. To avoid as much as possible this situation, efficient tools to facilitate the diagnosis of COVID-19 are needed.
To evaluate and compare prediction models to diagnose COVID-19 identified in a systematic review published recently using performance indicators such as discrimination and calibration measures.
A total of 1618 adult patients present at two Emergency Department triage centers and for whom qRT-PCR tests had been performed were included in this study. Six previously published models were reconstructed and assessed using diagnostic tests as sensitivity (Se) and negative predictive value (NPV), discrimination (Area Under the Roc Curve (AUROC)) and calibration measures. Agreement was also measured between them using Kappa’s coefficient and IntraClass Correlation Coefficient (ICC). A sensitivity analysis has been conducted by waves of patients.
Among the 6 selected models, those based only on symptoms and/or risk exposure were found to be less efficient than those based on biological parameters and/or radiological examination with smallest AUROC values (< 0.80). However, all models showed good calibration and values above > 0.75 for Se and NPV but poor agreement (Kappa and ICC < 0.5) between them. The results of the first wave were similar to those of the second wave.
Although quite acceptable and similar results were found between all models, the importance of radiological examination was also emphasized, making it difficult to find an appropriate triage system to classify patients at risk for COVID-19.
Coronavirus disease 2019 (COVID-19) is an infectious disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The first cases of COVID-19 have been identified in December 2019 in Wuhan, China and the World Health Organization (WHO) declared COVID-19 a global pandemic on March 11, 2020. To date, more than 180 million cases have been confirmed and more than 4 millions of deaths are to be deplored . Symptoms range from mild (nothing for asymptomatic patients, fever, cough) to severe (shortness of breath or difficulty in breathing) and can lead to hospitalisation and admission to intensive care units (ICUs) or even death . The need to increase capacity and reorganise health care departments has become rapidly apparent as the number of potentially infected patients is on the increase. Since the beginning of the pandemic, hospitals have been continuously overcrowded, with several observed waves of infected cases and hospitalisations. For example, in Belgium, nearly 7500 beds were occupied in hospitals, including more than 1400 in intensive care, for a maximum of 2800 beds, a capacity that was increased during this period, which is a record during the second wave in November 2020 . To avoid at maximum this overcrowding and to take care of the patients in the best possible way, Emergency Departments (EDs) need to have efficient tools to confirm the diagnosis of COVID-19.
Until now, the most common first-line screening test used as gold standard to diagnose case of COVID-19 remains the quantitative real-time reverse transcriptase polymerase chain reaction (qRT-PCR) even if other screening tools based, for instance, on saliva or deep throat sputum (DTS) seemed to show similar efficacy . This test has the advantage to be realised quickly with proving high sensitivity, but with observed false-negative results [5, 6], and can produce results in 3–4 h . Unfortunately, in practice and due to a high number of tests realised, the waiting time for the results is generally longer, up to more than 24 h in hospitals. As suggested by Soedarsono et al.  to prevent the false-negative results, a combination of qRT-PCR with clinical, radiological or serological examinations could further support the clinicians in the triage of patients at high risk of COVID-19.
Systematic reviews or validation of prediction and prognostic models have been already realised in other studies [8, 9]. In Wynants et al. , the authors presented 33 diagnostic models for predicting COVID-19 with discrimination and calibration measures, but, although suggested but not realised by the authors themselves, these models have not been validated in other data sets or by independent investigators. In a recent external validation by Gupta et al. , the authors focused on patients confirmed with COVID-19 and different outcomes including mortality, ICU admission or progression to severe COVID-19. Their primary objective was to detect patient at high risk of deterioration, not to focus on an efficient tool to classify patients at high risk of COVID-19. In both studies, patients from only one specific wave were enrolled.
Based on a new systematic review of recent literature, the objectives of our research are to evaluate and compare prediction models for diagnosis of COVID-19 using an independent dataset including several waves of the pandemic. Results from our systematic review are already available and published in Archives of Public Health and entitled “A systematic review of prediction models to diagnose COVID-19 in adults admitted to healthcare centers” . This paper presents the selected models that will be validated in the present study. The evaluation of these models will be based on criteria of performance indicators as discrimination and calibration measures. Several agreement indexes will be also computed for purposes of model comparison. Moreover, as mentioned above, the employed database incorporated a lot of information over two complete waves (from March to June 2020 and from September 2020 to January 2021) of COVID-19 infected cases in Belgium and allowed the construction of alternative models to perform sensitivity analyses. In so doing, the results achieved are to provide an overview of the latest models as a basis for decision making, to guide and advance further studies in COVID-19 model development, and to confirm and/or verify the conclusions from Wynants et al.  that all COVID-19 diagnostic models to date are poorly developed or useless. Finally, some easy-to-use models could be highlighted to help clinicians classify patients at high risk of COVID-19.
