Perform validation using a non-proportional hazards (exponential model) and compare it with a linear model to assess the stability of both models.2. Exclude short-term trials (<24 weeks) and conduct a direct comparison of the 24-week trial data. All analyses adhered to the Bucher method for indirect comparisons (14).

See Table 4.

See Table 5.

See Table 6.

1. Sensitivity analysis using non-proportional hazards model (exponential model)

To validate the primary analysis assumption of linear risk accumulation over time, we conducted sensitivity analyses with a non-proportional hazards model (exponential model). This model assumes that the incidence of adverse events (AEs) follows an exponential relationship over time, defined as: λ(t) = λ₀e^βt, where event rates were recalculated using maximum likelihood estimation and standardized to a 24-week timeframe.

Results: The adjusted pooled risk ratio (RR) was 1.07 (95% CI: 0.89–1.28) with no heterogeneity (I²) = 0%。Compared to the primary analysis under the linear assumption [pooled RR: 1.09 (0.95–1.25)], the difference was minimal (2%), and confidence intervals fully overlapped.

Conclusion: The exponential model yielded results highly consistent with the linear model, supporting the robustness of the linear risk accumulation hypothesis.

2. Subgroup analysis (trials with 24-week duration only)

To evaluate the impact of short-term trials (12- and 16-week durations) on pooled results, we analyzed data exclusively from 24-week trials.

Results: The RR for 24-week trials was 1.10 (95% CI: 0.92–1.30), aligning in direction and showing substantial overlap with the primary pooled RR [1.09 (0.95–1.27)].

Sensitivity analyses excluding short-term trials revealed a difference of <1% from the original pooled RR, indicating negligible influence of short-term data on conclusions.

Conclusion: The subgroup analysis confirmed the stability of primary findings, with no evidence of bias introduced by short-term trials.

3. ARD analysis

In all clinical trials, namilumab and dupilumab demonstrated no statistically significant differences (confidence intervals included 0). However, numerical trends toward higher rates of adverse events (AEs) were consistently observed in the experimental groups compared to the placebo groups (e.g., in the Shawn trial for namilumab, the absolute risk difference [ARD] reached 8.5%). These findings warrant careful interpretation in the context of clinical significance (Table 8).

1. Sensitivity analysis using non-proportional hazards model (exponential model)

In the sensitivity analysis using a non-proportional hazards model (exponential model), the hazard ratio was 0.61 (95% CI: 0.34–1.09), which aligns in direction and shows overlapping confidence intervals with the primary proportional hazards model result (0.79 [0.42–1.48]). Although the point estimate difference suggests a potentially more pronounced short-term risk reduction in the exponential model, no significant heterogeneity was observed between the two models within statistical uncertainty bounds (p = 0.32). These findings support both the robustness of the primary analysis and the validity of the proportional hazards assumption.

2. Subgroup analysis (trials with 24-week duration only)

In the subgroup analysis restricted to 24-week trials, the pooled risk ratio (RR) point estimate was slightly higher (0.81 vs. 0.79) but accompanied by a wider confidence interval (0.38–1.72 vs. 0.42–1.48). Crucially, the interval fully encompassed the original pooled estimate and did not cross the null line (RR = 1). This indicates that excluding short-duration trials did not materially alter the direction or significance of the conclusions.

3. ARD analysis

When interpreting absolute risk differences (ARDs), clinicians should consider both confidence intervals and baseline placebo rates. For example, nemolizumab’s SAE ARD of −3.3% corresponds to a number needed to treat (NNT) of 30 to prevent one SAE, but the 95% CI (−8.5 to +1.9%) indicates this estimate is compatible with both meaningful benefit and trivial harm (Table 9).

1. Sensitivity analysis using non-proportional hazards model (exponential model)

After adjustment using the non-proportional hazards model, the pooled risk ratio (RR) shifted from 0.89 to 1.21 (95% CI: 0.29–5.02), with substantial overlap in confidence intervals. This indicates that even under the assumption of exponentially increasing risk over time, the conclusions remained consistent, showing no statistically significant difference (p > 0.05).

