Development of Psoriasis Assessment Tools Among Patients in the CorEvitas Psoriasis Registry

Abstract

Background

Dermatologists would benefit from an easy to use psoriasis severity assessment tool in the clinic.

Objective

To develop psoriasis assessment tools to predict PASI and Dermatology Life Quality Index (DLQI) using simple measures typically collected in clinical practice.

Methods

Data included 33 605 dermatology visits among plaque psoriasis patients enrolled in the CorEvitas Psoriasis Registry (4/15/15-7/11/20). Performance (adjusted coefficient of determination [R²_adj], root mean square error [RMSE]) in predicting PASI and DLQI was assessed for 16 different linear regression models (specified a priori based on combinations of BSA, Investigator’s Global Assessment [IGA], itch, skin pain, patient global assessment, age, sex, BMI, comorbidity index, prior biologic use), and 2 stepwise selection models and 1 elastic net model based on 56 available variables. For each prediction model, concordance (sensitivity, specificity) of predicted PASI75, PASI90 and DLQI 0/1 with observed values was evaluated.

Results

Mean (SD) age, BSA, and PASI were 51 (14) years, 6 (11), and 4 (6), respectively; 46% were women, and 87% were biologic experienced. A model predicting PASI using BSA plus IGA performed best among a priori specified models (R²_adj = .72, RMSE = 2.93) and only marginally worse than models including additional variables (R²_adj range .64-.74, RMSE range 2.82-3.36). Models including IGA had the best concordance between predicted and observed PASI75 (sensitivity range 83-85%, specificity range 88-91%) and PASI90 (sensitivity range 76-82%, specificity range 94-98%). DLQI prediction was limited.

Conclusion

An assessment tool for psoriasis including BSA and IGA may be an ideal option to predict PASI in a clinic setting.

Introduction

There are several psoriasis disease severity clinical measurement tools available, but no instrument meets all validity criteria.^1-3 The Psoriasis Area Severity Index (PASI) is considered the gold standard for disease severity measurement in clinical trials. It estimates disease severity using a summary score based upon qualitative assessments of plaque redness and thickness, and body surface area (BSA) across four body sites. Yet the PASI has some limitations, including a lack of consensus on interpretability, low response distribution (especially in mild –to-moderate disease (BSA 3-10%), the BSA component is non-linear,^4-6 and is not useful in measuring non-plaque psoriasis subsets.^7-10 Further, it is time-consuming to ascertain, and is therefore rarely used in real-world practice.^5,6

Simpler measures are available such as BSA and the Investigator’s Global Assessment (IGA). Still, neither alone is likely sufficient. BSA captures skin involvement but not plaque qualities (eg, erythema, induration, scale), and the opposite is true of IGA. Additionally, none of these measures assess the patient’s perspective of disease, and treatment guidelines recommend including assessment of health-related quality of life (HRQoL).^11,12 Therefore, patient care would benefit from a psoriasis severity assessment tool that is easy to use in the clinic and gives a more comprehensive assessment of disease burden such that it is correlated with HRQoL measures, such as the Dermatology Life Quality Index (DLQI).

To fill this need, a prior study evaluated an optimal psoriasis assessment tool (OPAT™), based on the combination of BSA and simple patient reported outcome measures (PROM), as an alternative to PASI. ¹³ Findings suggested that a proxy PASI score calculated using only these inputs was predictive of an observed PASI assessment and DLQI, in 3800 patients from three UNCOVER randomized control trials. ¹⁴ However, this study comprised a relatively homogeneous, moderate-to-severe psoriasis population participating in clinical trials of ixekizumab, etanercept, or placebo. Thus, it is unclear how well the combination of BSA and a PRO can predict PASI or DLQI in the broader population of patients using systemic therapy in real-world dermatologic practices.

Therefore, to address this gap, the objective of our study was to develop an assessment tool for psoriasis utilizing data from a real-world cohort of psoriasis patients being treated with systemic therapies, representing a more heterogeneous population of patients seen in dermatology practice. Our goals were to (1) develop scales intended to enhance and ideally optimize the assessment of psoriasis that calculate “predicted PASI” and “predicted DLQI” based on BSA combined with other measures that are easily ascertained in a clinic setting, and (2) use predicted PASI and DLQI at therapy initiation and 6-months follow-up to calculate predicted PASI75, PASI90, and DLQI 0/1, and evaluate their agreement with these observed measures.

