Abstract
Introduction
Nemolizumab is a humanized monoclonal antibody that specifically inhibits the receptor for interleukin-31, the major pruritogen in atopic dermatitis (AD). While the patient profile associated with the early therapeutic response to nemolizumab in AD has been increasingly delineated, the characteristics of patients who benefit from long-term nemolizumab therapy remain unclear. This study aimed to identify early clinical features predicting long-term improvement in pruritus Visual Analog Scale (VAS) score among patients receiving nemolizumab.
Methods
This study included patients participating in two Japanese clinical trials (JP01 and JP02) who did not achieve a 50% improvement in pruritus VAS score (VAS50) at week 16 after initiation of nemolizumab treatment. Patients who achieved VAS50 at week 52 were classified as responders, and those who did not were nonresponders. A machine-learning decision tree model was developed using 213 candidate variables. Model performance was evaluated using the F1 score. Shapley additive explanation (SHAP) values were used to interpret variable importance.
Results
Among 118 eligible patients, the final model achieved an F1 score of 0.737. The two strongest predictors of long-term response were the absence of worsening of AD and/or the occurrence of other cutaneous disorders until week 16, and a weekly mean itch scale score < 2.1 at week 16. Patients meeting both criteria were highly likely to achieve VAS50 by week 52.
Conclusions
Machine-learning identified straightforward and clinically applicable predictors of long-term nemolizumab efficacy in patients who did not have early VAS improvement. These findings may support decision-making for treatment continuation beyond week 16.
Trial Registration Number
JapicCTI-173740 (JP01), and JapicCTI-183894 (JP02).
Supplementary Information
The online version contains supplementary material available at 10.1007/s13555-026-01768-0.
Keywords: Atopic dermatitis, IL-31, Machine learning, Nemolizumab, Post hoc analysis
Key Summary Points
| Why carry out this study? |
| Clinically, it is often difficult to determine whether nemolizumab should be continued in patients with atopic dermatitis (AD) who do not achieve sufficient pruritus improvement within the first 16 weeks of therapy. |
| Although nemolizumab is effective in reducing pruritus, predictors of long-term treatment response in patients with AD who do not show early improvement remain poorly defined. |
| Using machine learning methods, a post hoc analysis was performed on data from Phase III trials in Japan to explore the characteristics of patients who ultimately achieved a 50% improvement in pruritus on the Visual Analog Scale (VAS50) with long-term nemolizumab treatment. |
| What was learned from the study? |
| Among patients with AD who did not achieve VAS50 at week 16 of nemolizumab treatment, most responders who achieved VAS50 at week 52 did not experience worsening skin symptoms by week 16, and had a weekly mean itch scale score <2.1 at week 16. |
| Findings from the initial 16 weeks of nemolizumab treatment can help predict the response to long-term nemolizumab use for pruritus management. |
Introduction
Nemolizumab is a humanized monoclonal antibody that binds to interleukin-31 (IL-31) receptor A (IL-31RA) to inhibit the IL-31 signaling pathway, which plays a role in pruritus transmission [1, 2]. Phase III clinical trials on nemolizumab in patients with atopic dermatitis (AD) in Japan (JP01 and JP02) showed that administering 60 mg nemolizumab subcutaneously every 4 weeks along with topical corticosteroid (TCS) or topical calcineurin inhibitor (TCI) use resulted in a greater reduction in pruritus and improvement in the signs and symptoms of AD than a placebo plus a topical agent [3, 4]. By improving pruritus and skin lesions, nemolizumab improves patients’ quality of life [5]. In Japan, nemolizumab is approved to mitigate pruritus associated with AD, and to treat AD itself in several other countries.
Recent advances have led to rapid diversification in AD treatment, highlighting the need to select optimal therapeutic strategies tailored to each patient’s condition. Given that long-term treatment continuity is required for AD management, both patients and physicians seek early perceived efficacy and sustained long-term effectiveness. In the JP01 and JP02 trials, 34.3% (excluding those who received rescue therapy) and 45.5% of patients, respectively, achieved a 50% improvement in the pruritus Visual Analog Scale (VAS) score (VAS50) at week 16 [4]. Furthermore, with long-term nemolizumab continuation (68 weeks in JP01 and 52 weeks in JP02), 57.3% and 55.7% of patients achieved VAS50, respectively [4]. These findings demonstrate three distinct patterns of patient responses: those who achieve early therapeutic effects against pruritus (rapid responders), those who benefit from long-term treatment continuation (responders), and those who show no response to therapy (nonresponders). This trend was similarly observed in improvement in Eczema Area and Severity Index (EASI) scores [5], indicating heterogeneity in patient responses to nemolizumab treatment.