Presentation of the selected models of the systematic review
Thirteen articles were included in our systematic review “A systematic review of prediction models to diagnose COVID-19 in adults admitted to healthcare centers”  and all were performed in 2020. Each study proposed diagnostic models for COVID-19 based on socio-demographics, clinical symptoms, blood tests, or other characteristics that were compared to the qRT-PCR test. The number of variables included in the model varied from 4 to 15. The presence of fever appeared in 7 models, the blood value of eosinophils in 6 models, and C-reactive protein (CRP) in 5 models. Four studies included comorbidities, gender (male) or chest X-ray as a predictor in their models. Finally, age, cough, white blood cells (WBC) were significant predictors in three out of 13 studies and lymphocytes was present in two out of the 13 studies. It can be noted that some variables can be directly collected while others require more time for their investigation. Sample sizes varied from 100 to 172 754 subjects and most studies were conducted at a single site or institution. Most of the models were developed using logistic regressions. From these logistic regressions, some authors developed a score and derived cut-off values. Models such as XGBoost, random forest and machine learning were also applied. All presented classification measures, with a wide range of sensitivity and specificity values depending on the model and 12 presented a discrimination measure. All models performed well to identify patients at risk of COVID-19 but only one proceeds to an external validation. The risk of bias was estimated as low for all models using the PROBAST tool .
Among these 13 articles, six were kept in this study to calculate scores, cut-off values and fit models. The other articles were discarded due to missing information > 20% and/or the impossibility to calculate the score or to fit the model due to the methodology used and/or lack of information despite contacts with the authors as it will be explained in detail in the following sections. As mentioned in , it can also be noted that the collected variables were sometimes country-specific and cannot be obtained if the model is to be put into use in a setting other than the research context. They are studies from Vieceli et al. , Tordjman et al. , Kurstjens et al. , Aldobyany et al. , Nakakubo et al.  and Fink et al.  and are presented in detail in the Additional file 1: Appendix A1. For most of them, a score and cut-off values could be obtained but a binary logistic regression was only available for three studies [12, 13, 17]. A score and cut-off value had to be refitted due to missing information and another missing variable was replaced by its median value to fit the logistic regression model. For the score derived from Nakakubo et al., the two categories “moderate and high risk” were combined due to few subjects in the last category in the sample.
Data in the present study have been extracted from the Medical and Economic Information Service (SIME) of the University Hospital Center of Liège (CHU Liège) and included patients present at the two ED triage centers  of the CHU (Sart Tilman and Notre-Dame des Bruyères) with suspicion of COVID-19. Data were collected during the period from March 2, 2020, to January 31, 2021. The number of patients was 8033. This period primarily covered two complete waves of cases and patient admissions in Belgium : from March 2020 to June 2020 (wave 1) and from September 2020 to January 2021 (wave 2).
Socio-demographic information (age and gender) as well as comorbidities (cardiac disease, immunosuppression, renal failure), symptoms (fever, dry or wet cough, dyspnea, diarrhea), blood parameters (lactic acid dehydrogenase LDH, CRP, procalcitonin, lymphocytes or lymphocytes count ALC, basophils, ferritin, leukocytes, neutrophils or neutrophils count ANC), radiology exams, particularly chest X-ray results, were collected in the database. Socio-demographic information and clinical symptoms were factors easily available at ED’s admission whereas hospital diagnostic resources required a more important time-to-results. In addition, radiological resources were not recommended to all patients, as their clinical presentation could not require this type of work-up. The outcome was confirmed or unconfirmed COVID-19 case using a qRT-PCR. Two different qRT-PCR tests were used during these periods: one adapted from the protocol described by Corman et al. ; and a second was a commercial assay using the cobas® 6800 platform (Roche) . Patients for whom no qRT-PCR test was realised, aged < 18 years and for whom no biological parameters were not included in the analysis, representing 80% of the original dataset.
Eventually, 1618 patients (20% from the original database) were included in this study, with no pregnant women and with 32.1% positive cases to the qRT-PCR.
Results were expressed as numbers and frequencies for qualitative parameters and as mean and standard deviation (SD), median (P50) and interquartile range (IQR, P25-P75) and range (Min–Max) for quantitative parameters, globally and by groups, namely positive and negative confirmed COVID-19 patients. The normality of the distribution of the quantitative parameters was investigated using the mean-median comparison, the histogram and Quantile–Quantile plot and tested with the Shapiro–Wilk hypothesis test.
For all models and scores, discrimination was assessed by the Area Under the Receiver Operating characteristic Curve (AUROC). Values could range from 0 to 1 where AUROC of 0.5 suggests no discrimination, values from 0.7 to 0.8 are considered acceptable, from 0.8 to 0.9 as excellent, and more than 0.9 as outstanding . For models that provided a cut-off value, sensitivity (Se), specificity (Sp), positive and negative predictive values (PPV and NPV respectively) were also calculated with 95% confidence interval (95CI%).
