Clinical phenotyping uncovers heterogeneous associations between corticosteroid treatment and survival in critically ill COVID-19 patients

Abstract

Purpose

Disease heterogeneity in coronavirus disease 2019 (COVID-19) may render the current one-size-fits-all treatment approach suboptimal. We aimed to identify and immunologically characterize clinical phenotypes among critically ill COVID-19 patients, and to assess heterogeneity of corticosteroid treatment effect.

Methods

We applied consensus k-means clustering on 21 clinical parameters obtained within 24 h after admission to the intensive care unit (ICU) from 13,279 COVID-19 patients admitted to 82 Dutch ICUs from February 2020 to February 2022. Derived phenotypes were reproduced in 6225 COVID-19 ICU patients from Spain (February 2020 to December 2021). Longitudinal immunological characterization was performed in three COVID-19 ICU cohorts from the Netherlands and Germany, and associations between corticosteroid treatment and survival were assessed across phenotypes.

Results

We derived three phenotypes: COVIDICU1 (43% of patients) consisted of younger patients with the lowest Acute Physiology And Chronic Health Evaluation (APACHE) scores, highest body mass index (BMI), lowest PaO₂/FiO₂ ratio, and a 90-day in-hospital mortality rate of 18%. COVIDICU2 patients (37%) had the lowest BMI, were older and had higher APACHE scores and mortality rate (24%) than COVIDICU1. Patients with COVIDICU3 (20%) were the eldest with the most comorbidities, the highest APACHE scores, acute kidney injury and metabolic dysregulations, and the highest mortality rate (47%). These patients also displayed the most pronounced inflammatory response. Corticosteroid therapy started at day 5 [2–9] after ICU admission and administered for 5 [3–7] days was associated with an increased risk for 90-day mortality in patients with the COVIDICU1 and COVIDICU2 phenotypes (hazard ratio [HR] 1.59 [1.09–2.31], p = 0.015 and HR 1.79 [1.42–2.26], p < 0.001, respectively), but not in patients with the COVIDICU3 phenotype (HR 1.08 [0.76–1.54], p = 0.654).

Conclusion

Our multinational study identified three distinct clinical COVID-19 phenotypes, each exhibiting marked differences in demographic, clinical, and immunological features, and in the response to late and short-term corticosteroid treatment.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00134-024-07593-3.

Take-home message

By applying machine learning approaches on routine clinical parameters of thousands of critically ill patients affected with coronavirus disease 2019, we derived three distinct phenotypes, each displaying diverse immunological signatures. Only in patients with the hyperinflammatory phenotype, late and short-term corticosteroid therapy was not associated with worse survival.

Introduction

Currently, the immunomodulators dexamethasone (a corticosteroid) and tocilizumab (interleukin [IL]-6 receptor antagonist) are applied as standard-of-care in hospitalized patients suffering from coronavirus disease 2019 (COVID-19), because they were shown to reduce mortality in largely undifferentiated patient populations [1–3]. However, recent work suggests that this uniform treatment approach may be suboptimal. For instance, a retrospective study showed that corticosteroid treatment was associated with a survival benefit only in hyper inflamed COVID-19 patients, while increased mortality was observed in patients without signs of hyperinflammation [4]. Therefore, a “precision medicine” approach could further improve outcomes for COVID-19 patients, but this requires a deeper understanding of disease heterogeneity [5].

One method that is increasingly used to gain further insights into heterogeneity and could inform treatment responses is offered by identifying phenotypes: more homogenous subgroups within a larger heterogenous patient population. Phenotyping has a proven track record in other diseases, enabling more accurate diagnoses, selection of effective treatment approaches and predicting risk of developing (severe) disease, allowing for earlier and more effective interventions [6–8]. Several clinical COVID-19 phenotypes have been identified using a wide spectrum of methodologies [9–13]. Nevertheless, these studies have several shortcomings, including relatively small sample sizes, inclusion of patients from a single center or country, use of data from the early phase of the pandemic only, and lack of assessment of possible heterogeneous treatment effects.

In the present study, we derived, reproduced, and characterized new clinical phenotypes using multiple large cohorts of COVID-19 patients admitted to the intensive care unit (ICU) in several countries. Furthermore, we explored how the distribution of these phenotypes evolved over the course of the pandemic and assessed associations between corticosteroid treatment and survival across the phenotypes, aiming to enhance our understanding of COVID-19 heterogeneity and to pave the way for more targeted and effective therapeutic interventions tailored to individual patient needs.

Parts of this work have been presented at the European Society of Intensive Care Medicine LIVES 40 congress [14] and at the 42nd International Symposium on Intensive Care & Emergency Medicine [15].