Clinically, deciding whether to continue long-term treatment with nemolizumab for AD is based on its antipruritic effects. In Japan, the optimal use guidelines recommend evaluating treatment response within 16 weeks and discontinuing therapy if no sufficient clinical benefit is achieved [6]. However, many patients require continued administration to maintain disease remission, and thus, it may be appropriate to continue treatment if clinically justified [6]. Therefore, the ability to predict long-term effectiveness at an early stage is crucial not only for supporting clinical decision‑making in accordance with these recommendations but also for sustaining patients’ motivation for treatment and minimizing the overall economic burden. Recently, real-world evidence from patients with AD treated with nemolizumab has increased. Masuo et al. investigated the characteristics of patients who responded rapidly to nemolizumab by stratifying them by treatment responsiveness [7]. Similarly, Sugiyama et al. investigated the patient profiles of those in whom nemolizumab treatment is considered appropriate on the basis of clinical observations and cutaneous adverse events [8]. Therefore, aggregating and analyzing real-world clinical data are crucial for the development of effective AD therapies. However, data on the characteristics of responders and nonresponders who eventually exhibit therapeutic effects under long-term nemolizumab administration are scarce, as such insights are likely to emerge only through long-term patient monitoring.
Recent advances in machine learning (ML) methodologies for analyzing clinical data have drastically enhanced our ability to forecast patient outcomes, medication efficacy, and safety profiles. Although statistical predictive models have helped identify clinically relevant predictors [9], several studies have used more complex ML models to characterize patients with AD who are likely to benefit from biological agents and systemic treatments [10, 11]. Data analysis using ML enables the prediction of long-term efficacy from existing data without requiring extended patient monitoring. Because identifying characteristics of patients who are more likely to respond to long-term treatment with nemolizumab is critical, we aimed to develop an ML model to predict long-term efficacy (up to week 52) in patients who did not achieve VAS50 at week 16 and to identify early clinical indicators characterizing long-term responders in pooled data from the JP01 and JP02 Phase III trials.
Methods
Study Sample
This study included 211 patients with AD who were ≥ 13 years. These patients were administered 60 mg nemolizumab every 4 weeks for up to 52 weeks in two Japanese Phase III clinical trials: JP01 (JapicCTI-173740) and JP02 (JapicCTI-183894) [3, 4]. The patients had inadequate pruritic responses (score of ≥ 3 on a 5-level itch scale; range, 0–4, with higher scores indicating worse itching [12], and a VAS score for pruritus of 50 or more on the day of randomization) to treatments including medium-potency TCS or TCI for at least 4 weeks and oral antihistamines for at least 2 weeks. In the JP01 trial, concomitant AD medications (medium-potency TCS, TCI, and oral antihistamines) remained unchanged through week 16, with no limitations on their use thereafter until week 52. Patients who received rescue therapy due to AD exacerbations being included in the analysis. In the JP02 trial, no restrictions were placed on concomitant medications, except that patients who had received systemic steroids, immunosuppressants, phototherapy, desensitization therapy, or modulating therapy within 28 days prior to initiating nemolizumab were excluded.
Criteria for Responder Classification
Nemolizumab responders were defined as patients who achieved VAS50. Patients who met the VAS50 at week 16 were defined as “rapid responders” and were excluded from ML model development. Of the patients who did not achieve VAS50 at week 16, those who subsequently achieved VAS50 at week 52 were identified as “responders,” whereas those who did not were “nonresponders.” These latter two groups were included in ML model development. After 16 weeks of nemolizumab treatment, 121 of 211 patients failed to achieve VAS50. Data from 118 patients were included in the analysis after excluding 3 patients with missing values in at least one potential explanatory variable. Baseline characteristics of these 118 patients are presented in Table S1. The responder to nonresponder ratio at week 52 was 51:67 (Fig. 1).
Fig. 1.
