Comprehensive Statistical Approach to Investigate the Risk Factors Affecting the Occurrences of Cellulitis Episodes among Patients with Lymphedema

Niswatul Qona’ah; Martin Lukusa

doi:10.3121/cmr.2024.1897

Abstract

Background: Cellulitis is an infection of the skin and the tissues just under the skin. As any disease, cellulitis has various physiological and physical effects that deteriorate a patient’s quality of life. Luckily, cellulitis can be treated when dealt with in a timely fashion. Nonetheless, some patients may experience more than one episode of cellulitis or a recurrence of cellulitis that was previously cured. In fact, the occurrences of cellulitis episodes are believed to follow a statistical distribution. The frequency distribution of cellulitis episodes is scrutinized herein. We aimed to investigate the risk factors that affect the number of cellulitis episodes and the pattern of association between cancer types and cellulitis episodes by using analytical and visual approaches.

Methods: A statistical approach applying a two-part count regression model was used instead of the traditional one-part count model. Moreover, multiple correspondence analysis was used to support the finding of count regression models.

Results: The results of analysis of the sample from the National Cheng Kung University hospital in Taiwan revealed the mean age of patients was 58.7 ± 14.31 years old. The two-part regression model is conceptually and numerically better than the one-part regression model when examining the risks factors that affect cellulitis episodes. Particularly, we found the significant factors based on the best model are cellulitis history (; P value < 0.001), clinical stage of cancer (3) (; P value < 0.001), no cancer (; P value < 0.05), cancer of female reproductive organs (; P value < 0.05), breast cancer (; P value < 0.05), and age ≥ 60 years (; P value < 0.05). Multiple correspondence analysis approach found cancer types (breast and female reproductive organ), age ≥ 60 years, and cellulitis history were more likely to link to excess zero cellulitis or one cellulitis episode.

Keywords:

Lymphedema is a chronic condition caused by a failure in the lymphatic system, meaning lymphatic fluid is not draining from the body.¹ A typical medical health issue generally associated with lymphedema is cellulitis. The prevalence of cellulitis among patients with lymphedema in Asia has been previously discussed.² Cellulitis is an infection of the skin and the tissues just under the skin.^3,4 Like any disease, cellulitis has various physiological and physical effects that deteriorate a patient’s quality of life. Cellulitis is common in Taiwan.^5,6 Some cellulitis patients may experience more than one cellulitis episode, while others may experience a recurrence of the cellulitis.^7,8 The occurrences of cellulitis episodes are generally recorded as discrete and non-negative numbers or counts. Count data are distributed as non-negative integers that are intrinsically heteroskedastic, right skewed, and have a variance that may not equal the mean.⁹ The use of adequate models leads to reliable results and valid inference. Count regression one-part models including Poisson regression models, binomial regression models, negative binomial regression models, generalized Poisson regression models, etc.,¹⁰ occupy an important place in statistics and epidemiology literature. In particular, the two-part models including zero-inflated Poisson (ZIP) and zero-inflated negative binomial (ZINB) models are needed when there is a large proportion of zeros.¹¹ In Taiwan, cellulitis is still considered a rare disease, which means, among the population there are a large proportion who have never had cellulitis. On the other hand, there are patients who had cellulitis before but currently do not have it. This phenomenon inflates the proportion of zeros.¹² These statistical models enable us to investigate the risk factors that influence the occurrences of cellulitis episodes and to assess the strength of the relationship between cellulitis and cancer in patients from the National Cheng-Kung University (NCKU) hospital. So far, logistic regression and cox-regression have been most used for this purpose, and there are no studies that have used the count regression model.⁸

Multiple regression analysis was conducted to recognize the factors associated with Health-related Quality of Life (HRQoL).¹³ Multivariate linear and beta regression models¹⁴ were used to determine the strength of association and margin effect. Poisson regression and its derivatives have not yet been used intensively. In particular, the two-part model is not known in the study of cellulitis or lymphedema. Correspondence analysis is a graphical statistics method for analyzing the associations among factors in multivariate data.¹⁵ This approach enables the ability to uncover the pattern of associations between two, three, or more categorical variables. As far as we know, potential risk factors for cellulitis in patients with lymphedema have not yet been studied using a two-part model and the multiple correspondence analysis, conjointly. Hence, the objective of this study was to assess the risk factors for cellulitis in patients with lymphedema using a zero- inflated regression and multiple correspondence analysis in a cross-sectional study. Our findings can be used in the implementation of the development of guidelines for the diagnosis and management of cellulitis in Taiwan.

