Derivation of "specic population who could benet from Rosuvastatin": A secondary analysis on randomised controlled trial to uncover novel value of Rosuvastatin for precise treatment of ARDS

Background The high heterogeneity of ARDS contributes to paradoxical conclusions from previous investigations of rosuvastatin for ARDS. Identication of the population (phenotype) who could benet from rosuvastatin is a novel exploration for precise treatment of ARDS. Methods The patient population for this analysis consisted of unique patients with ARDS enrolled in the SAILS trial (Rosuvastatin vs. Placebo). Phenotypes were derived using consensus k means clustering applied to routinely available clinical variables within 6 hours of hospital presentation before receiving placebo or rosuvastatin. Kaplan–Meier statistic was used to estimate the 90 day cumulative mortality for screening specic population who could benet from rosuvastatin, with cut-off value as P <0.05. Results The derivation cohort included 585 patients with ARDS. Of the 4 derived phenotypes, phenotype 3 was identied as "specic population who could benet from rosuvastatin" since rosuvastatin resulted in a signicant reduction in 90 day cumulative mortality for ARDS (hazard ratio [HR] 0.29 [95% CI 0.09, 0.93]; P=0.027). Meanwhile, there were no signicant differences in baseline characteristics between those assigned to rosuvastatin and those assigned to placebo. Additionally, rosuvastatin markedly improved the free of cardiovascular failure (10.08±3.79 in Rosuvastatin group vs 7.31±4.94 in Placebo group, P=0.01) and coagulation abnormality (13.65±1.33 vs 12.15±3.77, P=0.02) to day 14 in phenotype 3. Patients classied as phenotype 3 exhibited but not limited to the relative higher platelet count (390.05±79.43×10^9/L), lower CRP (20.23±11.99μg/L) and Creat (1.42±1.08 mg/dl), compared with patients classied as other phenotypes. Besides that, rosuvastatin seemed to increase 90 day mortality for patients in phenotype 4 (HR 2.76[95% CI 0.09, 9.93], P=0.076), with its adverse effect on the reduction of free of renal failure to day 14(4.70±4.99 vs 10.17±4.69, P=0.01). Patients in phenotype 4 showed a relative severe illness baseline features particularly renal failure. Conclusions This secondary analysis of SAILS trial identied the specic population who can benet from rosuvastatin using machine learning applied to clinical variables at the time of hospital presentation, which uncovered a novel value of rosuvastatin for the treatment of ARDS, with validation clinical trials to be warranted to assess these further.


Introduction
Acute respiratory distress syndrome (ARDS) is a highly heterogeneous and complicated critical illness.
Despite advances in clinical management, the mortality of severe ARDS remain as high as 40-46%, due to the lack of targeted therapeutic protocols for distinct patients. To categorize ARDS for furthermore appropriate therapy is the critical unmet need for precise treatment and improvement for salvage rate of ARDS. [1][2] In consideration of rosuvastatin's anti-in ammatory effects and pathogenesis of ARDS(inadequate control of in ammatory responses in lung), rosuvastatin was attempted to be utilized in the treatment for ARDS in the last decade. [3][4][5][6][7] The previous studied demonstrated that rosuvastatin could improve outcomes of ARDS in animal models. [8][9][10] Unfortunately, a large multicenter randomized controlled trail (RCT) in 2014 conducted by Jonathon et al. (named SAILS trial) suggested that rosuvastatin therapy did not improve clinical outcomes in patients with ARDS. [11] The cordial reason to these paradoxical conclusions is the heterogeneity of ARDS. ARDS, as an overly broad de nition of syndrome, encompasses a vast, multidimensional array of clinical and biological features. Markedly differed from experimental animals, patients with ARDS actually consist of diverse phenotypes, which appear different clinical characteristics, immune status, biological processes and severity. Several investigations successfully classi ed ARDS to distinct subgroups via biomarkers or clinical features [12][13], which indicated that appropriate therapies for distinct patients may be the promising strategy for precise treatment in ARDS. Rosuvastatin, as an immunomodulatory intervention to attenuate in ammation, may just bene ts some speci c population.
Obviously, there is a robust need to explore the novel ARDS phenotype (who could bene t from rosuvastatin) to optimize therapeutic strategy for ARDS and reduce mortality furthermore. Fortunately, the Jonathon et al. had uploaded the original data of SAILS trial to ARDS-Net database, made it possible for us to perform secondary analysis to nd speci c population who could bene t from rosuvastatin. Thus, we aimed to derive ARDS phenotype by using unsupervised clustering algorithm, in order to uncover novel value of Rosuvastatin for precise treatment of ARDS.

