American Journal of Emergency Medicine 46 (2021) 38-44

Contents lists available at ScienceDirect

American Journal of Emergency Medicine

journal homepage:

Development and validation of a model for the Early prediction of the RRT requirement in patients with rhabdomyolysis

Chao Liu, MD ^a,b,1, Qian Yuan, MD ^c,1, Zhi Mao, MD, PhD ^d, Pan Hu, MD ^e, Rilige Wu ^f, Xiaoli Liu, MD ^g,

Quan Hong, MD ^a,b, Kun Chi, MM ^a,b, Xiaodong Geng, MD ^a,b, Xuefeng Sun, PhD ^b,?

^a Medical School of Chinese PLA, 28 Fuxing Road, Beijing, China

^b Department of Nephrology, Chinese PLA General Hospital, Chinese PLA Institute of Nephrology, State Key Laboratory of Kidney Diseases, National Clinical Research Center for Kidney Diseases, 28 Fuxing Road, Beijing, China

^c Beijing Xiaomi Mobile Software Co., Ltd., China

^d Department of Critical Care Medicine, Chinese PLA General Hospital, Beijing 100853, China

^e Department of Anesthesiology, The 920 Hospital of Joint Logistic Support Force of Chinese PLA, 650032 Kunming, Yunnan, China.

^f Medical Big Data Research Center, Chinese PLA General Hospital, Beijing, China.

^g School of Biological Science and Medical Engineering, Beihang University, 100191 Beijing, China.

a r t i c l e i n f o

Article history:

Received 31 December 2020

Received in revised form 16 February 2021 Accepted 3 March 2021

Keywords: Rhabdomyolysis Model

Renal replacement therapy eICU-CRD

MIMIC-III

a b s t r a c t

Introduction: Rhabdomyolysis (RM) is a complex set of clinical syndromes involving the rapid dissolution of skel- etal muscles. The early detection of patients who need Renal replacement therapy is very important and may aid in delivering proper care and optimizing the use of limited resources.

Methods: Retrospective analyses of the following three databases were performed: the eICU Collaborative Re- search Database (eICU-CRD), the Medical Information Mart for Intensive Care III (MIMIC-III) database and elec- tronic medical records from the First Medical Centre of the Chinese People’s Liberation Army General Hospital (PLAGH). The data from the eICU-CRD and MIMIC-III datasets were merged to form the derivation cohort. The data collected from the Chinese PLAGH were used for external validation. The factors predictive of the need for RRT were selected using a LASSO regression analysis. A logistic regression was selected as the algorithm. The model was built in Python using the ML library scikit-learn. The accuracy of the model was measured by the Area Under the Receiver Operating Characteristic Curve . R software was used for the LASSO regression anal- ysis, nomogram, concordance index, calibration, and decision and clinical impact curves.

Results: In total, 1259 patients with RM (614 patients from eICU-CRD, 324 patients from the MIMIC-III database and 321 patients from the Chinese PLAGH) were eligible for this analysis. The rate of RRT was 15.0% (92/614) in the eICU-CRD database, 17.6% (57/324) in the MIMIC-III database and 5.6% in the Chinese PLAGH (18/321). After the LASSO regression selection, eight variables were included in the RRT prediction model. The AUC of the model in the training dataset was 0.818 (95% CI 0.78-0.87), the AUC in the test dataset was 0.794 (95% CI 0.72-0.86), and the AUC in the Chinese PLAGH dataset (external validation dataset) was 0.820 (95% CI 0.70-0.86).

Conclusions: We developed and validated a model for the early prediction of the RRT requirement among patients with RM based on 8 variables commonly measured during the first 24 h after admission. Predicting the need for RRT could help ensure appropriate treatment and facilitate the optimization of the use of medical resources.

(C) 2021

1. Introduction

Abbreviations: APS III, acute physiology score III; AUC, area under the curve; MIMIC-III, Medical Information Mart for Intensive Care III; RM, rhabdomyolysis; SOFA, Sequential Organ Failure Assessment..

* Corresponding author.

E-mail addresses: [email protected] (Q. Yuan), [email protected] (X. Liu), [email protected] (X. Sun).

¹ Contributed equally.

Rhabdomyolysis (RM) is a complex set of clinical syndromes involv- ing the rapid dissolution of skeletal muscles [1,2]. RM was first de- scribed in the German medical literature in 1881 [3]. The common causes of RM include trauma, alcohol and drug abuse, muscle ischemia, infections, electrolyte and Metabolic disorders, Genetic disorders, fa- tigue, long-term immobilization and sudden changes in body tempera- ture [2,4,5]. These factors cause the breakdown of skeletal muscle fibers and the leakage of toxic cellular contents and myoglobin into the sys- temic circulation, which, in turn, causes a series of physiological and

https://doi.org/10.1016/j.ajem.2021.03.006

0735-6757/(C) 2021

biochemical disorders in the human body, manifesting as organ damage in other systems [6-8]. The prevalence of Acute kidney injury fol- lowing RM ranges from 13% to 50% [9]. The kidneys need adequate per- fusion pressure and fluid volume to eliminate the toxins [10].

