1 Introduction
Stevens–Johnson syndrome (SJS) and toxic epidermal necrolysis (TEN) are severe and rare cutaneous drug reactions, characterized by large epidermal detachment of necrotic epidermis and erosions of mucous membrane [
1]. SJS, SJS-TEN (SJS/TEN overlap), and TEN are defined by the extent of epidermal detachment, representing less than 10%, 10%–30%, and greater than 30% body surface area (BSA), respectively [
2]. The previously reported prevalence of SJS/TEN (SJS, SJS-TEN, and TEN) was 1.4 –12.7 cases per million person per year [
3,
4]. Collecting generalizable information to analyze patient prognosis is difficult because of the rarity of these diseases.
A severity-of-illness score for TEN (SCORTEN) was developed from a logistic regression model in 2000 by French scientists [
5]. It consists of seven parameters and should be determined within 24 h after admission (Table 1, SCORTEN). SCORTEN could be easily calculated to estimate the expected mortality. Therefore, it has been widely used in various populations from different countries, such as United States, Australia, Europe, and China [
5–
9]. However, not every parameter of SCORTEN is always systematically collected. Thus, auxiliary score (AS) was developed [
7]. It comprises three factors (Table 1, auxiliary score) and has been applied internationally but not validated in China. The ABCD-10 model was described recently by Noe
et al. by using a multi-institutional cohort of patients in the United States. This model includes five parameters (Table 1, ABCD-10 model) and confirmed the importance of severe renal dysfunction as an SJS/TEN mortality risk factor [
10]. However, it has not been used in other populations.
Besides prediction models, other relevant factors contribute to the development of diseases and are associated with mortality. In previous studies, advanced age was recognized as a risk factor. Mortality increased with age, especially over 40 years [
11]. Tuberculosis was also identified as an independent risk factor for mortality in India [
12]. Metabolic syndrome and gout were associated with increased mortality in Texas [
13]. Dialysis showed poor prognosis at a burn center in Germany, and hemodialysis was confirmed as an SJS/TEN mortality risk factor in the Asian population [
14,
15]. In addition, patients with SJS/TEN have severe symptoms, including fever, mucositis, and chest tightness. Pulmonary symptoms are commonly observed, and most patients with SJS/TEN usually experience deteriorating pulmonary function. Thus, close monitoring, including chest computed tomography (CT), for possible complications is advisable. A CT scan is usually implemented for early and intensive treatment [
16]. However, the association between pulmonary CT findings and prognosis of patients with SJS/TEN has not been investigated thus far. Therefore, variables from previous models and clinical information, such as medical comorbidities and CT scan results, were included in the present study. Although SCOTREN has been widely used in clinical practice, reanalysis using a data-driven approach based on Chinese cases could provide insights to make the model specific to China. Besides, the AS and ABCD-10 models were introduced to clinical application only recently. Therefore, validation in patients from different sources and countries is needed.
This study aimed to compare the predictive ability of SCORTEN, AS, and ABCD-10 models; evaluate the risk factors associated with mortality in SJS/TEN; and propose possible improvements in these scores in Chinese patients.
2 Methods
2.1 Data collection
From January 2008 to January 2019, 217 patients with a diagnosis of SJS, SJS/TEN, or TEN from the Department of Dermatology of Huashan Hospital, Fudan University, were included in this study. The diagnosis was made in accordance with the clinical data and skin biopsy. For improved skin care, most patients with SJS/TEN were admitted to the intensive care unit with reverse-isolation procedures.
The data analyzed were those recorded during the first 24 h in the intensive care unit, including demographics, medications, comorbidities, laboratory data, physical examination, and outcome. The data were collected independently by two investigators named Yanhong Shou and Lu Yang. Each CT was reviewed independently by two radiologists. Medical information was cross-checked by Yongsheng Yang and Jinhua Xu. This study was approved by the Ethics Committee of Huashan Hospital (reference number 2019-289).
2.2 Evaluation of SCORTEN, AS, and ABCD-10 models
SCORTEN includes seven clinical variables: (1) age>40 years, (2) presence of cancer or malignancy, (3) heart rate>120 beats per minute, (4) serum urea level>10 mmol/L, (5) serum bicarbonate level<20 mmol/L, (6) serum glucose level>14 mmol/L, and (7) involved BSA at day 1>10%. Each variable equals to one point in the score. The expected mortality rate was calculated using the following formula:
logit= −4.448+ 1.237 (SCORTEN)
p= exp (logit)/(1+ exp (logit)) [
5]
AS is a simplified form of SCORTEN; it includes TEN (>30% skin detachment), presence of cancer/malignancy, and age. Age was grouped into four categories as≤30, 31–55, 56–75, and≥75 years, with zero, one, two, and three points, respectively. The prediction of expected mortality was based on the following formula:
logit= -3.136+ 0.913 (AS)
p= exp (logit)/(1+ exp (logit)) [
7]
The ABCD-10 model was calculated by taking the sum in accordance with the following principles: one point each for age 50 years or older, skin detachment greater than 10% of BSA, and serum bicarbonate level lower than 20 mmol/L; two points for the presence of active/ongoing cancer; and three points for dialysis prior to admission. The expected mortality rate was based on
logit= −3.764+ 0.898 (ABCD-10)
p= exp (logit)/(1+ exp (logit)) [
10]
2.3 Statistical analysis
Statistical analysis was performed using Stata 15.1 (Stata Corp., Ltd., College Station, Texas, USA). Mean and standard deviation (SD) were used to describe continuous variables if normally distributed or median and interquartile range (IQR) if otherwise. A two-tailed
P<0.05 was considered significant. The area under the receiver operating characteristic curve (AUC) was used to illustrate the performance of these established scores and the scores refined in the present study [
17]. A prediction rule should have AUC greater than 0.50, and it is regarded to work acceptably if the AUC is greater than 0.70. A nonparametric approach was applied for the comparison of AUCs. In addition, the score’s accuracy of the predictions was evaluated using Brier score (BS) [
18]. Calibration was evaluated using Hosmer–Lemeshow goodness-of-fit statistic [
19].
