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

Since the outbreak of COVID-19 disease began in December 2019 in China, soaring cases of confirmed SARS-CoV-2 are pummeling the global health system. More than 91 million people have developed SARS-CoV-2 infection, and more than 2 million have died [1]. Critical illness from COVID-19 has constrained intensive care unit (ICU) material and human resources [2]. As of January 18, 2021 more than 2.5 million people in Spain have been infected with SARS-CoV-2 and more than 53,000 have died [3]. Short-term mortality reported rate ranges from 16 to 62% of patients admitted to ICU [4‐8]. The heterogeneity of patients that have been treated in China [4], Italy [8], USA [5‐7] or Spain [9, 10] may explain the wide variation of mortality rate due to their population characteristics, presence of comorbidities and healthcare systems. A recent international survey [11] reported significant practice variations in the management of severe COVID-19 patients, including differences at the regional, hospital, and patient level. Therefore, it is necessary to characterize phenotypes, by extending the enrolment of patients outside of one ICU site to multiple patients being treated in different hospitals. Allowing to adequately measure mortality-related factors adjusted by the inter-hospital variation to determine clinical outcomes.

Risk factors represent the most important approach when defining treatment of hospitalized patients as these measures can inform clinical courses most likely for a patient given their a priori risk. However, risk factors can also interplay differently when they are included in different patient clusters. A single model based on general risk factors (one-size-fits-all) might be limited for clinical data interpretation and application across sites. Different combinations of risk factors may naturally cluster into previously undescribed subsets or phenotypes that may have different risks for a high mortality rate and that may therefore help to determine the response to treatments in COVID-19. We hypothesize that the presence of well-defined phenotypes in COVID-19 could help to more appropriately identify patients at risk of ICU mortality than general models for the entire population considering that this disease results in a constellation of symptoms, laboratory derangement, immune dysregulation, and clinical complications.

The primary objective was to determine the presence of distinct clinical phenotypes using unsupervised clustering methods that were applied to the datasets available on ICU admission. The second objective was to assess which factors are independently associated with ICU mortality. The added value of this large-scale multicenter prospective study lies to discover phenotypes based on clinical data available at ICU admission that can help explain the variation in clinical results of COVID-19 disease in the ICU.

The main finding of our study is that among patients with COVID-19, 3 clinical phenotypes were derived using habitual clinical and laboratory variables at ICU admission. The ability of identifying phenotypes using a small set of variables is a crucial step towards clinical application and has important implications for possible differential treatment guided by phenotypes and validated prognostic scoring systems [18, 19].

Our C phenotype was associated with more than double the ICU mortality than each of the remaining two phenotypes. This C phenotype was characterized by the interplay of older age (> 65 year), a high severity (APACHE II > 15 and SOFA > 5), greater burden of risk factors (hematologic disease and coronary disease) and a higher likelihood of developing further complications (shock and AKI).

Previous studies have implemented clustering techniques to analyze various data sources relating to demographic, geographic, environment, and socioeconomic determinants of health and disease. There are studies that have evaluated treatment decisions and characterized clinical phenotypes associated with complications, ICU admission and mortality risk in critically ill COVID-19 patients. According to the Situation Report & Public Health Guidance published by Johns Hopkins University on March 19th, 2020, people over 60 and those with chronic health conditions are at the highest risk for COVID-19 complications [20]. To our knowledge, this is the first study with a high number of critically ill patients to analyze the presence of phenotypes in patients with SARS-CoV-2 infection. This multicenter cohort study of 2,022 critically ill patients found that 660 patients (32.6%) died at ICU discharge. Our ICU mortality rates was significantly lower as reported in Yang et al. [4] in Wuhan, China (61.5%), by Myers et al. [7] in California, USA (50.0%), by Arentz et al. [6] in Washington, USA (67%) and by Richardson et al. [5] in New York, USA (78%), but slightly higher to reported by Grasselli et al. [8] in Lombardy region, Italy (26%). These observed differences in ICU mortality could respond to different healthcare models and important practice variations in the management of severe COVID-19 patients [11], but it can also depend on the frequency of presentation of the different phenotypes.

In our study, a great variability in model performance and risk factors were observed during cross-validation to choose the best model to use. In addition, we use 2 different techniques for the selection of important variables, one of them is the “classic” approach dependent on the p-value, while the other, a “modern” statistical approaches, is more in line with the new recommendations [21]. Although the performance of the models was similar, the variables included in each of them are different. This could be related to the presence of a very heterogeneous patient population, which is revealed during random partitioning (80%/20%) validation of each model or by implementing 2 variable selection techniques. In this context, three clinical phenotypes of COVID-19 patients were derived using routinely available clinical data at ICU admission by an unsupervised cluster analysis. The phenotypes were multidimensional, differed in their demographics, laboratory abnormalities, patterns of organ dysfunction, and associated with ICU mortality. In addition, our phenotypes are not similar with groupings or phenotypes of patients performed so far considering only the presence of clinical complications [22, 23], or the type of ARDS [24]. Our COVID-19 phenotypes can be identified at the time of the ICU admission, and thus could be useful in facilitating early tailored therapy and improve prognosis.

