FormalPara Key Summary Points

Why carry out this study?

Atopic dermatitis (AD) is a leading cause of chronic disease burden in skin conditions. The impact of AD on patients’ quality of life (QoL) and day-to-day activities may be important with a strong economic component.

Our study sought to determine what factors predict QoL impairment as measured by the Dermatology Life Quality Index (DLQI) in adult patients with AD using a multivariate machine learning analysis of patient survey data.

What was learned from the study?

Activity impairment was identified as the single most important factor predicting high QoL burden (DLQI>10), followed by hospitalisation during the past year and flare type across several machine learning models. Interestingly, current body surface area involvement was not a strong predictor of AD-related QoL impairment.

Our findings illustrate the role that factors beyond the skin play in determining how patients perceive AD and emphasise the importance of considering the patient’s perspective when assessing AD severity alongside traditional clinical measures, such as extent of disease.

Introduction

Atopic dermatitis (AD) is a chronic, inflammatory skin disorder affecting 2–10% of adults in developed countries [1, 2]. Although long regarded as a paediatric disease, one in four adults with AD report adult onset [3]. These late-onset cases tend to show a higher psychosocial impact [4].

AD is a leading cause of chronic disease burden in skin conditions [5]. The impact of AD on patients’ quality of life (QoL) and day-to-day activities may be important with a strong economic component [6, 7]. Measurement of this burden has been defined as a key domain by the Harmonising Outcome Measures of Eczema (HOME) initiative, with the Dermatology Life Quality Index (DLQI) being the recommended instrument to measure AD-related QoL impairment for adults in AD clinical trials [8].

The DLQI is the most widely used [9, 10], dermatology-specific QoL instrument in the AD literature [11]. Evidence from DLQI assessment across multiple studies indicates that AD significantly impacts QoL [7, 12]. Up to 55% of patients report suffering at least a moderate impact on QoL relating to their AD [13]. Higher DLQI impairment is associated with higher disease severity based on patient-reported assessment tools such as the Patient-oriented Eczema Measure (POEM) [14], the Patient-Oriented-Scoring of Atopic Dermatitis (PO-SCORAD) [15, 16] or patient-assessed severity, underlining the importance of the patients’ perspective [11]. Interestingly, QoL assessed by the DLQI, along with other QoL measures, correlates only moderately with physician-assessed AD severity measures such as the Eczema Area and Severity Index (EASI) [17, 18], objective Scoring of Atopic Dermatitis (SCORAD) or lesion extent (‘rule of nines’) [19]. The DLQI is increasingly used in clinical practice [20, 21] and supports decision making on whether to initiate or continue systemic therapy [22,23,24,25]. Consequently, research is needed to better understand what drives QoL burden in patients with AD.

Machine learning has become an important tool with decision making applications [26, 27]. It has a greater ability to predict patient outcomes than conventional analytical frameworks [28]. In dermatology, machine learning may support physicians in managing their patients from diagnosis through to personalised therapies [26].

The current study aims to report data from a large, international, patient satisfaction and unmet need survey of patients with AD and to use a machine-learning approach, novel in AD research, to identify factors influencing AD-related QoL impairment measured by the DLQI.

Methods

Study Design

An international cross-sectional survey was conducted in ten countries (Australia, Belgium, Canada, France, Germany, Italy, Japan, the Netherlands, Spain and the United Kingdom) between July and September 2019. Participants were recruited via internet panels and completed a web-based survey, conducted by Hall and Partners alongside Toluna and Axanteus. This survey collected patient-reported data on demographics, disease characteristics and disease burden using established patient-reported outcomes (supplementary materials appendix 1: Atopic dermatitis patient satisfaction and unmet need survey).

Study Population

Participants needed to be adults aged 18–75 years (20–75 years in Japan) and have a dermatologist-confirmed diagnosis of AD (including immunologist in Australia). Patients were asked to rate their AD as mild, moderate or severe. Patients estimated the extent of their affected body surface area (BSA) by reporting the number of palms needed to cover areas affected by AD, where 1% BSA was estimated as one palm. Patients whose self-rated AD was mild and had less than 3% body surface area (BSA) involvement were excluded.