Logistic regression was performed for all variables, including different variables included in the SCORTEN, AS, and ABCD-10 models, to explore other predictive factors of mortality. All variables with a significance level of two-tailed
P<0.1 were included as candidates in the backward stepwise multivariable model. Given the relatively low number of events per variable, bootstrapping was used for model selection to improve efficiency and avoid instability [
20]. A two-tailed
P<0.05 was considered significant.
3 Results
3.1 Patient characteristics
From January 2008 to January 2019, 217 patients with a definite diagnosis of SJS/TEN were included in this study, including 118 males and 99 females with a mean (SD) age of 48.18 (1.25) years. The cohort has been described in detail previously [
21]. The median value of BSA on admission was 30.55% (IQR, 25.50%–35.60%). The actual mortality was 4.15% (9/217), including 2.04% in the patient group with SJS, 9.52% in the group with SJS-TEN, and 6.90% in the group with TEN.
3.2 Evaluation of SCORTEN, AS, and ABCD-10
The actual mortality was lower than the predictive ones at each SCORTEN level, but no significant difference was found between the actual and expected mortality rate in each SCORTEN (P>0.05, Table 2). The calibration of SCORTEN indicated good agreement between the expected and observed deaths (Hosmer–Lemeshow goodness-of-fit test, P = 0.69; Table 3). The ROC curve confirmed good discrimination of SCORTEN (AUC= 0.78; 95% CI 0.61–0.96, Fig. 1). The low BS (0.04, Table 3) showed excellent predictive value of the model.
AS predicted 39.1 deaths, higher than the actual mortality, as described in Table 4 (P = 0.00). The AUC value of lower than 0.7 indicated that AS did not work acceptably enough in discrimination (AUC= 0.69, 95% CI 0.53–0.85, Table 3, Fig. 1). This model showed that calibration (Hosmer–Lemeshow goodness-of-fit test, P = 0.25, Table 3) and accuracy (BS 0.07, Table 3) were not as good as SCORTEN.
The predicted mortality of the ABCD-10 model was overestimated by 2%, and the number of observed deaths was smaller than the predicted ones at each score level, with no significant difference (P>0.05, Table 5). The AUC value was 0.68 (95% CI 0.50–0.86), revealing 68% probability that a randomly selected patient who died could receive a higher ABCD-10 score than a randomly selected patient who survived. The ABCD-10 model indicated good predictive ability (BS 0.04, Table 3) and calibration (Hosmer–Lemeshow goodness-of-fit test, P = 0.55, Table 3).
3.3 Possible refinement
In the univariable analysis, BSA (>10%), heart rate (>120 bpm), serum urea level (>10 mmol/L), serum bicarbonate level (<20 mmol/L), dialysis, and the imaging appearance of pulmonary consolidation on CT were associated with a significantly increased risk in death (Table 6). This imaging appearance was described as pulmonary consolidation, which had high-density spots and tablets with irregular margins, indicating inflammatory changes. Those vital variables were included as candidates in the multivariable model. The pulmonary consolidation on CT remained as the only significant risk factor based on backward stepwise logistic regression (OR= 6.43, 95% CI 1.13–36.67, P = 0.04).
In the 16 cases with imaging appearance of pulmonary consolidation on CT, 13 patients were diagnosed with pulmonary infection, one diagnosis was combined with underlying disease chronic obstructive pulmonary disease, and two patients had no pulmonary disease (Table 7). The diagnosis of pulmonary infection was given according to clinical data, such as WBC count, results of microbiologic tests, and imaging signs of radiography or CT [
22]. In our study, pulmonary infection was not significantly associated with mortality (OR= 1.40, 95% CI 0.17–11.90,
P = 0.76). In addition, there were other five patients combined with pulmonary disorders. 3 patients had tuberculosis, asthma and lung cancer respectively, all confirmed non-pneumonia through admission CT scan. Other two patients were diagnosed with chronic obstructive pulmonary disease. This study indicated that other underlying pulmonary disorders did not make patients more vulnerable to the imaging appearance of pulmonary consolidation (OR= 3.75, 95% CI 0.39–36.01,
P = 0.25).