Method
This study was reviewed and approved by Institutional Ethics Committee of Zhongda Hospital.
Institutional Ethics Committee of Zhongda Hospital and conducted under several data use agreements.
The Ethical approval was shown in e- Figure 1. The data for the ARDSnet project were obtained under a waiver of informed consent and with authorization under the Health Insurance Portability and Accountability Act.

Patient Population
The patient population for this analysis consisted of unique patients with ARDS enrolled in the SAILS trial (Rosuvastatin vs. Placebo), which was published in 2014. The diagnostic criterion of ARDS in the SAILS trial was referenced to the Berlin de nition of ARDS in 2012 [1][2]. To eliminate in uence of immunosuppression on evaluation of rosuvastatin for ARDS, the patients were distinguished into 160 de nitely immunosuppressed patients and 585 other patients for respective analysis. The de nitely immunosuppressed patients included ARDS patients with the comorbidity of Acquired immune de ciency syndrome, Leukemia, Non-Hodgekins Lymphoma, cancer receiving chemotherapy, or patients receiving therapy of immune suppression in past 6 months. After excluding 160 de nitely immunosuppressed patients, 585 other patients were enrolled in the derivation cohort for further unsupervised clustering analysis.

Screen Clinical Features for Phenotyping
Based on the database of SAILS trial, we rstly extracted the available variables within the rst 6 hours of hospital presentation before receiving placebo or rosuvastatin, and excluded variables with missing rate greater than 10%. These clinical available characteristics included age, alanine aminotransferase, APACHE III score, aspartate aminotransferase, blood urea nitrogen, C-reactive protein, creatine kinase, creatinine, diastolic BP, Glasgow Coma Scale, height, heart rate, male sex, Paco2, Pao2:Fio2 mmHg, Pao2, platelet count, predicted body weight, respiration rate, serum albumin highest, serum albumin lowest, serum glucose lowest, shock at baseline, systolic BP, temperature, urineout and weight.
Furthermore, to screen candidate variables for identi cation of "speci c population who can bene t from rosuvastatin ", we conducted differential analysis by using t-test on clinical available variables between Rosuvastatin group and placebo group from survival patients, P <0.3 as the threshold value.

Statistical Methods
To derive the phenotypes, we rst assessed the candidate variable distributions, missingness, and correlation. Multiple imputation with chained equations was used to account for missing data. [14] In order to identify different phenotypes of ARDS, the consensus k means clustering through candidate variables was utilized to perform consistent clustering on 585 patients in derivation cohort. [15] The clustering was performed using 100 iterations, with each iteration containing 80% of samples. The optimal clustering strategy was determined by cumulative distribution function curves of the consensus score, clear separation of the consensus matrix heatmaps, characteristics of the consensus cumulative tribution function plots, and adequate pair wise-consensus values between cluster members.
To evaluate the effect of rosuvastatin for outcomes of ARDS in different subgroups, Kaplan-Meier statistic was used to estimate the 90 day mortality. Organ failure free days to day 14(day), free of cardiovascular failure to day14 (day), free of coagulation abnormality to day14 (day), free of hepatic failure to day 14(day), free of renal failure to day 14(day), ICU free days to day 28(day) and ventilator free days to day 28 were analyzed by means of analysis of variance. 28 day mortality, 60 day mortality and 90 day mortality were analyzed by chi-square test. P value less than 0.05 was set as threshold value to screen signi cant results.
To observe the clinical features variations in different phenotypes, means of analysis of variance and chisquare test were utilized to assess the continuous variables and dichotomous variable respectively, with cut-off value as P <0.05.
The brief analysis ow plots were illuminated in e- Figure 2.
Software and versions R x64 3.6.1 was conducted to process data, analyze data and plot diagrams.