Patient disposition
ML Model Development
ML model development was conducted following the flow chart in Fig. 2. This approach incorporated 213 variables as predictors, including patient demographics, such as age, sex, height, and weight; disease severity measures, including EASI score, pruritus VAS score, and quality of life-associated patient-reported outcome, subscores; and laboratory parameters, at baseline and week 16, as well as the differences from the baseline to week 16 (Table S2). Strongly correlated pairs of explanatory variables (correlation coefficients > 0.8) were delineated to avoid multicollinearity. Subsequently, network analysis was performed to calculate the degree centrality for each factor, and the factors with the highest centrality were eliminated. This process was repeated until the correlation coefficients among all factors were < 0.8. Factors with a variance of 0 were considered to contain no informational value and were excluded.
Fig. 2.
Process diagram for ML model development. ML machine learning
The patient dataset was randomly partitioned into an 80% training subset for ML model development and a 20% testing subset for performance assessment. A series of analyses confirmed the absence of specific bias in the VAS score distributions across the split datasets. A decision tree model was selected for the ML model [13], with the F1 score used to assess model performance. The F1 score, the harmonic mean of precision and recall, is especially relevant for evaluating models prioritizing both sensitivity and specificity [14, 15]. The model’s hyperparameters were optimized with GridSearchCV [16] within the range specified in Table S3. After an exhaustive search and stepwise selection [17], two explanatory variables were determined for use (Figure S1). The model was constructed using all combinations.
The entire process from data partitioning to model evaluation was iterated 10,000 times, extracting models with average F1 scores in the top 10% as “top models.” Shapley additive explanation (SHAP) values were used to quantitatively evaluate the contribution of explanatory variables in the predictive model to identifying potential responder patient profiles [18]. SHAP values have positive and negative values, with the magnitude of the absolute value reflecting the explanatory variable’s contribution to distinguishing between “responders” and “nonresponders.” The relationships among explanatory variables were investigated by averaging the SHAP values for each patient across specific combinations of top models and examining variation in SHAP values across explanatory variables. This led to redeveloping the decision tree model (the integrated model) based on the progression of SHAP values for each variable. All analyses were performed in Python 3.8.18 with the SHAP (v.0.45.0) and scikit-learn (v.1.3.0) packages [18, 19].
Ethics
The present study was a post hoc analysis of pooled data from two previously conducted Japanese Phase III clinical trials, Study-JP01 (JapicCTI-173740) and Study-JP02 (JapicCTI-183894). Because this study involved only secondary analysis of data from the original trials and did not involve any new recruitment, intervention, or contact with participants, additional ethical approval was not required. The original trials were conducted in accordance with the ethical principles of the Declaration of Helsinki, Good Clinical Practice guidelines, and all other applicable regulatory requirements. The trial protocols and related documentation were approved by the institutional review boards at each participating center, including the Sapporo Skin Clinic IRB. Written informed consent was obtained from all patients or their legal guardians prior to initiation of any study procedures and treatment. Publicly available ethics committee reference numbers/IRB numbers for the original trials could not be identified from the published reports.
Results
Pairs of Explanatory Variables with the Highest F1 Score
Table 1 summarizes the explanatory variable pairs with the highest average F1 scores among the “top models.” The pair with the highest predictive performance (F1 score = 0.823) consisted of “mean itch scale score at week 16” and “worsening of AD and/or the occurrence of other cutaneous disorders (AD/OTCD) up to week 16.” The various cutaneous disorders classified as OTCD in the JP01 and JP02 trials are listed in Table S4. Most of the top ten pairs included the mean itch scale score at week 16 or its change from baseline, indicating that itch severity at week 16 is a key determinant of long-term VAS50 achievement. Other variables frequently paired with itch-related indices included urinalysis protein level, high-density lipoprotein (HDL) cholesterol, alkaline phosphatase (ALP), EASI subscores, and insomnia severity index (ISI) subscores. Because the highest-performing model included the mean itch scale score at week 16 and AD/OTCD, subsequent analyses focused on these two variables.
Table 1.