Although RRT clears myoglobin from the bloodstream, thereby po-

tentially decreasing the amount of Renal damage, it involves the use of an extracorporeal circuit with many potential complications [2,11]. The initiation of RRT in clinical practice should not be based on the myoglobin or creatine kinase (CK) serum concentration but rather the degree of Renal impairment [2]. It is necessary to consider all co- occurring factors in RM patients and individualize the treatment plan. Therefore, the use of more variables to support the early detection of RM patients who need RRT is very important and may aid in delivering proper care and optimizing the use of limited resources.

In this study, we constructed a model based on the eICU Collabora- tive Research Database (eICU-CRD) [12], the Medical Information Mart for Intensive Care III (MIMIC III) [13] dataset and data collected from the First Medical Centre of the Chinese People’s Liberation Army Gen- eral Hospital (PLAGH) to identify patients who need RRT at the time of hospital admission.

1. Methods

We performed a longitudinal, multicenter, retrospective study based on the eICU-CRD (v2.0) [12], MIMIC-III (v1.4) [13,14] and PLAGH

databases.

• 1. Data sources

The eICU-CRD (v 2.0) is a multicenter, telehealth ICU database, and this freely available dataset includes data from more than 200,000 ICU patients at 208 ICUs in the United States from 2014 to 2015. The MIMIC-III (v1.4) database is a large, open-access, single-center dataset that includes data from more than 50,000 ICU patients at Beth Israel Deaconess Medical Center (Boston, MA, USA) between 2001 and 2012. The source hospital for the MIMIC-III database does not participate in the eICU program. To obtain permission to access the database, re- searchers must complete the National Institutes of Health’s web-based course called Protecting Human research participants (certification number 29493483). This study involved an analysis of third-party, anonymized, publicly available data from the eICU-CRD and MIMIC-III database with pre-existing institutional review board (IRB) approval. The PLAGH is a tertiary hospital integrating high levels of education and research in China. The data collected from the PLAGH were used for external validation.

• 1. Study population

All Adult ICU patients (age >= 18 years) diagnosed with RM according to the International Classification of Diseases (ICD-9) codes were con- sidered. The exclusion criteria were (1) a peak CK level less than 1000 UL/L; (2) a hospital stay less than 2 days; (3) patients with unknown outcomes; and (4) outliers based on professional knowledge, whether errors existed in the data that were obviously inconsistent with the ac- tual situation, and whether the value exceeded the theoretical range (exceeded the mean +- 3 standard deviations (SDs)) [15].

• 1. Data extraction

The patient data were extracted from the eICU-CRD and MIMIC-III databases using PostgreSQL Version 9.6. The occurrence of AKI was de- termined based on the classification of Kidney Disease: Improving Global Outcomes (KDIGO) guidelines [16]. The baseline characteristics were extracted within the first 24 h after patient admission.

The demographic parameters included age, sex, and ethnicity. The clinical parameters included heart rate (HR), mean arterial pressure

(MAP), respiratory rate (RR), temperature, and percutaneous oxygen saturation (SPO₂). The laboratory parameters included CK, peak CK, serum creatinine (Cr), Blood urea nitrogen , alanine transami- nase (ALT), aspartate aminotransferase , serum bicarbonate, serum chloride, serum potassium, serum sodium, hematocrit, hemo- globin, phosphate, lactate, red cell distribution width (RDW), albu- min, calcium, platelet count, and white blood cell count. The comorbidities included Acute Kidney Injury , acute hepatic fail- ure (AHF) and atrial fibrillation. The disease severity scores were the acute physiology score III (APS III) [17] and Sequential Organ Failure Assessment score [18]. The outcomes included the rate of RRT and in-hospital mortality. The etiology of RM was also collected and classified according to the methods described in the relevant liter- ature [1,6,19].

• 1. Statistical analysis

All calculations and analyses were performed using Python version

3.7.1 and R version 3.6.3. The continuous variables are presented as the means +- SDs or medians and interquartile ranges (IQRs). The cate- gorical variables are represented as the total number and percentage. The comparisons between the groups were performed using the Kruskal-Wallis test for the continuous variables or the chi-square test for the categorical variables and an analysis of variance. A two-tailed P-value <0.05 was considered statistically significant.

