key: cord-0866698-ap8lmte0 authors: Su, C.; Xu, Z.; Hoffman, K.; Goyal, P.; Safford, M. M.; Lee, J.; Alvarez-Mulett, S.; Gomez-Escobar, L.; Price, D. R.; Harrington, J. S.; Torres, L. K.; Martinez, F. J.; Campion, T. R.; Kaushal, R.; Choi, A. M. K.; Wang, F.; Schenck, E. J. title: Identifying organ dysfunction trajectory-based subphenotypes in critically ill patients with COVID-19 date: 2020-07-18 journal: nan DOI: 10.1101/2020.07.16.20155382 sha: ed86e5b29c5f9e8b298dab076658efebd85086cc doc_id: 866698 cord_uid: ap8lmte0 Rationale. COVID-19-associated respiratory failure offers the unprecedented opportunity to evaluate the differential host response to a uniform pathogenic insult. Prior studies of Acute Respiratory Distress Syndrome (ARDS) have identified subphenotypes with differential outcomes. Understanding whether there are distinct subphenotypes of severe COVID-19 may offer insight into its pathophysiology. Objectives. To identify and characterize distinct subphenotypes of COVID-19 critical illness defined by the post-intubation trajectory of Sequential Organ Failure Assessment (SOFA) score. Methods. Intubated COVID-19 patients at two hospitals in New York city were leveraged as development and validation cohorts. Patients were grouped into mild, intermediate, and severe strata by their baseline post-intubation SOFA. Hierarchical agglomerative clustering was performed within each stratum to detect subphenotypes based on similarities amongst SOFA score trajectories evaluated by Dynamic Time Warping. Statistical tests defined trajectory subphenotype predictive markers. Measurements and Main Results. Distinct worsening and recovering subphenotypes were identified within each stratum, which had distinct 7-day post-intubation SOFA progression trends. Patients in the worsening suphenotypes had a higher mortality than those in the recovering subphenotypes within each stratum (mild stratum, 29.7% vs. 10.3%, p=0.033; intermediate stratum, 29.3% vs. 8.0%, p=0.002; severe stratum, 53.7% vs. 22.2%, p<0.001). Worsening and recovering subphenotypes were replicated in the validation cohort. Routine laboratory tests, vital signs, and respiratory variables rather than demographics and comorbidities were predictive of the worsening and recovering subphenotypes. Conclusions. There are clear worsening and recovering subphenotypes of COVID-19 respiratory failure after intubation, which are more predictive of outcomes than baseline severity of illness. Organ dysfunction trajectory may be well suited as a surrogate for research in COVID-19 respiratory failure. 1 The COVID-19 pandemic has created an unprecedented opportunity to explore a large cohort of patients infected with a single pathogen thus providing a window to examine patient variability in response to a uniform insult. Despite this opportunity, distinct subphenotypes of severe-COVID-19 associated respiratory failure remain largely unexplored (1) (2) (3) . SARS-CoV-2 infection often leads to hypoxemic respiratory failure requiring treatment with mechanical ventilation which meets clinical and pathologic criteria for Acute Respiratory Distress Syndrome (ARDS) (4) (5) (6) . In COVID-19 respiratory failure, like other forms of ARDS, there is significant risk of morbidity and mortality. However, there is clear heterogeneity in outcomes, even in those treated with mechanical ventilation (4, 5, (7) (8) (9) . The baseline clinical characteristics and predictors of mortality of those requiring mechanical ventilation have been described (4, 7, 8, 10) . These studies offer some insight into a differential host response but are limited to characterizing patients at baseline. In prior studies of ARDS (11, 12) , unique subphenotypes have been described, which identify hyperinflammatory and hypoinflammatory populations with differential demographics, clinical characteristics, inflammatory markers and outcomes. These subphenotypes are primarily characterized by host response inflammatory markers and patterns of organ injury, but are agnostic of the type of insult or infection. In COVID-19, baseline risk stratification may be insufficient to characterize subphenotypes that accurately reflect the complexity of the disease arc (13) . Serial, temporally ordered, All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . Sequential Organ Failure Assessment (SOFA) (14) (15) (16) (17) and comprehensive Electronic Health Records (EHR) data are well suited to develop data-driven subphenotypes (18) , where the goal is to identify coherent patient groups with similar clinical courses. Dynamic time warping (DTW) (19) is a well-established technique for evaluating the similarities among temporal sequences (20, 21) . DTW is particularly well suited to evaluate longitudinal changes in organ dysfunction in COVID19. Characterizing a more complete representation of the disease course in COVID19 may offer insight into its pathophysiology. We used DTW to conduct a two staged post-intubation trajectory analysis of SOFA-based organ dysfunction in patients with COVID19 to identify unique subphenotypes. In order to understand the differential disease course, we then explored clinical and biologic features including demographics, comorbidities, clinical characteristics, inflammatory markers, and treatments predictive of these trajectories. This was a retrospective two staged modeling analysis on two cohorts of intubated COVID-19 patients. The overall workflow of our study is illustrated in Figure 1 . We used individual patient data from two New York Presbyterian (NYP) system hospitals located in New York city: the New York Presbyterian Hospital-Weill Cornell All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07. 16.20155382 doi: medRxiv preprint Medical Center (NYP-WCMC), an 862-bed quaternary care hospital, and the New York Presbyterian-Lower Manhattan Hospital (NYP-LMH), a 180-bed non-teaching academic affiliated hospital. Patients were admitted from Mar 3, 2020 to May 12, 2020. SARS-CoV2 diagnosis was made through reverse-transcriptase-PCR assays performed on nasopharyngeal swabs. The critical care response to the pandemic has been previously described (22) . The NYP-WCMC cohort was used as the development cohort to derive subphenotypes, and the NYP-LMH cohort was used for validation. The focus of this study was critically ill patients with COVID-19 who were treated with intubation (Supplemental Appendix 1). We collected all data from either the Weill Cornell-Critical carE Database for Advanced Research (WC-CEDAR), Weill Cornell Medicine COVID Institutional Data Repository (COVID-IDR), or via manual chart abstraction (REDCap). WC-CEDAR aggregates and transforms data from institutional electronic health records for all patients treated in ICUs in NYP-WCMC and NYP-LMH (23) . The COVID-IDR contains additional aggregate EHR data on all patients who were tested for SARS-CoV-2 at NYP-WCMC or NYP-LMH. The REDCap database contains high-quality manually abstracted data on all patients who tested positive for COVID-19 at NYP-WCMC or NYP-LMH (24) . In our analysis, the patient information incorporated included demographics, laboratory tests, vital signs, and respiratory variables obtained from WC-CEDAR, comorbidity information obtained All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07.16.20155382 doi: medRxiv preprint from the REDCap database, and medication data obtained from the COVID-IDR. Data analyzed were detailed in Supplemental Appendix 2. The SOFA score is the sum of six organ dysfunction subscores, including cardiovascular, central nervous system (CNS), coagulation, liver, renal, and respiration (14, 17) . In this study, the CNS, coagulation, liver, and renal subscores were derived according to the standard SOFA scoring system (14) . The respiration subscore was calculated using a combination of the traditional and modified scoring method (25) . The cardiovascular SOFA subscore was calculated with additional vasopressors according to a norepinephrine equivalency table, where phenylephrine and vasopressin were converted to a norepinephrine equivalency (26) . SOFA scores were derived every 24 hours from the time of intubation, and the worst score within that 24-hour data period was selected for each patient . (14) We included patients with positive results on viral RNA detection by real-time reverse transcriptase polymerase chain reaction (RT-PCR) test from nasopharyngeal swabs specimens and treated with mechanical ventilation at the ICU in NYP-WCM and NYP-LMH. We excluded patients who were less than 18 years old. Since our aim was to identify clinically meaningful organ dysfunction progression patterns of intubated patients, trajectories with low quality (20 (5.7%) patients missing over 50% SOFA All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07.16.20155382 doi: medRxiv preprint records) and outlier trajectories (10 (2.9%) patients with unchanged or heavily fluctuated within the 7-day window after intubation) were excluded from the analysis (Supplemental Appendix 3 and Figure E-1). SOFA scores were derived every 24 hours and post intubation 7 day SOFA trajectories were constructed for analysis. Missing values within