Methods

Cohorts

Phenotypes were derived in a large cohort from the Netherlands (n = 13,279 adult COVID-19 ICU patients admitted to 82 hospitals from February 2020 to February 2022). In this derivation cohort, the evolution of phenotype distribution and associated mortality over the course of the pandemic was also explored. The identified phenotypes were subsequently reproduced using a Spanish cohort (n = 6225 adult COVID-19 ICU patients admitted to 55 hospitals from February 2020 to December 2021), in which associations between corticosteroid treatment and survival across the different phenotypes were assessed as well (heterogeneity of treatment effect analysis). Phenotypes were immunologically characterized using three smaller single-center cohorts of adult COVID-19 ICU patients (Radboudumc, n = 194; Amsterdam UMC, n = 202; Jena, n = 314), which contained detailed and/or longitudinal data on immunological biomarkers not available in the other two cohorts. A schematic overview of the cohorts utilized and analyses conducted is provided in Fig. 1. An elaborate description of the cohorts, e.g. detailing in- and exclusion criteria, availability of variables, and assays used for biomarker measurements is provided in the Electronic Supplementary Material (ESM) Methods section.

Fig. 1.

Comprehensive overview of the analyses conducted on the different cohorts in the present manuscript

Statistical analyses

Clustering variable selection, phenotype derivation, and reproduction

Clustering variables were selected based on their availability within the derivation cohort, likelihood to be present in other cohorts for reproduction and intercorrelation (see ESM Methods section for details). As required for clustering analyses, missing data were imputed using chained random forests (see ESM Table S3) with predictive mean matching, after which log-transformation, scaling, and centering were performed on the new dataset.

Multiple methodologies were employed to determine the optimal clustering method and number of clusters (OPTICS plots [16], alluvial plots [17], and consensus clustering [cumulative distribution and matrix heatmap plots] [18], see ESM Methods section for details).

To assign the derived phenotypes to patients in the other cohorts, the Euclidean distance of all variables for each individual patient to the centroids of the three phenotypes in the derivation cohort was calculated, as previously described [6]. Subsequently, patients were assigned to the nearest phenotype, defined by the smallest cumulative Euclidean distance (see ESM Methods section for details).

Statistical testing

Differences in patient characteristics and crude outcomes between the phenotypes were tested using Kruskal–Wallis tests followed by Dunn’s post-hoc tests and chi-square tests. P values of pairwise comparisons were Bonferroni-corrected. Differences in patient characteristics and crude outcomes between patients who did or did not receive corticosteroids were tested using Wilcoxon rank-sum tests or chi-square tests. Survival analysis was performed using time-adjusted Cox proportional hazards regression models, taking into account the day on which corticosteroid treatment was initiated to prevent immortal time bias. Furthermore, in sensitivity analyses, models were adjusted for differences between centers (mixed effect analyses using the coxme package), disease severity (Acute Physiology And Chronic Health Evaluation [APACHE] II score) and age. To investigate whether there were heterogeneous associations between corticosteroid treatment and survival across the phenotypes, the interaction effect between phenotype and corticosteroid treatment was added to the Cox model. Subsequently, this model was tested against the model without the interaction effect with a likelihood ratio test. The applied threshold for significance was < 0.05 for two-sided tests. All analyses were performed using R 3.6.1 [19] (other nonstandard packages used are described in the ESM Methods section).

Results

Patients

Characteristics of all cohorts are listed in ESM Table S4. The derivation cohort consisted of 13,900 patients, of which 13,279 were eligible for analysis. Sixty-nine percent were male. Median [interquartile range, IQR] age was 64 [55–71] years, the APACHE II severity score was 16 [13–20], and 90-day in-hospital mortality was 26%. The Spanish reproduction cohort comprised 6225 patients, of whom 70% were male, age was 63 [54–71] years, the APACHE II score was 12 [9–15], and 90-day mortality was 31%. The smaller cohorts (Radboudumc, Amsterdam UMC and Jena) used for immunological characterization were comparable in terms of clinical characteristics and disease severity (ESM Table S4).

Selection of variables and model

Thirty-seven clinical variables were used for clustering, based on availability in the National ICU Registry of the Netherlands (NICE). The selection of variables collected in this registry is based on existing prognostic models (APACHE II/IV and Simplified Acute Physiology Score (SAPS) II) and relevant comorbidities. In addition, domain experts have determined which quality indicators (corrected for case mix) are most relevant for benchmarking. After evaluating intercorrelation (ESM Fig. S1), 21 variables remained for phenotype derivation (ESM Table S5). The smooth rise in reachability distance in the OPTICS plot indicated that a partitioning approach like consensus k-means clustering or its partitioning around medoids (PAM) equivalent was the most appropriate fit for this data compared to other methods like hierarchical clustering (ESM Fig. S2). The consensus clustering results of k-means and PAM approaches revealed that the former was a better fit for the present data set in combination with the most appropriate cluster size (ESM Fig. S3). Furthermore, three classes yielded a good fit, as the relative change under the cumulative distribution function suggested very little statistical gain from increasing the number of classes (ESM Fig. S3). The alluvial plot also supported a three-class model (ESM Fig. S4).

