Reliability and reproducibility of clinical phenotypes developed during the first wave of COVID-19: A validation study in critically ill patients from the second and third wave

doi:10.21203/rs.3.rs-2598565/v1

Background: During the first wave of the COVID-19 pandemic, different clinical phenotypes were published. However, none of them have been validated in subsequent waves, so their current validity is unknown. The aim of the study is to validate the unsupervised cluster model developed during the first pandemic wave in a cohort of critically ill patients from the second and third pandemic waves.

Methods: Retrospective, multicentre, observational study of critically ill patients with confirmed COVID-19 disease and acute respiratory failure admitted from 74 Intensive Care Units (ICU) in Spain. To validate our original phenotypes model, we assigned a phenotype to each patient of the validation cohort using the same medoids, the same number of clusters (n= 3), the same number of variables (n= 25) and the same discretisation used in the development cohort. The performance of the classification was determined by Silhouette analysis and general linear modelling. The prognostic models were validated, and their performance was measured using accuracy test and area under curve (AUC)ROC.

Results: The database included a total of 2,033 patients (mean age 63[53-92] years, 1643(70.5%) male, median APACHE II score (12[9-16]) and SOFA score (4[3-6]) points. The ICU mortality rate was 27.2%. Although the application of unsupervised cluster analysis classified patients in the validation population into 3 clinical phenotypes. Phenotype A (n=1,206 patients, 59.3%), phenotype B (n=618 patients, 30.4%) and phenotype C (n=506 patients, 24.3%), the characteristics of patients within each phenotype were significantly different from the original population. Furthermore, the silhouette coefficients were close to or below zero and the inclusion of phenotype classification in a regression model did not improve the model performance (accuracy =0.78, AUC=0.78) with respect to a standard model (accuracy = 0.79, AUC=0.79) or even worsened when the model was applied to patients within each phenotype (accuracy = 0.80, AUC 0.77 for Phenotype A, accuracy=0.73, AUC= 0.67 for phenotype B and accuracy= 0.66 , AUC= 0.76 for phenotype C )

Conclusion: Models developed using machine learning techniques during the first pandemic wave cannot be applied with adequate performance to patients admitted in subsequent waves without prior validation.

Trial Registration: The study was retrospectively registered (NCT 04948242) on June 30, 2021

Severe SARS-CoV-2 infection

Phenotypes

Risk factors

Prognosis

Machine Learning

validation

Phenotypes have been described as a way to characterize patients heterogeneous in different diseases (1–11). Patients with sepsis are amongst those patients with a very variable presentation including comorbid conditions, age, and inflammatory response (1–3). More recently, this approach has been also introduced and tested in COVID-19 patients (4–13). COVID-19 is a single pathogen disease with some specific organ failure damage mainly restricted to the lungs. Due to the different array of clinical presentation from asymptomatic disease to severe acute respiratory distress syndrome, many research groups, including large number of patients aimed to phenotype and subgroup patients in order to better characterize clinical trajectories and response to treatment (4–13).

During the first wave, patients with COVID-19 represented a huge challenge for healthcare systems due to the high demand of invasive mechanical ventilation and refractory hypoxemia. Our group (13), and others (4–12), found mainly three phenotypes based on the mortality rates finding some particular clinical characteristics associated with a guarded prognosis. We hypothesised that clinical phenotypes are not static models and clinical characteristics previously published will not avail to be implemented in other clinical settings and when patients are not within the same time frames (14,15). Based on these considerations, the aim of the study is to assess the reproducibility of the application of an unsupervised cluster model developed during the first pandemic wave, in a cohort of critically ill patients from the second and third pandemic waves.

Study design

This study is a pre-planned secondary analysis derived from a multi-centre, prospective, observational study consisting of a large-scale data source of critically ill patients to determine the reproducibility of the application of a machine learning model developed during the first COVID-19 pandemic wave to identify clinical phenotypes, when applied to a cohort of 2330 critically ill patients from the second and third pandemic waves (1 July 2020 to 31 July 2021). The study was retrospectively registered at Clinical-Trials.gov (NCT 04948242) on the 30th of June 2021. Recorded variables are shown in additional file (e-Table 1).

The study received approval by the Institution’s Internal Review Board (Comitè Ètic d’Investigació amb Medicaments -CEIm-IISPV- IRB# CEIM/066/2020). Local researchers-maintained contact with a study team member, and participating hospitals obtained local ethics committee approval. The study was carried out according to the principles of the Declaration of Helsinki and the Clinical Trials Directive 2001/20/EC of the European Parliament relating to the Good Clinical Practice guidelines.

We reported results in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (16).

Patients

Consecutive patients older than 16 years of age were eligible for participation if admitted to one of the 74 participating ICUs (72 from Spain, 1 from Ireland and 1 from Andorra) with acute respiratory failure and a COVID-19 diagnosis was confirmed by a positive reverse transcriptase-polymerase chain reaction for SARS-CoV-2 from upper or lower respiratory tract samples [17]. The follow-up of patients was scheduled until October 31, 2021, which confirmed ICU discharge or death whichever occurred first.

Outcomes: The primary outcome included all-causes of ICU mortality. Patients who were discharged alive from ICU were evaluated in the data as alive considering mortality was defined as any in-ICU death. All complications and outcomes were followed during ICU admission.

Data preparation

We first assessed the value distributions and missingness of the 25 candidate clinical variables. For data quality control, continuous variables with missing data > 20% were excluded of database. Missing data for continuous variables were imputed using R-package "missForest" for the statistical software R/CRAN. The imputation was applied to impute the missing values of D dimer (15%), Ferritin (15%), D-Lactate dehydrogenase (10%), Procalcitonin (10%), creatinine (10%), SOFA score (10%), APACHE II score (9%) and C-reactive protein (CRP) (5%). Categorical data (including ICU mortality) were available for all patients. The study definitions used in the present analysis are shown in the Additional file (Page 2).

Statistical analysis

No statistical sample size calculation was performed a priori, and sample size was equal to the number of patients admitted to the participant’s ICUs with confirmed COVID-19 during the study period. To describe baseline characteristics, the continuous variables were expressed as median (Q1-Q3 range]) and categorical variables as number of cases (percentage). For patient demographics and clinical characteristics, differences between groups were assessed using the

chi-squared test and Fisher’s exact test for categorical variables, and the Mann–Whitney U or Wilcoxon test for continuous variables.

1.- Original phenotype derivation

We originally derived phenotypes using the development cohort (13). More Specifically, an unsupervised clustering analysis was applied in a cohort population of 2022 critically il patients admitted to ICU during the first pandemic wave, and three different clinical phenotypes were derived: (1) Cluster A phenotype (severe disease); (2) Cluster B phenotype (critical disease) and (3) Cluster C phenotype (life-threatening disease). The characteristics of each phenotype are shown in Table 1 and more detailed information on the development of phenotypes is available in the original publication (13). An overview of the primary analysis plan is outlined in Figure 1.