2. Subgroup analysis (trials with 24-week duration only)

In the subgroup analysis restricted to 24-week trials, the RR suggested a marginal increase in risk (1.14), though the confidence interval (0.36–3.62) included the null value (RR = 1). The direction of this estimate contrasted with the primary pooled result (RR = 0.89). Potential explanations include:

Small-sample bias: Limited event counts (treatment group: 9 events vs. control: 4 events) led to unstable effect estimates.

Sources of heterogeneity: Shorter-duration trials (e.g., a 12-week trial with RR = 3.17) disproportionately influenced the pooled RR downward, though their own confidence intervals were extremely wide (e.g., 0.67–15.02), reflecting high uncertainty.

3. ARD analysis

In the Sonja trial for namilumab, a marginally higher rate of adverse events leading to treatment discontinuation (AELTD) was observed in the experimental group compared to the placebo group [absolute risk difference (ARD) = 2.9%]; however, the confidence interval (CI) included 0, indicating no statistically significant difference (Table 10).

Through Bayesian and Beta-binomial model sensitivity analyses, we observed that: The continuity correction may underestimate the true variance (e.g., for the edema risk ratio: 1.03 in the original model vs. 1.01 in the Bayesian method); RR: 1.01 (0.01–102.4). The beta-binomial model fails to converge in double-zero event studies (e.g., the edema data from the Nemolizumab group), highlighting methodological limitations in analyzing ultra-sparse data. Bayesian credible intervals spanning multiple orders of magnitude (e.g., edema 95% CrI: 0.01–102.4) demonstrate that statistical “non-significance” does not equate to clinical equivalence. Although sensitivity analyses demonstrate the robustness of conclusions for primary outcomes (e.g., AELTD risk), the extremely wide credible intervals for zero-event outcomes like edema underscore the need for clinical vigilance against extreme risks in single-study samples. Real-world data are essential to supplement these findings, particularly for long-term medication use. For outcomes dominated by zero events (e.g., edema), Bayesian models yield exceptionally broad credible intervals (RR = 1.01 [0.01–102.4]), reflecting inherent uncertainty in sparse event modeling, albeit directionally consistent with the primary results (Table 12).

Meta-regression identified baseline itch severity (β = 0.062 per 1-point increase, p = 0.027) and itch assessment tools (WI-NRS vs. PP-NRS: β = 0.127, p = 0.032) as independent modifiers of adverse event (AE) risk.

Stratified analyses: The WI-NRS group showed a nominally higher AE risk point estimate (RR = 1.17 vs. 1.13 with PP-NRS), though the between-group difference was not statistically significant (p = 0.28).

Sex effect: A higher proportion of female participants was weakly associated with reduced AE risk (β = −0.005 per 1% increase, p = 0.012).

After adjusting for baseline heterogeneity, indirect comparisons demonstrated a diminished risk difference between the two agents (adjusted RR = 1.05, 95% CI: 0.82–1.34).

This exploratory analysis based on the Bucher indirect comparison framework assessed safety profiles between nemolizumab and dupilumab. The relative risk (RR) of 1.11 (95% CI: 0.85–1.47) for overall adverse events (AEs), indicating no statistically significant difference between the two biologics. However, numerical differences were noted in absolute risk metrics. All observed numerical differences should be interpreted as hypothesis-generating signals rather than confirmatory evidence, given their lack of statistical significance and overlapping confidence intervals. Notably, the wide CI spanning the null value in the indirect RR comparison (0.85–1.47) underscores substantial uncertainty in comparative AE risks. Furthermore, the absence of severity-stratified data (e.g., CTCAE grading) precludes clinical interpretation of frequency-based outcomes. For instance, nemolizumab’s significant ARD may be driven by transient mild events (e.g., injection-site reactions), while dupilumab’s CI spanning benefit and harm could reflect heterogeneous risk profiles combining low-grade conjunctivitis with potential rare serious AEs. These exploratory findings do not establish clinically meaningful risk differentials. Treatment decisions should weigh individual patient factors against the unquantifiable uncertainty inherent in indirect comparisons.