Methods

Registry Overview

The CorEvitas Psoriasis Registry is a prospective, multicenter, observational disease-based registry launched in April 2015 in collaboration with the National Psoriasis Foundation, the design of which has been previously described. ¹⁵ Briefly, patients are recruited by participating dermatologists from private and academic practices in the US and Canada across 46 states and provinces. Patients are enrolled if they meet the following criteria: aged ≥18 years, psoriasis diagnosed by a dermatologist, and initiated or switched to a US Food and Drug Administration (FDA) approved systemic psoriasis treatment at or within 12 months of enrollment. Data is collected from both dermatologists and patients during routine clinical visits occurring at approximately 6-month intervals.

All participating investigators were required to obtain full board approval for conducting research involving human subjects. Sponsor approval and continuing review was obtained through a central IRB (IntegReview, Protocol number is Corrona-PSORIASIS-500). For academic investigative sites that did not receive a waiver to use the central IRB, approval was obtained from the respective governing IRBs, and documentation of approval was submitted to the Sponsor before initiating any study procedures. All registry patients were required to provide written informed consent prior to participating.

Study Population

This analysis included data from the 10 961 patients enrolled in the registry with plaque psoriasis and no history of pustular morphology from April 15, 2015 to July 11, 2020. These patients contributed 33 605 registry visits to the analyses, including both enrollment and follow-up visits.

Study Variables

Outcomes

PASI and DLQI were collected at all registry visits. PASI considers percentage of area affected by psoriasis and the severity of redness, thickness, and scaling of the skin, and is measured on a scale 0-72 where a higher score indicates greater severity. ¹⁶ The DLQI consists of 10 questions concerning patients’ perception of the impact of skin diseases on different aspects of their HRQoL over the last week. The 10 items include areas such as symptoms and feelings, daily activities, leisure, work or school, personal relationships, and the side-effects of treatment. ¹⁷ Respondents indicate the degree of problems in the past week using a 4-point Likert scale: 0 (not at all/not relevant), 1 (a little), 2 (a lot), and 3 (very much). A total DLQI score is calculated, ranging from 0-30 with higher scores indicating worse HRQoL.

Predictors of PASI and DLQI

BSA is the percent (0-100%) of the total area of the body affected by psoriasis. ² IGA is a five-point scale that provides a global clinical assessment of disease severity ranging from 0 to 4: 0 (clear), 2 (mild), 3 (moderate), and 4 (severe disease). ¹⁸ Patient reported outcome measures (PROMs) included itch/pruritis and skin pain reported on a Visual Analog Scale (VAS) of 0 (none) to 100 (very severe), and a Patient Global Assessment (PGA) for psoriasis reported on a VAS of 0 (very well) to 100 (very poor).

Other variables included age (years), sex, race (White vs non-White), body mass index (BMI, kg/m²), the Modified Charlson Comorbidity Index (mCCI) score, ¹⁹ and prior biologic experience (biologic-naïve vs biologic-experienced).

Statistical Analysis

The data from the 33 605 registry visits were partitioned into three mutually exclusive datasets to address study objectives. First, patient-visits at which a biologic therapy initiation occurred and had a subsequent 6-month follow-up visit were identified (3267 initiation/follow-up visit pairs; 6534 total visits), and a 40% random sample of visits with complete data (1303 visit pairs; 2606 total visits) was selected as the dataset to evaluate agreement between the predicted PASI and DLQI outcomes obtained using psoriasis assessment models and measured PASI and DLQI outcomes. Next, the patient-visits not selected for concordance analyses (n = 31 041) were partitioned via random sample into training (60%, n = 17 185 patient-visits) and testing (40%, n = 13 856 patient-visits) datasets for psoriasis assessment model development and internal validation, respectively.

Of the 17 185 patient visits in the psoriasis assessment model development training data set, 16 840 patient-visits with complete data were used to describe patient characteristics via means and standard deviations for continuous variables and counts and percentages for categorical variables. Additionally, these patients were used to describe the associations among psoriasis disease severity and PROMs by calculating Pearson correlations among BSA, PASI, IGA, itch, skin pain, PGA, and DLQI, which were visualized as a heatmap.