The data from the eICU-CRD and MIMIC-III datasets were merged for

further analysis. Variables with >40% missing values were excluded from further analysis, and the median of the overall population was used to interpolate the remaining missing data. The study cohort was randomly divided into the following two parts: 70% of the data were used for the model training, and 30% of the data were used for the model testing. The data collected from the Chinese PLAGH were used as an external validation dataset.

A LASSO regression analysis was used to select the variables that were predictive of the need for RRT. The model was built in Python using the ML library scikit-learn [20]. A logistic regression was selected as the algorithm. A propensity score matching (PSM) analysis was per- formed using Python. R software was used for the LASSO regression analysis, nomogram, concordance index, calibration, and decision and clinical impact curves.

1. Results
   1. Baseline characteristics of the included patients

The screening process is shown in Fig. 1. In total, 1259 patients with RM (614 patients from eICU-CRD, 324 patients from the MIMIC-III data- base and 321 patients from the PLAGH) were eligible for this analysis. We merged the data from the eICU-CRD and MIMIC-III datasets into one dataset, and this dataset was randomly divided into the following two parts: 70% of the data were used for the model training, and 30% of the data were used for the model testing (Table 1). Meanwhile, the data collected from the Chinese PLAGH were used as an external valida- tion dataset (Table 1). The RRT requirement rate in the training dataset was 16% (107/656), that in the testing dataset was 14.9% (42/282) and that in the external validation dataset was 5.6% (18/321).

The data from the eICU-CRD,MIMIC-III and PLAGH datasets are

shown in Table S1. Ninety-two patients (15.0%) in the eICU-CRD

dataset, 57 patients (17.6%) in the MIMIC dataset and 18 patients (5.6%) in the PLAGH dataset needed RRT. The in-hospital mortality rate was 10.1% (62/614) in the eICU-CRD dataset, 10.5% (34/324) in the MIMIC-III dataset and 3.7% (12/321) in the PLAGH dataset (Table S1). The distribution of the etiologies of RM in the eICU-CRD and MIMIC-III datasets are shown in Fig. S1. RM was most commonly observed in the following settings: trauma (188), metabolic and electro- lyte disorders (149), infections (114), alcohol abuse (87), myopathy

(80), seizures or convulsions (71), heat or dehydration (62) and hypothyroidism (43).

eICU-CRD (n=881); MIMIC-III (n=388); PLAGH (n=350)

Excluded age < 18 years

Patients diagnosed as RM according ICD-9 codes (n=1629): eICU-CRD (n=885); MIMIC-III (n=390); PLAGH (n=354)

eICU-CRD (n=675); MIMIC-III (n=351); PLAGH (n=337)

Excluded peak CK < 1000 UL/L

eICU-CRD (n=624); MIMIC-III (n=335) ); PLAGH (n=330)

Excluded hospital stay < 2d

Excluded outliers and unknown outcome

eICU-CRD (n=614); MIMIC-III (n=324); PLAGH (n=321)

ALL (n=1259)

Fig. 1. Flow chart of the patient Selection process. ICD-9, International Classification of Diseases, Ninth Revision; eICU-CRD, eICU Collaborative Research Database; RM, rhabdomyolysis; MIMIC-III, Medical Information Mart for Intensive Care III; CK, creatine kinase. Outliers: Based on professional knowledge, whether errors existed in the data that are obviously not consistent with the actual situation, and whether the value exceeded the theoretical range (exceeded the mean +- 3 standard deviations (SDs)); PLAGH, Chinese People’s Liberation Army General Hospital.

• 1. Model building and evaluation

Twenty-four variables measured at admission were included in the LASSO regression. The detail of variables that included in the LASSO re- gression are described in supplementary file 1. After the LASSO regres- sion selection (Fig. 2), eight variables were found to be predictors of the need for RRT (Table S2). These variables were age, Cr, CK, AST, albu- min, calcium, phosphate and atrial fibrillation (yes vs. no). The 8 vari- ables were used to build the model by a logistic regression in the training dataset. The area under the receiver operating characteristic curve of the model in the training dataset was 0.818 (95% CI 0.78-0.87), and the AUC in the testing dataset was 0.794 (95% CI 0.72-0.86). The AUC in the external validation cohort was 0.820 (95% CI 0.70-0.86) (Fig. 3). This model (AUC = 0.818) performed better than Cr (AUC = 0.747) and CK (AUC = 0.667) alone in predicting the need for RRT (Fig. S2). The correlations between the included variables are shown in Fig. S3. In addition, we conducted an analysis to determine whether fewer variables (6 or7) were predictive (Table S3). When we excluded one or two variables, the model’s Predictive performance de- creased, but for some variables, the degree of model performance deg- radation was not high. Whether such variables can be excluded requires further verification using an external validation cohort with a larger sample size.