The observed numerical difference in edema events with nemolizumab [absolute risk difference (ARD) = 0.062–0.190/100 person-weeks, +1.17%], corresponding to a number needed to harm (NNH) of 85, requires cautious interpretation within rigorous methodological constraints despite time-corrected sensitivity analyses. While the pathophysiological rationale suggests IL-31 receptor antagonists may theoretically induce fluid retention through neurovascular modulation pathways (16–19). The current evidence remains exploratory due to multiple limitations. Statistically, the substantial overlap in risk ratio confidence intervals [RR = 1.89, 95% CI 0.52–5.18; sensitivity analysis RR = 2.54 (0.30–21.43)] indicates susceptibility to type I/II errors. Methodological concerns include non-standardized fluid retention assessment protocols that risk detection bias amplification, challenges in differentiating random event clustering from true drug effects given low placebo-group event rates (0–2 cases/group) and treatment duration heterogeneity (12–24 weeks), and potential model misspecification arising from the linear time-effect assumption conflicting with nonlinear drug accumulation patterns. Clinically, even if accepting signal validity, even if a true risk exists, the estimated population-level risk elevation is minimal [ARD<0.2/100 person-weeks, number needed to treat (NNT) = 1,610–526], likely below clinical management significance.

This mechanistic-methodological paradox underscores the necessity for prospective validation through pre-specified fluid monitoring protocols, standardized treatment durations, and expanded sample sizes to resolve current evidentiary uncertainties.

Current evidence indicates that for nemolizumab, no conjunctivitis events were observed in the 24-week trial (0/187 vs. 1/95 in placebo), though 12-week data revealed a non-significant numerical difference with an upward trend in relative risk (RR = 1.59, 95%CI:0.28–8.93), potentially related to early exposure dynamics. The absolute risk difference (ARD) exhibited an asymmetric distribution. For dupilumab, exploratory analyses showed numerically elevated but statistically non-significant conjunctivitis risk at 24 weeks (ARD = +3.89%, 95%CI:-0.4% to +8.2%; RR = 2.01, 95%CI:0.29–13.77). The ultra-wide confidence intervals spanning both potential harm (upper bound: RR = 13.77) and protection (lower bound: RR = 0.29), coupled with the absence of statistical significance, strictly preclude causal inferences. While the observed numerical imbalance theoretically aligns with IL-4/IL-13 pathway inhibition effects on ocular mucosal immunity (20), this hypothesis-generating signal must be interpreted as an exploratory observation requiring validation in prospective trials with protocol-driven ophthalmic monitoring. These analyses collectively emphasize that numerical differences even those mechanistically plausible do not constitute confirmatory evidence of risk and should serve solely to inform future hypothesis-testing studies.

Nemolizumab showed heterogeneous AE risk across trials. Specifically, the most notable absolute difference was observed in Ständer 2025 trial (n = 282, ARD+8.9% with placebo AE rate 65.3%), while the paradoxical result in Ständer 2020 trial (n = 70, ARD-5.6% with placebo rate 66.7%) may reflect limited sample size. Although the upper CI limit suggests potentially clinically relevant differences in high-risk scenarios, it’s crucial to emphasize this estimate should not be extrapolated across trials.

The rarity of mechanism-related events (e.g., edema, conjunctivitis) led to imprecise ARD estimates. For example, nemolizumab’s edema ARD was +1.17% (95% CI: −1.24 to 3.58%), with a number needed to treat (NNH) of 85 to result in one additional edema case. The clinical relevance of such findings may depend on the severity of the outcome.

This study examined the independent effects of baseline pruritus scores and assessment tools on AE risks:

Pruritus intensity and risk association: Higher baseline pruritus scores (regardless of WI-NRS or PP-NRS) were associated with increased AE risks (+6.2% per unit), which may correlate with more severe skin barrier disruption and neuroimmune activation in patients with intense pruritus.