2.- Phenotype external validation

To evaluate the reproducibility, we validated our original phenotypes like the primary analysis. We assigned a phenotype to each patient of the validation cohort using the medoids obtained with the development cohort. The analysis was carried out using the same number of clusters (n= 3), the same number of variables (n= 25) and the same discretisation of variables used in the development of the original model (13). We then characterised the new population assigned to each phenotype by assessing their differences in demographics, clinical and laboratory variables, comorbidities, and mortality from the original phenotypes.

3.- Performance evaluation of phenotype classification: The performance of the classification was determined by:

a.- Silhouette analysis: The silhouette analysis measures how well an observation is clustered, and it estimates the average distance between clusters. The silhouette coefficient (Si) measures how similar an object “i” is to the other objects in its own cluster versus those in the neighbour cluster. Si values range from 1 to - 1: A value of Si close to 1 indicates that the object is well clustered. A value of Si close to -1 indicates that the object is poorly clustered, and that assignment to some other cluster would probably improve the overall results. Finally, the silhouette plot displays a measure of how close each point in one cluster is to points in the neighboring clusters.

b) General linear model (GLM): In addition, for the aims of confirming phenotypes performance, we developed a supervised predictive model (GLM) with validation dataset. GLM (original) was performed to investigate the association between baseline (on ICU admission) variables and ICU-mortality. The GLM model comprised factors of clinical interest and all significant covariates (p < 0.05) in the univariate analysis of ICU mortality and presence of collinearity was studied by variance inflation factors (VIF). The results are presented as odds ratios (OR) and 95% confidence intervals (CI).

To assess whether the new classification of patients according to phenotype improves the performance of the original GLM model, a new GLM model was run again now with the inclusion of the phenotype variable (Modified model) and finally the model was applied in patients within each phenotype (model A, B and C).

For all model internal validation, database was randomly split into two subsets: (a) a “training set” (70%), and (b) a “validation set” (30%). Models’ performance was examined using accuracy test, sensitivity(S), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV) and area under ROC curve (AUC). Data analysis was performed using R software (cran.r-project.org).

Patients’ characteristics at ICU admission

From July 1, 2020, to July 31, 2021, a total of 2,330 critically ill patients from 74 ICUs were enrolled in the present analysis. The median age was 63 (63-92) years, and 1643(70.5%) were men. A total of 1,630 (70.0%) patients had at least one coexisting comorbidity. Arterial hypertension (n=1123; 48.2%) and obesity (n= 942;40.4%) were the most frequently comorbid conditions reported. The severity of illness was intermediate according to the APACHE II (12 [9-16]) and SOFA (4 [3-6]) scores. PaO2/FiO2 ratio on the day of ICU admission was 124.8 (99-145) and 915 (39.3%) patients required high flow nasal cannula (HFNC). Only 26.5% (n=618) of patients required mechanical ventilation on ICU admission. The ICU crude mortality was 27.2% (n=634). As expected, patients who died were more severe, had a higher frequency of comorbidities and complications than those who survived. The clinical characteristics of the of the patients and laboratory results are shown in Table 2.

Comparing the validation population with the original cohort population (10) (Table 1), the validation population showed lower age, lower severity (APACHE) and organ dysfunction (SOFA) level, lower inflammatory status, and less development of complications such as shock and acute kidney injury (AKI). In contrast, the presence of hypertension, obesity, diabetes, chronic renal failure, and asthma were more frequent in the validation population. Ventilatory support was different between the 2 populations, with a decrease in the use of invasive mechanical ventilation (IMV) and an increase in the use of high-flow nasal cannulas (HFNC) and non-invasive ventilation (NIV) upon admission to the ICU in the validation population. Despite these characteristics, crude mortality in ICU was lower in the validation cohort but this difference did not achieve statistical significance (Table 1).

Application of unsupervised cluster analysis in validation population

The 25 variables considered as predictors in the development of the original model (10) were included in the new validation model, considering the same discretisation with respect to the original model. The variables categorized as independently associated with ICU mortality are shown in Table 2.

The application of unsupervised cluster analysis allowed the classification of patients in the validation population into 3 clinical phenotypes. Phenotype A (severe disease) included 1206 patients (59.3%), phenotype B (critical disease) included 618 patients (30.4%), while phenotype C (life-threatening disease) included the remaining 506 patients (24.3%). ICU mortality increased significantly from phenotype A (20.9%), B (29.4%) to C (39.5%, p<0.001 for all comparisons).

The 3 clinical phenotypes in a lower dimensional space are shown in additional file (e-Figure 1). The size and characteristics of the validation and original phenotypes in the 3-class model are shown in Table 1. The number of patients included in the validation phenotype A represented a significantly higher percentage of total of patients (59%) compared to the original phenotype A which only included 26.6% of the population (p<0.001). Validation phenotype A patients were less severe, with lower levels of inflammation, less development of shock and similar frequency of comorbidities (except for obesity) compared to the original phenotype A patients (Table 1 and Figure 2). Despite these differences, the crude ICU mortality of the validation phenotype A (20.9%) was not different from that of the original phenotype A (20.3%, p=0.77).

Patients classified within validation phenotype B represented a lower percentage of total of validation population (26.5%) respect of original phenotype B (30.8, p<0.001). Validation phenotype B patients were less severe, with lower levels of inflammation, less development of shock and similar frequency of comorbidities compared to the original phenotype B patients (Table 1 and Figure 2). Despite these characteristics, no differences in crude ICU mortality were observed between the two phenotypes (25.5% vs 29.4%, p=0.17 for phenotype B original and validation respectively). Finally, patients included in validation phenotype C represented a lower percentage of total of validation population (21.7%) respect of original phenotype C (42.5%, p<0.001). Validation phenotype C patients were less severe, with lower levels of inflammation and less development of shock upon ICU admission. In contrast, the presence of hypertension, obesity, diabetes, chronic renal failure, coronary disease, chronic obstructive pulmonary disease, and asthma were more frequent in the validation C phenotype (Table 1 and Figure 2). The ICU crude mortality was lower in validation C phenotype (39.5%) than original C phenotype (45.5%, p<0.01).

The determination of the Si coefficient (silhouette analysis) allowed us to observe a mean value for the overall population of -0.007, with mean values of 0.384, -0.341 and -0.556 for phenotypes A, B and C respectively (additional file e-Table 2 and e-Figure 2). These findings suggest that the classification performed is not adequate and most of the classified are either between two clusters (Si close to 0) or probably located in the wrong cluster (Si -0).