Table 1

Baseline characteristics of the three cohort.

Training dataset (n = 656) Testing dataset (n = 282) External validation dataset (n = 321)

Non-RRT (n = 549)

RRT (n = 107)

P

Non-RRT (n = 240)

RRT (n = 42)

P

Non-RRT (n = 303)

RRT (n = 18)

P

Clinical parameters Age (years)

57.0 [44.0,69.0]

51.0 [40.7,60.0]

0.001

59.0 [45.0,73.0]

51.0 [40.0,62.0]

0.015

28.0 [19.0,46.0]

41.0 [35.0,47.0]

0.012

Male, n (%)

356 (64.8)

75 (70.1)

0.350

160 (66.7)

29 (69.0)

0.901

251(82.8)

13 (77.8)

0.467

Vital signs

MAP (mmHg)

82.0 [74.4,88.0]

78.3 [72.0,85.0]

0.031

82.0 [72.0,88.0]

82.0 [76.0,91.0]

0.126

91.0 [83.0,99.0]

94.0 [90.0,110.0]

0.062

Heart rate (beats/min) Respiratory rate

93.0 [82.0,101.0] 20.0 [17.0,22.0]

96.0 [88.0,108.0] 20.0 [18.0,24.0]

<0.001 0.232

93.0 [84.0,102.0] 20.0 [18.0,23.0]

93.0 [85.0,103.0] 20.0 [18.0,22.0]

0.920 0.845

79.0 [72.0,89.0] 18.0 [18.0,19.0]

76.0 [72.0,87.0] 18.0 [18.0,19.0]

0.579 0.763

(beats/min) Temperature (?C)

37.0 [36.7,37.3]

37.0 [36.6,37.3]

0.973

37.0 [36.6,37.4]

37.0 [36.5,37.4]

0.817

36.6 [36.4,37.0]

36.8 [36.3,37.0]

0.716

Laboratory parameters

Creatinine (mg/dL)

1.5 [1.0,2.6]

3.5 [1.8,5.9]

<0.001

1.5 [0.9,2.9]

3.0 [1.9,5.9]

<0.001

1.0 [0.8,1.7]

4.9 [3.0,8.8]

<0.001

Creatine kinase (U/L)

4126.0

8794.0

<0.001

3625.0

13,325.8

<0.001

3511.0

10,691.0

0.008

[1742.0,10,576.0]

[3709.0,30,683.0]

[1624.0,10,417.0]

[3616.0,28,656.0]

[1486.0,11,653.0]

[4480.127641.0]

BUN (mg/dl)

27.0 [16.0,48.0]

38.0 [24.5,63.5]

<0.001

27.0 [16.0,51.0]

39.5 [22.2,55.8]

0.048

12.1 [7.3,24.8]

44.8 [35.6,68.9]

<0.001

ALT (U/L)

82.0 [41.0,294.6]

205.0 [74.0,472.0]

<0.001

72.0 [36.0,200.0]

141.0 [52.5342.8]

0.013

66.7 [29.1183.4]

39.8 [24.6125.5]

0.300

AST (U/L)

194.0 [85.0,589.8]

437.0

<0.001

173.0 [75.8541.5]

376.5 [98.8856.8]

0.015

122.8 [45.5477.1]

60.1 [26.2385.9]

0.104

[145.51273.0]

Albumin (g/dL)

3.2 [2.9,3.7]

3.1 [2.5,3.3]

<0.001

3.2 [2.8,3.6]

3.2 [2.6,3.2]

0.030

3.7 [3.2,4.1]

2.8 [2.5,3.4]

<0.001

Calcium (mg/dL)

8.3 [7.5,8.9]

7.5 [6.7,8.4]

<0.001

8.4 [7.6,9.0]

7.9 [6.7,8.8]

0.031

8.5 [7.8,9.0]

7.7 [7.0,8.1]

0.003

Phosphate (mg/dL)

3.9 [2.6,4.4]

5.4 [3.5,7.9]

<0.001

3.8 [2.7,4.2]

4.6 [3.7,6.9]

<0.001

3.3 [2.7,3.9]

4.4 [3.2,5.8]

0.007

Comorbidities, n (%) Acute kidney injury

368 (67.0)

100 (93.5)

0.031

162 (67.5)

36 (85.7)

0.028

67 (22.1)

3 (16.7)

0.772

Atrial fibrillation

56 (10.2)

18 (16.8)

0.070

29 (12.1)

5 (11.9)

0.823

4 (1.3)

1 (5.6)

0.224

Acute hepatic failure

12 (2.2)

4 (3.7)

0.312

8 (3.3)

2 (4.8)