Pairs of explanatory variables used in the “top models”
| Rank | F1 score | Pairs of explanatory variables | |
|---|---|---|---|
| Variable 1 | Variable 2 | ||
| 1 | 0.823 | Itch scale (mean)/week 16 | Worsening of AD and/or the occurrence of OTCD up to week 16 |
| 2 | 0.822 | Itch scale (mean)/difference | Urinalysis protein/week 0 |
| 3 | 0.813 | Itch scale (mean)/week 16 | Urinalysis protein/week 0 |
| 3 | 0.813 | Itch scale (mean)/difference | Worsening of AD and/or the occurrence of OTCD up to week 16 |
| 5 | 0.812 | Itch scale (mean)/week 16 | HDL cholesterol/difference |
| 6 | 0.811 | Itch scale (mean)/week 16 | Urinalysis protein/week 16 |
| 7 | 0.810 | Itch scale (mean)/week 16 | Alkaline phosphatase/difference |
| 8 | 0.809 | Itch scale (mean)/week 16 | ISI subscore (difficulty staying asleep)/week 0 |
| 9 | 0.806 | Itch scale (mean)/week 16 | Itch scale (mean)/difference |
| 10 | 0.805 | Itch scale (mean)/week 16 | EASI subscore (excoriation) total/week 0 |
Difference denotes the change in values from baseline (week 0) at week 16. Itch scale (mean)/week 16 denotes the weekly mean daily itch scale scores at week 16. AD atopic dermatitis, EASI Eczema Area and Severity Index, HDL high-density lipoprotein, ISI insomnia severity index, OTCD other cutaneous disorder
Relationship between SHAP Values and Explanatory Variables
The association between each variable and its SHAP value was examined to understand how the pairs of explanatory variables identified in the previous section affect the prediction of each responder. Figure 3 shows the relationship between the measured values of each variable, the corresponding number of patients, and the SHAP values. The presence of positive or negative SHAP values indicates that the corresponding explanatory variable contributed to classifying patients as responders or nonresponders.
Fig. 3.
Association between the SHAP value and observed values of each explanatory variable. Relationship between the itch scale score at week 16 (a) or AD/OTCD (b) and SHAP values in the model predicting responders and nonresponders. Each histogram shows the number of patients for each unit value of the explanatory variable. Red plots indicate the mean SHAP values ± standard deviation for each patient in the predictive model. AD/OTCD, worsening of atopic dermatitis and/or the occurrence of other cutaneous disorders; SHAP, Shapley additive explanations
Many patients presented with an itch scale score of approximately 2.1 or 3.1 at week 16; the shift from positive to negative SHAP values occurred at an itch scale score of 2.1, indicating that this threshold is a critical criterion in predictor of responders. Among the 69 patients with a mean itch scale score < 2.1, the SHAP values were > 0. However, among the 49 patients with a mean itch scale score ≥ 2.1, the SHAP values were < 0 (Fig. 3a), indicating a tendency to classify patients with an itch scale score < 2.1 as responders and those with an itch scale score ≥ 2.1 as nonresponders at week 52.
AD/OTCD, a categorical variable denoting exacerbation in AD or other cutaneous disorders, was defined as 0 for no exacerbation and 1 for exacerbation. Among the 24 patients who experienced worsening of skin symptoms (AD/OTCD = 1), the SHAP values were < 0, indicating the tendency to be classified as nonresponders. Conversely, among the 94 patients without deterioration in skin symptoms (AD/OTCD = 0), the SHAP values were > 0, suggesting the tendency to be classified as responders (Fig. 3b).
Relationship Between the Itch Scale Score at Week 16 and AD/OTCD
The influence of SHAP values when classification was performed using one variable within responders categorized by another variable was investigated to explore the interrelation between the 16-week itch scale and AD/OTCD. Initially, we investigated the relationship of AD/OTCD and SHAP values within groups of responders (n = 69) and nonresponders (n = 49) classified by itch scale severity at week 16. In both groups, AD/OTCD = 1 indicated a negative SHAP value (Fig. 4a, b), consistent with the trend observed in Fig. 3b. Subsequently, the association between the 16-week itch scale score and SHAP values was examined in groups classified by AD/OTCD values (comprising 94 responders and 24 nonresponders). In each group, an itch scale score < 2.1 corresponded to SHAP values > 0, indicating responder status, consistent with the pattern shown in Fig. 3a. Conversely, when the itch scale was ≥ 2.1, the SHAP values fell below 0, indicating a tendency toward nonresponder classification (Fig. 4c, d). These findings demonstrate that AD/OTCD independently influences patient classification, irrespective of the itch scale score at week 16.
Fig. 4.