Supervised predictive model validation.

a. Original GLM model

To further assess the robustness and usefulness of the new phenotypes developed, a GLM model was carried out. The 25 clinical and laboratory variables used for the clustering analysis were used as predictors in the original GLM (additional file e-Table 3 and e-Table 4). AKI (OR=2.5[1.9-4.4]), myocardial dysfunction (OR=2.2[1.4-3.3]), IMV (OR=1.9[1.4-2.6]), GAP-ICU (OR=1.08[1-03-1.12]), age (OR=1.03[1.02-1.05], RCP (OR=1.02[1.01-1.04] and PaO2/FiO2 (OR=0.99[0.99-1.0]) were variables associated with ICU mortality (additional file e-Figure 3). No collinearity was observed (additional file e-Table 5) and the performance of the model are show in Table 3 and additional file (e-Table 6 and e-Figure 4).

b. Modified GLM model with the inclusion of the phenotype classification.

When the phenotype variable was included in the model (modified GLM), it was observed that phenotype type was not associated with mortality, while the variables independently associated with mortality were the same as in the original GLM model. (Additional file e-Table 7 and e-Figure 5). No collinearity was observed (additional file e-Table 8) and the performance of the model are show in Table 3 and additional file (e-Table 9 and e-Figure 6).

c. GLM model in the A Phenotype population

The characteristics of the patients classified within phenotype A according to the evolution in ICU can be seen in additional file (e-Table 10). When the GLM model was applied in this population (additional file e-Table 11), it was observed that myocardial dysfunction (OR= 3.6[1.8-7.2]), AKI (OR= 2.9[1.8-4-6], age (OR= 1.03[1.02-1.05]) and SOFA score (OR= 1.01[1-1.01]) were variables associated with ICU mortality (additional file e-Figure 7). The performance of the model is show in Table 3 and additional file (e-Table 12 and e-Figure 8).

GLM model in the B Phenotype population

The characteristics of the patients classified within phenotype B according to the evolution in ICU can be seen in the additional file (e-Table 13). When the GLM model was applied in this population (additional file e-Table 14), it was observed that presence of more than 3 quadrants infiltrates in chest x-ray (OR= 6.5[1.1-37.7]), myocardial dysfunction (OR= 2.6[1.05-6.8]), AKI (OR= 2.0[1.08-3.8], arterial hypertension (OR= 1.9[1.04-3.7]), age (OR= 1.04[1.02-1.07]) and PCT (OR= 1.01[1-1.01]) were variables associated with ICU mortality (additional file e-Figure 9). The performance of the model is show in Table 3 and additional file (e-Table 15 and e-Figure 10).

GLM model in the C Phenotype population

The characteristics of the patients classified within phenotype C according to the evolution in ICU can be seen in the additional file (e-Table 16). When the GLM model was applied in this population (additional file e-Table 17), it was observed that AKI (OR= 4.4[2.4-8.2], IMV (OR=2.0[1.06-3.9]), Angiotensin Converting Enzyme Inhibitors (ACEI) (OR= 1.9[1.03-3.7]), GAP-ICU (OR=1.1[1.08-1.2]), age (OR=1.05[1.02-1.08]) and APACHE II (OR=1.03[1.01-1.05]) were variables associated with ICU mortality (additional file e-Figure 11) . The performance of the model is show in Table 3 and additional file (e-Table 18 and e-Figure 12).

Our main finding was that the application of an unsupervised clustering model developed during the first pandemic wave of COVID-19 on a set of patients belonging to the second and third waves has not performed an adequate classification of patients within 3 original phenotypes. These findings are since the “Si” Silhouette coefficients were close to or below zero and that the characteristics of the new patients within each phenotype were significantly different from the original phenotypes. In addition, the inclusion of phenotype classification in a regression model did not improve model performance over a standard model or even worsened when the model was applied to patients within each phenotype. This strongly suggests the rejection of our study hypothesis and alerts to the need to update models when new populations need to be considered.

Several authors have published the development of different immune and clinical phenotypes in patients with COVID-19 with a methodology similar to our study (4–12). Many of these studies have not validated the results (5–9,11) and some of them (4,10) have performed an external validation but with several limitations. Gutiérrez-Gutiérrez B et al (4) in an elegant study, which included 4035 consecutive adult patients admitted to 127 hospitals in Spain with COVID-19 (February 2 - March 17, 2020) deploying 3 clinical phenotypes and performed external validation in a cohort of 2226 consecutive adult patients admitted from February 25 and April 19, 2020. Although the performance of the classification of validation patients within each phenotype was adequate (AUC ROC > 0.80), validation is performed in the same period corresponding to the first pandemic wave. The clinical characteristics of patients with COVID-19 and their evolution may vary over time, as observed in our study, either by the consequences of the change in management of the intensive care unit or by the variants of SARS-CoV-2 could affect the patient's presentation, clinical course, and phenotypes of the patient. Therefore, these phenotypes should be validated again with populations of other pandemic waves.

To the best of our knowledge, only one study has validated phenotypes in 2 different waves. Wang X et al. (10) deploying 6 phenotypes in more than 20,000 patients with COVID-19 (not all hospitalised) during a period prior to August 28, 2020 (training sample - first wave) while validation was performed on patients collected between August 29 and October 31, 2020 (test sample - second wave). Although the authors conclude that the phenotypic latent class analysis (LCA) model was a better predictor of hospitalisation and progression to ICU admission or mortality than a risk assessment based solely on age and comorbidities, the concordance index in the test sample was as low as 0.56 and did not exceed 0.77 in any comparison. Suggesting, as observed in our study, that the non-inclusion of second wave patients in the original model results in inadequate model performance.

Patients affected by COVID-19, have great heterogeneity in their clinical presentation and high pathophysiological complexity. A thorough understanding of the complex biological systems present in these patients, requires conceptual and analytical strategies that respect this complexity. The identification of different clinical phenotypes emerges as a valid strategy that could provide guidance to personalise treatment and achieve the best outcomes (18). Unfortunately, machine learning methodologies for producing personalised diagnostics and therapeutics are still largely fragile, unexplainable, and often insufficiently validates.

As we observed in our study, although the application of an unsupervised clustering model allows us to classify patients into different clinical phenotypes with different characteristics and evolution among them. Patient characteristics differ significantly from the phenotypes corresponding to the first pandemic wave and their classification does not improve the performance of linear mortality models. The time frame of the study and restricted access to confounding factors involved in disease severity and clinical presentation need to be considered. In addition, morbidity and mortality could vary over time, either because of different treatments administered in the ICU, the strain circulating in a specific geographical area or the change in the recommended drug treatment for the disease. Finally, other recently described variants of SARS-CoV-2 (6) could affect the presentation, clinical course, and phenotypes of patients. All these confounding factors, or as many as possible, should be included in the model if errors in risk classification of patients are to be avoided.

