Models for predicting skin tears: A comparison

Abstract

A recently published model that predicted the risk of skin tears in older adults was compared with seven additional published models. Four models were excluded because of limitations in research design. Four models were compared for their relative predictive performance and accuracy using sensitivity, specificity, and the area under the curve (AUC), which involved using receiver‐operating characteristic analysis. The predictive ability of the skin tear models differed with the AUC ranging between 0.673 and 0.854. Based on the predictive ability, the selection of models could lead to different clinical decisions and health outcomes. The model, which had been adjusted for potential confounders consisted of five variables (male gender, history of skin tears, history of falls, clinical skin manifestations of elastosis, and purpura), was found to be the most parsimonious for predicting skin tears in older adults (AUC 0.854; 81.7% sensitivity; 81.4% specificity). Effective models serve as important clinical tools for identifying older individuals at risk of skin tears and can better direct more timely and targeted prevention strategies that improve health outcomes and reduce health care expenditure.

1. INTRODUCTION

Skin tears can be significant wounds that have been defined as “trauma‐induced partial or full thickness wounds, which primarily occur on the extremities of older persons with age‐related changes to the skin's structural and mechanical support properties, and are commonly associated with elastosis and/or ecchymosis.”1 Early identification of individuals at risk of skin tears could be facilitated using validated skin tear predictive models, which would enable better targeting of preventative strategies and resources to reduce the incidents of these injuries.

Predictive modelling is the process of data mining for the specific purpose of predicting an outcome that more accurately supports decision‐making and early allocation of preventive approaches.2, 3, 4 The availability of advanced statistical software enables a range of predictive modelling techniques to be applied to clinical datasets after simultaneously exploring multiple independent variables, which forecast their relative contribution to the desired outcome or dependent variables.5, 6 A comparison of models may afford a useful test for evaluating the reliability of predicting skin tear risk among older individuals, regardless of the health care setting. Predictive modelling does not apply a preconceived hypothesis to predict future outcomes, but allows for adjustment of known measured confounders and considers the interaction between variables.5 Statistical analysis is used to explore clinically important variables with ostensibly unimportant variables to show associations that may be both foreseeable or unforeseeable.7

2. METHOD

A detailed search of the literature was conducted between January 2000 and December 2019 of the following electronic databases: PubMed, Medline, CINAHL, Scopus, and Evidence Based Medicine Reviews to identify primary articles that had reported the development of statistical models, which predicted the risk of skin tears in older persons. The inclusion criteria were: English language, primary studies, predictive modelling, ageing, and skin tears. The predictive skin tear models were assessed for their performance using sensitivity, specificity, and the area under the curve (AUC) that involves using receiver‐operating characteristic (ROC) analysis.

3. DATA ANALYSIS

Multiple logistic regression was conducted to assess the independent contribution of individual variables in each study. The ROC curve computed the sensitivity and specificity of the various models in discrimination testing of aged‐care residents with and without skin tears, and a corresponding AUC provided a graphic illustration of the performance of the models.8 Model effectiveness was compared using the AUC to determine the model that best identified older individuals at risk and not at risk of skin tears. The AUC is widely used to clarify the performance of both individual models and for comparing models. Models with a higher ROC curves and a larger AUC values are considered to have a strong effect9 and are therefore better at predicting whether a person is at risk of a skin tear. The Meyers et al10 classification was used to interpret the AUC. Values for the AUC ranged from 0 to 1, with values less than 0.5 indicating no discrimination; 0.6 as poor discrimination; 0.7 as acceptable/good discrimination; 0.8 as very good discrimination; and 0.9 and above as excellent discrimination. The sensitivity and specificity of the AUCs were calculated to determine the proportion of skin tears that the various models correctly classified. Sensitivity and specificity ranged between 0 and 100% with higher values having a higher AUC to indicate better predictability.10

4. RESULTS

The search strategy identified eight skin tear predictive models, all of which have been published in peer‐reviewed journals since 2015. These results are reported in chronological order of year of publication.

Model 1 derived from a 3‐month prospective cohort study of 368 patients, aged 65 years or older, was undertaken by Sanada et al11 in a long‐term medical facility in Japan. Baseline data were collected on a wide range of factors including: pre‐existing skin tears, age, gender, length of hospital stay, Braden Scale scores, body mass index (BMI), immobility, paralysis, articular contracture, medications, comorbidities, and nutritional factors. Despite only a small number of patients (n = 14) having each sustained a single skin tear, multiple logistic analysis identified a history of skin tears (OR = 15.42, CI = 3.53‐67.13) and a 6‐point decrease in the total score of the Braden Scale (OR = 0.10, CI = 0.01‐0.83) to significantly predict the risk of skin tears.