Phenotype derivation

By employing the 21-variable and 3-class k-means model, three clinically distinct phenotypes emerged, coined COVIDICU1, 2, and 3, making up 43% (n = 5734), 37% (n = 4844), and 20% (n = 2701) of the cohort, respectively (Fig. 2a). Renal and metabolic markers and comorbidities contributed most towards the differences between the phenotypes, whereas sodium, temperature and respiratory rate contributed least (ESM Fig. S5). All demographic, clinical, routine laboratory, and outcome variables of the three phenotypes are listed in Table 1, with all parameters showing significant differences between phenotypes. Analysis of unimputed data yielded highly similar results (ESM Table S6). The COVIDICU1 phenotype was characterized by the largest proportion of females, the highest body mass index (BMI), arterial pH and albumin as well as the lowest age, fewest comorbidities and the lowest PaO₂/FiO₂ ratio (Fig. 2b–d). COVIDICU2 was defined by the fewest comorbidities, lowest BMI and glucose levels as well as the highest PaO₂/FiO₂ ratio. The COVIDICU3 phenotype had the highest proportion of males, most comorbidities, acute kidney injury (AKI) and metabolic derailments, the highest white blood cell counts and disease severity scores. Patients with this phenotype also exhibited the highest vasopressor and mechanical ventilation requirements. The APACHE II and SAPS II severity scores were very similar between COVIDICU1 and COVIDICU2, whereas the COVIDICU3 phenotype exhibited the highest scores (Fig. 2d). Nevertheless, there was a large overlap in disease severity scores between all phenotypes, indicating that the classification into phenotypes not simply recapitulated classification by these scores. The same applies to the PaO₂/FiO₂ ratio (Fig. 2d). 90-day in-hospital mortality was lowest for COVIDICU1 (18%), intermediate for COVIDICU2 (24%), and highest for COVIDICU3 (47%, Table 1 and Fig. 2e). Similar results were observed for ICU length-of-stay (Table 1). Analysis of 90-day in-hospital mortality on unimputed data revealed highly similar results (ESM Table S6 and ESM Fig. S6).

Fig. 2.

Phenotype characteristics and outcomes in the derivation cohort. a Distribution of the three clinical phenotypes, b Standardized mean difference (SMD) of all clustering variables per phenotype, illustrating how far each variable for that group is removed from the mean of the entire cohort. PF PaO₂/FiO₂ ratio, Creat creatinine, Temp temperature, Comor comorbidity, Bili bilirubin, WBC white blood cell count, HR heart rate, Gluc glucose, MAP mean arterial blood pressure, Thrombo thrombocytes, Resp respiratory, Bicarb bicarbonate, BMI body mass index, Alb albumin, c distribution for each phenotype of sex as well as requirement of vasopressive medication and mechanical ventilation, d density plots of the APACHE II and SAPS II severity scores and the PaO₂/FiO₂ ratio. The vertical lines indicate the median value for each phenotype, E) survival curve of in-hospital mortality for the three phenotypes

Table 1.

Patient characteristics and outcomes of the different phenotypes in the derivation cohort