Even if artificial intelligence and machine learning are claimed to achieve indisputable results, they depend on the algorithms being defined accordingly and considering all confounding factors. It must be recognised that a small change in the characteristics of the test cases can have a considerable impact on the results of the model. As proposed by Feng J et al. (19), machine learning (ML) and artificial intelligence (AI) algorithms have the potential to gain insights from clinical data and improve patient outcomes. However, these highly complex systems are sensitive to changes in the environment and can lose performance. Even after successful integration into clinical practice, ML/AI algorithms need to be continuously monitored and updated to ensure their long-term safety and effectiveness. Our results warn of the need for continuous validation of linear and non-linear models, as well as of the expiration of these models in the face of significant changes in the characteristics of the population.

In this context, many of the risk factors (20–22) and treatment measures (23,24) that clinicians are currently considering and administering to patients with COVID-19 were determined using patients from the first pandemic wave, without further validation. The importance of these variables may not be the same for patients from other pandemic waves, but, strikingly, the need to validate these important findings in new COVID-19 patient populations has not been raised.

Our study has several limitations that must be recognized. First, in the current study, we focus only on the data available on ICU admission, so we cannot exclude that a patient's clinical profile may change over time. However, our objective was to define the phenotypes at admission to the ICU because it is an opportune window to decide on eligibility for treatment. Second, we have only considered ICU patients and predominantly within a single country. Therefore, considering all the confounding variables that may vary between centres or regions, our results cannot be transferred without prior validation. Third, the observed clinical phenotypes could reflect different profiles of virus-host interactions, because of a different infectious viral load, a variable humoral and cellular immune response against SARS-CoV-2, or cellular receptor characteristics and expression, together with the genetic background of the host that have not been considered in our database. Although information on underlying immunological or virological mechanisms may be of great interest for phenotype development (5), at the time of patient admission to the ICU, these data are not available, therefore, rapid determination of clinical phenotypes retains its value and can help decide on treatment eligibility in clinical practice.

Our results suggest that classification or risk models developed using machine learning techniques during the first pandemic wave cannot be applied with adequate performance to patients admitted in subsequent waves. Our study alerts researchers to the need to validate the results and continuously update the models to ensure adequate performance.

ADRS: Acute distress respiratory syndrome

AKI: Acute kidney injury

APACHE II: Acute Physiology And Chronic Health Evaluation II

AUC: Area Under the ROC Curve

CKD: Chronic kidney disease

CI: Confidence interval

COPD: Chronic obstructive pulmonary disease

CRP: C- Reactive Protein

GLM: Generalized linear regression model

HFNC: High flow nasal cannula oxygen therapy

ICU: Intensive Care Unit

LDH: D-lactate dehydrogenase

LOS: Length of stay

MV: Mechanical ventilation

PAM: partition around medoids

PCA: Principal component analysis

PCT: Procalcitonin

ROC: Receiver Operator Characteristic curve

RT-PCR: real time polymerase chain reaction

SEMICYUC: Sociedad Española de Medicina Intensiva, Critica y Unidades Coronarias

SOFA: Sequential Organ Failure Assessment

WHO: World Health Organization

- Ethical Approval and Consent to participate

The study was approved by the reference institutional review board at Joan XXIII University Hospital (IRB# CEIM/066/2020) and each participating site with a waiver of informed consent.

- Consent for publication

Not applicable

- Availability of supporting data

The anonymized database collected for the study by the SEMICYUC, and the data dictionary that defines each field in the set, will be made available to reviewers if they consider it necessary prior confidentiality agreement.

- Competing interests

All authors declare that they have no conflicts of interest.

- Funding

This study was supported by the Spanish Intensive Care Society (SEMICYUC) and Ricardo Barri Casanovas Foundation (AR). The study sponsors have no role in the study design, data collection, data analysis, data interpretation, or writing of the report.

- Authors' contributions

RA, JG, FA, DE, SVJ, ZR, PE, SB, VM, JHM, AMA, CBA, dVOM, BJC, MIL, SCS, PLJC, EA, MLI and MB had substantial contributions to conception and design of the work.

RA, TS, GJ, DE and FA had substantial contribution for data acquisition.

RA, GJ, MLI, DE and MB had substantial contribution for data analysis and interpretation of data for the work.

RA, MLI, SVJ, DE and GJ drafting of the manuscript.

ZR, SB, AMA, CBA, SCS, AE and PLJC critically reviewed the draft manuscript.

The corresponding author (RA) had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. All authors approved the final version of the manuscript. The views expressed in this article are those of the authors and not necessarily those of the SEMICYUC.

- Acknowledgements

To Alexis Garduno for English manuscript edition and to the COVID-19 SEMICYUC Investigators listed in Appendix 2.