Model 2 stemmed from a 6‐month non‐matched case‐control study conducted by Lewin et al12 in a Western Australian tertiary hospital between 2008 and 2009. Data were collected from 151 cases (patients who developed a skin tear) and 302 controls (patients who did not develop a skin tear but who were admitted to the hospital within 1‐day of the case) aged 50 years or older. Variables were factors reported in the literature to be associated with the risk of skin tears. Multivariate regression analysis identified six parsimonious variables that significantly predicted the risk of skin tears. These variables comprised of one individual characteristic (inability to reposition independently) and five skin characteristics (presence of senile purpura [OR = 2.66, CI 1.47‐4.81], bruising [OR = 6.24, CI = 3.24‐12.01], haematoma [OR = 2.26, CI = 1.30‐3.94], evidence of healed skin tears [OR = 5.42, CI = 2.71‐10.83], and oedema [OR = 3.01, CI = 1.62‐5.61]).12

To externally validate these predictive variables, a prospective cohort study was conducted by Newall et al13 between August 2012 and September 2013 on 1466 Western Australian tertiary hospital patients aged 50 years or older. Subsequent analysis showed the model had high sensitivity (87%), but low specificity (36.1%) for identifying individuals at risk of skin tears. Model 3 resulted from the secondary analysis of data when the non‐matched case‐control study was combined with the prospective cohort study. Model 3 identified senile purpura, haematoma, previously healed skin tears, advanced age, and the ability to reposition independently were better predictors of skin tears.13

Model 4 related to a prospective cohort study conducted by Koyano et al14 of 149 patients aged over 65 years in a long‐term medical facility in Japan over an 8‐month period. Baseline dermal thickness and low‐echogenic pixel skin properties were measured on the forearm using a 20‐MHz ultrasound scanner. Collagen type IV, matrixmetalloproteinase‐2, and tumour necrosis factor‐alpha were measured using an epidermal skin blotting technique.15, 16 Adjusted hazard ratios (HRs) for the main confounders (age, gender, steroid use, history of skin tears, and Braden Scale score) were obtained using the Cox proportional hazard model (Model 4). Dermal thickness (HR = 0.52; CI = 0.33‐0.81) was identified as the only significant predictor of skin tears, with the cut‐off point for dermis thickness being 0.80 mm (AUC = 0.77; CI = 0.66‐0.88).

Model 5 pertained to a point prevalence survey by Bermark et al17 which was conducted in a Danish hospital among 202 patients over a 24‐hour period. Patients ranged in age between 19 and 99 years (mean = 70.7, SD = 16.5). Multiple logistic regression analysis indicated that previous skin tears (OR = 9.3, CI = 2.6‐33.4), ecchymosis (OR = 5.6, CI = 1.4‐23.2), and risk of falling (OR = 3.8, CI = 1.2‐12.0) were significantly associated with development of skin tears (Model 5).

Model 6 was derived from Rayner et al's18 (2019) prospective cohort study that was conducted on 173 residents aged over 65 years with Fitzpatrick skin types I‐IV,19 from four Western Australian residential aged‐care facilities, between January 2014 and June 2015. At baseline, Rayner et al18 evaluated individual and skin characteristics and used a range of non‐invasive technologies to objectively quantify morphological (colour, thickness, and elasticity) and physiological (transepidermal water loss, hydration, and pH) skin properties. Transepidermal skin proteins using skin blotting sampling was taken to measure collagen type IV, matrixmetalloproteinase‐2, and tumour necrosis factor‐alpha.15, 16 Multivariable analysis was conducted to predict the risk of skin tears at 6 months. Having adjusted for potential explanatory variables (age, gender, history of skin tears, history of falls, and BMI), Rayner et al18 produced a predictive model with a set of five baseline parsimonious variables that significantly and independently predicted the risk of skin tears in older adults. The variables included male gender (OR = 3.08, CI = 1.22‐7.77), history of skin tears (OR = 3.82, CI = 1.64‐8.90), history of falls (OR = 3.37, CI = 1.54‐7.41), clinical skin manifestations of elastosis (OR = 3.19, CI = 1.38‐7.38), and purpura (OR = 3.64, CI = 1.42‐9.35).18