Patients
A total of 745 patients met ARDS criteria were enrolled in nal analysis, with 379 patients in Rosuvastatin group and 366 patients in Placebo group. The age of patients investigated ranged from 18 to 89 (median 54) with 51% male. The mean Pao2:Fio2 level was 143.48mmHg (standard deviation [SD], 63.57mmHg) and the mean APACHE III score was 93.42 (SD, 20.15mmHg). Detailed baseline demographic and clinical characteristics were shown in supplemental electrical data- Table 1 (e-Table 1) and e- Table 2.
Derive ARDS phenotypes and identify speci c population who can bene t from rosuvastatin After differential analysis on clinical available variables, we nally screened serum glucose highest, Creactive protein and platelet count as candidate variables for further unsupervised clustering analysis, shown in e- Table 3.
After excluding 160 de nitely immunosuppressed patients, 585 patients were enrolled in derivation cohort. The consensus k means clustering models suggested that a 4-class model was the optimal t with the 4 phenotypes, since the clearest separation of the consensus matrix heatmap could be found in the 4-class model, shown in Figure 1.
According to Kaplan-Meier statistic analysis, phenotype 3 was identi ed as "speci c population who can bene t from rosuvastatin", shown in Meanwhile, there were no signi cant differences in baseline characteristics between those assigned to rosuvastatin and those assigned to placebo in phenotype 3 cohort. Baseline characteristics of patients in derived 4 phenotypes were shown in e- Table 4-7.
In phenotype 3 cohort, free of cardiovascular failure and coagulation abnormality to day 14 differed signi cantly between the two groups. Additionally, rosuvastatin resulted in a slight raise in ventilator free days to day 28 for ARDS. There were no signi cant between-group differences in any of the other outcomes. The above results were illuminated in Table 1.
For a better insight in patients who could bene t from rosuvastatin, we compared the clinical characteristics among different phenotypes. Patients classi ed as phenotype 3 appeared highest Platelet count (390.05±79.43×10^9/L) lowest CRP (20.23±11.99μg/L ) and lowest Creat (1.42±1.08 mg/dl).
Additionally, other distinct clinical characteristics in different phenotypes were described in Table 2. Indeed, phenotype 3 could be identi ed through our 4-class model.

Characteristics and outcomes in other phenotypes
Kaplan-Meier survival analysis indicated that rosuvastatin had no effect for ARDS in cohorts of other phenotypes. In phenotype 2 cohort, rosuvastatin appeared a slight reduction in free of hepatic failure to day 14. In addition, rosuvastatin presented a moderate reduction in free of renal failure to day 14 in phenotype 4 cohort. More details of characteristics and outcomes in other phenotypes were described in Table 1