Christopher W. Seymour, Jason N. Kennedy, Shu Wang, Chung-Chou H. Chang, Corrine F. Elliott, Zhongying Xu, et al. Derivation, Validation, and Potential Treatment Implications of Novel Clinical Phenotypes for Sepsis. JAMA. 2019;321(20):2003-2017. https://doi.org/10.1001/jama.2019.5791
Anousone Daulasim, Antoine Vieillard-Baron, Guillaume Geri. Hemodynamic clinical phenotyping in septic shock. Curr Opin Crit Care. 2021 1;27(3):290-297. https://doi.org/10.1097/MCC.0000000000000834.
Cereuil, A., Ronflé, R., Culver, A. et al. Septic Shock: Phenotypes and Outcomes. Adv Ther 39, 5058–5071 (2022). https://doi.org/10.1007/s12325-022-02280-5
Belén Gutiérrez-Gutiérrez, María Dolores del Toro, Alberto M Borobia, Antonio Carcas, Inmaculada Jarrín, María Yllescas et al.Identification and validation of clinical phenotypes with prognostic implications in patients admitted to hospital with COVID-19: a multicentre cohort study. Lancet Infect Dis 2021;21: 783–92 https://doi.org/10.1016/S1473-3099(21)00019-0
Thibault Dupont, Sophie Caillat-Zucman, Véronique Fremeaux-Bacchi, Florence Morin, Etienne Lengliné , Michael Darmon et al. Identification of Distinct Immunophenotypes in Critically Ill Coronavirus Disease 2019 Patients. CHEST 2021; 159(5):1884-1893 https://doi.org/10.1016/j.chest.2020.11.049
Nicolas de Prost, Etienne Audureau, Nicholas Heming, Elyanne Gault, Tài Pham, Amal Chaghouri et al. Clinical phenotypes and outcomes associated with SARS-CoV-2 variant Omicron in critically ill French patients with COVID-19. Nature Communications 2022; 13:6025 https://doi.org/10.1038/s41467-022-33801-z
Lovemore N. Sigwadhi, Jacques L. Tamuzi, Annalise E. Zemlinb, Zivanai C. Chapandukac, Brian W. Allwoodd, Coenraad F. Koegelenberg et al. Latent class analysis: an innovative approach for identification of clinical and laboratory markers of disease severity among COVID-19 patients admitted to the intensive care unit. JID Regions 2022;5:154–162. https://doi.org/10.1016/j.ijregi.2022.10.004
Juliana F. da Silva, Alfonso C. Hernandez-Romieu, Sean D. Browning, Beau B. Bruce, Pavithra Natarajan, Sapna B. Morris et al. COVID-19 Clinical Phenotypes: Presentation and Temporal Progression of Disease in a Cohort of Hospitalized Adults in Georgia, United States. Open Forum Infect Dis 2020 ;8(1):ofaa596. https://doi/10.1093/ofid/ofaa596.
Elizabeth R. Lusczek, Nicholas E. Ingraham, Basil S. Karam, Jennifer Proper, Lianne Siegel,Erika S. Helgeson et al. Characterizing COVID-19 clinical phenotypes and associated comorbidities and complication profiles. PLoS ONE 2021; 16(3): e0248956. https://doi.org/10.1371/journal.pone.0248956
Xiaofeng Wang, Lara Jehi, Xinge Ji, Peter J. Mazzone. Phenotypes and Subphenotypes of Patients With COVID-19. A Latent Class Modeling Analysis. CHEST 2021; 159(6):2191-2204. https://doi.org/10.1016/j.chest.2021.01.057
Charles R. Vasquez, Shruti Gupta, Todd A. Miano, Meaghan Roche, Jesse Hsu, Wei Yang et al. Identification of Distinct Clinical Subphenotypes in Critically Ill Patients With COVID-19. CHEST 2021; 160(3):929-943. https://doi.org/10.1016/j.chest.2021.04.062
Wenjing Ye, Weiwei Lu, Yanping Tang, Guoxi Chen, Xiaopan Li, Chen Ji et al. Identification of COVID-19 Clinical Phenotypes by Principal Component Analysis-Based Cluster Analysis. Front. Med. 2020; 7:570614. https://doi.org/10.3389/fmed.2020.570614
Alejandro Rodríguez, Manuel Ruiz‑Botella, Ignacio Martín‑Loeches, María Jiménez Herrera, Jordi Solé‑Violan, Josep Gómez et al. Deploying unsupervised clustering analysis to derive clinical phenotypes and risk factors associated with mortality risk in 2022 critically ill patients with COVID‑19 in Spain. Crit Care 2021; 25:63 https://doi.org/10.1186/s13054-021-03487-8
Rosenman M, He J, Martin J, Nutakki K, Eckert G, Lane K et al. Database queries for hospitalizations for acute congestive heart failure: flexible methods and validation based on set theory. J Am Med Inform Assoc 2014; 21:345–352. https://doi:10.1136/amiajnl-2013-001942
Madigan, D., P. B. Ryan, M. Schuemie, P. E. Stang, J. M. Overhage, A. G. Hartzema, et al. “Evaluating the impact of database heterogeneity on observational study results.” Am J Epidemiol. 2013 ;178(4):645-51. https://doi.org/10.1093/aje/kwt010
Vandenbroucke JP, von Elm E, Altman DG, Gøtzsche PC, Mulrow CD, Pocock SJ, Poole C, Schlesselman JJ, Egger M; STROBE Initiative. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): explanation and elaboration. PLoS Med. 2007 Oct 16;4(10):e297. https://doi.org/10.1371/journal.pmed.0040297
World Health Organization (WHO). Laboratory testing for 2019 novel coronavirus (2019-nCoV) in suspected human cases [accessed Jan 22, 2023]. Available in https://www.who.int/publications/i/item/10665-331501
Moreno G, Ruiz-Botella M, Martín-Loeches I, Gómez Álvarez J, Jiménez Herrera M, Bodí M et al. A differential therapeutic consideration for use of corticosteroids according to established COVID-19 clinical phenotypes in critically ill patients. Med Intensiva (Engl Ed). 2023 Jan;47(1):23-33. https://doi.org/10.1016/j.medine.2021.10.016
Jean Feng, Rachael V. Phillips, Ivana Malenica, Andrew Bishara , Alan E. Hubbard, Leo A. Celi, Romain Pirracchio. Clinical artificial intelligence quality improvement: towards continual monitoring and updating of AI algorithms in healthcare. npj Digital Medicine 2022; 5:66 ; https://doi.org/10.1038/s41746-022-00611-y
Fei Zhou, Ting Yu, Ronghui Du, Guohui Fan, Ying Liu, Zhibo Liu, Jie Xiang et al.Clinical course and risk factors for mortality of adult inpatients with COVID-19 in Wuhan, China: a retrospective cohort study. Lancet 2020; 395: 1054–62 https://doi.org/10.1016/S0140-6736(20)30566-3
Hai-Han Yu , Chuan Qin , Man Chen , Wei Wang , Dai-Shi Tian. D-dimer level is associated with the severity of COVID-19. Thromb Res 2020 ;195:219-225. https://doi:10.1016/j.thromres.2020.07.047
Huan Han, Qingfeng Ma , Cong Li , Rui Liu , Li Zhao , Wei Wang et al. Profiling serum cytokines in COVID-19 patients reveals IL-6 and IL-10 are disease severity predictors. Emerg Microbes Infect. 2020 Dec;9(1):1123-1130. https://doi.org/10.1080/22221751.2020.1770129
Chi Zhang, Zhao Wu, Jia-Wen Li, Hong Zhao, Gui-Qiang Wanga. Cytokine release syndrome in severe COVID-19: interleukin-6 receptor antagonist tocilizumab may be the key to reduce mortality. Int J Antimicrob Agents. 2020; 55(5): 105954. https://doi.org/10.1016/j.ijantimicag.2020.105954
The RECOVERY Collaborative Group. Dexamethasone in Hospitalized Patients with Covid-19. N Engl J Med 2021; 384(8): 693-704. https://doi.org/10.1056/NEJMoa2021436

Table 1: Patient characteristics of the original and validation cohort and within each phenotype obtained by applying the original model in the validation population.