	COVIDICU1 (n = 5734, 43%)	COVIDICU2 (n = 4844, 37%)	COVIDICU3 (n = 2701, 20%)	p value
Sex, male	3731 (65%)	3394 (70%)	2043 (76%)	< 0.001
BMI, kg/m2	31.1 [27.8–35.2]	26.7 [24.6–29.4]	29 [26–32.9]	< 0.001
Normal (< 25)	439 (8%)	1459 (30%)	517 (19%)	< 0.001
Overweight (25 to < 30)	1928 (34%)	2289 (47%)	1036 (38%)	< 0.001
Class 1: (30 to < 35)	1845 (32%)	819 (17%)	669 (25%)	< 0.001
Class 2: (35 to < 40)	906 (16%)	160 (3%)	300 (11%)	< 0.001
Class 3: (> 40)	558 (10%)	53 (1%)	140 (5%)	< 0.001
Age, years	59 [50–67]	66 [58–73]	68 [61–73]	< 0.001
APACHE II score	15 [13–18]	15 [13–18]	21 [17–25]	< 0.001
APACHE IV score	55 [46–64]	57 [47–68]	76 [65–89]	< 0.001
SAPS II	29 [22–36]	34 [27–40]	45 [37–52]	< 0.001
AIDS	0 (0%)	4 (0%)	5 (0%)	0.008
Cardiovascular insuffiency	53 (1%)	38 (1%)	80 (3%)	< 0.001
Chronic dialysis	0 (0%)	0 (0%)	62 (2%)	< 0.001
Chronic renal insuffiency	32 (1%)	33 (1%)	461 (17%)	< 0.001
Cirrhosis	11 (0%)	14 (0%)	29 (1%)	< 0.001
COPD	463 (8%)	342 (7%)	363 (13%)	< 0.001
Diabetes mellitus	1169 (20%)	545 (11%)	1277 (47%)	< 0.001
Hematological malignancy	35 (1%)	94 (2%)	82 (3%)	< 0.001
Immunological insufficiency	361 (6%)	397 (8%)	430 (16%)	< 0.001
Neoplasm	33 (1%)	20 (0%)	38 (1%)	< 0.001
Respiratory insufficiency	219 (4%)	185 (4%)	151 (6%)	< 0.001
Mechanical ventilation	3239 (56%)	3048 (63%)	2059 (76%)	< 0.001
Comorbidity scorea	0.41 (0.63)	0.34 (0.58)	1.08 (0.90)	< 0.001
PaO2/FiO2 ratio	80 [64–106]	116 [86–164]	98 [74–137]	< 0.001
No ARDS (> 300 mmHg)	33 (1%)	208 (4%)	86 (3%)	< 0.001
Mild ARDS (> 200–≤ 300 mmHg)	122 (2%)	421 (9%)	154 (6%)	< 0.001
Moderate ARDS (> 100–≤ 200 mmHg)	1380 (24%)	2005 (41%)	954 (35%)	< 0.001
Severe ARDS (≤ 100 mmHg)	3864 (67%)	1590 (33%)	1291 (48%)	< 0.001
PaCO2 (mmHg)	36 [33–42]	37 [32–44]	41 [33–49]	< 0.001
Respiratory rate (max), breaths/min	34 [30–40]	30 [26–36]	32 [28–38]	< 0.001
Vasoactive medication	2319 (40%)	2519 (52%)	1898 (70%)	< 0.001
Hematocrit (min)	0.39 [0.36–0.42]	0.37 [0.34–0.40]	0.36 [0.32–0.40]	< 0.001
Heart rate (max), beats/min	99 [88–111]	96 [84–108]	110 [95–127]	< 0.001
Mean arterial pressure (min), mmHg	68 [62–75]	64 [58–70]	61 [55–68]	< 0.001
Mean arterial pressure (max), mmHg	112 [102–125]	105 [95–117]	105 [94–120]	< 0.001
Acute renal failure	87 (2%)	106 (2%)	634 (23%)	< 0.001
Creatinine (max), µmol/L	68 [57–82]	70 [59–86]	136 [99–202]	< 0.001
Blood urea nitrogen, mg/dL	20 [15–26]	20 [15–26]	37 [27–53]	< 0.001
Urinary output, L	1.76 [1.30–2.40]	1.40 [1.01–1.98]	1.30 [0.80–1.90]	< 0.001
Bilirubin, µmol/L	8 [6–11]	9 [6–12]	8 [6–13]	< 0.001
Sodium (max), mmol/L	139 [137–142]	138 [136–141]	139 [136–142]	< 0.001
Potassium (max), mmol/L	4.3 [4–4.5]	4.1 [3.9–4.4]	4.7 [4.3–5.1]	< 0.001
Glucose (max), mmol/L	11.4 [9.2–15]	9.2 [7.6–11.4]	14.4 [11.2–18.4]	< 0.001
pH (min)	7.46 [7.41–7.49]	7.44 [7.38–7.48]	7.33 [7.26–7.40]	< 0.001
Bicarbonate (max), mmol/L	27 [25–29]	26 [24–28]	23 [21–26]	< 0.001
Albumin (min), g/L	31 [27–34]	26 [23–29]	27 [23–30]	< 0.001
White blood cell count (max), × 109/L	9.4 [6.9–12.4]	9.2 [6.7–12.3]	11.4 [8.1–15.9]	< 0.001
Thrombocytes (min), × 109/L	256 [201–321]	231 [176–297]	214 [160–280]	< 0.001
Temperature, °C	37.5 [37–38.1]	37.9 [37.2–38.8]	37.7 [37.1–38.5]	< 0.001
ICU length-of-stay survivors, days	8 [5–16]	10 [5–20]	13 [6–29]	< 0.001
ICU length-of-stay nonsurvivors, days	17 [9–26]	17 [9–25]	12 [5–21]	< 0.001
ICU mortality	958 (17%)	1024 (21%)	1163 (43%)	< 0.001
Hospital length-of-stay survivors, days	18 [12–31]	22 [14–37]	28 [16–48]	< 0.001
Hospital length-of-stay nonsurvivors, days	21 [13–30]	20 [13–30]	16 [8–2]	< 0.001
In-hospital mortality	1037 (18%)	1156 (24%)	1273 (47%)	< 0.001
28-day in-hospital mortality	738 (13%)	836 (17%)	1029 (38%)	< 0.001
90-day in-hospital mortality	1035 (18%)	1148 (24%)	1263 (47%)	< 0.001

Underlined parameters were used for clustering. Data are presented as median [interquartile range], mean (standard deviation), or number (%). Unimputed data are listed, so not all percentages add up to 100% due to missingness in some of the variables (see Table S3). P values were calculated by Kruskal Wallis tests or chi-square tests across all phenotypes

AIDS acquired immunodeficiency syndrome, APACHE II/IV Acute Physiology and Chronic Health Evaluation II/IV, COVID-19 coronavirus disease 2019, BMI body mass index, COPD chronic obstructive pulmonary disease, ARDS acute respiratory distress syndrome, ICU intensive care unit, SAPS II simplified acute physiology score 2