Compliance with Ethics Guidelines

Participants were recruited under a market research code of conduct. The research was carried out in compliance with national laws protecting respondents’ personal data and with guidance from the European Society for Opinion and Market Research, the European Pharmaceutical Marketing Research Association and the British Healthcare Business Intelligence Association. Participants provided informed consent, consent to be included in the analysis, and consent for reporting of adverse events as part of their survey return and were provided with a small payment (approximately $30) to compensate them for the time spent answering the survey.

As outlined above, the research described in our study was conducted under the auspices of a market research code of conduct rather than in a clinical setting. While the 1964 ICH Declaration of Helsinki does not apply, all consents received were freely given and participants could withdraw from any/all aspects of the research whenever they liked.

Assessment of AD-Related Quality of Life: Dermatology Life Quality Index (DLQI)

Respondent’s AD-related QoL impact was assessed using the DLQI. The DLQI comprises 10 items across six domains: symptoms and feelings; daily activities; leisure; work and school; personal relationships; and treatment assessed over the last week. Each question is rated from 0 (not affected) to 3 (very much affected), resulting in total scores from 0 to 30 [9]. Validated score band descriptors define 0–1 as having no effect, 2–5 having a small effect, 6–10 having moderate effect, 11–20 having a very large effect, and 21–30 having an extremely large effect on patients’ lives [29].

Predictive Variables

Machine learning models were applied to predict high levels of QoL impairment, defined as DLQI > 10 versus DLQI ≤ 10. Predictor variable selection from the information available in the survey was based on medical knowledge and the scientific literature. On the basis of findings from the predictive model, we visualized the relationship between dichotomized DLQI and important factors to further characterize our findings. The following potentially relevant predictors assessed in the survey were included:

Demographics

Information on age and gender.

Hospitalisation

Hospitalisation was defined as whether a patient had been hospitalised due to their AD within the last year (yes/no).

Disease extent

Patients estimated the extent of their affected BSA by reporting the number of palms needed to cover areas affected by AD, where 1% BSA was estimated as one palm. BSA involvement was rated: currently; best in the last year; and worst in the last year.

Activity impairment

Activity impairment (other than work) due to AD was measured using the Work Productivity and Activity Impairment Questionnaire: Atopic Dermatitis instrument (WPAI-AD) [30].

Flare characteristics

Flare characteristics included in this analysis were: no flares; seasonal flares; occasional mild flares; frequent moderate flares; severe, frequent flares; and chronic severe flares. The questionnaire also asked about flare frequency over the previous year. Flare triggers were grouped into three categories: environmental (e.g. climate, allergens such as pollen, pet or fragrances), psychological (e.g. stress at work or at home or lack of sleep) and intrinsic (e.g. dry skin, hormonal changes or being ill).

Affected body regions

Affected areas were reported on the basis of the question ‘Where on your body do you usually have AD?’ Respondents could select multiple locations which were grouped during analysis into head/neck, arms/hands, core, and legs/feet.

Treatments

The survey asked respondents about current treatments for their AD, including topical corticosteroids (TCS), topical calcineurin inhibitors (TCI), systemic corticosteroids, systemic immunosuppressants (including cyclosporine A, methotrexate, azathioprine, and mycophenolate mofetil), and biologics (including dupilumab). Topical therapy use was assessed as was systemic therapy use, both as monotherapy or in combination with topical therapy.

Statistical Analyses

The results of the patient survey including socio-demographic, disease and treatment information were characterized descriptively. Machine learning models were applied to the data with dichotomised DLQI as the response variable to identify factors most predictive of AD-related QoL burden Data were split randomly into training (75% sample) and testing (25% sample) subsets. Only patients without any missing data were entered in the machine learning analyses. All models were fitted on the training set and validated using repeated cross validation with 50 repeats and 5 splits. The trained and validated model parameters of each model were used on the test set to ensure unbiased prediction. Optimal parameter tuning was performed by using grid search algorithms in the train function of caret package of the statistical program, R32.