The imaging appearance of pulmonary consolidation on CT was added to the model to improve the AUC. The variables in SCORTEN, AS, and ABCD-10 and the variable of pulmonary consolidation on CT formed the refined models. The AUCs of these models were higher than those of previous models. However, they were not significant (0.79 vs. 0.78, P = 0.54; 0.74 vs. 0.69, P = 0.18; 0.74 vs. 0.68, P = 0.20, Table 3, Fig. 2). In addition, the refined AS and ABCD-10 models had similar discrimination compared with the original SCORTEN (0.74 vs. 0.78, P = 0.23; 0.74 vs. 0.78, P = 0.30). Both models suggested good predictive ability (BS, 0.04 and 0.04) and calibration (Hosmer–Lemeshow goodness-of-fit test, P = 0.67 and P = 0.57, respectively), as shown in Table 3.
4 Discussion
In this cohort of 217 patients from China, the predictive ability of the SCORTEN, AS, and ABCD-10 models was validated. SCORTEN represented good discrimination, calibration, and accuracy. The imaging appearance of pulmonary consolidation on CT was observed to be a considerable variable. Furthermore, the refined AS or ABCD-10 model could be an alternative in China.
SCORTEN has been an acknowledged and well-established severity-of-illness score starting from its initial publication in 2000. The score could be easily calculated to predict SJS/TEN-related mortality [
10,
23,
24]. In the present study, the observed AUC of SCORTEN was 0.78, which was smaller than the 0.82 reported by Garin
et al. [
5]. Previous studies have proven that SCORTEN worked well for patients with SJS-TEN and those with score values of 4 and 5 [
5]. SCORTEN was also commended for its applicability to different patients, especially those with serious conditions or few symptoms [
25,
26]. In the present study, SCORTEN tended to overestimate the number of deaths for patients, with a score of 0–5.
As for AS and ABCD-10, the number of expected deaths at each level was higher than the observed. Although AS and ABCD-10 may be less accurate than SCORTEN in patients, their calculation was simple. The AS and ABCD-10 models had advantages of reduced laboratory and vital parameters, and they did not rely on knowledge of specific clinical information.
The prevalence of risk factors in different patient populations differs. Previous studies have recognized additional risk factors not included in the original SCORTEN model, including tuberculosis, metabolic syndrome, gout, and dialysis [
12–
15]. In the present study, the imaging feature of pulmonary consolidation on CT, measured at the time of admission by using CT scan, was found to be an independent predictor of in-hospital mortality. The refined models were formed using the variables in the original SCORTEN, AS, and ABCD-10 models and the variable of pulmonary consolidation on CT. Comparison of the refined AS and ABCD-10 models with original SCORTEN showed that they were similar in discrimination. In the population, this imaging appearance may be a crucial mortality risk factor to consider. Besides SCORTEN, the refined AS or ABCD-10 model could serve as an alternative.
In the 16 cases with imaging appearance of pulmonary consolidation on CT, three were not diagnosed with pulmonary infection. Therefore, this imaging appearance at admission did not equal to pulmonary infection, and it may represent pulmonary inflammatory reactions due to SJS/TEN. This imaging appearance could also be a precaution of rigorous clinical monitoring. Pulmonary infection is a common comorbidity of SJS/TEN [
25,
27], and patients with pulmonary infection have increased risk of acute respiratory distress syndrome [
28]. CT scan combined with clinical data at admission could help screen patients with pulmonary infection potentially in need of antibiotics, especially when some patients were treated with systemic glucocorticoids. Precautions to prevent death could be implemented if antibiotics are necessary.
This study has inevitable limitations. First, it was a single-center study and had only nine death cases. In the cohort, many patients with SJS/TEN who were treated for serious primary diseases in other departments were not included, and dermatologists were required to hold consultations. In addition, the cohort included patients with skin problem as the first symptom of SJS/TEN, and the number of cases with serious complications was low. Thus, the SCORTEN of 3 (
n = 13), 4 (
n = 5), and 5 (
n = 1) accounted for a small minority of SJS/TEN cases (19/217, 8.7%) in the cohort, while the median percentage of BSA denuded was higher than that of previous studies [
4,
10]. However, the present findings were consistent with others in terms of the AUC of SCORTEN, and no significant difference was found between the actual and the expected mortality rate in each SCORTEN. Second, data during the first 24 h after admission were collected, whereas another study evaluated the score at different times and concluded that the discriminatory power of SCORTEN increased on days 3 and 5 after admission [
29]. Third, as an observation study, heterogeneity may exist in the assessment of epidermal detachment as it was defined by different physicians. In addition, the treatment in this cohort was not standardized due to the lack of evidence on the best therapy for patients with SJS, SJS-TEN, and TEN. Other described SJS/TEN cohorts share these same limitations.
In summary, SCORTEN had the best performance when evaluating the severity of SJS/TEN in this cohort. This study demonstrated the association between the imaging feature of pulmonary consolidation on CT and increased mortality, and patients with this imaging appearance need rigorous clinical monitoring. Furthermore, refined AS and ABCD-10 models could be alternatives when SCORTEN variables were not available in Chinese patients.
PDF
(339KB)