Variable ^a	Overall		Phenotype C (life-threatening disease)		Phenotype B (critical disease)		Phenotype A (severe disease)
	ORIGINAL cohort (n=2017)	VALIDATION cohort (n=2330)	ORIGINAL cohort (n=857)	VALIDATION cohort (n=506)	ORIGINAL cohort (n=623)	VALIDATION cohort (n=618)	ORIGINAL cohort (n=537)	VALIDATION cohort (n=1206)
General characteristics and severity of illness
Age , mean (IQR),years	64(55-71)	63 (53-92)**	66(58-72)	66 (59-73)	63(53-71)	62(53-70)	63 (53-70)	62(52-70)
Male, n(%)	1419(70.3)	1643(70.5)	626(73.0)	377 (74.5)	416(66.8)	428 (69.3)	377(70.2)	838(69.5)
APACHE II, mean (IQR) ^b	13(10-17)	12(9-16)***	17(14-22)	15 (11-19)***	13(10-16)	12(9-16)*	12(9-16)	11(8-15)*
SOFA, mean (IQR) ^c	5(3.7)	4.0(3-6)***	7(6-8)	6(4 – 8)**	5(3-7)	4(3-6)*	4(3-5)	3(2-4)*
Laboratory findings
D-Lactate dehydrogenase, mean (IQR), U/L	537(417-707)	463.0 (358-584)**	670(554-929)	525 (401-632)*	477(378-570)	466(349-595)	474(372-564)	432(347-548)
White blood cell, mean (IQR), x10⁹	8.8(6.2-12.2)	9.3(6.7-12.9)*	10(6.9-13.6)	10(7.1-14.4)	8.5(6-11.7)	9.8(7-13.3)*	7.7(5.8-10.2)	8.9(6.4-12)*
Serum Creatinine, mean (IQR), mg/dL	0.88(0.7-1.1)	0.80 (0.6-1.0)	0.99(0.76-1.36)	0.93 (0.71-1.27)	0.80(0.66-1.00)	0.77(0.64-0.95)	0.80(0.66-1.01)	0.79(0.65-0.97)
C-Reactive Protein, mean (IQR), mg/mL	15.5(9.1-24.3)	11.3(6.3-17.9)**	18(10-26)	13(7-20)***	14(9-22)	11 (6.4-17)**	14(8-20)	11(6-17)**
Procalcitonin , mean (IQR), ng/mL	0.3(0.1-2.0)	0.19(0.09-1.17)**	0.5(0.2-1.3)	0.25 (0.11, 0.80)*	0.2(0.1-0.5)	0.17(0.08-0.48)	0.2(0.1-0.6)	0.18(0.08-0.4)
Serum lactate, mean (IQR), mmol/L	1.5(1.1-2.0)	1.4 (1.1-1.7)*	1.6(1.2-2.2)	1.5(1.1-1.9)*	1.4(1.0-1.9)	1.4(1.1-1.9)	1.5(1.1-1.9)	1.4(1.0-1.8)
D dimer, mean (IQR), ng/mL	1593(720-3790)	995 (589-2115) ***	2260(1009-4894)	1300[685-3932)***	1319(634-3548)	1059(629-2648)*	1090(580-2100)	861(522-1641)**
Ferritin, mean (IQR), ng/mL	1600(1290-2240)	1381(947-1807)***	1800(1416-2377)	1487(949-1962)**	1554(1271-1936)	1370[963-1839)*	1538(1280-1899)	1345(937-1709)*
Coexisting condition and Comorbidities
Arterial hypertension, n(%)	932(46.2)	1123 (48.2)***	548(63.9)	421 (83.2)***	173(27.8)	169(27.3)	211(39.3)	533(44.2)
Obesity (BMI>30), n(%) ^d	653(32.3)	942 (40.4)***	294(34.3)	214 (42.3)***	200(32.1)	224(36.2)	159(29.6)	504(41.8)***
Diabetes , n(%)	418(20.7)	599 (25.7)***	198(23.1)	188 (37.2)***	108(17.3)	122(19.7)	112(20.9)	289(24.0)
Coronary arterial disease, n(%)	124(6.1)	152 (6.5)	48(5.6)	54 (10.7)**	41(6.6)	29(4.7)	35(6.5)	69(5.7)
COPD, n(%)	148(7.3)	180 (7.7)	73(8.5)	65 (12.8)*	38(6.1)	39(6.3)	37(6.9)	76(6.3)
Chronic renal disease , n(%) ^e	85(4.2)	153 (6.6)***	44(5.1)	60 (11.9)***	10(1.6)	27(4.4)	31(5.8)	66(5.5)
Hematologic disease, n(%)	72(3.5)	66 (2.8)	30(3.5)	21 (4.2)	22(3.5)	16(2.6)	20(3.7)	29(2.4)
Asthma, n(%)	120(5.9)	156 (6.7)*	34(4.0)	32 (6.3)	45(7.2)	45(7.3)	41(7.6)	79(6.6)
HIV, n(%)	5(0.2)	9 (0.4)	2(0.2)	3 (0.6)	1(0.2)	2(0.3)	2(0.4)	4(0.3)
Pregnancy , n(%)	4(0.2)	12 (0.5)	0(0.0)	1 (0.2)	3(0.5)	2(0.3)	1(0.2)	9(0.7)
Autoimmune disease, n(%) ^f	74(3.6)	53 (2.3)	36(4.2)	13 (2.6)	18(2.9)	19(3.1)	20(3.7)	21(1.7)
Chronic heart disease , n(%) ^g	57(2.8)	82 (3.5)	26(3.0)	32 (6.3)***	10(1.6)	14(2.3)	21(3.9)	36(3.0)
Neuromuscular disease , n(%)	16(0.8)	12 (0.5)	8(0.9)	4 (0.8)	5(0.8)	2(0.3)	3(0.6)	6(0.5)
Oxygenation and ventilator support
High Flow nasal cannula, n(%)	376(18.6)	915 (39.3)***	27(3.2)	26 (4.9)	3(0.5)	2(0.3)	345(64.2)	887 (73.6)***
Non-invasive ventilation , n(%)	141(6.9)	187 (8.0)**	50(5.8)	32 (6.3)	26(4.2)	66 (10.7)***	65(11.9)	89 (7.4)**
Invasive mechanical ventilation , n(%)	1173(58.1)	679 (33.4)***	694(81.0)	364 (71.9)***	475(76.2)	312 (50.5)***	3(0.6)	3 (0.2)
PaO2/FiO2, mean (IQR)	132(96-163)	124(99-145)**	126(88-155)	125(97-151)	165(144-212)	127(106-150)**	111(82-133)	122(97-141)*
Complications
Shock , n(%) ^h	904(44.8)	474 (20.3)***	652(76.1)	309 (61.1)***	196(31.5)	84(13.6)***	56(10.4)	81 (6.7)***
Acute kidney dysfunction , n(%) ⁱ	579(28.7)	512 (22.0)**	350(40.8)	181 (35.8)	118(18.9)	113(18.3)	111(20.7)	218 (18.1)
Myocardial dysfunction, n(%) ^j	169(8.3)	177 (7.6)	96(11.2)	56 (11.1)	43(6.9)	41(6.6)	30(5.6)	80(6.6)
ICU crude mortality , n(%)	657(32.6)	634 (27.2)	389(45.4)	200 (39.5)*	159(25.5)	182 (29.4)	109(20.3)	252 (20.9)

All comparison between overall and clusters. *p<.05; ** p<.01 ; ***p<.001 , others comparison p>.01

Abbreviations: IQR, interquartile range; APACHE II, Acute Physiology and Chronic Health Evaluation II; SOFA, Sequential Organ Failure Assessment; BMI, body mass index; COPD, Chronic obstructive pulmonary disease; HIV, human immunodeficiency viruses; PaO2/FiO2, Partial pressure arterial oxygen/ fraction of inspired oxygen

^aCorresponds to minimum or maximum value, as appropriate, within 12 hours of ICU admission. The variables in this Table were no transformed for your comparison.