Rayner et al20 conducted additional multivariable analysis to understand the inclusion of elastosis and purpura into the risk prediction model. Three individual variables (ageing, gender, and smoking), three clinical skin variables (uneven skin pigmentation, cutis rhomboidalis nuchae, and history of actinic keratosis) and one transepidermal skin property variable (collagen type IV) was found to predict the risk of skin elastosis.20 Cutaneous manifestation of elastosis of the dorsal forearm was associated with photoage‐related skin changes and increased skin stiffness. Photoage‐related skin changes that manifested on exposed skin surfaces altered underlying mechanical properties, increased skin stiffness, and contributed to the risk of skin tears through loss of tissue flexibility. Conversely, four individual characteristics (age, history of skin tears, history of falls, antiplatelet therapy) and three skin properties (pH, subepidermal low echogenicity band of the dorsal forearms, and skin thickness) were found to predict the risk of purpura.20 Purpuric lesions of the dorsal forearm were associated with age‐related skin changes and decreased skin thickness, possibly from loss of dermal collagen. The reduction in skin thickness most likely impaired the skin's structural integrity and ability to resist mechanical forces and contributed to the risk of skin tears.

Model 7, a point prevalence study by Van Tiggelen et al21 (2019), was conducted across 10 Belgian nursing homes over a 6‐month period. Data from patient records and skin observations were collected from 795 residents with a mean age of 85 years (SD: 8.6). In total, 24 residents with 28 skin tears were identified at time of the assessment. Multiple logistic regression analysis indicated that age (OR = 4.03, CI = 1.29‐12.61), history of skin tears (OR = 3.83, CI = 1·30‐11.32), chronic use of corticosteroids (OR = 2.96, CI = 1.06‐8.53), dependency for transfers (OR = 3.74, CI = 1.09‐13.31), and use of adhesives/dressings (OR = 7.05, CI = 2.74‐18.14) were significantly associated with skin tears (Model 7).

Model 8, a cross‐sectional descriptive correlational study of 140 patients aged over 65 years from a Singapore hospital was conducted by Soh et al22 (2019) over a one‐month period. The authors used a modified International Skin Tear Advisory Panel Risk Assessment Pathway and determined that the mean skin tear risk score was 7.19 (SD = 2.40). Multiple linear regression analysis identified six variables that predicted the risk of skin tears in this population. They included age of 75 to 84 years (β = −0.28), age 65 to 74 years (β = −0.26), caregiver‐dependence (β = −0.26), BMI between 18.5 and 23 (β = 0.24), BMI < 18.5 (β = −0.17), and having dementia (β = 0.17).

While eight models were identified, the Koyano et al,14 the Bermark et al,17 the Van Tiggelen et al,21 and Soh et al22 models were excluded from further assessment owing to the type of analysis or research design used. The Koyano et al14 model (Model 4) used survival analysis to calculate the specific effects of skin properties on skin tear occurrence.23 Three models, Bermark et al17 (Model 5), Van Tiggelen et al21 (Model 7), and Soh et al22 (Model 8) used a prevalence study research design that made it difficult to establish the temporal relationship between exposure and the occurrence of skin tears. Only prospective cohort studies were included in this comparison study to minimise the problem of determining the cause and effect, which is associated with prevalence or cross‐sectional studies.24

Detailed assessments of four published models that were based on prospective research and which, predicted the risk of skin tears in older adults were assessed.11, 12, 13, 18 The four recently published articles used logistic regression to combine multiple predictors to estimate the risk of a skin tear injuries in older individuals.11, 12, 13, 18 Baseline data of the four models are presented in Table 1.

Table 1.

Baseline population characteristics for the skin tear models

	Sanada et al11	Lewin et al12	Newall et al13	Rayner et al18
Study design	Prospective cohort	Case‐control	Prospective cohort	Prospective cohort
Number of individuals assessed	368	Cases 151 Controls 302	1466 (data combined with Lewin et al12)	173
Health care setting	Long‐term medical facility	Tertiary hospital	Tertiary hospital	Four aged‐care facilities
Geographical location of study	Japan	Western Australia	Western Australia	Western Australia
Gender Number of males Number of females	94 274	Not reported Not reported	793 781	50 123
Age of population (years)	81‐92	>50	50‐95+	65‐107
Fitzpatrick skin types	Not specified	Not specified	Not specified	I‐IV
Variables adjusted for in the analyses	Age Gender	Not reported	Not reported	Age Gender History of skin tears History of falls body mass index (BMI)

Three models used a prospective cohort design,11, 13, 18 while one used a non‐matched case‐control study design.12 One model was based on research conducted in Japan,11 and three sampled Western Australia aged persons.12, 13, 18 In total, the researchers assessed 2460 individuals aged 50 years and over from two tertiary hospitals12, 13 and five long‐term care facilities.11, 18

Only two of the studies defined a skin tear.11, 18 Sanada et al11 and Rayner et al18 both cited Payne and Martin's25 (p. 19) definition which was “a traumatic wound occurring principally on the extremities of older adults, as a result of friction alone or shearing and friction forces which separate the epidermis from the dermis (partial thickness wound) or which separate both the epidermis and the dermis from underlying structures (full thickness wound).” The four models identified between two and six variables that significantly predicted skin tears in older adults. In total, six individual characteristics (age, history of falls, history of skin tears, inability to reposition, low Braden Scale score, and gender) and six clinical skin characteristics (bruising, elastosis, haematoma, oedema, purpura, and senile purpura) were identified as predictors of the risk of skin tears. The four skin tear models, number of predictors, individual characteristics, and clinical characteristics are presented in Table 2.