0.648

20 (6.6)

4 (22.2)

0.127

Severity of illness APS III

57.0 [42.0,72.0]

76.0 [63.0,95.0]

<0.001

58.0 [41.0,74.0]

71.0 [59.0,87.0]

<0.001

-

-

-

SOFA

6.0 [3.0,8.0]

9.0 [7.0,12.0]

<0.001

5.0 [3.0,8.0]

10.0 [7.0,12.0]

<0.001

-

-

-

Outcomes, n (%) Hospital mortality, n

50 (9.1)

20 (18.7)

0.006

20 (8.3)

6 (14.3)

0.245

11 (3.6)

1 (5.6)

0.506

(%)

ALT, alanine aminotransferase; AST, aspertate aminotransferase; BUN, blood urea nitrogen; MAP, mean arterial pressure; RRT: renal replacement therapy.

A B

24 23 23 23 22 22 21 19 17 16 11 8 6 6 4 3 0

Binomial Deviance

0.80

0.85

0.90

Coefficients

0.0

8 8 8 8 6 0

0.2

0.4

1

6

3

42

8

7

5

Fig. 2. Variable selection using the least absolute shrinkage and selection operator (LASSO) binary logistic regression model. (A) LASSO coefficient profiles of the 24 baseline features.

0.70

0.75

-0.4

-0.2

-8	-7	-6 -5	-4	-3	-2	-7	-6	-5	-4	-3	-2
		Log???						Log Lambda

(B) Tuning parameter (?) selection in the LASSO model used 5-fold cross-validation via minimum criteria variable selection.

• 1. Nomogram establishment

The nomogram was established based on the 8 variable values, which were used to predict the RRT requirement of RM patients (Fig. 4). The nomogram had a concordance index (c-index) of 0.818 (95% CI 0.76-0.85) for predicting the RRT requirement in the cohort. The calibration curves showed that the predicted rates were consistent with the actual results observed in the cohorts (Fig. S4). The decision curve and clinical impact curve showed that the nomogram had a supe- rior standardized net benefit over Cr and CK in terms of the patients’ outcome (Fig. 5).

• 1. PSM analysis

In our study, the in-hospital mortality rate in the RRT group (17.4%, 26/149) was higher than that in the non-RRT group (8.9%, 70/789). The patients in the RRT group had more severe disease than those in the non-RRT group (SOFA, 9.0 [7.0, 12.0] vs. 6.0 [3.0, 8.0], P < 0.001) (APS

III, 76.0 [61.8, 94.2] vs. 55.0 [41.0, 74.0], P < 0.001). Therefore, we per-

formed a PSM analysis (Table S4). After the PSM analysis, the SOFA score (9.0 [7.0, 12.0] vs. 9.0 [6.0, 12.0], P = 0.592) and APS III score

(75.0 [62.8, 92.0] vs. 71.0 [57.0, 92.0], P = 0.301) of the patients in the RRT group were similar to those of the patients in the non-RRT group. The in-hospital mortality rate in the RRT group was 17.4% (26/149),

and that in the non-RRT group was 16.8% (25/149) (P = 0.995). How- ever, the patients in the RRT group had more complications, such as atrial fibrillation and AKI (Table S4).

• 1. Risk factors related to mortality in the RRT group

The ordinary least squares (OLS) analysis showed that atrial fibrilla- tion and a lower Hematocrit level were risk factors related to mortality among the RM patients with RRT (Table S5).

1. Discussion

In this study, we developed and validated a clinical model for the prediction of the need for RRT among RM patients. The performance of this model was satisfactory based on the AUCs. The 8 variables re- quired for the prediction of the need for RRT are generally readily avail- able at hospital admission. If the model indicates that a patient will need RRT, more aggressive treatment can be initiated.

RM can be caused by many factors and is a common, life-threatening condition that develops alone or in combination with an underlying acute condition [21,22]. A Multicenter retrospective study enrolled 387 patients with severe RM (CK > 5000 U/L), and 103 patients (26.6%) needed RRT [21]. An observational study investigating RM in the ICU showed that the RRT rate was 17.0% (58/342) among patients

Fig. 3. Predictive performance evaluation. Model discrimination was assessed by the AUC. The AUC of the model in the training dataset (A) is 0.818 (95% CI 0.78-0.87), the AUC derived from testing the dataset (B) is 0.794 (95% CI 0.72-0.86), and the AUC derived from the external validation cohort (C) is 0.820 (95% CI 0.70-0.86). AUC, area under the receiver operating characteristics curve.