Materials and Methods

The sample under investigation is a retrospective study data which was collected in the NCKU hospital located in Taiwan from 2014 to 2021. This sample contains various details about the patients who had clinical visits. These medical records were obtained from 242 patients. The sample size needed to reach a particular power was computed using PASS 2024, version 24.0.2. for a Poisson regression (count response Y versus X’s). The over-dispersion parameter is assumed to be 1 (no over-dispersion is assumed), with 90% power, the number of needed subjects will be minimum 212. According to Signorini (1991),16 the formula to obtain the sample size for Poisson regression can be written as

(1) (1)

where N is the sample size, ϕ is a measure of over-dispersion, μ_T is the mean exposure time, Z is the standard normal deviate, R² is the square of the multiple correlation coefficient when the covariate of interest regressed on the other covariates. This sample size was used to test the null hypothesis that β₁=0 versus the alternative β₁=B1. Figure 1 shows the process of sample selection. This sample is particularly made up of cancers patients. Nearly half of them had breast cancer, followed by female reproductive cancer types. The distribution of cancer types agrees with the facts that there are more females than males in this sample. Interesting is that the mean age of patients found is not very far from previous studies where proportion of males was larger than females. Hsu et al.⁵ found that the weather (chillness) was not an important risk factor. More importantly, they pointed out that it was due to location. These data comprised patients from a southern city (Tainan). Hsu et al.⁵ used the Taiwan data base.

Figure 1.

Flowchart of sample selection

The outcome variable denoted by (Y) is the number of cellulitis episodes. The independent variables (X) consist of cellulitis history (X₁), clinical stage (X₂), chillness (X₃), type of cancer (X₄), and age (X₅). Variables are presented in Table 1. The number of cellulitis episodes (Y) is count outcome variable, which takes 0,1,2,…. Cellulitis history (X₁) is a binary variable where 1 means a patient experienced cellulitis in the past, and 0 means they never experienced it. Clinical Stage (X₂) has a categorical scale that is classified into two categories including under the 3rd (1st or 2nd) stage and the 3rd stage. Chillness (X₃) is a binary variable where 1 means yes and 0 means “no”. Cancer type (X₄) is a categorical variable. It is recorded into four categories including no cancer (no related cancer), female reproductive organ (cervical, endometrial, and ovarian) cancers, breast cancer, and other types of cancer. Age variable (X₅) has a numeric scale, originally. It is recorded as a binary variable such as X₅=0 refers to age < 60 years and X₅=1 refers to age ≥ 60 years. The statistical method we used to further analyze including regression models for count data (one-part models and two-part models) and Multiple Correspondence Analysis is explained in following sections.

View this table:

Table 1:

Measurement of Variables

Regression Models for Count Data

According to Wooldridge (2010)¹⁷ and Schober & Vetter (2021),¹⁸ a general regression model applied to count data may be described

(2) (2)

where x=(X₁, X₂,… ,X_p ) is the vector of potential risk factors, and θ = (α₀, α₁, … , α_p ) is the coefficient estimates.

One-Part Count Regression Models

The traditional and popular model for count is the regular Poisson (RP). However, in other situations where RP shows limitations, Quasi-Poisson (QP) or Negative binomial (NB) can be used. In particular, When the mean of cellulitis events is believed to be different from its variance, this phenomenon is called overdispersion.^19,20 Briefly, we introduced the two most popular one-part count regression models in the supplementary material.

Two-Part Count Regression Models

Popularized by Lambert in1992,¹¹ the zero-inflated (ZI) models are mixtures of the zero-mass model and a specific count model. The ZI models are designed to handle the excessive proportion of zeros and overdispersion in the outcome count variable.²¹ In all ZI models zeros are generated from two distinct processes. The first process generates the structural zeros from the zero-mass function whereas, the second one generates the random zeros from a Poisson distribution, negative binomial, generalized Poisson models, etc. We introduce the most popular including the ZI Poisson (ZIP) and ZI negative binomial (ZINB). The two-part models are presented in the supplementary materials.

The estimation procedure of θ for one-part models and two-part models are carried in R packages. More details are found in the supplementary. To choose the best model fit, the Akaike information criterion (AIC) can be used to compare various model fits. The AIC is calculated by AIC = 2k – 2ln (L), where k is number of parameters and L is likelihood. This criterion helps us choose whether the models have optimal parsimony. The model with the smallest AIC implies the best fit.

Multiple Correspondence Analysis

The correspondence analysis (CA) is a graphical method for analyzing the associations among factors in multivariate data.²² It enables us to uncover the pattern of associations between two, three, or more categorical variables by using a contingency table.²³ In practice, when there are more than two categories, CA becomes multiple correspondence analysis (MCA).²⁴ This study investigates the pattern and strength of the association between three categories including cellulitis numbers (A), types of cancer (B), and cellulitis history (C). For instance, Sourial et al.²⁵ presented the use of MCA in epidemiology. The ideas behind the CA steps are explained by Greenacre and Clarke.²³ The contribution of one cell to the total χ² statistics is very useful in establishing the nature of dependency. This statistic is used to determine whether the rows (category A) and columns (category B) are independent from one another. Phrased differently, χ² statistics test whether there is a statistically significant dependence between category A and category B. A closely related measure to χ² in CA is known as the inertia of the table, and given by

(3) (3)

where n is the total number of rows times columns of a contingency table. Thus, χ² = n × Total inertia. In practice, when the number of categorical risk factors is greater than three, the interpretation of MCA patterns is complicated. More details of MCA are provided in the supplementary material.