For models where information was available to calculate outcome probabilities, model calibration was assessed by means of the Brier score, values can range from 0 for a perfect model to 0.25 for a non-informative model [21, 22] and represents a measure of accuracy, and by calibration of predicted probabilities versus observed probabilities using LOESS- smoothed plot. Results were reported with calibration slopes and intercept (calibration-in-the-large). A perfect calibration slope is equal to 1 while slopes < 1 indicate an underestimation of low risk and overestimation of high risk and slopes > 1 means underestimation of high risk and overestimation of low risk. The estimated regression intercept represents the overall miscalibration, where 0 indicates good calibration, > 0 denotes an average underestimation, and < 0 denotes an average overestimation . For models where information for intercept was missing, we calculated the intercept using the model linear predictors as an offset term as suggested by Gupta et al. .
A sensitivity analysis was also conducted to compare Se, Sp, PPV, NPV but also discrimination and calibration measures for each selected model when using the complete data set and data set where patients were excluded from (1) wave 1 and between the two waves and (2) wave 2 and between the two waves.
Agreement between models was tested by means absolute and relative measures. For continuous scores, a pairwise comparison using Bland–Altman (BA) with limits of agreement (LOA) and two-way fixed IntraClass Correlation Coefficient (ICC (A, 1) ) with 95CI%. Values less than 0.5 are indicative of poor reliability, values between 0.5 and 0.75 indicate moderate reliability, values between 0.75 and 0.9 indicate good reliability, and values greater than 0.90 indicate excellent reliability . As scores had different value range, they were rescaled (mean/standard deviation) for these calculations. For binary or categorical score, Cohen’s Kappa was computed . Values > 0.6 indicates substantial agreement .
If a maximum 20% of the information to calculate a score or to fit a model was unobtainable from the data, calculation was based on the available variables . Scores and possible cut-off values were refitted to the actual number of variables. To fit models, missing variables were replaced by the mean, or the median value given in the original article. Where more than 20% of variables were missing, the score/model was discarded from this study.
The amount of missing data varied from 0.2% to 63%. Multiple imputation using the Fully Conditional Specification (FCS) method  was applied and all statistical analyses, diagnostic values, discrimination, calibration and agreement, were realised on the 60 generated data sets. The Rubin’s rules  was applied to pool the obtained results.
Results were significant at the 5% critical level (p < 0.05). The statistical analyses were carried out using SAS (version 9.4 for Windows) statistical package and R (version 4.0) with particular packages rms , CalibrationCurve , BlandAltmanLeh , and multiagree  and more common iir  and psych .
Description of the data
All variables used in this study are presented in Table 1 and correspond to those included in the 6 selected models. Among the 1618 patients presenting to the two ED triage centers and enrolled in this study, 54.6% were men with a median age of 73 (IQR: 62–82) years, 80.1% had an abnormal radiology and 32.1% were positive to the qRT-PCR test. Information about comorbidities, symptoms and biological parameters are also provided, for all patients and by results of qRT-PCR test. Highest percentages of missing values appeared in the information on comorbidities and symptoms. For some scores/models, transformation of biological parameters was made due to differences in measure units from the in original studies. Presence of fever, dry cough and abnormal result of chest X-ray were more present in positive patients. For biological parameters, median values of LDH, CRP and ferritin were higher in positive patients while median values of leukocytes and neutrophils were lower.
Diagnostic tests results
By applying the six selected models on the employed dataset, the results revealed that the smallest value for Se was given by the cut-off value derived from the score from Nakakubo et al.  but, globally, all studies provided, according to the choice of the cut-off value, quite similar results. For studies providing only one cut-off value, i.e. Vieceli et al.  and Nakakubo et al. , the values for Se and NPV were higher (> 0.7) than those for Sp and PPV (< 0.7). When several cut-offs values were proposed, namely in Aldobyany et al. , Tordjman et al.  and Kurstjens et al. , Se and PPV were higher for smaller values of cut-off while Sp and NPV were higher for greater values. Cut-offs values from 1 to 3 in Tordjam et al.  gave results > 0.8 for Se and NPV. Same remarks could be made for cut-off values from 2 to 5 in Kurtsjens et al. . Finally, multiple imputation gave robust results when compared with those from the whole sample (Table 2).