Psoriasis Assessment Model Development

To construct psoriasis assessment models to calculate predicted PASI and DLQI, separately, we specified a priori 16 linear regression models based on different combinations of the following predictors: BSA, IGA, itch, skin pain, PGA, age, gender, BMI, mCCI, and biologic experience. Models were specified by starting with a model with BSA alone, followed by models adding only IGA, or each PROM (itch, skin pain, PGA), additional models adding interactions of BSA with IGA or each PROM, and finally models adding other covariates. Three additional prediction models were constructed utilizing variable selection methods (not considering interactions) for the previously listed predictors: two linear regressions using stepwise selection procedures (forward selection and backwards elimination) via the Akaike information criterion (AIC), and one regularized elastic net regression to select from all available variables plus interactions of BSA with IGA and PROMs. Elastic net models included a shrinkage term to reduce overfitting in the development set to improve the predictive ability; this penalty shrinks coefficients towards zero, and automatically conducts variable selection. Thus, in total there were 19 different prediction models constructed. Models were run in the training dataset, and the predictive performance of all 19 models was evaluated in the testing dataset by calculating the adjusted coefficient of determination (R²_adj), root mean square error (RMSE), and mean absolute error (MAE). Ninety five percent confidence intervals (95% CI) for these estimates were approximated by replicating the sample splitting, model development, and model testing 400 times, and utilizing the 2.5th and 97.5th percentiles as the lower and upper bounds, respectively. Further, as a sensitivity analysis, mixed-effects model variants of the models were created by adding random effects for patient identification to adjust for multiple observations per patient.

Concordance of Predicted and Measured PASI and DLQI Outcomes

The dataset created for evaluating concordance of the psoriasis assessment model that calculated and observed PASI and DLQI outcomes was used to compare predicted values of PASI and DLQI with measured values for each of the 19 models. For each patient, predicted PASI and DLQI were calculated at both the biologic initiation and 6-month follow-up visits, separately, using each of the 19 prediction models from the development phase. Using these predicted values, predicted PASI75, PASI90 and DLQI 0/1 at 6-month follow-up were calculated. Patients achieved PASI75/90 if their PASI decreased by 75%/90% from baseline to follow-up. Patients achieved or maintained DLQI 0/1 if their DLQI was ≤1 at follow-up. Additionally, observed PASI75, PASI90, and DLQI 0/1 were calculated using observed PASI and DQLI at initiation and follow-up. The observed and predicted values from each prediction model were compared using sensitivity (SE), specificity (SP), negative predictive value (NPV), and positive predictive value (PPV).

To evaluate psoriasis assessment models constructed utilizing a population of patients with severe psoriasis, all models were repeated and applied to a restricted test set of patients with PASI ≥12, BSA ≥10, and moderate-to-severe IGA (3-4).

Ethics

This study was carried out in accordance with the Declaration of Helsinki. All participating investigators were required to obtain full board approval for conducting noninterventional research involving human subjects with a limited dataset. Sponsor approval and continuing review was obtained through a central Institutional Review Board (IRB), the New England Independent Review Board (NEIRB; no. 120160610). For academic investigative sites that did not receive a waiver to use the central IRB, full board approval was obtained from the respective governing IRBs and documentation of approval was submitted to CorEvitas, LLC prior to the initiation of any study procedures. All patients in the registry were required to provide written informed consent and authorization prior to participating.

Results

Among the patient-visits with complete data in the psoriasis assessment model training dataset, mean (SD) age, PASI, and DLQI were 51 (14) years, 4 (6), and 4 (5) respectively; 46% were female, 81% were White, and 87% were biologic-experienced (Table 1). PASI was strongly correlated with BSA (r = .78) and IGA (r = .66), and moderately correlated with DLQI (.46), PGA (.44), itch (.47), and skin pain (.43) (Figure 1). DLQI was moderately correlated with IGA (.50), PGA (.57), itch (.59), skin pain (.59), and BSA (.41). Other strong correlations were observed among PGA, itch, and skin pain: .66 for PGA and itch, .61 for PGA and skin pain, and for .71 itch and skin pain.

Table 1.