^aCalculated by adding one point for the presence of each of the following comorbidities: AIDS, cardiovascular insufficiency, chronic dialysis, chronic renal insufficiency, cirrhosis, COPD or respiratory insufficiency, diabetes mellitus, hematologic malignancy, immune insufficiency, and metastatic neoplasm

Phenotype reproduction

The COVIDICU phenotypes were subsequently externally reproduced in a Spanish cohort (n = 6225). This approach revealed a highly similar distribution, with 37%, 48%, and 15% of patients assigned to COVIDICU1, COVIDICU2, and COVIDICU3, respectively (ESM Fig. S7a and ESM Table S7). Characteristics of the three phenotypes were also comparable to the derivation cohort, as was the distribution of APACHE II scores and 90-day mortality (22%, 30%, and 56%, respectively, ESM Fig. S7b-d and ESM Table S7).

Evolution of phenotype distribution and associated mortality over the course of the pandemic

In the derivation cohort, three distinct COVID-19 waves were identified in (Fig. 3). The first wave (March–June 2020) was characterized by a sharp spike in admissions per week with very high numbers, while the subsequent waves were more prolonged, with fewer admissions per week. In the first wave, the COVIDICU2 phenotype was most prevalent, whereas this shifted to the predominance of the (lower mortality) COVIDICU1 phenotype in the later waves. In accordance, overall mortality was lower in the second and third waves, also because the proportion of patients with COVIDICU3 remained relatively stable throughout the pandemic.

Fig. 3.

Evolution of phenotype distribution and associated mortality over the course of the pandemic in the derivation cohort (n = 13.279). ICU admissions per week (dates provided as MM–YY), colored by phenotype. The phenotype proportions in each week are depicted at the bottom. Three major waves were identified, indicated by the colored lines underneath the admission number bars. The 90-day in-hospital mortality rate for each phenotype and the overall mortality rate are shown at the top of the plot

Immunological characterization of the different phenotypes

The phenotype characteristics in the three smaller cohorts, Radboudumc (n = 194), Amsterdam UMC (n = 202), and Jena (n = 314), used for immunological characterization were largely similar to that of the derivation and reproduction cohorts (ESM Figs. S8, S9 and S10; ESM Tables S8, S9, and S10). Patients with the COVIDICU3 phenotype exhibited the highest levels of cytokines and routine laboratory inflammatory markers during the first three days of admission (Fig. 4a). Furthermore, principal component analysis (PCA) performed on 37 inflammatory biomarkers measured within the first 48 h of ICU admission revealed that the COVIDICU3 phenotype showed the most pronounced separation from the other two phenotypes for the panels “coagulation”, “epithelial”, “inflammation and organ damage”, and “pro-inflammatory cytokines and chemokines” (Fig. 4b–e; an overview of all individual markers across the three phenotypes in this cohort is provided in ESM Fig. S11). For these panels, the differences between COVIDICU3 and the other two phenotypes were predominantly driven by levels of thrombomodulin and plasminogen activator inhibitor-1 (PAI-1), RAGE and tenascin C, RAGE and granzyme B, and IL-13 and Rantes, respectively. Less separation between COVIDICU3 and the other two phenotypes was observed for the panels “anti-inflammatory cytokines and chemokines”, “anti-viral response”, and “endothelial” (ESM Fig. S12). The COVIDICU1 and COVIDICU2 phenotypes showed large overlap in all panels (ESM Fig. 4b–e, Fig. S12).

Fig. 4.

Immunological characterization of the phenotypes in different cohorts. a Heatmap of all available cytokines and routine laboratory inflammation markers in the Radboudumc cohort (n = 194). Biomarker data were obtained during the first 3 days of ICU admission. If multiple measurements were available within this period, the mean value was used. Color indicates the scaled value of the plasma concentration of the different markers. CRP C-reactive protein, PCT procalcitonin, IL interleukin, TNF tumor necrosis factor, IFN interferon, IP interferon gamma-induced protein, MIP macrophage inflammatory protein, MCP monocyte chemoattractant protein, RA receptor antagonist, b principal component analysis (PCA) plot of patients of the Amsterdam UMC cohort (n = 202) based on the markers in the ‘coagulation’ panel measured within 48 h of ICU admission, c PCA plot of patients of the Amsterdam UMC cohort (n = 202) based on the markers in the ‘epithelial’ panel measured within 48 h of ICU admission, d PCA plot of patients of the Amsterdam UMC cohort (n = 202) based on the markers in the ‘inflammation and organ damage’ panel measured within 48 h of ICU admission, e PCA plot of patients of the Amsterdam UMC cohort (n = 202) based on the markers in the ‘pro-inflammatory cytokines and chemokines’ panel measured within 48 h of ICU admission; see Table S1 for composition of the panels, f plasma concentrations of the proinflammatory cytokines TNF, IL-8, and IP-10, as well as the anti-inflammatory cytokine IL-10 during the first 15 days of ICU admission in the Radboudumc cohort (n = 194). Data are presented as geometric mean (GM) and 95% confidence interval (CI), g plasma concentrations of the routine laboratory inflammatory markers CRP, D-dimer and PCT, and the pro-inflammatory cytokine IL-6 during the first 15 days of ICU admission in the Radboudumc cohort. Data are presented as GM and 95% CI, h plasma concentrations of CRP, D-dimer, PCT, and IL-6 during the first 14 days of ICU admission in the Jena cohort (n = 314). Data are presented as GM and 95%CI