Discrimination and calibration measures
Discrimination and calibrations results are given in Table 3. Values of AUROC for all scores or models were acceptable (> 0.7) or excellent (> 0.8) excepted for the score derived from Aldobyany et al.  where the AUROC was equal to 0.60. Brier score were similar (0.18–0.19) and showed poor accuracy. All models showed however good calibration with calibration intercept equal to 0. Models derived from Vieceli et al.  and Tordjam et al.  tended to overestimated high risk (calibration slope < 1) while model derived from Fink et al.  tented to underestimate it (calibration slope > 1). Calibration plots were depicted in Fig. 1 and confirmed these findings. Values of AUROC seemed to be robust after multiple imputation according to the median and IQR values of scores in each group, positive and negative cases, in the sample.
Agreement between scores (rescaled) and binary were also assessed. Results are presented in Table 4 for ICC and Bland–Altman and in a heatmap (Fig. 2) for the Kappa. Except between the scores and cut-off scores derived from Vieceli et al.  and Kurstjens et al.  where ICC and Kappa were higher or equal to 0.5, the other scores and cut-off scores showed poor agreement. Indeed, ICC varied from to 0.16 to 0.44, Kappa from 0.02 to 0.46. Higher values of Kappa (> 0.6) were only observed between the cut-off score derived from the same original score. Even if the mean differences were close to 0, the limits of agreement (LOA) were very large according to the fact that score were rescaled in order to be compared.
Finally, to conduct a sensitivity analysis and to check if results remained stable according to different groups of patients, diagnostic tests results, discrimination and calibration measures were also calculated for each wave of patients. Results are given in the Additional file 1: Appendix A2 for Se, SP, PPV and NPV and in Additional file 1: Table A3 for discrimination and calibration measures. All calculations remained stable over time. Moreover, the same observations as those made when all the patients were taken into account in the analyses were still valid.
The performance of the studied scores and models
All AUROC values obtained in this study were close but generally smaller to those mentioned in the original paper. For example, in Fink et al. , original AUROC value was equal to 0.85 (0.81–0.90) which was a better result than in this study (0.78 (0.75–0.80)). It could be explained by the missing information about one variable (highest FiO2) to reconstruct the logistic regression model and that had to be replaced by the median value mentioned in the original article. Despite another missing information to calculate score derived from Aldobyany et al. , the obtained AUROC value in this study was similar. The missing information concerned the exposure risk and counted for 3 points out of 11. Globally, results remained acceptable but were excellent or outstanding in the original articles. For the binary regression models, all showed a good calibration with calibration intercept equal to 0 but two models tended to overestimate high risk while one tended to underestimate it. The Brier scores showed poor accuracy but they did not always perform well to evaluate clinical utility of diagnostic performance of these prediction models as mentioned in . When comparing diagnostic tests results for these two studies, again, quite similar but smaller values for Se/NPV and/or Sp/PPV were found when compared to the results of the original articles. A cut-off value between 1 and 3 could be suggested to maximize the value of Se in  and, for the same reason, a cut-off of 3 in . Unfortunately, no comparison could be made with results derived from Nakakubo et al. , because the original article proposed to classify patients in three groups (low, moderate and high risk). The last two groups were collapsed in this study and showed encouraging results for Se and NPV. All these results nuance or even contradict the critical remark of Wynants et al.  on the non-utility of this kind of model.
The relative importance of different types of predictive variables
The variables used in the different scores are described in details in Additional file 1: Appendix A4. As mentioned above, except for one score, all models presented in this study showed acceptable (> 0.70) or excellent (> 0.80) values for AUROC. The score with the smallest value (0.60) was derived from Aldobyany et al. , the only article that did not use biological parameters or results from radiological examination, highlighting the importance of these parameters. Unfortunately, only a score was given, and no calibration measures could be calculated. The model derived from Tordjman et al. , which was also very simple with only four biological parameters and no radiological exam, showed an AUROC value equal to 0.74 and good and quite similar calibration results when compared to the other most sophisticated models where the result from chest X-ray was used. However, it tended to overestimate high risk. These findings showed that a clinical triage of patients based only on symptoms and/or risk exposure is less efficient than one based on biological parameters and/or radiological examination. This confirms the suggestion cited in  about the importance of a combination of qRT-PCR with clinical, radiological or serological in the triage of patients to prevent false-negative results. It also confirms the difficulty to build a diagnostic model that would be simple, effective and based on information immediately available upon hospital arrival or at the triage stage. If anamnestic data are easily and rapidly available, biological analysis are time consuming and even more radiological findings. A balance must be found between the accuracy and the time needed to calculate the prediction that should be faster that qRT-PCR method. The future solution could potentially come from innovative strategies combining point-of-care testing and artificial intelligence-driven models as described recently by Soltan et al. .