a trajectory were imputed based on the last observation carried forward (LOCF) strategy. A two-staged subphenotyping method was performed to derive SOFA trajectory subphenotypes ( Figure 1 ). In the first stage, we used baseline SOFA to group patients with a similar upfront risk of death (17) , as additive organ dysfunction has previously been identified to be associated with poor outcomes in COVID19 (8) . We partitioned the patients into three baseline severity strata (mild, intermediate, and severe) according to their SOFA scores within the first 24 hours after intubation. The SOFA score cut-offs were set to 0-10, 11-12, and 13-24 in order to obtain a balanced distribution of patients across the three strata. In the second stage, we identified the subphenotypes with similar 7-day SOFA progression patterns. Dynamic Time Warping (DTW)(19) was adopted to evaluate the similarities between pairwise patient SOFA trajectories within each baseline stratum and then hierarchical agglomerative clustering (HAC) (27) was performed on these similarities to derive the similar patient clusters as trajectory subphenotypes. DTW can account for the differences among the evolution heterogeneity among the temporal curves and is thus able to evaluate their similarity more robustly. (19) The optimal numbers of subphenotypes were determined by clear All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. To validate these findings, we replicated these subphenotypes from the NYP-LMH cohort. We analyzed 30-day all-cause mortality as the primary outcome for patients within each phenotype. Successful extubation or need for tracheostomy within 30 days after intubation were secondary outcomes. We examined the associations between patient characteristics and clinical variables and the identified trajectory subphenotypes, to see if there are early markers that can discriminate between them. Patient characteristics we investigated included demographics, comorbidities, medications prescribed within the window from 3-day before to 5-day after intubation, and blood type (29) . Laboratory test results included: complete blood count, basic metabolic panel, liver function tests, coagulation profile and inflammatory markers including d-dimer, fibrinogen, ferritin, erythrocyte sedimentation rate, lactic acid, troponin, lactate dehydrogenase, creatine kinase, procalcitonin and Creactive protein. Vital signs included: GCS, mean arterial pressure and temperature, urine output. Respiratory variables included: P/F ratio, FiO2, Pao2, PaCO2, PH, PEEP, All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07.16.20155382 doi: medRxiv preprint peak inspiratory pressure, plateau pressure, driving pressure, static compliance, minute ventilation, ventilatory ratio, and tidal volume indexed to ideal body weight at day 1 and day 3 post-intubation. Univariate statistical tests were performed in those association analyses. Specifically, one-way analysis of variance (ANOVA, with Tukey HSD post hoc test), Kruskal-Wallis test (with Dunn post hoc test), student's t-test, Mann-Whitney test, Chisquare test, and Fisher's exact test have been used whenever appropriate. The pvalues were then corrected for multiple testing using false discovery rate (FDR) estimation. Analysis of covariance (ANCOVA) for the between-strata/subphenotypes comparisons was also applied based on the generalized linear model (GLM) with adjustment on age at baseline. All statistical tests were performed with Python 3.7 based on statsmodels package 0.11.1. We trained a random forest model with the trajectory subphenotypes as targets and the patient clinical characteristics at specific time points after intubation as input predictors to define if these trajectory subphenotypes can be predicted early. Our implementation was with Python 3.7 based on scikit-learn package 0.22.2. Candidate predictors included demographics, comorbidities, medications prescribed around the intubation event, SOFA subscores, laboratory tests, vital signs, and respiratory variables as described above. Prediction performances were measured by area under the receiver operating characteristics (AUC-ROC). The importance of predictors was visualized as a heatmap to demonstrate their contributions on subphenotype prediction. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07.16.20155382 doi: medRxiv preprint The study is approved by the IRB of Weill Cornell Medicine with protocol number 20-04021909. Table 1 . For the NYP-WCMC cohort, patients were first partitioned into mild, intermediate, and severe strata based on the SOFA scores within one day after intubation, consisting of 76 (23.29%), 116 (36.48%), and 126 (39.62%) patients, respectively; while for the NYP-LMH validation cohort, the three strata consist of 10 (11.90%), 35 (41.67%), and 39 (46.43%) patients, respectively. As shown in Table 1 , the patients in both NYP-WCMC and NYP-LMH cohorts exhibit additive patterns of post intubation baseline organ All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . Figure 2 demonstrates the individual averaged SOFA curves for patients in the two subphenotypes across all strata: a worsening subphenotype of which SOFA score increased within the 7-day observation window, and a recovering subphenotype of which SOFA score improved. The clinical characteristics of these subphenotypes were summarized in Table 2 . Overall, there was no marked difference in terms of demographics, comorbidity burden, and pattern of organ dysfunction (distribution of SOFA subscores and total score) between the worsening and recovering subphenotypes within each baseline severity stratum at baseline. This suggests that, though the subphenotypes varied in 7-day organ dysfunction progression patterns, they have similar clinical status immediately after intubation. We further investigated All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07.16.20155382 doi: medRxiv preprint medications prescribed within each subphenotype and didn't find significant signal as well (Supplemental Table E-3 ). In addition, clinical characteristics and medications of the subphenotypes re-derived in the NYP-LMH validation cohort were summarized in Supplemental Tables E-2 and 4. Statistics of 30-day post-intubation clinical primary and secondary outcomes (mortality, extubation, and tracheostomy) of subphenotypes were illustrated in Figure 2A and severe stratum 50.0% vs. 14.8%, p<0.001). There was no significant difference of 30day tracheostomy detected between the subphenotypes. Importantly, the recovering subphenotype within the severe baseline stratum had a lower mortality risk compared to the worsening subphenotypes at mild and intermediate baseline strata. The trajectory subphenotypes derived in the NYP-LMH validation cohort had similar trends in all three clinical outcomes within the 30-day window after intubation (see Figure 2A and Supplemental Figure E-3B) . Across all baseline strata, the All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Vital signs, laboratory variables, and respiratory variables were evaluated to identify early-stage markers predictive of the two-stage classification. First of all, the three baseline strata of the NYP-WCMC cohort were observed to be well separated by a series of clinical variables in addition to the differential organ dysfunction pattern noted above (Supplemental Table E (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . pressure (PEEP), plateau pressure, minute ventilation, static compliance, driving pressure, FiO2, and ventilator ratio. As shown in Supplemental Table E-8, most markers identified within the NYP-WCMC cohort showed consistent signals within the NYP-LMH subphenotypes, even though some significance vanishes, as the confidence intervals were wide. We trained random forest models for predicting the worsening and recovering trajectory subphenotypes within each baseline stratum according to the early stage marker values. Overall, as shown in Supplemental Figure E Generally, predictor importance varied as the progress of time. Models trained on day 1-3 after intubation were observed to involve more contributions from the laboratory tests, vital signs, respiratory variables than other predictors; SOFA subscores, especially All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . cardiovascular, CNS, and renal subscores showed relatively higher importance over models trained on day 4 or 5 data within the intermediate and severe strata. Age contributed to day 1-3 prediction to some extent, while other demographics, medications and comorbidities showed weak importance in prediction. We finally assessed ABO/RH blood type distribution across the subphenotypes (Supplemental Tables E-9 and 10). Overall, there is no significant signal detected. In this study, we identified novel trajectory subphenotypes of COVID-19 patients with an objective machine learning approach. The subphenotypes we identified are based on organ dysfunction trajectory over 7-days following intubation, which is different from existing data-driven subphenotyping methods that focus on patient data at a specific timestamp (12, 30, 31) . The use of novel methodology, in addition to the