Final model fit was provided by applying the training dataset on the test data to give an unbiased model estimate. This methodology was applied to eight machine learning procedures: logistic regression, stochastic gradient boosting, neural network, random forest, decision tree, sparse linear discriminant analysis, bagged decision tree and gradient boosted trees algorithm (See Supplemental Fig. 1).

Models were selected on the basis of their predictive performance by assessing prediction accuracy measures including positive predictive value (PPV), negative predictive value (NPV), sensitivity, specificity and overall accuracy. DeLong’s test compared models’ receiver operating characteristic (ROC) curves. Critical p-value was taken to be < 0.05 for a two-sided test.

To assess the contribution of each feature to predictive performance, we computed the variable importance value of each predictor for the corresponding model and visualized it. The variable importance measures were scaled 0–100.

Results

Patient Demographics and Disease Characteristics

In total, 2314 patients from 10 countries (Australia, n = 201; Belgium, n = 194; Canada, n = 250; France, n = 250; Germany, n = 250; Italy, n = 252; Japan, n = 252; the Netherlands, n = 165; Spain, n = 250, United Kingdom, n = 250) completed the survey (Table 1).

Table 1 Demographics and disease characteristics of survey respondents
Full size table

Mean age at AD diagnosis (Table 1) was 20.2 years and mean disease duration was 19 years. Patients suffered from long-standing disease (mean 19 years), with half of the patients first diagnosed during adulthood. Mean BSA involvement was 5.9%. Accordingly, 87% of patients had a BSA > 3% and < 10%, indicative of mild disease, while 13% had a BSA ≥ 10%, consistent with moderate-to-severe disease. Only 31.7% of patients rated their AD severity as mild, while the majority perceived their disease as moderate (58.1%) or severe (10.2%). The most commonly reported involved body regions were the face (43%; Fig. 1a), the scalp (40%), and the neck (37%). Genital AD involvement was reported in 11% of participants. Participants experienced on average 7.5 flares and were hospitalised because of their AD on average 2.7 times in the previous year (Table 1). The mean age of participants was 39 years and 74.5% reported full-/part-time employment.

Fig. 1
figure 1

Sites affected by atopic dermatitis and patient reported DLQI score. DLQI Dermatology Life Quality Index

Full size image

QoL Impairment in Patients with AD Based on DLQI Assessment

Mean DLQI indicated a moderate-to-large effect on QoL (Table 1). Overall, 35% of patients reported DLQI ≤ 5 consistent with a small/no impact on QoL, while 44% had DLQI > 10, indicative of very-large-to-extremely-large impact on QoL (Fig. 1b).

Prediction of QoL Impairment in Patients with AD

The predictive analysis comprised 2098 respondents without any missing data information. Of these, 1192 (56.8%) had DLQI ≤ 10 and 906 (43.2%) DLQI > 10. All applied machine learning models showed high accuracy, PPV and NPV with a range of 74–78%, 70–75% and 77–81%, respectively (Table 2). Accuracy measures were consistent across models, indicating robustness and model independence of the data on a large scale. The logistic regression model, the random forest and the neural network were selected (Fig. 2) on the basis of cross-model predictive performance and DeLong’s test.

Table 2 Accuracy of assessed modelling
Full size table
Fig. 2
figure 2

Importance matrices for factors associated with DLQI ≥ 10 by modelling methodology – logistic regression model, random forest model and neural network. DLQI Dermatology Life Quality Index, WPAI AI Work Productivity and Activity Impairment Questionnaire Activity Impairment subscale, BSA body surface area; Conventional Sys Conventional systemic treatment. All factors are indexed against the most important factor and were scaled 0–100

Full size image

On all three selected models, the activity impairment subscale of the WPAI-AD (WPAI AI) was the most important factor for very-large-to-extremely-large impacts of AD on QoL (DLQI > 10). Hospitalisation during the past year was also highly ranked across models. Flare type ranked third on both the logistic regression model and the random forest and fourth on the neural network. Cross-model ranking was less consistent for other attributes. BSA coverage (best), the most highly ranked of the three BSA factors on the neural network, only ranked eighth. Correspondingly, BSA (current) was highest ranked.