Agreement between the scores and/or model
Finally, the scores, although calculated from some identical parameters and all showing acceptable results for diagnostic tests, discrimination and calibration, showed poor agreement between them. Indeed, ICC varied from to 0.16 to 0.44 while Kappa coefficients from 0.02 to 0.46. Moreover, the limit of agreement in the Bland–Altman analyses was very large. Poor agreement between the cut-off values could be explained by the fact that either the sensitivity values were very good at the expense of the specificity values or vice versa, and thus, the objectives, to maximize Se or Sp, were not the same for all cut-off scores. Another explanation could be the different predictor variables included in the scores and/or models, predictors with various clinical meaning. As mentioned by Gilbert et al. , the multiple predictive scores described currently in the literature present an important heterogeneity of the variables used (clinical, biological, radiological) related either to their recommended time-to-results, the availability of the data or resources in the concerned setting where the score was developed. In accordance, one important point that remains partially unanswered is the generalizability of these scores. It can be noted that the “best” value for agreement (ICC and Kappa > 0.5) was observed in two studies that have these two common parameters, LDH and results from chest X-ray, two parameters already recognised as important in the detection of COVID-19 .
The robustness of the results
A sensitivity analysis by waves of patients, i.e. corresponding to the periods from March 2020 to June 2020 (wave 1) and from September 2020 to January 2021 (wave 2) , was performed. Results remained stable. However, in the original studies, all patients were recruited in the period corresponding to wave 1, depending on the country where the study was conducted. Even though the characteristics of the patients seemed to be different between the different periods (Additional file 1: Appendix Table A4), recent studies did not show any age or comorbidity differences between patients hospitalized during the first and second waves, although they pointed to a shorter hospitalization period in the second wave [44, 45]. The derived scores, cut-off scores and models could be considered robust according to the different periods of the pandemic.
Strengths and limitations
This study evaluated and compared prediction models for diagnosis of COVID-19 identified through a systematic review of recent literature using performance indicators and agreement indexes, which, to the best of our knowledge, had not been done before. Moreover, the available database embodied a lot of information over two complete waves (from March to June 2020 and from September 2020 to January 2021) of cases in Belgium and allowed us to construct different models and to perform sensitivity analyses. Nevertheless, this study presents also certain important limitations. Indeed, from the original sample (8033), only 1618 (20%) patients were enrolled in the study due to the lack of information concerning the results from qRT-PCR test, biological parameters and radiological information. This information was necessary to calculate scores and/or fit models derived from the selected studies because (1) the result of qRT-PCR test was the outcome and (2) the other variables were predictors present in almost all the score and/or models selected in . Despite this, the number of included patients met the sample size rules-of-thumb that suggests at least 100 events and 100 non-events [46, 47]. Secondly, roughly 60% of missing values were observed for comorbidities and symptoms. This is explained by the fact that, even if ED nurses and physicians were aware of the data collection, they were unfortunately overwhelmed by the number of patients. However, despite this amount of missing data, the results obtained in this study after multiple imputation seemed to be robust when compared to with those observed in the sample. To our knowledge, regarding robustness, similar observation was not mentioned in . Another limitation was the inability to calculate score or to fit model for all articles from our previous systematic review. Indeed, more than 20% information like known exposure, Visual Analog Scale pain, Sequential Organ Failure Assessment or ethnicity required in [48,49,50,51,52,53] was not recorded in the database. Moreover, the modelling approach used in  could not be reproduced. So, only 6 articles out of 13 (46%) could be selected in this study, 5 presented score and one or more cut-off values, and only 3 presented results from binary regression models. Future external validation studies could address this issue of heterogeneity in model development and predictors by employing a more proactive and prospective approach in data collection. In so doing, more models could be externally validated with sufficient data, hence robust evidence being yielded. Finally, as already mentioned in , the choice of the studies included in the systemic review presents several biases like biases inherent in each selected studies but also specific to the systematic review like the database searched, the limitation of published with peer-reviewed studies and, of course, the period of time where this systematic review took place. Indeed, the COVID-19 pandemic has continued to evolve since then with new scientific and medical advances. That’s why, as suggested in , carrying out a living systematic review would be optimal.
All derived scores, cut-off scores and models showed quite acceptable and similar results in terms of diagnostic tests, discrimination and calibration measures. Moreover, the values of the different measures calculated, although lower than those of the original articles, were still close and lead to similar conclusions. Despite this fact, poor agreement was found between the different derived scores and cut-off scores. Two scores had an advantage over others such that COVID-19 diagnosis could be calculated from rapid diagnostic as comorbidities or symptoms and/or blood sample. Nevertheless, these two models had the lowest but though acceptable values for discrimination and calibration measures, highlighting the importance of radiological examination to obtain more efficient models, which entails difficulties in specifying an easy-to-use tool to help clinicians to classify patients at risk of COVID-19.
Availability of data and materials
Anonymized data are available upon reasonable request from the corresponding author.