robust size of our cohort, ensure that the identified trajectory based subphenotypes are less likely to suffer from cognitive bias (13) and are likely to be temporally stable (32) . More concretely, we adopted a divide and conquer approach to identify the subphenotypes. Prior research has identified that additive organ dysfunction is predictive of increased mortality in COVID-19 associated ARDS (8) . Therefore, we divided the patients into three different baseline strata (mild, intermediate and severe) according to additive All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . SOFA based organ dysfunction. We identified two salient trajectory subphenotypes within each stratum. Importantly, the baseline demographics, comorbidities and pattern of organ dysfunction did not differ between the worsening and recovering subphenotypes at each stratum. This suggests the existence of differential progression pathways that are irrespective of baseline risk factors for severe disease. This finding is unique compared to other subphenotyping projects as we are including a more complete picture of the disease course (12, 30, 31) . It also highlights the temporal heterogeneity of COVID-19 and the importance of avoiding prognostication based on early post intubation clinical characteristics. We found that the worsening subphenotypes in the baseline mild and intermediate strata showed an even higher risk of death compared to the recovering subphenotype within the baseline severe stratum (Figure 3) . Indeed, there is an urgent need to understand the pathophysiology of progressive non-pulmonary organ dysfunction in this disease. We assessed the differences between a broad range of laboratory tests, vital signs, and respiratory variables in the worsening and recovering subphenotypes. Importantly, basic laboratory tests and inflammatory markers were differentially associated with the worsening and recovering subphenotypes over time, which suggests that there is value in clinically following markers such as D-dimer, LDH, ferritin, procalcitonin and C-reactive protein. In the mild stratum, markers from the regular blood panel such as total white blood cell count and neutrophil counts, while inflammatory markers, such as ferritin and LDH, differentiate in the severe stratum. Laboratory tests, vitals and inflammatory markers in the intermediate stratum were less able to All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07. 16.20155382 doi: medRxiv preprint distinguish between the worsening and recovering subphenotypes. However, differences emerged over a longer time horizon (e.g., day 3). This further highlights the dynamic nature of COVID-19 and the difficulty in early prognosis in the critically ill population, despite severely deranged baseline organ dysfunction and inflammatory markers. We built multivariable prediction models for the identified trajectory Interestingly, over the course of the first 7 days following intubation, liver failure remained rare. At different points in the course, inflammatory markers such as creatine kinase and D-dimer predicted worsening and recovering subphenotypes. Our study was conducted on the two NYP system hospitals. Woresning and recovering SOFA subphenotypes, clinical characteristics, and outcomes from the validation cohort was consistent with the original subphenotypes. Although, due to the All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. Second, we did not use the progression of inflammatory markers such as Creactive protein, D-dimer or ferritin, which are known risk factors for this disease, to identify the subphenotypes. Nor did we stratify patients based the severity of respiratory failure alone. Instead, we chose to see how these factors interacted with traditional organ dysfunction, as most patients with COVID19 die from multisystem organ failure and not refractory respiratory failure (8, 9) . Third, differentiating trajectory subphenotypes in this critically ill population was difficult, as AUC-ROC metrics of prediction modeling using data at day 1 post-intubation were around 0.7. By restricting our analysis to a very high-risk population, we All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07.16.20155382 doi: medRxiv preprint decreased the discriminative power of many of our biomarkers to predict outcomes. All patients were high risk. However, we have added to our understanding of patients with critical COVID-19, by documenting the natural history of organ dysfunction in this population. Future research efforts, with novel biomarkers, are needed to predict worsening and