Points Creatinine Creatine kinase Phosphate AST

Albumin Calcium Age

AF (no or yes) Total Points

RRT requirement

0 10 20 30 40 50 60 70 80 90 100

0 2 4 6 8 10 12 14 16 18 20 22 24

0 50000 150000 250000 350000

0 2 4 6 8 10 12 14 16 18

0 2000 6000 10000 14000 18000

5	4.5 4	3.5 3	2.5 2 1.5 1
18	14	10 8	6 4 2

90 75 60 45 30 15

1

0

0 20 40 60 80 100 120 140 160 180 200 220 240

0.01 0.1 0.3 0.5 0.70.8 0.9 0.95

Fig. 4. Nomogram predicting RRT in patients with RM. AF, atrial fibrillation; AST, aspartate aminotransferase; RM, rhabdomyolysis; RRT, renal replacement therapy.

with CK levels >1000 U/L [23]. In our study, the RRT rate was 15.9% (149/938). Therefore, establishing a model to identify patients who will need RRT is important.

Large studies specifically dedicated to severe RM are rare. In this study, we extracted data from the eICU-CRD and MIMIC-III databases and merged the data into one dataset for further analysis. In total, 938 patients with RM (614 patients from the eICU-CRD and 324 patients from the MIMIC-III database) were eligible for this analysis. In addition, we collected data from the PLAGH as an external validation dataset. Eight variables (age, Cr, CK, AST, albumin, calcium, phosphate and atrial fibrillation (yes vs no)) were used to establish the model. The AUC was 0.818. We further established a nomogram to clearly visualize the pre- diction model and used a decision curve analysis [24] to assess the po- tential clinical impact of this model. Furthermore, we validated our model in the dataset from the PLAGH, and the AUC was 0.820. The pa- tients in the PLAGH are all Asian. The rate of RRT in the PLAGH was 5.6%, which may indicate that this model is generalizable to other areas worldwide.

A similar study (2371 patients in the derivation cohort and 1397 pa- tients in the validation cohort) reported the development of a risk pre- diction score (McMahon Score) for kidney failure or death in RM patients (age >= 18 years and CK > 5000 U/L) [25]. The independent pre- dictors of the Composite outcome were age, a female sex, the etiology of rhabdomyolysis, and the initial creatinine, Creatine phosphokinase, phosphate, calcium, and bicarbonate values. The C statistic of the predic- tion model was 0.82 (95% CI, 0.80-0.85) in the derivation cohort and

0.83 (0.80-0.86) in the validation cohort. The included variables dif- fered from those in our model possibly due to the following reasons. First, the patients included in the previous study were from Massachu- setts General Hospital (derivation cohort) and Brigham and Women’s Hospital (validation cohort), while in our study, the patients were all treated in the ICU of various hospitals. The study populations differed between the two studies. Second, the outcome of the previous study was kidney failure or death, but in our study, the outcome was only the need for RRT. Third, the CK level in the patients included in the

previous study was above 5000 U/L, but in our study, the CK level was above 1000 U/L.

Whether RRT can reduce mortality is still controversial [26-28]. One systematic review concluded that despite the improvement in the myo- globin, creatinine, and electrolyte levels in RM patients treated with RRT, the mortality rates remained unchanged [29]. A retrospective study analyzed the effect of RRT in patients with Heat stroke (compli- cated with RM and AKI), and the results showed that RRT can remove myoglobin, support multiOrgan function, and modulate systemic in- flammatory response syndrome (SIRS), but the impact on patient out- come still needs validation in larger randomized controlled trials [30].

In our study, after the PSM analysis, the in-hospital mortality rate in the RRT group (26/149) was similar to that in the non-RRT group (25/ 149). However, the patients in the RRT group had longer hospital stays and more complications, such as atrial fibrillation and AKI. RRT may not improve mortality for several reasons. First, the timing of the initiation of RRT may not be appropriate. Second, when patients have more complications, RRT alone cannot improve the outcome. Therefore, the early detection of RM patients who need RRT is very important and may aid in delivering proper care and optimizing the use of limited re- sources. In addition, we identified the risk factors related to mortality in the RRT group to facilitate the identification of patients at high risk of mortality.