95% Confidence interval
Area under the Roc curve
University Hospital Center
Coronavirus disease 2019
Intensive care unit
Lactic acid dehydrogenase
Limits of agreement
Negative predictive value
Percentile 50, median
Negative predictive value
Quantitative real-time reverse transcriptase polymerase chain reaction
Severe acute respiratory syndrome coronavirus 2
Medical and Economic Information Service
White blood cell
World Health Organization
WHO Coronavirus (COVID-19) Dashboard|WHO Coronavirus (COVID-19) Dashboard With Vaccination Data [Internet]. [cited 2021 Apr 20]. Available from: https://covid19.who.int/.
Coronavirus [Internet]. [cited 2021 Apr 9]. Available from: https://www.who.int/health-topics/coronavirus#tab=tab_3.
COVID-19-Situation épidémiologique|Coronavirus Covid-19 [Internet]. [cited 2021 Apr 20]. Available from: https://covid-19.sciensano.be/fr/covid-19-situation-epidemiologique.
Khiabani K, Amirzade-Iranaq MH. Are saliva and deep throat sputum as reliable as common respiratory specimens for SARS-CoV-2 detection? A systematic review and meta-analysis. Am J Infect Control [Internet]. 2021 Mar [cited 2021 Apr 9]; Available from: https://linkinghub.elsevier.com/retrieve/pii/S0196655321001401.
Soedarsono S, Febriani A, Hasan H, Widyoningroem A. Management of severe COVID-19 patient with negative RT-PCR for SARS-CoV-2: role of clinical, radiological, and serological diagnosis. Radiol Case Reports [Internet]. 2021 Mar [cited 2021 Apr 9]; Available from: https://linkinghub.elsevier.com/retrieve/pii/S1930043321001874.
Arevalo-Rodriguez I, Buitrago-Garcia D, Simancas-Racines D, Zambrano-Achig P, Del Campo R, Ciapponi A, et al. False-negative results of initial RT-PCR assays for COVID-19: a systematic review. PLoS ONE. 2020;15(12):e0242958.
What tests could potentially be used for the screening, diagnosis and monitoring of COVID-19 and what are their advantages and disadvantages? - The Centre for Evidence-Based Medicine [Internet]. [cited 2021 Apr 9]. Available from: https://www.cebm.net/covid-19/what-tests-could-potentially-be-used-for-the-screening-diagnosis-and-monitoring-of-covid-19-and-what-are-their-advantages-and-disadvantages/.
Wynants L, Van Calster B, Collins GS, Riley RD, Heinze G, Schuit E, et al. Prediction models for diagnosis and prognosis of COVID-19: systematic review and critical appraisal. BMJ. 2020;369:26.
Gupta RK, Marks M, Samuels THA, Luintel A, Rampling T, Chowdhury H, et al. Systematic evaluation and external validation of 22 prognostic models among hospitalised adults with COVID-19: an observational cohort study. Eur Respir J. 2020;56(6). https://doi.org/10.1183/13993003.03498-2020.
Locquet M, Diep AN, Beaudart C, Dardenne N, Brabant C, Bruyère O, et al. A systematic review of prediction models to diagnose COVID-19 in adults admitted to healthcare centers. Arch Public Health. 2021;79:1–15. https://doi.org/10.1186/s13690-021-00630-3.
Wolff RF, Moons KGM, Riley RD, Whiting PF, Westwood M, Collins GS, et al. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann Intern Med. 2019;170(1):51.
Vieceli T, de OliveiraFilho CM, Berger M, Saadi MP, Salvador PA, Anizelli LB, et al. A predictive score for COVID-19 diagnosis using clinical, laboratory and chest image data. Brazilian J Infect Dis. 2020;24(4):343–8.
Tordjman M, Mekki A, Mali RD, Saab I, Chassagnon G, Guillo E, et al. Pre-test probability for SARS-Cov-2-related infection score: the PARIS score. PLoS ONE. 2020;15:e0243342.
Kurstjens S, van der Horst A, Herpers R, Geerits MWL, Kluiters-De Hingh YCM, Göttgens EL, et al. Rapid identification of SARS-CoV-2-infected patients at the emergency department using routine testing. Clin Chem Lab Med. 2020;58(9):1587–93.
Aldobyany A, Touman A, Ghaleb N, Alsaggaf R, Murtaza N, Hamada A, et al. Correlation between the COVID-19 respiratory triage score and SARS-COV-2 PCR test. Front Med. 2020;7. Available from: https://pubmed.ncbi.nlm.nih.gov/33365322/.
Nakakubo S, Suzuki M, Kamada K, Yamashita Y, Nakamura J, Horii H, et al. Proposal of COVID-19 clinical risk score for the management of suspected COVID-19 cases: a case control study. BMC Infect Dis. 2020;20(1).
Fink DL, Khan PY, Goldman N, Cai J, Hone L, Mooney C, et al. Development and internal validation of a diagnostic prediction model for COVID-19 at time of admission to hospital. QJM An Int J Med. 2020;114:699–705.