Demographics and Clinical Characteristics Among Patient-Visits in the CorEvitas Psoriasis Registry; Patient-Visits With Complete Data in the Psoriasis Assessment Model Development Training Data Set.

Total	N = 16840*
Age (years), mean (SD)	51.4 (14.4)
Gender – Female, n (%)	7707 (45.8%)
Race - White n (%)	13618 (81.0%)
BMI (kg/m2), mean (SD)	31.0 (7.5)
BMI (kg/m2) - ≥30 (obese)	8202 (48.7%)
Psoriasis duration (years), mean (SD)	16.7 (13.9)
mCCI, mean (SD)	1.3 (.6)
BSA (% involvement), mean (SD)	5.8 (10.9)
BSA - categorical
Mild disease [0, 3]	9518 (56.5%)
Moderate disease [3, 10]	4079 (24.2%)
Severe disease [10, 20]	1888 (11.2%)
Very severe disease [20 100]	1355 (8.0%)
PASI mean (SD)	3.5 (5.6)
PASI >10	n = 16840
n (%)	1677 (10.0%)
IGA
Clear	4364 (25.9%)
Almost clear	3716 (22.1%)
Mild	3739 (22.2%)
Moderate	4064 (24.1%)
Severe	957 (5.7%)
Number of previous biologic therapies
0	2106 (12.5%)
1	6256 (37.1%)
2+	8478 (50.3%)
DLQI (score: 0-30), mean (SD)	4.4 (5.5)
DLQI - categorical
No Effect at all	7417 (44.2%)
Small effect	4503 (26.8%)
Moderate effect	2576 (15.3%)
Very large effect	1925 (11.5%)
Extremely large effect	373 (2.2%)
Patient global assessment, mean (SD)	27.7 (28.8)
Patient overall itch/pruritis (VAS range 0-100), mean (SD)	27.4 (31.6)
Patient overall skin pain (VAS range 0-100), mean (SD)	16.4 (25.8)

Includes only patient-visits with complete data among the 17 185 visits in the training dataset; SD, Standard Deviation; BMI, Body Mass Index; mCCI, modified Charlson Comorbidity Index (The modified Charlson comorbidity index is calculated by giving 1 point to the following conditions: myocardial infarction, congestive heart failure, peripheral vascular disease, TIA and/or stroke, chronic obstructive pulmonary disease (COPD), peptic ulcer disease, diabetes, leukemia, lymphoma, solid tumor, and liver disease); BSA, Body Surface Area; PASI, Psoriasis Area Severity Index; IGA, Investigator’s Global Assessment; DLQI, Dermatology Life Quality Index; VAS, Visual Analogue Scale (Range 0-100).

Figure 1.

Heatmap of Pearson correlations of BSA, PASI, itch, skin pain, PGA, and DLQI among the 16 840 patient-visits in the psoriasis assessment model development training dataset.

Psoriasis Assessment Model Development

For the a priori specified models predicting PASI, BSA alone explained 62% of the variance in PASI (R²_adj = .62; 95% CI = .60, .64). Predictive performance improved only slightly with the addition of a PRO (itch, skin pain, PGA) to the model (all R²_adj = .64, RMSE range 3.32-3.34) (Table 2). The model with BSA and IGA performed marginally better compared to that with BSA alone (R²_adj = .72; RSME = 2.93). For all the a priori selected models, including additional variables (interaction terms, age, gender, mCCI, BMI, prior biologic use) did not improve predictive performance compared to models with only two variables. Models constructed using forward, backward, and elastic net selection methods all performed like each other, and performance was better than BSA plus PROM models, but like the BSA plus IGA model.

Table 2.

Predictive Performance of Regression Models for Predicting PASI and DLQI Using the Psoriasis Assessment Model Development Testing Data Set of 13 856 Patient-Visits.