Table 2.

Predictors included in the skin tear models

Variables	Sanada et al11	Lewin et al12	Newall et al13 data combined with Lewin et al12	Rayner et al18
No. of predictors identified	2	6	5	5
Individual characteristics
Age			✓
History of falls				✓
History of skin tears	✓	✓	✓	✓
Inability to reposition		✓	✓
Low Braden Scale score	✓
Gender				✓
Clinical characteristics
Bruising		✓
Elastosis				✓
Haematoma		✓	✓
Oedema		✓
Purpura (2‐20 mm)				✓
Senile purpura		✓	✓

A history of skin tears was the only predictor to be included in every model (Table 2). Both Lewin et al12 and Newall et al13 also reported inability to reposition, senile purpura, and haematoma as independent predictors of skin tears. Newall et al13 was the only model to identify age to be a predictor of age‐related skin tears.

The ROC curve and corresponding AUC were used to graphically illustrate and compare the performance of each of the skin tear risk prediction models (Figure 1). Model 6 by Rayner et al18 was subsequently evaluated to determine the importance of the adjusted cofounding variables (age, gender, history of skin tears, history of falls, and BMI) on the prediction of skin tears. This model is now referred to as Models 6a and 6b. Model 6a evaluated the influence of the adjusted variables. Conversely, Model 6b gauged the impact of not adjusting for cofounding variables on predicting the risk of skin tears.

Figure 1.

Comparison of receiver‐operating characteristic (ROC) curves for the five skin tear models

The areas under the ROC curves for each model are listed in Table 3. Both Figure 1 and Table 3 demonstrate that various models differed significantly in terms of sensitivity, specificity, and AUC (0.673‐0.854) in the prediction of skin tears.

Table 3.

Comparison of the area under the receiver‐operating characteristic (ROC) curve and associated 95% confidence interval for the age‐related skin tear risk prediction models

Model	Risk of skin tears
	Area under the curve (AUC)
Model 6a	0.854 (0.797‐0.912)
Model 6b	0.820 (0.756‐0.885)
Model 2	0.673 (0.592‐0.753)
Model 3	0.678 (0.599‐0.758)
Model 1	0.715 (0.639‐0.792)

The Rayner et al18 model (Model 6a) also presented the opportunity to compare genders and the predicted risk of skin tears in individuals aged over 65 years with and without purpura and skin elastosis. Purpura referred to the presence of non‐inflammatory, non‐palpable, blue or purplish skin macules that ranged in diameter from 2 to 20 mm. Conversely, elastotic skin was described as coarse, thickened, scaly, dry, and rigid textured that occurred across exposed skin surfaces.18

Figures 2 and 3 indicate an increased risk of skin tears associated with age for both females (Figure 2) and males (Figure 3) with and without purpura, and with and without elastosis, using the adjusted Rayner et al18 model (Model 6a).

Figure 2.

Predicted risk of skin tears at 6 months for females with clinical manifestations of purpura and elastosis

Figure 3.

Predicted risk of skin tears at 6 months for males with clinical manifestations of purpura and elastosis

5. DISCUSSION

This paper compares the relative performance of published age‐related skin tear risk prediction models in terms of their prediction capacity, ease of measuring variables, and practicality regarding time and expense for use in clinical practice.11, 12, 13, 14, 17, 18, 21, 22 In total, eight published skin tear models were identified, four were excluded because of limitations in research design, only one model had been validated in an external population13 and none were routinely used in the clinical settings. A comparison of the remaining four age‐related skin tear risk prediction models showed significant differences in terms of sensitivity and specificity.11, 12, 13, 18 Rayner et al's18 model (Model 6a), which had adjusted for potential confounders, identified five clinical variables that are measurable at the bedside and appeared to be reliable for predicting older individuals at risk of a skin tear.