1. Limitations

There were some limitations in this study. First, the external valida- tion was based on a subset of patients who were all Asian, while the two US datasets included a mix of different ethnicities. This limitation hin- ders extrapolating these findings to other mixed ethnicity ICU patient groups. Second, this study included a modest sample size in the training dataset. Future multicenter studies with large sample sizes are needed to explore the applicability of this model. Third, the patients included in the training and testing datasets were treated in the ICU, and discrep- ancies in characteristics between patients treated in the ICU and those

A B

Cr CK

MODEL

All None

Number high risk

Number high risk with event

0.2

0.3

Number high risk (out of 1000)

600

800

1000

0.0 0.2 0.4 0.6 0.8 1.0

High Risk Threshold

C D

Number high risk

Number high risk with event

Number high risk

Number high risk with event

Number high risk (out of 1000)

600

800

1000

Number high risk (out of 1000)

600

800

1000

0.0 0.2 0.4 0.6 0.8 1.0

High Risk Threshold

Net Benefit

0.1

0.0

0

200

400

1:100 1:5 2:5 3:4 4:3 5:2 5:1 100:1

Cost:Benefit Ratio

0.0 0.2 0.4 0.6 0.8 1.0

0

200

400

High Risk Threshold

0.0 0.2 0.4 0.6 0.8 1.0

0

200

400

High Risk Threshold

1:100 1:5 2:5 3:4 4:3 5:2 5:1 100:1

Cost:Benefit Ratio

1:100 1:5 2:5 3:4 4:3 5:2 5:1 100:1

Cost:Benefit Ratio

Fig. 5. Decision curve and clinical impact curve of the nomogram for RRT in RM patients. The model had superior standardized net benefit over Cr and CK. CK, creatine kinase; Cr, creatinine.

treated in the general ward may decrease the generalizability of the model. Finally, CK may not be routinely checked in all patients seeking treatment at a hospital; therefore, several patients may have developed RM without being diagnosed.

1. Conclusions

In summary, we developed and validated a model for the early pre- diction of the need for RRT among patients with RM based on 8 variables commonly measured during the first 24 h after admission. Predicting the need for RRT could help ensure appropriate treatment and the opti- mization of the use of medical resources.

Supplementary data to this article can be found online at https://doi. org/10.1016/j.ajem.2021.03.006.

Ethical approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Availability of sata and material

The authors confirm that all data underlying the findings are fully available without restriction. All relevant data are provided in the paper and its Supporting Information files.

Competing interests

The authors have no potential conflicts of interest with respect to the research, authorship, and/or publication of this article to declare.

Credit author statement

All authors contributed to the design of the study concept. Chao liu, Qian Yuan and Xuefeng Sun contributed to collect data, develop models, and drafted the manuscript. Zhi Mao, Pan Hu, Rilige Wu and Xiaoli Liu

contributed to further analyze and interpret data. Chao Liu, Rilige Wu, Qian Yuan, Kun Chi, Xiaodong Geng and Quan Hong contributed to clean the data and algorithm programming. Chao liu, Qian Yuan and Xuefeng Sun contributed to the statistical analysis. Chao liu, Zhi Mao and Xuefeng Sun contributed to edit and approve the manuscript. All authors revised the manuscript draft and approved the final version for submission.

Funding

This study was supported by five grants 1. National Natural Science Foundation of China (81870463), 2. Fund of Chinese PLA 13th Five- Year Plan for Medical Sciences (BLB19J009), 3. Major Research Plan of the National Natural Science Foundation of China (92049103), 4. Foster- ing Fund of Chinese PLA General Hospital for National Distinguished Young Scholar Science Fund(2019-JQPY-002), and 5. National Key Re- search and Development Project (2018YFE0126600).

This manuscript has been edited and proofed by a professional En- glish translation service.

Declaration of competing interest

The authors declare no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Acknowledgements

Not applicable.