Saegerman C, Gilbert A, Donneau A-F, Gangolf M, Diep AN, Meex C, et al. Clinical decision support tool for diagnosis of COVID-19 in hospitals. PLoS ONE. 2021;16(3):e0247773. https://doi.org/10.1371/journal.pone.0247773.
Corman VM, Landt O, Kaiser M, Molenkamp R, Meijer A, Chu DK, et al. Detection of 2019 novel coronavirus (2019-nCoV) by real-time RT-PCR. Eurosurveillance. 2020;25(3):2000045. https://doi.org/10.2807/1560-7917.ES.2020.25.3.2000045.
Hosmer Jr. DW, Lemeshow S, Rodney XS. Applied Logistic Regression, 3rd Edition | Wiley [Internet]. 2013 [cited 2021 Apr 8]. 528 p. Available from: https://www.wiley.com/en-us/Applied+Logistic+Regression%2C+3rd+Edition-p-9780470582473.
Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology. 2010;21:128–38.
Lindhiem O, Petersen IT, Mentch LK, Youngstrom EA. The importance of calibration in clinical psychology. Assessment. 2020;27:840–54.
Stevens RJ, Poppe KK. Validation of clinical prediction models: what does the “calibration slope” really measure? J Clin Epidemiol. 2020;1(118):93–9.
McGraw KO, Wong SP. Forming inferences about some intraclass correlation coefficients. Psychol Methods. 1996;1(1):30–46.
Koo TK, Li MY. A guideline of selecting and reporting intraclass correlation coefficients for reliability research. J Chiropr Med. 2016;15(2):155.
Measuring Agreement: Models, Methods, and Applications|Wiley [Internet]. [cited 2021 Apr 20]. Available from: https://www.wiley.com/en-us/Measuring+Agreement%3A+Models%2C+Methods%2C+and+Applications-p-9781118078587.
Hallgren KA. Computing inter-rater reliability for observational data: an overview and tutorial. Tutor Quant Methods Psychol. 2012;8(1):23.
Aguayo GA, Donneau AF, Vaillant MT, Schritz A, Franco OH, Stranges S, et al. Agreement between 35 published frailty scores in the general population. Am J Epidemiol. 2017;186(4):420–34.
Liu Y, De A. Multiple imputation by fully conditional specification for dealing with missing data in a large epidemiologic study. Int J Stat Med Res. 2015;4(3):287–95.
Rubin DB, Wiley J, York N, Brisbane C, Singapore T. Multiple imputation for nonresponse in surveys. In: Rubin DB, editor. Hoboken: Wiley. https://doi.org/10.1002/9780470316696.
Harrell E, Frank M. Package “rms” Title Regression Modeling Strategies [Internet]. 2021 [cited 2021 Apr 8]. Available from: https://github.com/harrelfe/rms.
GitHub-BavoDC/CalibrationCurves: The package with the val.prob.ci.2 code, formerly called the val.prob.ci.2 package. [Internet]. [cited 2021 Apr 8]. Available from: https://github.com/BavoDC/CalibrationCurves/.
GitHub-cran/BlandAltmanLeh: This is a read-only mirror of the CRAN R package repository. BlandAltmanLeh—Plots (Slightly Extended) Bland-Altman Plots [Internet]. [cited 2021 Apr 20]. Available from: https://github.com/cran/BlandAltmanLeh.
Vanbelle S. Comparing dependent kappa coefficients obtained on multilevel data. Biometrical J. 2017;59(5):1016–34.
Matthias Gamer A, Matthias Gamer M. Package “irr” Title Various Coefficients of Interrater Reliability and Agreement [Internet]. 2019 [cited 2021 Apr 20]. Available from: https://www.r-project.org.
Package “psych” Title Procedures for Psychological, Psychometric, and Personality Research [Internet]. 2021 [cited 2021 Apr 20]. Available from: https://personality-project.org/r/psych/.
Vieceli T, de Oliveira Filho CM, Berger M, Saadi MP, Salvador PA, Anizelli LB, et al. A predictive score for COVID-19 diagnosis using clinical, laboratory and chest image data. Brazilian J Infect Dis. 2020;24(4):343–8.
Aldobyany A, Touman A, Ghaleb N, Alsaggaf R, Murtaza N, Hamada A, et al. Correlation between the COVID-19 respiratory triage score and SARS-COV-2 PCR test. Front Med. 2020;7: 605689.
Tordjman M, Mekki A, Mali RD, Saab I, Chassagnon G, Guillo E, et al. Pre-test probability for SARS-Cov-2-related infection score: the PARIS score. PLoS ONE. 2020;15: e0243342.