^bAPACHE II score to the severity of illness, the score is obtained by adding the following components 1) 12 clinical and laboratory variables each with a score range of 0 to 4 points (APS). The APS is determined from the worst physiologic values during the initial 24 h after ICU admission, 2) age with a range 0 to 6 points and 3) Chronic health points if the patients has history of severe organ system insufficiency or is immunocompromised assign 5 points if the patients is no operative or emergency postoperative and 2 points for elective postoperative patients with a total score range of 0 to 71.

^c SOFA score corresponds to the severity of organ dysfunction, reflecting six organ systems each with a score range of 0 to 4 points (cardiovascular, hepatic, hematologic, respiratory, neurological, renal), with a total score range of 0 to 24,

^d Defined as a body mass index (calculated as weight in kilograms divided by height in meters squared) of 30 or greater.

^e Baseline eGFR< 60 on at least two consecutive values at least 12 weeks apart prior or hemodialysis

^fincluded acute leukemia, myelodysplastic syndrome and Lymphomas.

^g According to the New York Heart Association (NYHA) Functional Classification III and IV

^hDefined as patients in whom adequate fluid resuscitation therapy are unable to restore hemodynamic stability and need any dose of vasopressor drugs.

ⁱDefine as an abrupt and sustained (more than 24 hours) decrease in kidney function and categorized according to RIFLE criteria

^j Define as an acute decrease in ejection fraction (EF) with dilatation of ventricles observed in echocardiography upon ICU admission

Table 2: Characteristics of 2330 COVID-19 critically ill patients included in the study. The data are shown with the discretisation of the variables used to carry out the study

Variable^a,l	Overall	Survival	Non survival	p-value
General characteristics and severity of illness
No. of patients (%)	2330 (100.0)	1696 (72,8)	634 (27.2)	------
Age, median (Q1-Q3), years Discretization, No. (%) 16 – 56 57 – 66 67 – 71 >71	63 (63-92) 681 (29.2) 761 (32.7) 269 (11.5) 619 (26.6)	61 (51-69) 587(34.6) 583(34.4) 175(10.3) 351(20.7)	69 (61-75) 94(14.8) 178(28.1) 94(14.8) 268(42.3)	<0.001 <0.001
Male, No. (%)	1643(70.5)	1179 (69.5)	464 (73.2)	0.093
APACHE II, median (Q1-Q3)^b Discretization, No. (%) 1 – 12 13 – 14.6 14.7 – 16 > 16	12(9-16) 1039 (44.6) 525 (22.5) 140 (6.0) 626 (26.9)	11 (8-15) 869 (51.2) 370 (21.8) 94 (5.5) 363 (21.4)	14 (11-18) 170 (26.8) 155 (24.4) 46 (7.3) 263 (41.5)	<0.001 <0.001
SOFA, median (Q1-Q3)^c Discretization, No. (%) 0 – 3.9 4 – 7 7.1 – 8 > 8	4.0(3-6) 975 (41.8) 910 (39.1) 311 (13.3) 134 (5.8)	4 (3-5) 822 (48.5) 625 (36.9) 186 (11.0) 63 (3.7)	5 (4-7) 153 (24.1) 285 (45.0) 125 (19.7) 71 (11.2)	<0.001 <0.001
Shock at ICU admission, n (%)	474 (20.3)	293 (17.3)	181 (28.5)	<0.001
Quadrant infiltrates in chest x-ray, median (Q1-Q3) Discretization, No. (%) 1 quadrant 2 quadrants 3 quadrants 4 quadrants	3 (2-4) 119 (5.1) 573 (24.6) 541 (23.2) 1089 (46.7)	3 (2-4) 95 (5.6) 433 (25.5) 404 (23.8) 758 (44.7)	4 (2-4) 24 (3.8) 140 (22.1) 137 (21.6) 331 (52.2)	<0.001 0.019
PaO2/FiO2 at ICU admission, median (Q1-Q3) Discretization, No. (%) 0 – 158 > 158	124.8(99-145) 1913 (82.1) 417 (17.9)	128 (106 -141) 1369 (80.7) 327 (19.3)	110 (83- 136) 544 (85.8) 90 (14.2)	<0.001 0.005
Laboratory findings
D-Lactate dehydrogenase, median (Q1-Q3), U/L Discretization, No. (%) 0 – 467 > 467	463.0 (358-584) 1188 (51.0) 1142 (49.0)	437 (343-555) 953 (56.2) 743 (43.8)	523 (402-660) 235 (37.1) 399 (62.9)	<0.001 <0.001
White blood cell, median (Q1-Q3), x10⁹ Discretization, No. (%) 0 - 13.6 > 13.6	9.3(6.7-12.9) 1815 (77.9) 515 (22.1)	9.2 (6.6-12.5) 1371 (80.8) 325 (19.2)	10.0 (7.0-14.6) 444 (70.0) 190 (30.0)	<0.001 <0.001
Serum Creatinine, median (Q1-Q3), mg/dL Discretization, No. (%) 0 – 0.9 0.91 – 1.13 >1.13	0.80 (0.6-1.0) 1493 (64.1) 405 (17.4) 432 (18.5)	0.78 (0.6-0.9) 1172 (69.1) 286 (16.9) 238 (14.0)	0.90(0.7-1.2) 321 (50.6) 119 (18.8) 470 (74.1)	<0.001 <0.001
C-Reactive Protein, median (Q1-Q3), mg/mL Discretization, No. (%) 0 – 20 > 20	11.3(6.3-17.9) 1853 (79.5) 477 (20.5)	10.9 (5.9-17.3) 1383 (81.5) 313 (18.5)	12.9(6.9-20.2) 470 (74.1) 164 (25.9)	<0.001 <0.001
Procalcitonin, median (Q1-Q3), ng/mL Discretization, No. (%) 0 – 0.17 0.18 – 0.74 > 0.74	0.19(0.09-1.17) 1112 (47.7) 790 (33.9) 428 (18.4)	0.17 (0.08-0.43) 874 (51.5) 570 (33.6) 252 (14.9)	0.25 (0.11-0.82) 238 (37.5) 220 (34.7) 176 (27.8)	<0.001 <0.001
Serum lactate, median (Q1-Q3), mmol/L Discretization, No. (%) 0 – 2.13 >2.13	1.4 (1.1-1.7) 1981 (85.0) 349 (15.0)	1.3 (1.0-1.8) 1469 (86.6) 227 (13.4)	1.6 (1.2-2.0) 512 (80.8) 122 (19.2)	<0.001 0.001
D dimer, median (Q1-Q3), ng/mL Discretization, No. (%) 0 – 3000 > 3000	995 (589-2115) 1875 (80.5) 455 (19.5)	900 (540-1750) 1432 (84.4) 264 (70.0)	1340 (730-4167) 443 (69.9) 191 (30.1)	<0.001 <0.001
Ferritin ng/mL, median (Q1-Q3), Discretization, No. (%) 0 – 1610 1611 – 1740 1741 – 1810 >1810	1381(947-1807) 1523 (65.4) 162 (7.0) 72 (3.1) 573 (24.6)	1300 (905-1700) 1187 (70.0) 104 (6.1) 40 (2.4) 365 (21.5)	1590 (1088-1940) 336 (53.0) 58 (9.1) 32 (5.0) 208 (32.8)	<0.001 <0.001
Coexisting condition and Comorbidities
Arterial Hypotension, No.(%)	1123 (48.2)	741 (43.7)	382 (60.3)	<0.001
Obesity, No. (%)^d	942 (40.4)	683 (40.3)	259 (40.9)	0.83
Diabetes, No.(%)	599 (25.7)	391 (23.1)	208 (32.8)	<0.001
Coronary arterial disease, No.(%)	152 (6.5)	83 (4.9)	69 (10.9)	<0.001
COPD , No.(%)	180 (7.7)	110 (6.5)	70 (11.0)	<0.001
Chronic renal disease , No.(%)^e	153 (6.6)	85 (5.0)	68 (10.7)	<0.001
Hematologic disease , No.(%)^f	66 (2.8)	32 (1.9)	34 (5.2)	<0.001
Asthma, No.(%)	156 (6.7)	114 (6.7)	42 (6.6)	1.00
HIV, No.(%)	9 (0.4)	6 (0.4)	3 (0.5)	0.96
Pregnancy , No.(%)	12 (0.5)	12 (0.7)	0	0.07
Autoimmune disease, No.(%)	53 (2.3)	35 (2.1)	18 (2.8)	0.33
Chronic heart disease , No.(%)^g	82 (3.5)	38 (2.2)	44 (6.9)	<0.001
Neuromuscular disease , No.(%)	12 (0.5)	11 (0.6)	1 (0.2)	0.25
Ventilator support
Oxygen mask , No.(%)	610 (26.2)	450(26.5)	160 (25.2)	0.52
High Flow nasal cannula, No.(%)	915 (39.3)	720 (42.5)	195 (30.8)	<0.001
Non-invasive ventilation , No.(%)	187 (8.0)	129 (7.6)	58 (9.1)	0.25
Invasive mechanical ventilation, No.(%)	618 (26.5)	353 (20.8)	265 (41.8)	<0.001
Complications and outcome
Acute kidney dysfunction, No. (%) ^h	512 (22.0)	236 (13.9)	276 (43.5)	<0.001
Bacterial Community-acquired co-infection, No. (%)	207 (8.9)	121 (7.1)	86 (13.6)	<0.001
Ventilator associated pneumonia, No. (%)	666 (28.6)	409 (24.1)	257 (40.5)	<0.001
Myocardial dysfunction, No. (%)	177 (7.6)	67 (4.0)	110 (17.4)	<0.001
ICU crude mortality, No. (%)	634 (27.2)	NA	NA	------