In longitudinal biomarker analyses, proinflammatory cytokines tumor necrosis factor (TNF), IL-8 and IP-10 and the anti-inflammatory cytokine IL-10 all showed a similar pattern over time across all phenotypes, with the highest concentrations at ICU admission which gradually declined afterward (Fig. 4f). Concentrations were consistently highest in patients with the COVIDICU3 phenotype, followed by the COVIDICU2 and COVIDICU1 phenotypes. These findings are corroborated by longitudinal data of routine inflammatory parameters C-reactive protein (CRP), D-dimer, procalcitonin (PCT), and ferritin, which also declined over time for all phenotypes and were highest for COVIDICU3 during the first 15 days of admission (Fig. 4g, h; ESM Fig. S13). The cytokine IL-6 also declined during ICU admission in patients with the COVIDICU1 and COVIDICU2 phenotypes, but remained high in patients with the COVIDICU3 phenotype (Fig. 4g, h). Furthermore, the most pronounced lymphopenia was observed for the COVIDICU3 phenotype (ESM Fig. S13).

Phenotype-specific responses to corticosteroid treatment

Out of the 3146 patients of the reproduction cohort who were admitted before the publication of the preliminary results of the RECOVERY trial on dexamethasone in June 2020 [1], 62 were excluded due to missing data (either on corticosteroid use or mortality). Furthermore, 1332 patients were excluded because they already received corticosteroids on the ward or within the first 24 h of ICU admission, which may have influenced phenotype assignment. This left 1752 patients eligible for this analysis (a flowchart is provided in ESM Fig. S14 and patient characteristics are listed in ESM Table S11). The main corticosteroids used were dexamethasone, methylprednisolone and hydrocortisone (ESM Table S11). Corticosteroid therapy was started at a median [IQR] of 5 [2–9] days after ICU admission and was administered for 5 [3–7] days. Phenotype distribution and characteristics were largely comparable between patients treated with corticosteroids and those who were not (ESM Table S12). In the undifferentiated patient population, crude 90-day mortality was comparable between patients who received corticosteroids and those who did not (34% vs. 33%, p = 0.808). However, in time-adjusted cox proportional hazard models (accounting for immortal time bias), corticosteroid therapy was associated with an increased risk for 90-day mortality (hazard ratio [HR] 1.60 [1.35–1.90], p < 0.001). Adjustment for differences between centers, APACHE II score, and age did not relevantly change this association, and results were similar for 28-day mortality (ESM Table S13). There was a significant interaction effect between phenotype and corticosteroid treatment for 90-day mortality (p = 0.010), illustrating heterogeneity of treatment effect across the phenotypes. Heterogeneity of treatment effect for 90-day mortality was not present when splitting the cohort according to two commonly used CRP cutoffs (< 75 [n = 296] or ≥ 75 [n = 1312] mg/L [20]: p = 0.386 for interaction; < 150 [n = 722] or ≥ 150 [n = 886] mg/L [21, 22]: p = 0.152 for interaction; CRP data were missing for 144 patients). We subsequently evaluated associations between corticosteroid treatment and the risk for 90-day mortality within each phenotype. This analysis revealed that corticosteroids therapy was associated with harm in patients with the COVIDICU1 and COVIDICU2 phenotypes (HR 1.59 [1.09–2.31], p = 0.015 and 1.79 [1.42–2.26], p < 0.001, respectively), but not in patients with the COVIDICU3 phenotype (HR 1.08 [0.76–1.54], p = 0.654). Again, sensitivity analyses taking into account differences between centers, APACHE II scores, and age did not substantially influence these findings, and results for 28-day mortality were also comparable to those found for 90-day mortality (ESM Table S13).

Discussion

In the present work, we derived three distinct clinical phenotypes with different outcomes in a large cohort of 13,279 critically ill COVID-19 patients from all ICUs in the Netherlands. Subsequentially, these phenotypes were externally reproduced, leveraging another large cohort of COVID-19 ICU patients from Spain. We employed three different, smaller cohorts from two countries to immunologically characterize these phenotypes, revealing that the COVIDICU3 phenotype, clinically characterized by advanced age, the largest proportion of males, multiple comorbidities, AKI and metabolic derailments, displayed the most profound inflammatory signature. Our investigation also shows that short-term corticosteroid therapy initiated late after ICU admission is associated with worse 90-day survival in an undifferentiated patient population, but not in patients with the COVIDICU3 phenotype. However and importantly, corticosteroid treatment was not randomized, which may have introduced bias.