The four models identified a range of individual and clinical variables that predicted the risk of skin tears in older adults. Individual variables included age, gender, history of skin tears, history of falls, low Braden Scale score, and inability to reposition independently (Table 2). Clinical variables that predicted skin tears comprised: bruising, elastosis, haematoma, oedema, purpura, and senile purpura (Table 2). These clinical variables are relatively simple to assess, although some appear to be stronger predictors of skin tears than others. A comparison of the models showed that a history of skin tears was a common predictor in each model (Table 2) and an important predictor of skin tears overall. The inclusion of a history of skin tears was not surprising, as age‐related and photoage‐related skin changes are likely to alter the mechanical properties of skin leaving it vulnerable to skin tear injury.18

The four models differed considerably in their predictive ability (Figure 1 and Table 3). Very good discrimination was found for the Rayner et al's18 model, both with (Model 6a) and without adjustment (Model 6b) for potentially confounding variables (age, gender, history of skin tears, history of falls, and BMI). The evaluation found Lewin et al's12 model showed poor discrimination, even though it had the greatest number (n = 6) of predictors, and was not superior to the Sanada et al11 model, which demonstrated good discrimination despite having the least number (n = 2) of predictors. The inclusion of additional predictors may not be warranted as discrimination power has shown to increase with the addition of a predictor after the first predictor has been inserted into a model, but reduces after the inclusion of four or five predictors.26 Likewise, the Newall et al13 model with five predictors showed poor discrimination compared with the Sanada et al11 model with its two predictors.

To date, the Lewin et al's12 model is the only model externally validated, and was found to have high sensitivity (87%), but low specificity (36.1%) for identifying older adults at risk of skin tears.13 The lack of validation of the other three models could relate to the relative recency of predictive modelling in skin tear research.

The model by Rayner et al18 emerged from a prospective cohort study that had extensively collected data about individual characteristics, skin characteristics, morphological and physiological skin properties, and transepidermal skin proteins using skin blotting. This model identified five baseline variables (male gender, history of skin tears in the previous 12 months, history of falls within the preceding 3 months, cutaneous manifestations of elastosis, and purpura [≤20 mm]) to be robust predictors of skin tears among Australians aged 65 years or older. A history of falls, purpura (<20 mm), and clinical skin elastosis was not assessed in the other three earlier models. The predicted risk of skin tears increases with age for both females and males with clinical manifestations of purpura and elastosis (Figures 2 and 3). Purpura and elastosis are therefore considered important clinical predictors of skin tear risk in an Australian elderly population with Fitzpatrick skin types I‐IV.18 The Rayner et al18 predictive model, which had adjusted for cofounding variables, has clinical potential for increased utility among elderly Caucasian populations regardless of care setting, because of its simplicity and practicability. The model should inform evidence‐based clinical decisions leading to implementation of more timely and targeted preventive interventions with concomitant reduction in age‐related skin tears and health care costs.

This comparison paper had several strengths. A comprehensive search of the literature was undertaken to identify models that predicted the risk of skin tears in older adults, and their performance was evaluated. Only prospective cohort studies were included as they provided evidence for the temporal sequence for skin tears and are important factors when assessing causation and predicting risk. Prospective cohort studies also increase the generalisability of findings because of the population base nature of the research.

Although a definition of skin tears was not included in two of the models, this omission was not considered a serious limitation as the authors made reference to ageing skin in their study.12, 13 A consensus on skin tear definition, however, would facilitate better comparison of predictive models and improve their translation into clinical practice.

A limitation of this paper was the comparison of skin tear risk prediction models only published in English journals, which may have underestimated the number of skin tear models published to date and identification of other relevant variables. Differences in study design and methodological rigour may have limited comparisons between the various skin tear predictive models. A final limitation may be the inability to generalise the models across different geographical locations, diverse clinical care settings, or younger cohorts without further testing.

6. CONCLUSION

This paper compared the performance of models designed to identify older individuals at risk of skin tears. The strength of predictors, the ease and practicality of measuring variables, and parsimony are important determinants of general clinical usage of models that predict the risk for skin tears. The paper found that the Lewin et al's12 model was the only predictive model that had been externally validated, while the efficacy of the other models is unknown beyond their respective use in the authors' health establishments. The findings of this paper demonstrated that Rayner et al's18 model (Model 6a), which had adjusted for cofounding variables, had very good discrimination for predicting skin tears. The model includes observed clinical characteristics and readily obtainable data that have the potential to be used by any health care provider, regardless of their level of expertise. Further testing of the performance of all these predictive models is required across different age‐related and geographical populations and health care settings. In addition, this paper, identified several clinical variables that would be warranted for inclusion in the design of any future studies which aim to research new skin tear models and their performance.

Rayner R, Carville K, Leslie G, Dhaliwal SS. Models for predicting skin tears: A comparison. Int Wound J. 2020;17:823–830. 10.1111/iwj.13340