Data Analysis

Descriptive Statistics

The total sample size is 242 patients with lymphedema. Among all patients, the distribution of gender shows that 18% were males, while 82% were females. The actual age was distributed as follows: minimum age was 4 years, first quartile was 50 years, the median was 59 years, the third quartile was 68 years, and the maximum age was 94 years. Moreover, the mean age was 58.7 ± 14.31 years. Based on the histogram of the actual age (Figure 2), the age group variable distributed shows 43% of patients were above 60 years and 57% were below 60 years. Among all patients, 77% had cancer, and 23% suffered from something other than cancer. This group was classified as related to cancer. Among cancer patients, 47% had breast cancer, and nearly 13% had cervical or ovarian cancer. Other cancers including endometrial, colon, prostate, liposarcoma, rectal, laryngeal, etc. share the remaining 40%. Lymphedema patients were classified into three clinical stages; hence, the clinical stage variable showed that 62.4% of people with cancer were either in stage 1 or stage 2 cancer, and 37.6% were in stage 3. Among the 242 patients, 75% did not have a history of cellulitis; whereas, 28.10% have did have. The study by Collazos et al.²⁶ found that female patients were generally older and had higher rates of edema or lymphedema, which are predisposing factors for cellulitis, compared to male patients. This indicates cellulitis may be more common in older female populations. Additionally, female patients experienced higher rates of recurrent cellulitis, suggesting being female, especially at an older age, may be an independent risk factor for the condition. Consequently, the high proportion of female and cancer patients likely reflects the epidemiology of cellulitis, where older women and those with lymphatic disorders are at increased risk.

Figure 2.

Histogram of Age

The distribution of cellulitis numbers variable (Y) revealed 71.90% have never experienced occurrence of the cellulitis event, while 9.1% have experienced one cellulitis episodes. In addition, 5.36%, 1.34%, and 5.36% have experienced two, three, and four cellulitis episodes, respectively. We truncated the cellulitis outcome variable to take zero, one, two, three, and four. Since the zero (0) value has the largest proportion, we determined the cellulitis variable has an excess zero. The visualization of data distribution can be seen in Figure 3.

Histogram of cellulitis number (a) before truncation, (b) after truncation

Figure 3.

Histogram of cellulitis number (a) before truncation, (b) after truncation

Figure 3 shows the distribution of cellulitis variable is likely to be a Poisson with excess zero. The value of 12 seems like an outlier and might influence the analysis result and, thus, we truncated this value to 4. The left-hand panel (Figure 3a) is the histogram of cellulitis number before truncation, and the right-hand panel (Figure 3b) is the histogram after truncation. In effect, the truncation does not make a major change in data distribution.

In this analysis, we used some specific independent variables to build models for the cellulitis number. The chosen independent variables are cellulitis history, clinical stage, chillness, cancer type, and age. The following section provides the description of each independent variable. All independent variables are in the form of categorical variables, which are described in Table 2.

View this table:

Table 2:

Descriptions of independent variables

Figure 4a provides a visualization of the relationship between cellulitis history and cellulitis number. To obtain a more clear figure, we use transform Y, the number of cellulitis to Y + 1. The boxplot shows there is a relationship between cellulitis history and cellulitis number. Hence, in the further analysis, we use cellulitis history as a potential independent variable with the reference category 0 (have no cellulitis history).

Figure 4.

Cellulitis Number versus (a) cellulitis history, (b) clinical stage, (c) chillness, (d) cancer type, (e) age, (f) age recorded.

Figure 4b reveals there is a difference in cellulitis number between patients who were in the 3rd stage and those patients who were still under the 3rd stage. In effect, the patients who had reached the 3rd stage were more likely to have a larger number of cellulitis episodes than the patients who are under the 3rd stage. Thus, the clinical stage is a potential risk factor of cellulitis number. We can use the clinical stage as another potential independent variable in the next analysis (with the patients under 3rd stage as the reference category).

Chillness is another variable used as potential risk factor. Patients who have cellulitis are more likely to have chillness as a symptom (Figure 4c). We included this independent variable in the next analysis with the reference category as patients who have no chillness.

Another potential risk factor to consider in modelling the number of cellulitis episodes is cancer type. Cellulitis number is likely to be influenced by the type of cancer. Due to the fact that some cancer types are less represented, we have recorded all types into four groups as described in Table 2. The relationship between cancer type and cellulitis number is shown in Figure 4d. We can see there are differences in cellulitis numbers across cancer types. For instance, female patients with “female reproductive organ related cancers” such as cervical, endometrial, and ovarian are more likely to experience cellulitis episodes than those having other types of cancer. Thus, this category is set as the reference group in the regression analysis.

Figures 4e and 4f seem to reveal a relationship between age and cellulitis number. Based on Figure 4e, we can see that a large number of cellulitis episodes is related with patients who are around age 60 years. Hence, we can divide the age factor into two categories: < 60 years and ≥ 60 years, with the proportion as presented in Table 2. Figure 4f shows the difference in cellulitis number between these two categories of age.²⁷ Patients who are < 60 years are more likely to have a large number of cellulitis episodes. We used this category as a reference in the regression analysis. We have used box plots to select potential variables as risk factors for cellulitis episodes. In other studies, it is possible to use existing literature to motivate the choice of specific risk variables. We will describe this in the discussion section.