Head and neck involvement was the highest ranked body region on the logistic regression model and the random forest. Core was the highest ranked on the neural network. Amongst current treatments, topicals as monotherapy had the highest impact on patient’s QoL. While gender ranked fifth on the logistic regression model, age and gender were not highly ranked on any other model.

Characterisation of Activity Impairment as Single Most Important Factor Impacting AD-Related QoL

Activity impairment was the most important factor determining impaired AD-related QoL across the three selected models. We aimed to further examine this finding. In those with DLQI ≤ 10, median percentage activity impairment was 20%, while in those with DLQI > 10, median percentage impairment was 60% (Fig. 3a).

Fig. 3
figure 3

Characterisation of variable importance matrix factors by DLQI status. DLQI Dermatology Life Quality Index, BSA body surface area

Full size image

Hospitalisation for AD as a Predictor of AD-Related QoL

The proportion of patients who were hospitalized in the past year due to their AD was higher for those with higher DLQI impairment (Fig. 3b). Those with DLQI ≤ 10 had been hospitalized in 7% of cases. This percentage was 32.6% for those with DLQI > 10.

Severe Frequent Flares as Predictor of AD-Related QoL

The percentage of patients with AD experiencing frequent severe flares was higher in those with DLQI > 10 (25.5%) than those with DLQI ≤ 10% (7.1%), while similar percentages of those experiencing frequent, moderate flares were observed across both DLQI groups (39.8% and 37%, respectively). Occasional, mild flares were more frequent in the DLQI ≤ 10% group (25.2%% versus 11.6%) while there were only modest differences in seasonal flares (30.7% in DLQI ≤ 10 versus 23.1% in DLQI > 10).

Current Treatment with Topical Monotherapy as Predictor for Impaired QoL in Patients with AD

More than two-thirds (71.4%) of those with DLQI ≤ 10 were treated with topical therapies alone. While this proportion was lower amongst those with DLQI > 10, 44.8% of respondents were still managed in this way. With respect to systemic therapies, more respondents with a DLQI > 10 (30.6%) were managed with systemics than those with a DLQI ≤ 10 (18.9%), though such patients represented only a small proportion of either group.

BSA Involvement: An Inconsistent Predictor of AD-Related QoL Impairment

Mean BSA in the DLQI > 10 group (7.1%) was only modestly higher than in those with DLQI ≤ 10% (4.6%). Thus, when stratifying using the standard BSA cut-off of BSA ≥ 10% [31], while 16% of those with DLQI > 10 reported moderate-to-severe BSA, such patients still represented 9.6% of those with DLQI ≤ 10 (Fig. 3e). In both DLQI groups, patients with BSA < 10% represented the vast majority. This indicates that even mild disease, based on skin extent, could be associated with a large decrement in QoL for patients, and underlines our finding that current BSA alone is a weak predictor of AD-related QoL impairment.

Discussion

This is the first study to apply machine-learning methods to data from a survey of patients with AD to identify predictive factors for AD-related QoL impairment. We found that nearly half (44%) of our sample experienced a high level of burden as measured by the DLQI with a further 21% reporting moderate burden. Burden was largely independent of both clinical measures of disease extent, such as BSA, and patient self-reported disease severity. The vast majority of those sampled (86.7%) had an affected BSA > 3% and < 10% while 90% rated their disease severity as mild or moderate.

We assessed various factors including disease extent measured by BSA involvement as well as patient-reported outcomes, all of which might contribute to QoL impairment. Activity impairment outside of the workplace was identified as the single most important factor identifying patients with high AD-related QoL impairment across all three predictive models. In those patients, the impact of their AD on their ability to perform their usual activities was most pronounced. These findings show that activity impairment as measured by the WPAI correlates with DLQI and with AD disease severity [32, 33]. Lifestyle limitations and avoidance of social situations are frequently reported by patients with AD [7]. Higher levels of activity impairment have been associated with severity of symptoms such as sleep impairment, as well as mental health symptoms such as anxiety and depression [34]. The current results highlight the broader impact of AD on patients’ everyday lives and show that this factor is the single most important factor contributing to the high QoL burden for patients with AD. For patients, AD burden goes far beyond skin signs.