Differences in assessment tools: The AE risk point estimate was higher in the WI-NRS group (dupilumab trials) compared to the PP-NRS group (nemolizumab trials). This discrepancy might arise because WI-NRS captures “worst-itch moments,” potentially reflecting short-term inflammatory fluctuations, whereas PP-NRS measures “average intensity,” possibly aligning more closely with chronic pathological burden.

Baseline heterogeneity across trials (e.g., variability in prurigo duration) may influence outcomes in indirect comparative analyses.

A critical limitation of this study is the absence of CTCAE (Common Terminology Criteria for Adverse Events) severity grading across all included trials. The inability to stratify adverse events by severity (e.g., mild, moderate, severe) significantly constrains the clinical interpretation of safety profiles. For instance: Frequency-Severity Discrepancy: Equivalent overall AE rates between dupilumab and nemolizumab may mask divergent clinical impacts. Dupilumab’s observed conjunctivitis trend (RR = 2.01) might primarily involve mild, self-limiting cases (e.g., Grade 1: transient irritation), whereas nemolizumab’s edema signals (RR = 1.64) could represent Grade 2–3 events requiring therapeutic intervention. Such distinctions are critical for risk–benefit assessments in vulnerable populations (e.g., patients with preexisting ocular surface disease or cardiorenal comorbidities). Misinterpretation of Safety Signals: Severe but infrequent AEs (e.g., dupilumab-associated systemic infections or nemolizumab-induced angioedema) may be underestimated in frequency-based analyses, while high-frequency mild AEs (e.g., injection-site reactions) could overstate perceived risks. This creates a false equivalence in safety comparisons. Impact on Clinical Decision-Making: Without severity data, clinicians cannot prioritize interventions based on AE criticality. For example, a higher incidence of mild conjunctivitis may be clinically acceptable if balanced against lower risks of severe edema, but this trade-off remains unquantifiable in the current evidence.

Proposed solutions for future trials: Mandate CTCAE-Aligned Grading: All mechanism-specific AEs (e.g., conjunctivitis, edema) should be reported with CTCAE severity grades (Grades 1–5) and duration-adjusted event rates (e.g., events per 100 person-weeks by grade). Develop Composite Metrics: Integrate frequency and severity into weighted scores [e.g., AE severity index = Σ (Grade × Events)/Person-Time] to better reflect cumulative burden. Report Clinical action ability: Explicitly categorize AEs by their management requirements (e.g., “self-resolving,” “requiring topical therapy,” “leading to hospitalization”).

Extrapolating short-term nemolizumab data to 24-week exposure assumes non-proportional hazards (i.e., constant hazard ratios over time), which may overestimate differences if treatment effects vary with duration.

While meta-regression adjusted for pruritus assessment tools (WI-NRS vs. PP-NRS), residual confounding may persist due to unmeasured dimensional heterogeneity (e.g., “worst-itch moments” vs. “average intensity”). Additionally, shorter disease duration in WI-NRS trials (mean 5.4–5.7 years vs. 7.6–8.8 years in nemolizumab trials) could confound the relationship between disease chronicity and AE risks.

PN is a chronic, debilitating condition characterized by intractable itch, sleep disruption, and reduced quality of life. Patients often have severe disease refractory to conventional therapies. In such high-need populations, higher ARD thresholds for adverse events (AEs) may be acceptable if accompanied by robust efficacy (e.g., sustained itch reduction or lesion clearance).

Trials of dupilumab and nemolizumab demonstrate clinically significant efficacy: Dupilumab: WI-NRS improvement ≥4-point (60% dupilumabvs 18.4% placebo) (8). Nemolizumab: PP-NRS ≥ 4-point reduction (56.3% nemolizumab vs20.9% placebo) (9). FDA considers endpoints reflecting direct patient benefit (e.g., itch reduction, functional improvement) as critical to offset AE risks.

Overall AEs: The ARD for nemolizumab ranged from +1.0% to +8.5% across trials. In PN, this risk may be acceptable given the high unmet need.

Conjunctivitis (dupilumab: ARD 0–3.89%) typically involves mild, reversible cases, whereas edema (nemolizumab: ARD +1.17%) may require monitoring in cardiorenal comorbidities. These risks are generally manageable with routine care.