Regression Models for Cellulitis Number

Traditional Count Regression Models

One-part models consist of regular Poisson (RP) model, Quasi-Poisson (QP) model, and negative-binomial (NB) model. In the one-part model, we regress the number of cellulitis episodes as the dependent variable on five predictors including cellulitis history, clinical stage, chillness, type of cancer, and age, the detail of which are shown in Table 1. When the Poisson model is used, the regression model can be written as

(4) (4)

where η_i (X_i) is the linear predictor, X = (1, CellulitisHistory, ClinicalStage, Chillness, CancerType, Age) is the design matrix, β = (β₀, β₁, β₂, β₃, β₄, β₅) are regression coefficients. Note that β_0 is the intercept while β₁, β₂, β₃, β₄, β₅ are the regression slopes, with β₁= (β_1,1, β_1,2), β₂ = (β_2,1, β_2,2), β₃ = (β_3,1, β_3,2 ), β₄ = (β_4,1, β_4,2, β_4,3, β_4,4), β₅ = (β_5,1, β_5,2 ) are vectors of parameter. Table 3 provides the results of the fitting RP, QP, and NB models to the Cellulitis data. The Generalized Linear Models (GLM) setting is easily implemented in R via the GLMs function by specifying the link function.²⁸ The coefficient estimates are , which each ’s quantify the contributions of each predictor in (4). These contributions are quantified as .

View this table:

Table 3:

One-part and two-part models

Interpretation of One-Part Models

The overall result of RP, QP, and NB model fits shown in Table 3 are very similar. The RP, and QP model fits have identical regression coefficients, but they differ in terms of their standard errors since the QP model can account for overdispersion.²⁹ Based on the RP model, all of the independent variables have a significant effect on the cellulitis number. However, in the QP model, cancer type “no” is not statistically significant, since the 95% confident interval containing value 1, as well as in NB model. It means there is no significant difference in cellulitis number between no cancer and other cancer types. Based on all three models, cellulitis history, clinical stage, and chillness have a very significant effect on the number of cellulitis episodes. For instance, based on NB model fit compared with patients who have no cellulitis history, patients who have cellulitis history are 3.29 = exp(1.19) times more likely to be associated with an increasing number of cellulitis events.

Figure 5 shows the interval coefficient estimate of the three models. We can see the coefficient estimate is very close among the three models. The interval estimate means the standard error of the NB model seems slightly larger compared with RP and QP. Whereas, the RP model seems to have a standard error slightly smaller than the others, it causes the RP model to underestimate some cases due to the assumption of equidispersion. The NB and RP models overcome this dispersion problem by adding dispersion parameter (τ). In this case, NB has a smaller AIC (see Table 3), which means this model is more suitable to model cellulitis numbers than the RP model.

Figure 5.

Coefficient Plot One Part Model

Two Part Models

To account for the excess of zeros, we applied two part models. Two part models consist of the ZIP model and ZINB model. In every two-part model, there are count-part and zero-part models. Recall the regression coefficents in two-part models are α = (α₀, α₁,…, α_q) and β = (β₀, β₁,…, β_p ) such as β and α are involved in the count part and zero-part, respectively. Table 3 shows the coefficient estimates of the ZIP and ZINB models. We can see these two models have similar coefficient values. The count-part model for ZIP and ZINB can be written as below (see Table 3).

(5) (5)

On the other hand, the zero-part model can be written as

(6) (6)

Table 3 reveals the two models reached similar results in nature. This condition is also shown in Figure 6. The coefficients and standard error of the ZIP and ZINB model look similar. All regression coefficients in the count part and zero-part are significant, except chillness is not significant.

Figure 6.

Coefficient Plot Two-Part Model

Interpretation of Two-Part Models

Based on the count part of the ZIP and ZINB models, the clinical stage has a very significant effect on the cellulitis number. Compared with patients who are under 3rd stage, patients who had already been in the 3rd stage are 5.29 = exp(1.67) times more likely to be associated with an increasing number of cellulitis episodes. There is only one variable that has a significant effect in zero part, i.e., cellulitis history. Compared with patients who have no cellulitis history, patients who have a cellulitis history are 0.12 = exp (−2.14) times less likely to have zero (0) number of cellulitis episodes. Figure 7 shows an MCA plot that visualizes the association between cellulitis numbers (A), types of cancer (B), and cellulitis history (C).

Figure 7.