Patients with the COVIDICU1 phenotype characterized as the young, previously healthy overweight patient with single organ respiratory failure (i.e. the worst PaO₂/FiO₂ ratio of all phenotypes) exhibited the lowest levels of inflammatory markers among the three phenotypes and had the best chance of survival. These patients likely required ICU care due to ventilation challenges associated with obesity but exhibit relatively minor signs indicating failure of other organs. Together with their young age and relatively few comorbidities, this may provide an explanation for their favorable outcome. Late and short-term corticosteroid treatment was associated with an increased risk for 90-day mortality in these patients. Patients with the COVIDICU2 phenotype were less obese and had better pulmonary function but were markedly older and showed more pronounced hemodynamic instability in comparison to COVIDICU1 patients. Inflammatory markers of these patients fell in between those observed in the COVIDICU1 and COVIDICU3 phenotypes. This intermediate phenotype also translated to the survival rate, which was slightly worse than that of the COVIDICU1 phenotype but significantly better than for patients with the COVIDICU3 phenotype. In COVIDICU2 patients, late and short-term corticosteroid treatment was also associated with an increased risk for 90-day mortality. Patients with the COVIDICU3 phenotype were the oldest and most comorbid, often had AKI, and exhibited the most pronounced inflammation, not only on ICU admission but also in the weeks afterward. Plausibly related to this more pronounced inflammatory state, corticosteroid treatment was not associated with worse 90-day survival for patients with this phenotype only, even when administered late into their ICU admission and for a relatively short period of time. The more pronounced inflammatory state in patients with the COVIDICU3 phenotype may also explain the hyperglycemia observed in these patients, as inflammation is a well-known cause of insulin resistance [23].

We restricted our heterogeneity of treatment analysis to the period before the publication of the preliminary results of the RECOVERY trial [1], because virtually all patients received dexamethasone afterwards. Furthermore, the same variants of SARS-CoV-2 were dominant in this period, precluding possible bias by varying virulence [24]. Also, other advances in the prevention or treatment for COVID-19 patients that were discovered later on in the pandemic did not yet apply (e.g. vaccination, anti-coagulation regimens, mechanical ventilation adjustments, fluid management, etc.). Nevertheless, corticosteroid treatment was not randomized in this observational study, possibly introducing bias by indication. Although characteristics at ICU admission were largely similar between patients who received this type of treatment and those who did not, treatment was initiated relatively late after ICU admission. Therefore, it appears plausible that mainly patients who might have deteriorated or at least did not improve received corticosteroids. The late administration likely explains the discrepant results between our study and RCTs showing beneficial effects, where corticosteroid treatment was either initiated on the ward (RECOVERY [1]) or at ICU admission (REMAP-CAP [3]). Another factor that may be involved is the duration of treatment, as patients in our study were treated for a median of 5 days while the aforementioned studies applied a 10- and 7-day course, respectively [1, 3]. However, a recent meta-analysis demonstrated that corticosteroid treatment for 6 days or less was associated with a greater risk reduction for mortality than more prolonged treatment [25].

There have been other efforts on phenotyping of COVID-19 patients. In the early stages of the pandemic a single-center study of 483 COVID-19 patients discerned a hyperinflammatory and a hypoinflammatory subtype, with trends towards improved survival upon corticosteroid treatment in the former subgroup and worse survival in the latter [26]. As in ours, corticosteroid treatment was not randomized in that study. Furthermore, the analysis was not restricted to patients who received corticosteroids after variables for phenotype derivation were collected. As such, corticosteroid treatment may have had (major) effects on phenotype assignment, as it is well-known to alter many clinical and laboratory parameters. This precludes a clear interpretation of its effects. Another study in approximately 2000 Spanish patients found no heterogeneous associations between non-randomized corticosteroid treatment and survival across three phenotypes in the unadjusted analysis [27]. However, after adjusting for 20 variables in a regression model, a survival benefit for the phenotype that showed signs of hyperinflammation emerged. In our view, adjustment is not required, as patients sharing the same phenotype inherently do not considerably differ from each other, which is also apparent from our analyses, in which adjustment for age and disease severity did not relevantly change the results. It may, therefore, rather introduce bias, especially in light of the large number of adjustment variables used in the aforementioned study [27]. Furthermore, variables for phenotype derivation were collected within 24 h of ICU admission, while corticosteroids treatment was defined as administration within 48 h prior to until 24 h after ICU admission [27]. Again, this likely had a major impact on phenotype assignment.

The heterogeneous association between late corticosteroid therapy and survival together with a clearly dysregulated underlying biological mechanism (i.e. more pronounced inflammation) in COVIDICU3 patients illustrates the potential value of clinical phenotyping for precision medicine approaches. The wide availability of the used clustering variables across cohorts from different countries illustrates the feasibility and applicability of the identified phenotypes across various healthcare settings. Furthermore, clinical phenotyping proved to be superior compared with stratification by the inflammatory biomarker CRP, possibly because it captures more complexity, and was not affected by adjustment for well-established prognostic factors. Although its added value above prognostic enrichment still requires prospective evaluation, clinical phenotyping could therefore also represent a fruitful predictive enrichment strategy for future randomized intervention trials, especially in complex syndromes often encountered in the ICU. In such a scenario, routinely available clinical and laboratory data collected from the electronic health record to determine a patient’s phenotype could be used as an additional in- or exclusion criterium for study enrolment. This may be particularly promising for immunomodulatory interventions, as our data reveal that clinical phenotypes are highly correlated to immunological parameters. As such, clinical phenotyping may obviate the need for complicated and time-consuming immunological measurements that are rarely available beyond specialized centers.