Association between SHAP value and observed values for one explanatory variable within subgroups of responders categorized by another variable. Relationship between AD/OTCD and SHAP values in groups with mean itch scale scores of < 2.1 (a) and ≥ 2.1 (b) at week 16. Relationship between the itch scale score at week 16 and SHAP values in populations with an AD/OTCD of 0 (c) or 1 (d). Each histogram shows the number of patients for each unit value of the explanatory variable. Red plots indicate the mean SHAP values ± standard deviation for each patient in the predictive model. AD/OTCD worsening of atopic dermatitis and/or the occurrence of other cutaneous disorders, SHAP Shapley additive explanations
Table 2 shows the breakdown of responders/nonresponders across the four groups classified by the 16-week itch scale score and AD/OTCD, along with the prediction accuracy of responses by patient number. Patients were predicted to be nonresponders when AD/OTCD = 1, regardless of their itch scale scores at week 16. When AD/OTCD = 0, the likelihood of being a responder was predicted by the itch scale score at week 16.
Table 2.
Relationship between the composition of patient groups classified by two explanatory variables and the prediction of responders
| Pred.* (accuracy, %) Number of patients (responders:nonresponders) |
Itch scale score (mean) at week 16 | ||
|---|---|---|---|
| < 2.1 | ≥ 2.1 | ||
| AD/OTCD | = 0 |
Responders (67.9%) 56 (38:18) |
Nonresponders (78.9%) 38 (8:30) |
| = 1 |
Nonresponders (69.2%) 13 (4:9) |
Nonresponders (90.9%) 11 (1:10) |
|
0 no exacerbation, and 1 with exacerbation. AD/OTCD worsening of atopic dermatitis and/or the occurrence of other cutaneous disorders
*Pred. designates the most numerous responder type within each of the classified cohorts
Decision Tree Model for Predicting VAS50 at Week 52
The decision tree model (integrated model) was reconstructed on the basis of the relationships between the explanatory variables. Figure 5a presents the classification of 118 patients based on the integrated model, and Table 3 presents its predictive accuracy. In the final integrated model, patients were classified as nonresponders if AD/OTCD = 1. AD/OTCD = 0, they were classified as responders provided that the itch scale score at week 16 was < 2.1. The F1 score was 0.737, which was 0.086 lower than that of the “top models” (0.823), which was based on the mean 16-week itch scale score and AD/OTCD values (Table 1).
Fig. 5.
Constructed model and the breakdown of its predictive accuracy. a Decision tree model (the integrated model) for predicting the achievement of VAS50 at week 52 in patients who did not achieve VAS50 at week 16 after nemolizumab administration. The progression of percent change in the VAS score from baseline for patients predicted to be responders (b) or nonresponders (c) and the distribution in patient counts at week 52 based on observed values (d, e). Blue represents responders, whereas red indicates nonresponders. Lighter shades of each color denote populations with observed discrepancies between the predicted (predicted) and actual (true) responder status. AD/OTCD worsening of atopic dermatitis and/or the occurrence of other cutaneous disorders, Non nonresponder, Res responder, VAS pruritus on the Visual Analog Scale score, VAS50 50% improvement in VAS score
Table 3.
Predictive accuracy of the integrated model
| Accuracy | 0.737 | |
|---|---|---|
| Responders | Nonresponders | |
| Precision | 0.679 | 0.790 |
| Recall | 0.745 | 0.731 |
| F1 score | 0.710 | 0.760 |
Longitudinal changes in pruritus VAS scores within each group and the distribution of percent change in the pruritus VAS score from baseline to week 52 were analyzed to investigate the characteristics of patients with discrepancies between predicted and actual responder status (Fig. 5b–e). Patients incorrectly predicted as responders but observed to be nonresponders typically showed intermediate VAS improvements (~40–50%) that did not reach the VAS50 threshold. Conversely, patients predicted to be nonresponders who were actually responders did not show consistent distinguishing features. Overall, these findings indicate that patients most likely to achieve VAS50 at week 52 are those who show no worsening of skin symptoms up to week 16 and maintain a mean itch scale score < 2.1 at that time, reinforcing the utility of early clinical indicators to guide long-term nemolizumab therapy.
Discussion
This study aimed to develop an ML model using two clinical trial datasets for nemolizumab to predict whether patients with AD who did not achieve VAS50 at week 16 will go on to demonstrate improvements at week 52 with nemolizumab treatment. In Japan, the guidelines for the use of nemolizumab recommend considering treatment discontinuation if efficacy is not observed 16 weeks after treatment initiation [6]. The insights obtained from our results may help inform decisions about whether to continue nemolizumab therapy.