MCA plot

Interpretation of Multiple Correspondence Analysis

From Figure 7, we can see the percentage of variance explained by the first and second principal axis are 18.2% and 14.5%, respectively. In addition, MCA principal components with its inertia, proportions of explained variances, and cumulative proportion of variance explained are given in the table in the supplementary material. The cumulative proportion of variance explained suggests we need seven dimensions to reach a minimum 80% of total variance. It indicates the variables do not have a very high correlation for each other. The MCA plots reveal the following association patterns: (i.) people who have no cellulitis history (CellulitisHist_0), in 1st and 2nd clinical stage (ClinicStage_under3), and type of cancer: breast and others more likely to have 0 number of cellulitis events (NBcellulitis_0); (ii.) people who have cellulitis history (CellulitisHist_1), in the 3rd clinical stage (ClinicStage_3), type of cancer: female reproductive cancer (female_reprod), and aged ≥ 60 years are more likely to have 1 number of cellulitis events (NBcellulitis_1); (iii.) people who have cellulitis number more than 1 (2, 3, and 4) are rare cases, so the points of this category have a long distance from the other variables.

Models Fit Evaluation

Model evaluation using AIC criteria is presented in Table 3. Accordingly, ZIP model seems to have the smallest AIC followed by ZINB, NB, and RP. QP (does not have an AIC value). However, QP can overcome the overdispersion issue in the model. Based on AIC values, the two-part model is better than the one-part model. Thus, it is important to model the zero-part as well as the count part for our data set. It is interesting to see that all the potential independent variables selected have a significant effect on the number of cellulitis in the one part models as well as in the two-part models. Cellulitis history only can better explain the number of cellulitis in the zero part. We can see that ZIP has a minimum AIC (ie, 323.20). Hence, for modeling the number of cellulitis events, we can see that two-part models are better than one-part models.

Discussion

The AICs from the one-part models including the RP model and NB model are the AICs 352.69 and 338.69, respectively. The smaller the AIC, the better the model. The choice of the NB model instead of RP was logical, since the mean of cellulitis events was smaller than its variance. The regression model fit revealed that among the risk factors that affected a significant number of cellulitis episodes, cellulitis history had a positive effect on outcome; its coefficient is 0.856. Compared with patients without cellulitis history, patients with cellulitis history were exp(0.856) = 2.354 times more likely to experience occurrence of cellulitis episodes. The regression coefficient of clinical stage factor (3rd stage) was 1.7678, which indicated a positive association with number of cellulitis episodes. Compared with those bellow stage 3, patients in stage 3 were 5.86 times more likely to experience occurrence of cellulitis episodes.

Cancer-type included no cancer (no cancer related), female reproductive organs, breast, and others. We can see the regression coefficients of breast cancer and female_reprod cancer are 1.2219 and 1.5861, respectively. For patients with no cancer, the regression coefficient was not significant for NB regression. Compared with “other cancers,” those with breast cancer were 3.394 times and those with female reproductive organ cancer were 4.885 times more likely to experience cellulitis episodes. Factor chillness was a positive association with cellulitis number. Patient with chillness were 5.33 times more likely compared with those who did not have chillness. Age factor had a significant association with number of cellulitis events. Compared with those < 60 years, patients ≥ 60 years were 55 times likely to experience cellulitis episodes. Figure 5 provides a comparison of three one-part regression model fits.

The two-part model fits are about the ZIP fit and ZINB fit. Their corresponding AIC are 323.1942 and 325.1942, respectively. We can see that both are smaller than those from the one-part regression models. In addition, the concept of zero-inflated model seems to be appropriate for the cellulitis data set. In effect, the concept of excess of zeros suggests that the event zero is generate by two distinct sources: the structural zeros (false zeros) and random zeros (true zeros). Obviously, there were patients who have never experienced an occurrence of cellulitis episodes (Y = 0). On the other hand, there were patients who had been cured from cellulitis and had not yet experience any recurrence of the cellulitis (Y = 0). Similar to one-part models, the regression coefficients of clinical stage (3rd) was positive and had a significant effect on Y > 0. Compared to the patients in stage 1and 2, those in stage 3, were 5.2969 times likely to influence Y > 0. The chillness factor was fairly significant contrarily to the one-part model result. Cancer-type included no cancer, female reprod, breast, and others were positively associated with Y > 0. Explicitly, compared with “other cancers,” patients with no cancer, female_reprod, and breast cancers were 3.5873, 3.5158 and 3.1745, respectively. Factor age > 60 years had a negative significance and association with cellulitis number. This result is similar to the one-part model. The overdispersion parameter of ZINB was not significant. That is why ZINB and ZIP yielded the same values, because the ZINB reduced to the ZIP model.

Cellulitis history was used in the zero-part instead of the count part. We can see that cellulitis history had a negative sign −2.1429. That means cellulitis history reducing the probability of a patient to have 0 cellulitis. This makes sense, because people with cellulitis history are more likely to experience recurrence of cellulitis (Y > 0). Figure 5 provides a plot of regression coefficients. It enables comparison the CIs of the coefficients. We can see that the CIs are very similar due to their similarities in value.