		PASI			DLQI
Model	Predictors	± ${\mathrm{R}}_{\mathrm{a}\mathrm{d}\mathrm{j}}^2$ (95% CI)†	RMSE (95% CI)	MAE (95% CI)	± ${\mathrm{R}}_{\mathrm{a}\mathrm{d}\mathrm{j}}^2$ (95% CI)	RMSE (95% CI)	MAE (95% CI)
1	BSA	.62 (.60, .64)	3.43 (3.32, 3.53)	1.97 (1.93, 2.01)	.17 (.16, .18)	4.90 (4.82, 5.00)	3.61 (3.58, 3.65)
2	BSA + IGA	.72 (.71, .73)	2.93 (2.84, 3.03)	1.47 (1.43, 1.50)	.27 (.26, .29)	4.60 (4.50, 4.68)	3.24 (3.20, 3.28)
3	BSA + IGA + BSA * IGA	.74 (.72, .75)	2.85 (2.75, 2.96)	1.44 (1.40, 1.48)	.28 (.26, .29)	4.58 (4.49, 4.67)	3.23 (3.19, 3.27)
4	BSA + IGA + age + gender + mCCI + BMI + prior biologic experience	.72 (.71, .73)	2.92 (2.82, 3.02)	1.51 (1.47, 1.54)	.29 (.28, .30)	4.54 (4.46, 4.63)	3.18 (3.14, 3.23)
5	BSA + itch	.64 (.63, .66)	3.32 (3.21, 3.42)	1.85 (1.81, 1.88)	.40 (.38, .41)	4.18 (4.09, 4.26)	2.85 (2.81, 2.89)
6	BSA + Itch + BSA * itch	.64 (.63, .66)	3.32 (3.21, 3.42)	1.85 (1.81, 1.89)	.40 (.38, .41)	4.18 (4.09, 4.26)	2.84 (2.80, 2.88)
7	BSA + Itch + age + gender + mCCI + BMI + prior biologic experience	.65 (.63, .67)	3.29 (3.18, 3.39)	1.83 (1.79, 1.86)	.40 (.39, .42)	4.15 (4.07, 4.24)	2.82 (2.78, 2.86)
8	BSA + skin pain	.64 (.62, .66)	3.33 (3.22, 3.43)	1.88 (1.84, 1.92)	.40 (.38, .41)	4.17 (4.10, 4.26)	2.91 (2.87, 2.95)
9	BSA + skin pain + BSA * skin pain	.64 (.62, .66)	3.36 (3.22, 3.44)	1.88 (1.84, 1.92)	.40 (.38, .42)	4.16 (4.09, 4.25)	2.88 (2.85, 2.92)
10	BSA + skin pain + Age + gender + mCCI + BMI + prior biologic experience	.64 (.63, .66)	3.30 (3.19, 3.41)	1.85 (1.81, 1.89)	.41 (.39, .42)	4.14 (4.06, 4.22)	2.87 (2.83, 2.90)
11	BSA + PGA	.64 (.62, .66)	3.34 (3.23, 3.44)	1.87 (1.83, 1.90)	.37 (.35, .38)	4.28 (4.19, 4.37)	2.90 (2.86, 2.94)
12	BSA + PGA + BSA*PGA	.64 (.62, .66)	3.34 (3.23, 3.45)	1.87 (1.83, 1.91)	.37 (.35, .38)	4.28 (4.19, 4.37)	2.90 (2.86, 2.95)
13	BSA + PGA + age + gender + mCCI + BMI + prior biologic experience	.64 (.62, .66)	3.31 (3.20, 3.42)	1.85 (1.80, 1.88)	.38 (.36, .39)	4.25 (4.16, 4.33)	2.87 (2.83, 2.91)
14	BSA + IGA + itch + skin pain + PGA	.72 (.71, .74)	2.92 (2.82, 3.02)	1.48 (1.45, 1.52)	.47 (.46, .49)	3.91 (3.83, 3.99)	2.57 (2.53, 2.61)
15	BSA + IGA + itch + skin pain + PGA + age + gender + mCCI + BMI + prior biologic experience	.72 (.71, .74)	2.90 (2.81, 3.01)	1.52 (1.48, 1.55)	.48 (.46, .49)	3.89 (3.81, 3.97)	2.55 (2.52, 2.60)
16	BSA + IGA + BSA * IGA + Itch + BSA * itch + skin pain + BSA * skin pain + PGA + BSA * PGA + Age + gender + mCCI + BMI + prior biologic experience	.74 (.72, .75)	2.82 (2.73, 2.93)	1.48 (1.45, 1.52)	.48 (.46, .50)	3.88 (3.80, 3.96)	2.55 (2.52, 2.60)
17	Variables‡ selected from stepwise forward selection	.72 (.71, .74)	2.91 (2.82, 3.02)	1.52 (1.48, 1.55)	.48 (.46, .49)	3.89 (3.81, 3.97)	2.55 (2.52, 2.60)
18	Variables‡ selected from stepwise backward elimination	.72 (.71, .74)	2.91 (2.82, 3.02)	1.52 (1.48, 1.55)	.48 (.46, .49)	3.89 (3.81, 3.97)	2.55 (2.52, 2.60)
19	Variables‡ selected from elastic net	.74 (.73, .75)	2.82 (2.73, 2.93)	1.48 (1.45, 1.52)	.48 (.46, .50)	3.88 (3.80, 3.96)	2.56 (2.52, 2.60)