The prevalence of COVIDICU1 and COVIDICU2 alternated throughout the COVID-19 waves, while COVIDICU3 proportions remained relatively stable. COVIDICU2 was particularly prevalent during the initial wave, while COVIDICU1 was dominant in subsequent waves. Advances in preventive and pre-ICU care for COVID-19 patients may account for this. For instance, widespread vaccination campaigns in Europe were initiated in the winter of 2020. Furthermore, publication of the RECOVERY trial findings during the summer of 2020 likely also plays a major role, because and as alluded to before, most patients on oxygen received corticosteroid treatment on the ward from then on [1, 28]. This may have led to a different category of patients that did not respond and ultimately required ICU care. The fact that healthcare professionals gained experience with treating COVID-19 as the pandemic progressed, and other treatment protocols (i.e. ventilation, anti-coagulation, fluid management, etc.) were increasingly standardized across hospitals and countries may also have contributed to different patients ending up on the ICU [1, 29, 30].

Several limitations of our work deserve attention. First, and related to the aforementioned non-randomized use of corticosteroids, no pre-defined treatment protocol for this type of therapy was in place, representing another potential source of bias. Second, we were unable to include COVID-19 variant data in our analysis, as they were not available in any of the cohorts. Similarly, information on vaccination status or previous COVID-19 infections was lacking. Third, it is pertinent to acknowledge that cluster membership is not binary. Patients associate with a specific phenotype to a certain degree, which may be less pronounced for one patient than for another with the same phenotype. Thus, in practice, it could be envisioned that an algorithm informs a physician about the likelihood that a patient belongs to a specific phenotype, rather than merely providing the phenotype that the patient most closely resembles. This could facilitate more nuanced decision-making. Finally, except for the reproduction cohort, only in-hospital mortality data were available. However, as only very few patients died after hospital discharge in the reproduction cohort, this does not appear to introduce relevant bias.

In conclusion, our study highlights the value of clinical phenotyping to resolve heterogeneity among critically ill COVID-19 patients. This could inform future trial design and treatment decisions to improve clinical outcomes of critically ill patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

NB 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. Concept and design: PP, MK. Acquisition, analysis, or interpretation of data: All authors. Drafting of the manuscript: NB, MK. Critical revision of the manuscript for important intellectual content: All authors. Statistical analysis: NB, AM, RvA, EdB, JK. Administrative, technical, or material support: AM, RvA, EdB, EJK, AJ, DvL, DS, DTR, NFdK, LvG, FBM, FB. Supervision: PP, MK. All authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by a Clinical Research Award from the European Society of Intensive Care Medicine (ESICM, awarded to MK) and The Netherlands Organisation for Health Research and Development (ZonMw) COVID-19 Programme in the bottom-up focus area 1 “Predictive diagnostics and treatment” for theme 3 “Risk analysis and prognostics” (project number 10430 01 201 0011: IRIS). The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Data availability

For the NICE derivation cohort, data are only available upon reasonable request after approval by the (scientific) board of NICE after submission of a proposal with signed data access agreement. For the other cohorts, data are available upon reasonable request.

Declarations

Conflicts of interest

The authors declare no conflicts of interest.

Ethics approval

This study involving human participants was in accordance with the ethical standards of the institutional and national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. For data obtained in the derivation cohort: In accordance with Dutch legislation and compliance with the European General Data Protection Regulation, there is no need to obtain consent when anonymous data are used. For data obtained in the reproduction cohort: The study received approval from the Institution’s Internal Review Board (Comité Ètic d’Investigació Clínica, registry number HCB/2020/0370). Local researchers maintained contact with a study team member and participating hospitals obtained local ethics committee approval. For data obtained in the Radboudumc cohort: Data collection and blood sampling were performed as part of a cohort study, which was carried out in accordance with the applicable rules concerning the review of research ethics committees and informed consent in the Netherlands. All patients or legal representatives were informed about the details of this cohort study and could decline to participate. For data obtained in the Amsterdam UMC cohort: Data collection was approved by the Amsterdam UMC ethics committee (AUMC 2020_065). Written informed consent was obtained from all patients or their legal representatives with an additional option of deferred consent using an opt-out method for incapacitated patients. For data obtained in the Jena cohort: Data collection was carried out as part of a cohort study on hepatic dysfunction in COVID-19, which was approved by the ethics committee of Jena University Hospital (2020-1797). As only data from routine care were used in this study, there was no need for informed consent according to state law.