Regarding MCA, there are many dimensions with low magnitudes. It requires up to dim 7 to reach 80% of the variance explained. The eigenvalue and the squared eigenvalue (inertia) for each dimension is below 1. The reason is probably due to the fact that the sample size is not large and there are many factors with many levels. In addition, the correlation among factors is probably low. Overall, we can see that a kind of cloud made of cellulitis history, clinical stage 3, age > 60 years, and Y = 0, Y = 1, and cellulitis > 2 seems to be far from the cloud. We can see that clinical stage < 3, cellulitis hist (0) tend to be close to cellulitis number 0. To support our findings, Park et al.,³⁰ found cellulitis history was positively associated with a recurrence of cellulitis. Similarly, Santer et al.³¹ revealed that cellulitis history was increasing the possibility of experiencing recurrences of cellulitis. Using a multiple linear regression, Surrun et al.,³² found age < 74 years, breast cancer, renal impairment, and cellulitis history were positively associated with length of hospital stay (in days) for Singaporean patients suffering from cellulitis. Teerachaisakul et al.,2 used logistic regression to investigate risk factors for cellulitis in patients with lymphedema. Among factors affecting the presence of cellulitis is clinical stage. This supports our results on the effect of clinical stage. They found primary stage was associated with cellulitis. In our study, we found stage 3 was more related to number of cellulitis episodes.

Risk Factors

Many studies have pointed out that the diagnosis of cellulitis is not an easy task. Cellulitis symptoms are complexed and can mimic the symptoms of other diseases.³³ The diagnosis seems to rely on the history and other challenging facts. Cellulitis history has been found in many studies to be an important risk factor for various types of cellulitis. For instance, Raff and Kroshinsky in 2016⁸ and Chand et al. in 2022¹³ reported on risk factors and the global burden of cellulitis. Beside pointing out that cellulitis was associated with more excess volume, fat, and lean arm mass among females with breast cancer, Jørgensen et al.¹⁴ also found that cellulitis history was associated with more excess fat and lean mass among patients with breast cancer-related lymphedema. Al-Niaimi and Cox³⁴ pointed out that the relationship between cellulitis and lymphedema was a vicious cycle where each episode of cellulitis further damages the lymphatic system, leading to a degree of secondary lymphedema, which in turn constitutes an increased risk for cellulitis. This aspect supports the fact that clinical was found to be a significant risk factor in our study. Accordingly, about a quarter of patients with lymphedema will have at least one episode of cellulitis, similar to Jørgensen et al.¹⁴ and Keeley and Riches35 also highlighted the connection between cancers-lymphedema-cellulitis. In effect, lymphedema can be caused by cancer treatments or the cancer itself. On the other hand, lymphedema and cellulitis constitute a vicious circle of cause-effect.³⁴ With regard the association between lymphedema and cellulitis, Cho et al.³⁶ found out that comparing those who had early stage lymphedema, participants who had late stage lymphedema had statistically significant worse scores for symptom body image and HRQoL due to cellulitis. This statement supports our finding that patients in lymphedema clinical stage 3 were more likely to experience more cellulitis episodes that those in stage below 3.

Demographic Variables

With regards to the distribution of age and cellulitis, when studying the management of cellulitis in patients with lymphedema, Keeley and Riches³⁵ reported the distribution of patients revealed that 86% were females and 34% were males; and the average age was about 60 years old. This distributions of age and cellulitis is quite close to what we have in our study. Hsu et al.⁵ found that the distribution of cellulitis in the Taiwan population was 51.7% males compared to 48.2% females. Their sample was collected from the database of the Taiwanese public health insurance system. In our study, there were 81.4% females compared to 18.6% males. This makes our study unique in the sense that these data showed females were more comfortable than males going to that clinic. These are social behaviors we can observe when patients choose physician clinics. If the physician is female, it is possible that most females will go for visits, especially with regard to their health issues. In addition, the distribution of cancer types is another supporting evidence why there are more females than males in this particular hospital clinic data. On the other hand, while investigating the risk of cellulitis in cirrhotic patients in Taiwan, Lin et al.’s sample had more males than females. The larger proportion of males than females might be due to the fact that cirrhosis of the liver is distributed more among males than females. In their study, gender was a significant factor, with males 1.10 times likely to have cellulitis than females. Contrary to our study, gender was not a significant risk factor for the reasons given regarding the distribution of the factor of interest in the population.

Conclusion

This study brought together the one-part model and two-part count regression models to analyze the occurrence of cellulitis episodes among lymphedema patients. Moreover, an additional analysis based on multiple correspondence analysis was conducted to investigate the association between the number of cellulitis episodes with clinical stage and cancer types. We found the probability of cellulitis Y > 0 was associated with clinical stage (3rd), chillness, age > 60 years, breast cancer, female reproductive parts related cancers, and no cancer. We found that cellulitis history was associated with number of cellulitis episodes. In particular, people who have a cellulitis history should be aware and take serious precautions, because cellulitis is likely to be recurrent. The multiple correspondence analysis was been used to provide more information to the regression analysis results: cellulitis hist beast cancer, female_ reprod, and clinical stage, which confirmed the association found in the regression analysis. Future studies will focus on using a zero-inflated Bernoulli, since cloud seems to send away cellulitis (Y > 2, Y > 3, Y > 4, etc). The choice of two-part models over one-part models is more conceptual. A future study can include additional factors such as treatments, blood culture, etc.