The profound effect of activity impairment on QoL may be linked to the unremitting, unpredictable nature of AD. This would be congruent with flare type as an additional important predictor of higher levels of AD-related QoL impairment. Our respondents reported on average 7.5 flares per year, described mostly as frequent and moderate or as frequent and severe in those with poor QoL. Similarly, it has been reported that patients with AD were experiencing flares for more than one-third of the year, with flares impacting daily activities, limiting social interactions and leading to depression and feelings of frustrations and helplessness [35].

AD-related hospitalisation in the last year was reported by 18% of respondents. While this may seem high given the disease severity found in our sample, it is not very different from that seen in the international literature. Eckert and colleagues using data from a national survey found 0.3 (SD 1.0) mean hospitalisations in the past 6 months [36]. Such AD-related hospitalisation was associated with higher levels of AD-related QoL impairment across all our selected models. This conforms with recent research showing a significant increase in emergency room visits and hospitalisations with DLQI decrements [34].

An important finding of our analysis was that extent of BSA involvement did not strongly predict QoL impairment. Traditionally, BSA extent has been important in determining AD disease severity and in guiding treatment decisions. Our findings indicate that the association between BSA extent and QoL is limited, with many patients reporting low BSA extent still experiencing higher levels of QoL burden and rating their disease as being more severe than their BSA would indicate. Recently, lesion location was shown to significantly impact DLQI, especially if visible areas were affected [37]. In our analysis, affected body region was not among the most important factors impacting QoL despite the fact that the face was the most commonly affected area. One explanation could be different timeframes for assessing DLQI and body region involvement, with the DLQI measuring QoL in the past week while BSA was measured at various points over the previous year.

Our analysis is unique in its multivariate approach using machine learning and predictive models to assess AD-related QoL burden using the DLQI. This approach can assist health care providers to identify patients with high burden and to synthesize all the information gathered concerning the patient and their disease with the goal of optimizing clinical decision making and treatment choices for patients with AD.

Strengths and Limitations

Machine learning algorithms aim to achieve high predictive accuracy compared with standard regression models. All machine learning models applied in the study performed well, showing that the results of our study are robust and model independent. Incorporating multivariate factors into the models rather than only looking at univariate data or bivariate associations between two factors increases the likelihood of a good prediction. Data-driven approaches such as this study provide support to physicians on decision making in their routine clinical practice.

While the DLQI is a widely used, validated QoL outcome measure and its use could aid the generalisability of our findings, it has limitations [38]. It might not capture all QoL dimensions, with emotional aspects of wellbeing under-represented [10]. Additionally, some individual components include ‘not relevant’, which scores the same as ‘not at all’. This might lead to an underestimation of QoL in those responding ‘not relevant’ [39]. In our sample, the only category with a sizeable number of ‘not relevant’ responses related to work and study, which might have led to underestimation of this aspect [40].

Study limitations include a potential for selection and information bias for the survey, the self-reported nature of the instruments, and the confirmation by a managing clinician of an AD diagnosis with risks for recall and reporting biases. Other limitations include timeframe differences (e.g. DLQI assessed within the previous week, while affected locations being assessed over the previous year) and that some potential sources of AD QoL burden were not investigated, such as itch intensity or severity.

Conclusions

AD is an unremitting, unpredictable condition that imposes a considerable burden on patients. Our work shows that activity impairment and hospitalisation were associated with higher levels of AD-related QoL impairment, while traditional clinical decision making tools such as BSA were less predictive of this burden. Our findings illustrate the role that factors beyond the skin play in determining how patients perceive AD and emphasise the importance of considering patient-perspective disease impact when assessing AD severity alongside traditional clinical measures.