Disscussion
In this secondary analysis of SAILS trial, 4 phenotypes of ARDS were derived through routinely available clinical variables at the time of hospital presentation. These phenotypes were multidimensional, heterogeneous in their demographics, clinical characteristics, several laboratory abnormalities, effect of rosuvastatin therapy, and differed with traditional patient classi cations such as direct or indirect lung injury, patterns of organ dysfunction, or severity of ARDS. In phenotype 3, rosuvastatin began to exhibit bene ts for patients of ARDS, compared with placebo. This conclusion highlighted the importance of characterizing the heterogeneity of ARDS and early goal-directed therapy.
To our knowledge, the current study rstly identi ed the speci c population who can bene t from rosuvastatin, which could improve therapeutic system in ARDS to reduce mortality furthermore, with validation clinical trials to be warranted to assess these further. These patients exhibited but not limited to the relative higher platelet count (390.05 ± 79.43 × 10^9/L), lower CRP (20.23 ± 11.99 µg/L ) and Creat (1.42 ± 1.08 mg/dl), compared with other patients of ARDS. These patients probably suffered from a relative slight infection, and might bene t from rosuvastatin for its anti-in ammatory effect could restore cardiovascular failure rapidly. Indeed, current study indicated that rosuvastatin resulted in an obvious improvement in free of cardiovascular failure to day14, (7.31 ± 4.94 in Placebo vs 10.08 ± 3.79 in rosuvastatin, P = 0.01). Phenotype 3 could be identi ed through our machine learning constructed 4-class model rapidly. This model could be utilized to identify speci c population who can bene t from rosuvastatin at the time of patient presentation to the emergency department, and thus could be useful with regard to early treatment and enrollment in clinical trials. Only routinely available data were used in the clustering models, and the phenotypes were derived from a large observational cohort to ensure generalizability.
Rosuvastatin may improve in ammatory responses possibly via modulation on platelet-dependent mechanism, which might be potential treatment pathogenesis of rosuvastatin on the novel phenotype for ARDS. It was well known that platelet plays an important role in neutrophil mediated lung injury. [16][17] The present study indicated that patients classi ed as phenotype 3 exhibited relative high platelet counts. Additionally, in these patients, rosuvastatin signi cantly improved coagulation abnormality of ARDS, compared with placebo. Therefore, we hypothesized that platelets might be involved in the pharmacological mechanism of Rosuvastatin on speci c patients of ARDS, with validation experiments to be warranted to assess these related mechanism.
Rosuvastatin might be harmful for patients with de nite immunosuppression. Rosuvastatin was attempted to be utilized in patients with ARDS mainly due to rosuvastatin's anti-in ammatory effects.
However, the infection is the main risk factor of ARDS, and it has been veri ed that was patients with immunosuppression had worse outcomes since their weak immunity hardly eliminated pathogens. [18][19] Therefore, as immunosuppressed effect of rosuvastatin, it could not bene t such patients. This study similarly exhibited a trend that patients with de nite immunosuppression had a worse outcome when receiving rosuvastatin probably, shown in Fig. 1A.
Rosuvastatin seems to be harmful for patients classi ed as phenotype 4. Survival curves of phenotype 4 illuminated a trend that rosuvastatin resulted in a reduction in 90 day survival rate of ARDS, despite of the less rigorous con dence interval ( There are several limitations to the present study. Indeed, current analysis on treatment × phenotype interactions is largely limited by sample size. Therefore, these novel proof-of-concept ARDS phenotypes could be incorporated prospectively in future study designs that subsequently validate effect of rosuvastatin for ARDS.
[20] In addition, for the limitation of clinical correlation analysis, the further basic experiments should be conducted to sequentially research elaborate mechanisms of rosuvastatin in ARDS indicated by our analyses.

Conclusion
In this secondary analysis of SAILS trial from patients with ARDS, speci c population who can bene t from rosuvastatin was identi ed through routinely available clinical variables at the time of hospital presentation, which uncovered a novel value of Rosuvastatin for the treatment of ARDS, with validation clinical trials to be warranted to assess these further.

Declarations
Ethical Approval and Consent to participate Not available.

Consent for publication
Not available.
Availability of data and materials Not available.

Competing interests
The authors declare that they have no competing interests.    Figure 1 The consensus matrix heatmaps of consensus k means clustering. Figure 1A showed the sample distribution of 4 phenotypes after consensus k means clustering. Figure 1B-I illuminated consensus matrix heatmaps of different subgroup number (k = 2, 3, 4, 5, 6,7,8,9), respectively. It can be found that when k=4, the model exhibited the clearest separation of the consensus matrix heatmap.