†The entire process of sample splitting, model development, and model testing was replicated 400 times; 95% confidence intervals were calculated using the 2.5th and 97.5th percentiles from these distributions; ‡Models 17-18 selected from the following variables: BSA, Itch, Skin Pain, PGA, Age, Female, mCCI, BMI, Biologic-naive and IGA. Model 19 selected from all potential variables in model 17-18 as well as the following interactions: BSA *IGA, BSA*Itch, BSA*Skin Pain and BSA*PGA. Models 17 and 18 selected BSA, Skin Pain, PGA, Itch, IGA, Female and Biologic-naive. Model 19 selected BSA, Skin Pain, PGA, Itch, IGA, Female, Biologic-naive, age, mCCI, BMI, BSA*IGA, BSA*Itch, BSA*Skin Pain and BSA*PGA (all possible variables). ±Adjusted R², adjusted coefficient of determination; RMSE, root mean square error; MAE, mean absolute error; PASI, Psoriasis Area Severity Index; DLQI, Dermatology Life Quality Index; BSA, Body Surface Area; IGA, Investigator’s Global Assessment; mCCI, modified Charlson Comorbidity Index; BMI, Body Mass Index; **Models 1-16 have been pre-specified. Calibration for each model was assessed with calibration intercept (95% CI) and slope (95% CI) and is provided in supplemental materials. **Models 17-19 considered all aforementioned variables, and only a subset of variables were selected as determined by stepwise forward selection (using AIC), stepwise backward elimination (using AIC), and a regularized elastic net regression (α = .839, log(λ) = -9.47 for PASI; α = .285 and log(λ) = -22.32 for DLQI). Hyperparameters for the elastic net model were optimized via 10-fold cross-validation. The mixture parameter (α) combines the weights of the L1 (LASSO) and L2 (Ridge) penalties, and the shrinkage parameter (λ) indicates the extent to which regression coefficients are shrunk toward zero to reduce variation in prediction.

For models predicting DLQI, BSA alone explained only 17% of the variance in DLQI (R²_adj = .17; 95%CI = .16, .18), and while the addition of IGA improved performance, it was still poor (R²_adj = .27; RSME = 4.60). Models with BSA plus a PROM performed better, yet still explained only a low percent of the variance in DLQI (all R²_adj = .37-.40, RMSE range 4.17-4.28). The best performing models included the model containing IGA and all other PROs (all R²_adj = .47, RMSE = 3.91), and those constructed using selection methods, though test statistics still suggested poor predictive performance (all R²_adj = .48, all RMSE = 3.89) (Table 2).

Concordance of Predicted and Measured PASI and DLQI Outcomes

When comparing observed PASI75 and PASI90 to those predicted based on each of the constructed prediction models, specificity and PPV were high across all models, though sensitivity and NPV varied. Models including IGA had the best agreement, with SE and SP exceeding 82% for PASI75 (Models 2-4: SE range = 82.7-85.1, SP range = 87.7-90.7; Models 14-19: SE range = 82.8-85.6, SP range = 88.3-90.0) and exceeded 76% for PASI90 (Models 2-4: SE range = 80.2-81.9, SP range = 94.1-97.9; Models 14-19: SE range = 76.2-80.9, SP range = 93.4-96.0) (Table 3).

Table 3.