References

1. Bosch X, Poch E, Grau JM. Rhabdomyolysis and acute kidney injury. N Engl J Med. 2009;361(1):62-72.
2. Petejova N, Martinek A. Acute kidney injury due to rhabdomyolysis and renal re- placement therapy: a critical review. Crit Care. 2014;18(3):224.
3. Zutt R, van der Kooi AJ, Linthorst GE, Wanders RJ, de Visser M. Rhabdomyolysis: re- view of the literature. Neuromuscul Disord. 2014;24(8):651-9.
4. Panizo N, Rubio-Navarro A, Amaro-Villalobos JM, Egido J, Moreno JA. Molecular mechanisms and novel therapeutic approaches to Rhabdomyolysis-induced acute kidney injury. Kidney Blood Press Res. 2015;40(5):520-32.
5. Bessa Jr O. Alcoholic rhabdomyolysis: a review. Conn Med. 1995;59(9):519-21.
6. Zimmerman JL, Shen MC. Rhabdomyolysis. Chest. 2013;144(3):1058-65.
7. Luetmer MT, Boettcher BJ, Franco JM, Reisner JH, Cheville AL, Finnoff JT. Exertional Rhabdomyolysis: a retrospective population-based study. Med Sci Sports Exerc. 2020;52(3):608-15.
8. Elterman J, Zonies D, Stewart I, Fang R, Schreiber M. Rhabdomyolysis and acute kid- ney injury in the injured war fighter. J Trauma Acute Care Surg. 2015;79(4 Suppl 2): S171-4.
9. Simpson JP, Taylor A, Sudhan N, Menon DK, Lavinio A. Rhabdomyolysis and acute kidney injury: creatine kinase as a prognostic marker and validation of the McMahon score in a 10-year cohort: a retrospective observational evaluation. Eur J Anaesthesiol. 2016;33(12):906-12.
10. Bellomo R, Kellum JA, Ronco C. Acute kidney injury. Lancet. 2012;380(9843): 756-66.
11. Chavez LO, Leon M, Einav S, Varon J. Beyond muscle destruction: a systematic review of rhabdomyolysis for clinical practice. Crit Care. 2016;20(1):135.
12. Pollard TJ, Johnson AEW, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU collaborative research database, a freely available multi-center database for critical care research. Sci Data. 2018;5:180178.
13. Johnson AE, Stone DJ, Celi LA, Pollard TJ. The MIMIC code repository: enabling repro- ducibility in critical care research. J Am Med Inform Assoc. 2018;25(1):32-9.
14. Johnson AE, Pollard TJ, Shen L, Lehman LW, Feng M, Ghassemi M, et al. MIMIC-III, a freely accessible critical care database. Sci Data. 2016;3:160035.
15. Zheng CF, Liu WY, Zeng FF, Zheng MH, Shi HY, Zhou Y, et al. Prognostic value of platelet-to-lymphocyte ratios among critically ill patients with acute kidney injury. Crit Care. 2017;21(1):238.
16. Khwaja A. KDIGO clinical practice guidelines for acute kidney injury. Nephron Clin Pract. 2012;120(4):c179-84.
17. Knaus WA, Wagner DP, Draper EA, Zimmerman JE, Bergner M, Bastos PG, et al. The APACHE III prognostic system. Risk prediction of hospital mortality for critically ill hospitalized adults. Chest. 1991;100(6):1619-36.
18. Vincent JL, Moreno R, Takala J, Willatts S, De Mendonca A, Bruining H, et al. The SOFA (sepsis-related organ failure assessment) score to describe organ dysfunction/fail- ure. On behalf of the working group on sepsis-related problems of the European So- ciety of Intensive Care Medicine. Intensive Care Med. 1996;22(7):707-10.
19. Cabral BMI, Edding SN, Portocarrero JP, Lerma EV. Rhabdomyolysis. Dis Mon. 2020; 66(8):101015.
20. Abraham A, Pedregosa F, Eickenberg M, Gervais P, Mueller A, Kossaifi J, et al. Ma- chine learning for neuroimaging with scikit-learn. Front Neuroinform. 2014;8:14.
21. Candela N, Silva S, Georges B, Cartery C, Robert T, Moussi-Frances J, et al. Short- and long-term renal outcomes following Severe rhabdomyolysis: a French multicenter retrospective study of 387 patients. Ann Intensive Care. 2020;10(1):27.
22. Liu C, Yuan Q, Mao Z, Hu P, Chi K, Geng X, et al. The top 100 most cited articles on rhabdomyolysis: a Bibliometric analysis. Am J Emerg Med. 2020;38(9):1754-9.
23. El-Abdellati E, Eyselbergs M, Sirimsi H, Hoof VV, Wouters K, Verbrugghe W, et al. An observational study on rhabdomyolysis in the intensive care unit. Exploring its risk factors and main complication: acute kidney injury. Ann Intensive Care. 2013;3(1): 8.
24. Kerr KF, Brown MD, Zhu K, Janes H. Assessing the clinical impact of risk prediction models with decision curves: guidance for correct interpretation and appropriate use. J Clin Oncol. 2016;34(21):2534-40.
25. McMahon GM, Zeng X, Waikar SS. A Risk prediction score for kidney failure or mor- tality in rhabdomyolysis. JAMA Intern Med. 2013;173(19):1821-8.
26. Rachoin JS, Weisberg LS. Renal replacement therapy in the ICU. Crit Care Med. 2019;

47(5):715-21.

1. Tandukar S, Palevsky PM. Continuous renal replacement therapy: who, when, why, and how. Chest. 2019;155(3):626-38.
2. Ronco C. Extracorporeal therapies in acute rhabdomyolysis and myoglobin clear- ance. Crit Care. 2005;9(2):141-2.
3. Zeng X, Zhang L, Wu T, Fu P. Continuous renal replacement therapy for rhab- domyolysis. Cochrane Database Syst Rev. 2014;6:CD008566.
4. Zhang L, Kang Y, Fu P, Cao Y, Shi Y, Liu F, et al. Myoglobin clearance by continuous venous-venous haemofiltration in rhabdomyolysis with acute kidney injury: a case series. Injury. 2012;43(5):619-23.