recovering subphenotypes at an earlier time point in those with respiratory failure. Fourth, the surge conditions in New York City during the study period could affect the study. Care may have been influenced by the surge conditions during this difficult time. However, all patients were cared for in a critical care environment and despite the massive patient burden, the all cause 30-day mortality was 25.9%. In a population of critically ill patients with COVID-19 respiratory failure, there are distinct worsening and recovering organ dysfunction trajectory subphenotypes. Worsening status was predictive of poor outcomes in all strata regardless of baseline severity. These findings highlight the importance of supportive care for sequential organ failure in addition to respiratory failure in this disease. Trajectory based subphenotypes offer a potential road map for understanding the evolution of critical illness in COVID-19. We call for further analysis. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . https://doi.org/10.1101/2020.07.16.20155382 doi: medRxiv preprint l d s t r a t u m ( B a s e l i n e S O F A 0 -1 0 ) I n t e r m e d i a t e s t r a t u m ( B a s e l i n e S O F A 1 1 -1 2 ) S e v e r e s t r a t u m ( B a s e l i n e S O F A 1 3 -2 (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . M i l d s t r a t u m I n t e r m e d i a t e s t r a t u m S e v e r e s t r a t u (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 18, 2020. . COVID-19-associated acute respiratory distress syndrome: is a different approach to management warranted? Lung Recruitability in COVID-19-associated Acute Respiratory Distress Syndrome: A Single-Center Observational Study Subphenotyping ARDS in COVID-19 Patients: Consequences for Ventilator Management Associated Respiratory Failure Respiratory Pathophysiology of Mechanically Ventilated Patients with COVID-19: A Cohort Study Covid-19 in Critically Ill Patients in the Seattle Region -Case Series Epidemiology, clinical course, and outcomes of critically ill adults with COVID-19 in New York City: a prospective cohort study Investigators ftS-C. Factors Associated With Death in Critically Ill Patients With Coronavirus Disease 2019 in the US Risk Factors Associated With Mortality Among Patients With COVID-19 in Intensive Care Units in Clinical course and risk factors for mortality of adult inpatients with COVID-19 in Wuhan, China: a retrospective cohort study Development and validation of parsimonious algorithms to classify acute respiratory distress syndrome phenotypes: a secondary analysis of randomised controlled trials Subphenotypes in acute respiratory distress syndrome: latent class analysis of data from two randomised controlled trials The perils of premature phenotyping in COVID: a call for caution The SOFA (Sepsis-related Organ Failure Assessment) score to describe organ dysfunction/failure The early change of SOFA score as a prognostic marker of 28-day sepsis mortality: analysis through a derivation and a validation cohort Evaluation of SOFA-based models for predicting mortality in the ICU: A systematic review Serial Evaluation of the SOFA Score to Predict Outcome in Critically Ill Patients Deep phenotyping: Embracing complexity and temporality-Towards scalability, portability, and interoperability Dynamic Time Warping. Information Retrieval for Music and Motion Classification of surgical processes using dynamic time warping Data-Driven Subtyping of Parkinson's Disease Using Longitudinal Clinical Records: A Cohort Study Hospital Preparedness for COVID-19: A Practical Guide from a Critical Care Perspective Secondary Use of Patients' Electronic Records (SUPER): An Approach for Meeting Specific Data Needs of Clinical and Translational Researchers Clinical Characteristics of Covid-19 in New York City Comparison of the Spo2/Fio2 Ratio and the Pao2/Fio2 Ratio in Patients With Acute Lung Injury or ARDS Angiotensin II for the Treatment of Vasodilatory Shock Data clustering: a review CLUSTISZ: A Program to Test for the Quality of Clustering of a Set of Objects Derivation, Validation, and Potential Treatment Implications of Novel Clinical Phenotypes for Sepsis Phenotypic clusters within sepsis-associated multiple organ dysfunction syndrome Identification of pathophysiological patterns for triage and respiratory support in COVID-19 High risk of thrombosis in patients with severe SARS-CoV-2 infection: a multicenter prospective cohort study We thank the following Weill Cornell Medicine medical students for their contributions to (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.