Percent Agreement of Predicted PASI75, PASI90 and DLQI 0/1, Calculated Based on 19 Different Regression Models, and Observed Outcomes Among 1303 Patient-Visits With a Biologic Initiation and Subsequent 6-Month Follow-Up Visit.

	PASI75				PASI90				DLQI 0/1
Model	SE a	SP b	PPV c	NPV d	SE a	SP b	PPV c	NPV d	SE a	SP b	PPV c	NPV d
1	44.7	96.5	94.0	58.5	12.1	99.0	88.2	63.8	94.8	28.4	81.8	61.8
2	84.1	87.7	89.4	81.7	81.9	94.1	89.8	89.1	84.4	59.5	87.6	52.9
3	82.7	90.7	91.7	80.9	80.3	97.9	96.1	88.6	83.8	60.1	87.7	52.2
4	85.1	89.4	90.8	82.9	80.2	95.1	91.2	88.3	83.6	58.9	87.4	51.3
5	63.3	94.0	92.9	67.5	34.1	97.7	90.3	70.0	83.9	71.3	90.8	56.7
6	62.5	94.2	93.0	67.0	32.5	97.9	90.9	69.5	83.7	71.6	90.9	56.5
7	66.5	90.9	90.0	68.8	38.9	96.9	88.7	71.5	83.8	70.2	90.5	56.0
8	56.9	93.8	91.9	63.8	21.3	97.9	86.8	66.2	89.3	67.1	90.2	64.9
9	56.9	93.8	91.9	63.8	21.3	97.9	86.8	66.2	88.9	69.8	90.9	65.0
10	62.2	91.8	90.3	66.4	33.0	97.7	89.9	69.8	90.0	66.3	90.1	66.1
11	60.5	94.2	92.8	65.9	30.4	97.8	89.8	68.9	83.9	71.3	90.8	56.7
12	61.2	94.2	92.9	66.3	32.3	97.8	90.3	69.4	84.0	70.9	90.7	56.8
13	64.2	92.7	91.5	67.8	37.2	96.7	87.9	70.9	84.5	69.9	90.5	57.0
14	84.1	88.5	90.0	81.9	80.9	93.8	89.2	88.6	87.1	77.6	93.0	64.0
15	85.4	88.3	89.9	83.1	80.0	93.4	88.4	88.1	87.5	75.9	92.5	64.1
16	82.8	90.0	91.1	81.0	76.2	96.0	92.2	86.5	87.4	75.3	92.3	63.7
17	85.6	88.4	90.1	83.2	79.6	93.8	89.1	87.9	87.5	75.9	92.5	64.1
18	85.6	88.4	90.1	83.2	79.6	93.8	89.1	87.9	87.5	75.9	92.5	64.1
19	82.9	90.0	91.1	81.1	76.4	95.8	92.0	86.6	87.4	75.3	92.3	63.7

^aSE, sensitivity (%).

^bSP, specificity (%).

^cPPV, positive predictive value (%).

^dNPV, negative predictive value (%).

When comparing observed vs predicted DLQI 0/1, sensitivity and PPV were consistently high for the different prediction models, though specificity and NPV were low. Predicted DLQI 0/1 based on models that included PROMs (itch, skin pain, PGA) tended to have the best agreement with observed DLQI. Those that included IGA and all three PROs having the highest agreement (Models 14-16: SE range = 87.1-87.5, SP range = 75.3-77.6) (Table 3).

In the subset of patients with severe psoriasis, predictive performance of psoriasis assessment models was poor compared to the full cohort. The R²_adj for models predicting PASI did not exceed .45 and RMSE = 8.07 (model including all possible covariates), and for predicting DLQI the best performing model (including all possible covariates) yielded an adjusted R² = .34 and RMSE = 5.48 (Supplemental Table 1).

Discussion

In our study among patients from real-world clinical practices, psoriasis assessment models including up to 14 variables were the most predictive of PASI, though these models performed only marginally better than a model including only BSA and IGA, which explained greater than 70% of the variation in PASI. When predicting PASI response based on PASI calculated using the psoriasis assessment models, models that included IGA had the best agreement with observed PASI response, achieving the best balance between both high (∼88%) sensitivity and specificity. The ability of models to predict DLQI in our cohort was limited, with the best performing models including several variables and explaining less than 50% of variation in DLQI. While sensitivity for identifying observed DLQI 0/1 based on predicted DLQI was high (∼83%) for all models, specificity was lower, particularly for those models with only one or two variables (∼70%). Models including multiple PROMs had the best agreement with DLQI 0/1.