Acknowledgement

We thank Chien-liang Ho, MDH, PhD, from the Department of Surgery at the National Cheng Kung University Hospital, for providing the data set.

Footnotes

Disclosures: The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Author Contributions Statement
ML and NQ equally contributed in the data analysis and discussion sections. NQ wrote the program used to analyze the data. NQ also prepared the list of references. ML wrote the introduction, conclusion and methodology sections. ML and NQ both assisted in editing the final version of this manuscript.

Received December 12, 2023.
Revision received June 27, 2024.
Accepted July 1, 2024.

References

1.
↵
Mortimer P, Levine G. Let’s Talk Lymphoedema: The Essential Guide to Everything You Need to Know. Chicago: Elliott & Johnson; 2017.
Google Scholar
2.
↵
Teerachaisakul M, Ekataksin W, Durongwatana S, Taneepanichskul S. Risk factors for cellulitis in patients with lymphedema: a case-controlled study. Lymphology. 2013;46(3):150-156.
Google Scholar
3.
↵
Raff AB, Kroshinsky D. Therapy for Cellulitis. JAMA. 2016;316(19):2047. doi:10.1001/jama.2016.15613
Google Scholar CrossRef
4.
↵
Swartz MN. Clinical practice. Cellulitis. N Engl J Med. 2004;350(9):904-912. doi:10.1056/NEJMcp031807
Google Scholar CrossRef PubMed Web of Science
5.
↵
Hsu RJ, Chou CC, Liu JM, The association of cellulitis incidence and meteorological factors in Taiwan. Epidemiol Infect. 2019;147:e138. doi:10.1017/S0950268819000323
Google Scholar CrossRef
6.
↵
Lee CY, Tsai HC, Kunin CM, Lee SS, Chen YS. Clinical and microbiological characteristics of purulent and non-purulent cellulitis in hospitalized Taiwanese adults in the era of community-associated methicillin-resistant Staphylococcus aureus. BMC Infect Dis. 2015;15:311. doi:10.1186/s12879-015-1064-z
Google Scholar CrossRef
7.
↵
Boettler MA, Kaffenberger BH, Chung CG. Cellulitis: A Review of Current Practice Guidelines and Differentiation from Pseudocellulitis. Am J Clin Dermatol. 2022;23(2):153-165. doi:10.1007/s40257-021-00659-8
Google Scholar CrossRef
8.
↵
Raff AB, Kroshinsky D. Cellulitis: A Review. JAMA. 2016;316(3):325-337. doi:10.1001/jama.2016.8825
Google Scholar CrossRef PubMed
9.
↵
Hilbe JM. Modeling Count Data. Cambridge, MA: Cambridge University Press; 2014.
Google Scholar
10.
↵
Cameron AC, Trivedi PK. Regression Analysis of Count Data. 2nd ed. Cambridge, MA: Cambridge University Press; 2013.
Google Scholar
11.
↵
Lambert D. Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing. Technometrics. 1992;34(1):1-14. https://doi.org/10.1080/00401706.1992.10485228
Google Scholar CrossRef Web of Science
12.
↵
Ridout M, Hinde J, Demétrio CG. A score test for testing a zero-inflated Poisson regression model against zero-inflated negative binomial alternatives. Biometrics. 2001;57(1):219-223. doi:10.1111/j.0006-341x.2001.00219.x
Google Scholar CrossRef PubMed Web of Science
13.
↵
Chand S, Rrapi R, Ko LN, Risk Factors Predicting Cellulitis Diagnosis in a Prospective Cohort Undergoing Dermatology Consultation in the Emergency Department. Cutis. 2022;110(3):122-125. doi:10.12788/cutis.0602
Google Scholar CrossRef
14.
↵
Jørgensen MG, Hermann AP, Madsen AR, Cellulitis Is Associated with Severe Breast Cancer-Related Lymphedema: An Observational Study of Tissue Composition. Cancers (Basel). 2021;13(14):3584. doi:10.3390/cancers13143584
Google Scholar CrossRef
15.
↵
Sourial N, Wolfson C, Bergman H, A correspondence analysis revealed frailty deficits aggregate and are multidimensional. J Clin Epidemiol. 2010;63(6):647-654. doi:10.1016/j.jclinepi.2009.08.007
Google Scholar CrossRef PubMed
16.
Signorini DF. Sample size for poisson regression. Biometrika 1991;78:446-450. https://doi.org/10.1093/biomet/78.2.446
Google Scholar CrossRef Web of Science
17.
↵
Wooldridge JM. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press; 2010.
Google Scholar
18.
↵
Schober P, Vetter TR. Count Data in Medical Research: Poisson Regression and Negative Binomial Regression. Anesth Analg. 2021;132(5):1378-1379. doi:10.1213/ANE.0000000000005398