We analyzed 118 patients with AD who participated in two Japanese Phase III trials (JP01 and JP02), had continuously received 60 mg of nemolizumab for 52 weeks, and did not achieve VAS50 by week 16. Although the conditions for concomitant medication use differed between JP01 and JP02, the proportions of rapid responders and responders were comparable across trials (Fig. 1), suggesting that these restrictions had a minimal impact on achieving VAS50.
During model construction, explanatory variable pairs with high F1 scores were initially narrowed down (Table 1). Among the top ten explanatory variable pairs, the mean itch scale score at week 16 or the change from weeks 0 to 16 consistently ranked highly, indicating the importance of the itch score at week 16 in predicting the pruritus VAS score at week 52. Notably, the predictive importance of the week 16 itch scale should be interpreted with caution, because this measure could partly reflect the persistence of symptom trajectory rather than an independent prognostic factor. In addition to AD/OTCD, clinical laboratory values, including urinalysis, protein, HDL cholesterol, and ALP, were evaluated as potential explanatory variables for the final model. Although these values are not exclusive to AD, they are associated with pruritus. For instance, patients with chronic kidney disease often experience pruritus due to skin dryness and uremic toxin accumulation, with elevated urinalysis protein levels and low HDL cholesterol [20, 21]. Additionally, elevated ALP levels indicate primary biliary cholangitis, a condition that causes pruritus in most patients [22]. HDL cholesterol levels negatively correlate with pruritus symptom burden in patients with primary biliary cholangitis [23]. Nonetheless, when constructing the predictive model for responders versus nonresponders to nemolizumab, the SHAP value distribution for these variables was limited. Despite high F1 scores, they were not included as explanatory variables.
The integrated model identified responders to nemolizumab as patients with no worsening of skin symptoms by week 16 after treatment and a pruritus score of < 2.1 at the same time point. Of the 51 responders, 38 (74.5%) matched these characteristics. In cases with skin symptom exacerbation (AD/OTCD = 1), patients were predicted to be nonresponders regardless of the itch scale score at week 16, with 20 of 25 patients (80.0%) being classified as nonresponders. According to multicenter retrospective studies, certain patients treated with nemolizumab have been observed to develop distinct clinical skin symptoms that tend to occur early in the treatment and are not significantly associated with baseline disease severity or biomarkers [24]. Although discontinuation of nemolizumab may be necessary in severe cases, mild symptoms can often be managed with strong potency agents. In fact, 25 of 90 rapid responders (patients who achieved VAS50 at week 16) (27.8%) experienced an exacerbation of skin symptoms by week 16, indicating that worsening skin conditions by week 16 with nemolizumab do not predict a lack of early improvement in VAS scores.
This study has several limitations. First, the analysis was based exclusively on Japanese Phase III clinical trial populations, and therefore, its external validity remains uncertain, particularly in non-Japanese patients or in those with previous failure of biologic treatment. Second, as an exploratory post hoc analysis of controlled clinical trials, the findings may not fully generalize to real-world practice, and the strict inclusion criteria, limited sample size, and post hoc design raise the possibility of selection bias and overfitting despite repeated model training and internal validation. Third, the AD/OTCD variable was based on investigator-reported adverse events within a controlled trial environment, which does not reflect the lower visit frequency and less granular documentation in routine clinical care. Moreover, the heterogeneous nature of “other cutaneous disorders” may introduce variability (details in Table S4). Fourth, the model only applies to patients who do not achieve VAS50 at week 16 and therefore does not inform decision-making for rapid responders, who should not discontinue treatment even if they experience transient skin worsening. Finally, to enhance interpretability and clinical usability, the integrated model was simplified to include only two variables. The model predicts the likelihood of improvement in pruritus, rather than comprehensive AD control; future models should incorporate multidimensional outcomes, including EASI, patient-reported outcomes, and biomarker responses.