Our findings are consistent with those among patients from the UNCOVER clinical trials ¹³ which found that a proxy PASI score calculated using only two simple data inputs (eg, BSA and PROMs) was strongly predictive of an observed PASI assessment. The prior study found its models (OPAT™) including BSA plus one PROM (itch, skin pain or PGA) had sensitivities and specificities around 90% for identifying both PASI75 and PASI90. In our study, the models that included one PROM (itch, skin pain, or PGA) had somewhat lower predictive performance for PASI, and our predicted PASI outcomes had lower agreement with observed values. These findings may be due to the UNCOVER trials including a population with more severe disease compared to the CorEvitas Psoriasis Registry, which enrolls patients regardless of disease severity. The UNCOVER studies inclusion criteria required patients have BSA ≥10%, static physician global assessment≥3, and PASI≥12. Indeed, mean PASI in the previously published OPAT™ analysis was about 20 compared to 3.5 in the current study. Furthermore, correlations of PASI with BSA and PROMs were stronger in the UNCOVER cohorts, perhaps reflecting a more heterogeneous patient population in the CorEvitas Registry. Further, assessments may be more tightly controlled in randomized trials vs registry patients in clinical practice. Nevertheless, our findings indicate that an OPAT developed and applied in a cohort more representative of the psoriasis population, using data ascertained by dermatologists in a clinic setting, performed well for predicting PASI and identifying PASI response. Unlike the UNCOVER OPAT™ models, our psoriasis assessment models considered IGA as a predictor and found models including BSA plus IGA predicted PASI better than those including BSA plus a PROM. Although IGA is not a PROM, it is relatively easy to collect in the clinic, and our data suggest that its combination with BSA may provide a superior predictor of PASI than BSA alone, or BSA combined with a PROM. Our findings support prior studies that have found that BSA*IGA (or Physician’s Global Assessment) is highly correlated with PASI score, further suggesting that BSA*IGA may be an ideal assessment of disease severity in clinical practice as it is more practical than PASI, and is sensitive for measuring psoriasis severity in clinical trials^20-22 and registry patients. ²³

Similar to our study, regression models in the UNCOVER study did not predict DLQI (R² ˜ .65) as well as PASI. In our study, while models including BSA plus IGA were the most predictive of PASI, these models performed worse for predicting DLQI compared to those including PROMs. This finding is unsurprising as itch and skin pain capture disease symptoms that directly impact patient HRQoL and PGA reflects patient perception of disease impact, whereas IGA considers only plaque appearance. Our data suggest that a model including BSA plus all three of PROMs performed only marginally better at predicting DLQI than those including just one PRO, and models including additional variables did not improve prediction.

Our study has some limitations. Generalizability to the general population of psoriasis patients may be limited since registry dermatologists participate voluntarily and choose which of their patients to enroll, or if certain types of patients (eg, sicker, or healthier) agree to enroll. Additionally, the prediction models we developed require validation using other datasets prior to implementation. Nevertheless, our findings are likely more generalizable to typical patients seen in clinical practice vs those generated from clinical trials. Furthermore, the CorEvitas Registry included nearly 11 000 patients with over 33 000 visits at which an extensive battery of variables was collected, including measures not routinely available in other real-world data sources such as claims databases.

Conclusions

Our study developed what we consider prototypes of prediction models using BSA and IGA to predict PASI. External testing and further research are required to validate these models before they can be used in clinical practice. If and when a model is externally validated, our findings suggest that a prediction model including BSA and IGA may be ideal for a psoriasis assessment to predict PASI in the clinical setting. DLQI is more challenging to predict, though a psoriasis assessment tool with BSA plus at least one PROM may provide the best option. Once validated in other real-world patient cohorts, a psoriasis assessment tool including a few simple measures may provide dermatologists with a measure of disease severity that has advantages over PASI, as it is more easily implemented in clinical practice and might provide additional information on patient quality of life.

ORCID iD

Ryan W. Harrison https://orcid.org/0000-0003-4575-006X