Only routinely available clinical and laboratory data were used in the clustering models, and the phenotypes were derived from a large observational multicenter cohort to ensure generalizability. Importantly, we have observed that the variables associated with the ICU mortality varied between the global model and the models developed for each phenotype. The discrimination power (AUC) of A and B phenotypes models improved in comparison to the global model. However, for the C phenotype (severe COVID-19 disease), the performance of the model was not superior respect of the global model. The C phenotype was most strongly correlated with abnormal values of biomarkers as well as clinical features of cardiovascular dysfunction, AKI and subsequently a higher ICU mortality. Although the AUC for C phenotype is lower, the relationship between sensitivity and specificity in C phenotype model might be more appropriate. Specificity can sometimes be more important than sensitivity, because confirming that a person does not have the event under study (survival) is more important than detecting if a person has it.

Recently, several authors have proposed different clinical phenotypes of COVID-19 patients [22‐24]. Rello et al. [22] speculated that COVID-19 has five phenotypic presentations based on physiological and clinical features from published studies. Garcia-Vidal et al. [23] describe the main clinical complications of hospitalized patients with COVID-19 through classification into three pattern groups (inflammatory, co-infection and thrombotic). However, as the authors acknowledge, the cut-off points of the different biomarkers for defining phenotypes have been arbitrary and not scientifically supported. Finally, Gattinoni et al. [24] proposing two phenotypes for COVID-19 patients, (1) “Type L” characterized by high compliance and low lung recruitablity and (2) “Type H” with low compliance and high lung recruitability as a two “extremes” of a spectrum of respiratory failure in COVID-19 pneumonia. Despite the importance related to clinical experience in each of these approaches, none of these studies have been developed through a machine learning process to determine phenotypes nor have they been tested for validation.

Hypoxemia has been proposed as a marker of severity for the differentiation of phenotypes [22, 24]. In our study, the PaO₂/FiO₂ relationship at ICU admission was an independent risk factor for ICU mortality in overall multivariate analysis (as a continuous or dichotomized variable), but was only closely associated with ICU mortality in phenotype C. Other variables such as advanced age, serum D-dimer values and the development of AKI were variables more strongly related to ICU mortality in all subgroups or phenotypes analyzed than PaO₂/FiO₂ at ICU admission.

Our results should be interpreted in the context of the study limitations. First, although phenotypes were found to be generalizable in our population, risk factors and characteristics of clinical phenotypes were derived initially from data at ICU admission of multicenter observational study in Spain. However, these risk factors are similar to those that have been reported by other investigators [4‐8]. The cross-validation carried out and the high discrimination observed for each of the models built for phenotypes, suggests their applicability to other populations, but it should be examined considering the high variability observed in patients with COVID-19 and in the support measures applied. Second, because missing data were common for some variables included in the clustering models, multiple imputation was used in the primary analysis. However, variables with high missing values were excluded and the missing threshold used was reported elsewhere [18]. Third, only routinely available clinical data at ICU admission were used to identify risk factors and clinical phenotypes, and the inclusion of other data related to clinical evolution of patients in the ICU could change risk factors or phenotype assignments. However, our objective was to study early risk factors and phenotypes at ICU admission that may allow for early treatment implementation and as a result improve patient outcome. Fourth, although IL-6 is an excellent severity biomarker, we have not been able to include this biomarker in the models because more than 50% of patients had no IL-6 determination upon ICU admission. Although the inclusion of IL-6 in models could modify or improve their performance, we do not consider it appropriate to impute a large number of missing data. In addition, if IL-6 is a biomarker not usually available its inclusion in the models would not have practical application. Finally, we did not collect data on ethnicity or socioeconomic factors. These factors may play a role in the prevalence of pre-existing comorbidities and mortality due to COVID-19. Our findings should be interpreted within the context of the study population and its generalizability to other populations warrant further investigation.

To our knowledge this is the largest study that describe different phenotypes of patients with confirmed COVID-19 that were admitted to ICU to date. We not only characterized three novel clinical phenotypes, but extended findings outside of a single site ICU by characterizing the association of comorbidities with clinical phenotype and the association of clinical phenotypes with clinical outcomes. Different risk factors for the global population and clinical phenotypes were observed, possibly due to the heterogeneity of patients, which may limit the application of a single predictive model for all patients with COVID-19. Further research is needed to determine the application of these phenotypes in clinical practice, in other patient’s population and for clinical trial design.