Conclusions
Applying ML, early clinical factors were identified that predict mid- to long-term improvements in pruritus VAS scores among patients treated with nemolizumab. Our findings show that the absence of worsening cutaneous symptoms through week 16 and an itch scale score < 2.1 at week 16 are robust yet straightforward indicators of subsequent long-term response. These predictors may enable clinicians to make more informed decisions about continuing nemolizumab administration beyond week 16 in patients who do not initially demonstrate VAS50. Although these results may not be universally generalizable, they underscore the importance of ongoing monitoring and individualized treatment decisions for patients receiving nemolizumab. Future studies that incorporate real-world datasets and broader clinical outcomes are essential to refine predictive tools and identify patients who may benefit most from long-term nemolizumab therapy.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
The authors would like to thank the patients and their families, the investigators, and the project team members at Maruho Co., Ltd., Osaka, Japan.
Author Contributions
Tomoya Maeda, Kotaro Iwasaki, and Noriaki Kaneda designed the study. Tomoya Maeda was responsible for conducting the study. Tomoya Maeda, Kotaro Iwasaki, and Masashi Takizawa performed data analysis and interpretation. Tomoya Maeda, Kotaro Iwasaki, Masashi Takizawa, Kenji Kabashima, and Makoto Kawashima wrote and critically reviewed the manuscript. All authors approved the final manuscript for submission.
Funding
Maruho Co., Ltd., Osaka, Japan, supported this study and provided funding for the journal’s Rapid Service Fee.
Data Availability
The datasets generated during and/or analyzed during the current study are not publicly available due to the protection of personal information.
Declarations
Conflict of Interest
Makoto Kawashima received personal fees from Maruho Co., Ltd., and Sato Pharmaceutical Co., Ltd. Tomoya Maeda, Kotaro Iwasaki, Noriaki Kaneda, and Masashi Takizawa are employees of Maruho Co., Ltd. Kenji Kabashima has received consulting fees, honoraria, grant support, and/or lecture fees from AbbVie, Amgen, Eli Lilly, Kyowa Kirin, Japan Tobacco, LEO Pharma, Maruho, Mitsubishi Tanabe, Ono Pharmaceutical, Pfizer, Procter & Gamble, Sanofi, Regeneron, Taiho, and Torii Pharmaceutical.
Ethical Approval
The present study was a post hoc analysis of pooled data from two previously conducted Japanese Phase III clinical trials, Study-JP01 (JapicCTI-173740) and Study-JP02 (JapicCTI-183894), and involved no new recruitment, intervention, or contact with participants. Accordingly, additional ethical approval for the present analysis was not required. The original clinical trials were conducted in accordance with the ethical principles of the Declaration of Helsinki, Good Clinical Practice guidelines, and all other applicable regulatory requirements. The trial protocols and related documentation were approved by the IRBs at each participating center, including the Sapporo Skin Clinic IRB. Written informed consent was obtained from all patients or their legal guardians prior to initiation of any study procedures and treatment. Publicly available ethics committee reference numbers/IRB numbers for the original trials could not be identified from the published reports.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Nemoto O, Furue M, Nakagawa H, et al. The first trial of CIM331, a humanized antihuman interleukin-31 receptor A antibody, in healthy volunteers and patients with atopic dermatitis to evaluate safety, tolerability and pharmacokinetics of a single dose in a randomized, double-blind, placebo-controlled study. Br J Dermatol. 2016;174:296–304. [DOI] [PubMed] [Google Scholar]
- 2.Oyama S, Kitamura H, Kuramochi T, et al. Cynomolgus monkey model of interleukin-31-induced scratching depicts blockade of human interleukin-31 receptor A by a humanized monoclonal antibody. Exp Dermatol. 2018;27:14–21. [DOI] [PubMed] [Google Scholar]
- 3.Kabashima K, Matsumura T, Komazaki H, Kawashima M, Nemolizumab-JP01 Study Group. Trial of nemolizumab and topical agents for atopic dermatitis with pruritus. N Engl J Med. 2020;383:141–50. [DOI] [PubMed] [Google Scholar]
