SENTINOLD and the Challenge of Meaningful Risk Prediction in Melanoma

1.

Sentinel lymph node biopsy (SLNB) remains one of the most important prognostic procedures in melanoma. Although SLNB provides valuable staging information and may influence decisions regarding adjuvant therapy, identifying which patients are most likely to benefit from the procedure remains challenging, particularly among older adults. Elderly patients frequently present with biologically aggressive melanomas and worse melanoma‐specific survival, yet paradoxically demonstrate lower rates of sentinel node positivity than younger patients. Age‐related comorbidities, competing mortality risks, and procedural considerations further complicate decision‐making [1].

In this issue of the International Journal of Dermatology, Martínez Campayo and colleagues present the SENTINOLD nomogram, a prediction model developed specifically for patients aged 75 years and older [2]. By focusing on a population that has historically been underrepresented in melanoma studies, the authors address an important clinical gap. Their work represents another step toward individualized risk assessment and more nuanced discussions regarding SLNB in elderly patients.

The development of SENTINOLD also raises a broader question: what evidence should be required before predictive models are incorporated into routine clinical practice?

Historically, melanoma clinicians have welcomed new nomograms as potential tools to improve SLNB decision‐making. The Melanoma Institute Australia (MIA) SLNB nomogram, for example, was initially viewed as a major advance because it demonstrated improved discrimination compared with earlier models and performed well in several validation cohorts [3]. Yet subsequent population‐based validation studies revealed a more nuanced picture. Although overall discrimination remained acceptable, clinical utility proved more limited at the lower risk thresholds where SLNB decision‐making is most uncertain [4]. The experience highlighted an important lesson: statistical improvement alone does not necessarily translate into meaningful improvements in patient care (Table 1).

TABLE 1.

Beyond the C‐statistic: Proposed framework for evaluating melanoma prediction models intended to guide SLNB decision‐making.

Domain	Key question	Why it matters	Example limitation
Statistical performance	Can the model distinguish patients with and without SLN metastasis?	Measures overall predictive performance (e.g., AUC/C‐statistic).	Good discrimination may not translate into better SLNB decision‐making.
Calibration	Do predicted probabilities match observed outcomes?	Accurate risk estimates are essential near clinical decision thresholds	Underestimation of SLN positivity risk may lead to inappropriate avoidance of SLNB.
Clinical utility	Does use of the model improve decision‐making compared with current practice?	Determines whether predictions meaningfully change patient management.	Decision‐curve analyses may show little advantage over a “biopsy all” strategy at relevant risk thresholds.
Reproducibility	Are predictor variables measured consistently across institutions and pathologists?	Reliability of model inputs directly affects performance in clinical practice.	Mitotic rate, lymphovascular invasion, regression, and Clark level exhibit interobserver variability.
External validation	Does the model perform similarly in independent populations?	Ensures generalizability beyond the development cohort.	Performance may decline when applied to different patient populations, healthcare systems, or pathology practices.
Clinical impact	Does adoption improve outcomes, reduce unnecessary procedures, or enhance shared decision‐making?	Represents the ultimate test of clinical value.	Improved prediction may not meaningfully change SLNB utilization or patient outcomes.

Note: The experience with melanoma nomograms demonstrates that model evaluation should extend beyond statistical discrimination to include calibration, clinical utility, reproducibility of predictor variables, external validation, and ultimately clinical impact. A model that performs well in a development cohort may still have limited value if risk estimates are poorly calibrated, predictor variables are inconsistently measured, or use of the model does not improve clinical decision‐making or patient outcomes.

Accordingly, a favorable C‐statistic should be viewed as a starting point rather than the endpoint of model evaluation. The more important question is whether a prediction model improves clinical decision‐making. Experience with melanoma nomograms has shown that a model may achieve respectable overall discrimination while providing limited benefit if risk estimates are insufficiently accurate around the thresholds used to guide SLNB discussions. Accurate risk estimation at these decision points may ultimately be more important than modest improvements in overall model performance.

Beyond discrimination and clinical utility, an additional consideration is the reproducibility of the clinicopathologic variables used for prediction. Many melanoma nomograms incorporate histopathologic features such as mitotic activity, lymphovascular invasion, regression, Clark level, and histologic subtype, several of which are known to exhibit interobserver variability [5]. As models incorporate larger numbers of such variables, cumulative variability may affect performance across institutions and healthcare systems, underscoring the importance of rigorous external validation.

This challenge is not unique to any individual nomogram. Rather, it reflects a broader limitation of clinicopathologic prediction models. Additional variables may improve apparent model performance but can also introduce new sources of variability that affect generalizability. For this reason, predictive models should be held to the same evidentiary standards regardless of whether they are commercial products or freely available nomograms. Accessibility does not diminish the need for rigorous validation.

Several priorities remain for melanoma prediction models, including rigorous external validation, demonstration of clinical utility at relevant decision thresholds, assessment of robustness across diverse pathology practices, and prospective evaluation of real‐world impact.

The SENTINOLD study represents an important contribution to melanoma care and a valuable framework for individualized risk assessment in older adults. More broadly, it highlights both the promise and the challenges of contemporary prediction modeling. As melanoma risk stratification evolves, the goal is not simply to develop more sophisticated algorithms, but to create tools that are accurate, reproducible, clinically useful, and ultimately capable of improving patient outcomes in real‐world practice.

Funding

The author has nothing to report.

Conflicts of Interest

A.M. has received research funding from SkylineDx and is listed as an inventor on skin cancer‐related technologies owned by Mayo Clinic.

Linked Articles

This article is linked to Martínez Campayo et al.'s paper. To view this article, visit https://doi.org/10.1111/ijd.70413.