Assel M, Sjoberg DD, Vickers AJ. The Brier score does not evaluate the clinical utility of diagnostic tests or prediction models. Diagnostic Progn Res. 2017;1(1):1–7. https://doi.org/10.1186/s41512-017-0020-3.
Soltan AAS, Yang J, Pattanshetty R, Novak A, Yang Y, Rohanian O, et al. Real-world evaluation of rapid and laboratory-free COVID-19 triage for emergency care: external validation and pilot deployment of artificial intelligence driven screening. Lancet Digit Heal. 2022;4(4):e266–78.
Gilbert A, Ghuysen A. Triage in the time of COVID-19. Lancet Digit Heal. 2022;4:e210–1. https://doi.org/10.1016/S2589-7500.
Martha JW, Wibowo A, Pranata R. Prognostic value of elevated lactate dehydrogenase in patients with COVID-19: a systematic review and meta-analysis. Postgrad Med J. 2021;1–6. Available from: http://pmj.bmj.com/.
Contou D, Fraissé M, Pajot O, Tirolien JA, Mentec H, Plantefève G. Comparison between first and second wave among critically ill COVID-19 patients admitted to a French ICU: no prognostic improvement during the second wave? Crit Care. 2021;25(1):1–4. https://doi.org/10.1186/s13054-020-03449-6.
Iftimie S, Lopez-Azcona AF, Vallverdu I, Hernandez-Flix S, De Febrer G, Parra S, et al. First and second waves of coronavirus disease-19: a comparative study in hospitalized patients in Reus, Spain. PLoS One. 2021;16. Available from: /pmc/articles/PMC8011765/.
Collins GS, Ogundimu EO, Altman DG. Sample size considerations for the external validation of a multivariable prognostic model: a resampling study. Stat Med. 2016;35(2):214–26. https://doi.org/10.1002/sim.6787.
Steyerberg EW. Clinical prediction models: a practical approach to development, validation, and updating (Ser. Statistics for biology and health). Berlin: Springer; 2019.
Bar S, Lecourtois A, Diouf M, Goldberg E, Bourbon C, Arnaud E, et al. The association of lung ultrasound images with COVID-19 infection in an emergency room cohort. Anaesthesia. 2020;75(12):1620–5.
McDonald SA, Medford RJ, Basit MA, Diercks DB, Courtney DM. Derivation with internal validation of a multivariable predictive model to predict COVID-19 test results in emergency department patients. Acad Emerg Med. 2020;28(2):206–14.
Huang D, Wang T, Chen Z, Yang H, Yao R, Liang Z. A novel risk score to predict diagnosis with coronavirus disease 2019 (COVID-19) in suspected patients: a retrospective, multicenter, and observational study. J Med Virol. 2020;92(11):2709–17.
Sung J, Choudry N, Bachour R. Development and validation of a simple risk score for diagnosing COVID-19 in the emergency room. Epidemiol Infect. 2020;148:1–17.
Callejon-Leblic MA, Moreno-Luna R, Del Cuvillo A, Reyes-Tejero IM, Garcia-Villaran MA, Santos-Peña M, et al. Loss of smell and taste can accurately predict COVID-19 infection: a machine-learning approach. J Clin Med. 2021;10(4):570.
Gupta-Wright A, Macleod CK, Barrett J, Filson SA, Corrah T, Parris V, et al. False-negative RT-PCR for COVID-19 and a diagnostic risk score: a retrospective cohort study among patients admitted to hospital. BMJ Open. 2021;11(2): e047110.
Plante TB, Blau AM, Berg AN, Weinberg AS, Jun IC, Tapson VF, et al. Development and external validation of a machine learning tool to rule out COVID-19 among adults in the emergency department using routine blood tests: a large, multicenter, real-world study. J Med Internet Res. 2020;22(12): e24048.
Ethics approval and consent to participate
All procedures carried out in this study were in accordance with the ethical standards of the institutional and national responsible committee on human experimentation and the Helsinki Declaration of 1964 and its later amendments or equivalents.The study design was approved in principle by the Ethical Committee of the University of Liege before it started while the official approval with reference number 2021-013 was received on 26 January 2021. Informed consent was exempted by the Ethical Committee of the University of Liège for this type of study due to the anonymity of data retrieval.
Consent for publication
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
: Table A1. Presentation of the selected articles. Table A2. Sensitivity, specificity, predictive negative and positive value for scores from selected articles with cut-off values. Table A3. AUROC and calibration measures for scores and formulas from selected articles with cut-off values. Table A4. Descriptive analysis for all parameters used in scores/formulas, globally and by time period.
About this article
Cite this article
Dardenne, N., Locquet, M., Diep, A.N. et al. Clinical prediction models for diagnosis of COVID-19 among adult patients: a validation and agreement study. BMC Infect Dis 22, 464 (2022). https://doi.org/10.1186/s12879-022-07420-4