- 4.Kabashima K, Matsumura T, Komazaki H, Kawashima M, Nemolizumab JP01 and JP02 Study Group. Nemolizumab plus topical agents in patients with atopic dermatitis (AD) and moderate-to-severe pruritus provide improvement in pruritus and signs of AD for up to 68 weeks: results from two Phase III, long-term studies. Br J Dermatol. 2022;186:642–51. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Kabashima K, Matsumura T, Komazaki H, Kawashima M, Nemolizumab-JP01 Study Group. Nemolizumab improves patient-reported symptoms of atopic dermatitis with pruritus: post hoc analysis of a Japanese Phase III randomized controlled trial. Dermatol Ther (Heidelb). 2023;13:997–1011. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Saeki H, Ohya Y, Arakawa H, et al. English version of clinical practice guidelines for the management of atopic dermatitis 2024. J Dermatol. 2025;52:e70–142. [DOI] [PubMed] [Google Scholar]
- 7.Masuo Y, Yonekura S, Nakajima S, Sawada Y, Kabashima K. Clinical factors predicting nemolizumab response in atopic dermatitis. J Allergy Clin Immunol Glob. 2025;4:100457. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Sugiyama A, Yano-Takamori A, Kojima K, et al. Characteristics of suitable cases for treatment with nemolizumab based on clinical findings and cutaneous adverse events. J Dermatol. 2025;52:740–3. [DOI] [PubMed] [Google Scholar]
- 9.Ferrucci S, Casazza G, Zussino M, Tavecchio S, Marzano AV, Tedeschi M. Predictive factors of early response to dupilumab in patients with moderate-to-severe atopic dermatitis. J Clin Med. 2023;12:6575. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Wu JJ, Hong CH, Merola JF, et al. Predictors of nonresponse to dupilumab in patients with atopic dermatitis: a machine learning analysis. Ann Allergy Asthma Immunol. 2022;129:354–9. [DOI] [PubMed] [Google Scholar]
- 11.Ashizaki K, Ishikawa T, Nomura Y. Residual facial erythema in atopic dermatitis patients treated with dupilumab stratified by machine learning. J Eur Acad Dermatol Venereol. 2024;38:2286–94. [DOI] [PubMed] [Google Scholar]
- 12.Kawashima M, Nakagawa H. Olopatadine hydrochloride in children: evidenced efficacy and safety for atopic dermatitis treatment in a randomized, multicenter, double-blind, parallel group comparative study. Nishinihon J Dermatol. 2011;73:278–89. [Google Scholar]
- 13.Breiman L, Friedman J, Olshen RA, Stone CJ. Classification and regression trees. Belmont: Wadsworth; 1984. [Google Scholar]
- 14.van Rijsbergen CJ. Information retrieval. London: Butterworths; 1979. [Google Scholar]
- 15.Manning CD, Raghavan P, Schütze H. Introduction to information retrieval. Cambridge: Cambridge University Press; 2008. [Google Scholar]
- 16.Javed Mehedi Shamrat FM, Ghosh P, Sadek MH, Kazi MA, Shultana S. Implementation of machine learning algorithms to detect the prognosis rate of kidney disease. IEEE international conference for innovation in technology (INOCON), Bengaluru. 2020. pp. 1–7.
- 17.Jacobs MC. Regression trees versus stepwise regression. UNF graduate theses and dissertations. 1992. pp. 145. https://digitalcommons.unf.edu/etd/145
- 18.Lundberg SM, Lee SI. A unified approach to interpreting model predictions. 31st conference on neural information processing systems (NIPS 2017), Long Beach, CA. 2017. pp. 4768–77.
- 19.Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: machine learning in python. J Mach Learn Res. 2011;12:2825–30. [Google Scholar]
- 20.Verduzco HA, Shirazian S. CKD-associated pruritus: new insights into diagnosis, pathogenesis, and management. Kidney Int Rep. 2020;5:1387–402. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Gembillo G, Siligato R, Cernaro V, et al. Monocyte to HDL ratio: a novel marker of resistant hypertension in CKD patients. Int Urol Nephrol. 2022;54:395–403. [DOI] [PubMed] [Google Scholar]
- 22.Mayo MJ, Carey E, Smith HT, et al. Impact of pruritus on quality of life and current treatment patterns in patients with primary biliary cholangitis. Dig Dis Sci. 2023;68:995–1005. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Kaps L, Grambihler A, Yemane B, et al. Symptom burden and treatment response in patients with primary biliary cholangitis (PBC). Dig Dis Sci. 2020;65:3006–13. [DOI] [PubMed] [Google Scholar]
- 24.Sasaki W, Saito R, Suzuki K, et al. Clinical characteristics and risk factors for cutaneous manifestations associated with nemolizumab in atopic dermatitis: a multicenter retrospective study in Japan. J Dermatol. 2026;53:399–409. [DOI] [PMC free article] [PubMed]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets generated during and/or analyzed during the current study are not publicly available due to the protection of personal information.