Google Scholar CrossRef
19.
↵
Eiji Nakashima. Some Methods for Estimation in a Negative-Binomial Model. Annals of the Institute of Statistical Mathematics. 1997;49(1):101-115. https://doi.org/10.1023/A:1003114706239
Google Scholar
20.
↵
Piegorsch WW. Maximum likelihood estimation for the negative binomial dispersion parameter. Biometrics. 1990;46(3):863-867.
Google Scholar CrossRef PubMed Web of Science
21.
↵
Lukusa TM, Lee SM, Li CS. Review of Zero-Inflated Models with Missing Data. Current Research in Biostatistics. 2017;7(1):1-12. DOI:10.3844/AMJBSP.2017.1.12
Google Scholar CrossRef
22.
↵
Agresti A. An Introduction to Categorical Data Analysis. 3rd ed. New York: Wiley; 2013.
Google Scholar
23.
↵
Greenacre MJ, Clarke RT. Theory and Applications of Correspondence Analysis. London: Academic Press; 1984.
Google Scholar
24.
↵
Khangar NV, Kamalja KK. Multiple Correspondence Analysis and its Applications. Electronic Journal of Applied Statistical Analysis. 2017;10(2):432-462.
Google Scholar
25.
↵
Sourial N, Wolfson C, Zhu B, Correspondence analysis is a useful tool to uncover the relationships among categorical variables. J Clin Epidemiol. 2010;63(6):638-646. doi:10.1016/j.jclinepi.2009.08.008
Google Scholar CrossRef PubMed
26.
↵
Collazos J, de la Fuente B, de la Fuente J, Sex differences in hospitalized adult patients with cellulitis: A prospective, multicenter study. Int J Infect Dis. 2021;104:584-591. doi:10.1016/j.ijid.2021.01.044
Google Scholar CrossRef
27.
↵
Vijh R, Shah R, Andany N. A 60-year-old woman with recurrent cellulitis. CMAJ. 2019;191(9):E257-E259. doi:10.1503/cmaj.180686
Google Scholar CrossRef Full Text PubMed
28.
↵
Cheek PJ, McCullagh P, Nelder JA. Generalized Linear Models. 2nd ed. Boca Raton, FL: Chapman & Hall; 1990.
Google Scholar
29.
↵
Shahzad A, Amin M, Emam W, Faisal M. New ridge parameter estimators for the quasi-Poisson ridge regression model. Sci Rep. 2024;14(1):8489. doi:10.1038/s41598-023-50085-5
Google Scholar CrossRef
30.
↵
Park SI, Yang EJ, Kim DK, Jeong HJ, Kim GC, Sim YJ. Prevalence and Epidemiological Factors Involved in Cellulitis in Korean Patients With Lymphedema. Ann Rehabil Med. 2016;40(2):326-333. doi:10.5535/arm.2016.40.2.326
Google Scholar CrossRef
31.
↵
Santer M, Lalonde A, Francis NA, Management of cellulitis: current practice and research questions. Br J Gen Pract. 2018;68(677):595-596. doi:10.3399/bjgp18X700181
Google Scholar CrossRef Full Text PubMed
32.
↵
Surrun SK, Ahmad MT, Afzal S, Tan BH, Siriwong W, Chapman RS. The Risk Factors and Clinical Characteristics of Cellulitis: A Hospital-Based Case-Control Study in Singapore. Journal of Health Research. 2009;23(2):81–86. https://he01.tci-thaijo.org/index.php/jhealthres/article/view/156376
Google Scholar
33.
↵
Keller EC, Tomecki KJ, Alraies MC. Distinguishing cellulitis from its mimics. Cleve Clin J Med. 2012;79(8):547-552. doi:10.3949/ccjm.79a.11121
Google Scholar CrossRef Abstract/Full Text PubMed
34.
↵
Al-Niaimi F, Cox N. Cellulitis and lymphoedema: a vicious cycle. Journal of Lymphoedema. 2009;4(2):2–6. https://woundsinternational.com/wp-content/uploads/sites/8/2023/02/content_11173.pdf.
Google Scholar
35.
↵
Keeley V, Riches K. Cellulitis treatment for people with lymphoedema: UK audit. Journal of Lymphoedema. 2009;4(2):17–17–8, 20, 22–3 passim. https://woundsinternational.com/wp-content/uploads/sites/8/2023/02/content_11169.pdf.
Google Scholar
36.
↵
Cho KA, Cheng MH, Shih WM, Chen SC. Factors associated with health-related quality of life in gynaecologic cancer survivors with lower limb lymphedema: a cross-sectional study in Taiwan. BMC Womens Health. 2023;23(1):200. doi:10.1186/s12905-023-02340-0
Google Scholar CrossRef
37.
Lin MN, Tsai CC, Hung TH, Tsai CC. The risk of cellulitis in cirrhotic patients: a nationwide population-based study in taiwan. Gut Liver. 2012;6(4):482-485. doi:10.5009/gnl.2012.6.4.482
Google Scholar CrossRef PubMed

PDF

Previous Next

Comprehensive Statistical Approach to Investigate the Risk Factors Affecting the Occurrences of Cellulitis Episodes among Patients with Lymphedema

Abstract