Abbreviations: Q1-Q3: 1^st quartile – 3^rd quartile; APACHE II, Acute Physiology and Chronic Health Evaluation II; SOFA, Sequential Organ Failure Assessment; BMI, body mass index; COPD, Chronic obstructive pulmonary disease; HIV, human immunodeficiency viruses; PaO2/FiO2, Partial pressure arterial oxygen/ fraction of inspired oxygen

^a Corresponds to minimum or maximum value, as appropriate, within 12 hours of ICU admission. The variables in this Table were no transformed for your comparison.

^b APACHE II score to the severity of illness, the score is obtained by adding the following components 1) 12 clinical and laboratory variables each with a score range of 0 to 4 points (APS). The APS is determined from the worst physiologic values during the initial 24 h after ICU admission, 2) age with a range 0 to 6 points and 3) Chronic health points if the patients have history of severe organ system insufficiency or is immunocompromised assign 5 points if the patients is no operative or emergency postoperative and 2 points for elective postoperative patients with a total score range of 0 to 71.

^c SOFA score corresponds to the severity of organ dysfunction, reflecting six organ systems each with a score range of 0 to 4 points (cardiovascular, hepatic, hematologic, respiratory, neurological, renal), with a total score range of 0 to 24,

^d Defined as a body mass index (calculated as weight in kilograms divided by height in meters squared) of 30 or greater.

^e Baseline eGFR< 60 on at least two consecutive values at least 12 weeks apart prior or haemodialysis

^f included acute leukaemia, myelodysplastic syndrome and Lymphomas.

^gAccording to the New York Heart Association (NYHA) Functional Classification III and IV

^h Define as an abrupt and sustained (more than 24 hours) decrease in kidney function and categorized according to RIFLE criteria

Table 3: Models (GLM) performance comparison

Point estimates	Original GLM	Modified GLM	Phenotype A GLM	Phenotype B GLM	Phenotype C GLM
Accuracy	0.79	0.78	0.80	0.73	0.66
AUC ROC	0.79	0.78	0.77	0.67	0.76
S	0.81	0.81	0.84	0.77	0.72
Sp	0.66	0.65	0.55	0.57	0.60
PPV	0.92	0.92	0.93	0.89	0.73
NPV	0.43	0.42	0.29	0.36	0.55

AUC: area under curve, S: sensibility, Sp: Specificity, PPV: positive predictive value, NPV: negative predictive value

No competing interests reported.

Reliability and reproducibility of clinical phenotypes developed during the first wave of COVID-19: A validation study in critically ill patients from the second and third wave

Status:

Version 1

Abstract

Figures

Introduction

Material And Methods

Results

Discussion

Conclusion

Abbreviations

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1