ScoreNet: an interpretable neural network that turns eight baseline measurements into a risk score in advanced melanoma, trained on five trials and validated on two others
A team centred on the US Food and Drug Administration (FDA) used an "interpretable" neural network to build a simple prognostic score, a weighted sum of eight measurements taken before treatment, in patients with advanced melanoma. The score was built on 2,711 patients from five trials that supported drug approvals, then validated on two other trials, and it sorts patients into low, moderate or high risk of reduced survival. The idea is attractive because it yields a tool usable at the bedside; what remains to be seen, in the full text we could not read, is how much better it does than classical scores.
The context
In advanced melanoma (unresectable or metastatic), prognosis varies enormously from one patient to another, even on the same treatment. Clinicians have known markers: LDH level (lactate dehydrogenase, an enzyme whose elevation generally reflects a large tumour burden), the ECOG performance status (a 0-to-5 scale describing the patient's general condition), the presence of liver metastases, or PD-L1 expression measured by immunohistochemistry (a stain showing whether the tumour expresses a protein targeted by some treatments). But these markers are mostly assessed one by one, or combined in classical statistical models.
The problem is twofold. On one side, a deep learning model can combine many variables, but it behaves as a black box that nobody wants to use to classify a patient. On the other, a points-based score, of the "add a few numbers" kind, is readable but is often built by hand or by regression, with strong assumptions about the shape of relationships. ScoreNet tries to hold both ends: let a neural network learn the weights, but constrain it to produce, in the end, a readable additive score.
The method
The task. This is prognostic prediction: from variables measured before treatment, estimate the risk of death (overall survival, OS) and of progression or death (progression-free survival, PFS), then split patients into three groups, low, moderate and high risk.
The model. ScoreNet is described as an "end-to-end", "interpretable" neural network. In this context, interpretable means that the network's output reduces to a weighted sum of input variables, which can then be converted into points. We could only read the abstract and metadata of the early-access version: the depth of the network, how weights are turned into points, the loss function for survival data and the handling of missing values are therefore not described here.
The data. Training uses 2,711 patients with advanced melanoma, from five clinical trials that supported marketing authorisations. These are randomised trials whose patients are selected by strict inclusion criteria: a more homogeneous and fitter population than a routine clinic. Validation is done on two separate trials. The names of the trials, their respective sizes and dates are not in the abstract.
The eight variables of the score. Sum of the longest diameters of target lesions (a standardised measure of tumour burden on imaging), LDH, the systemic immune-inflammation index (a ratio computed from blood cells), haemoglobin, ECOG status, liver metastasis, pulse rate, and PD-L1 status by immunohistochemistry.
The comparator. The abstract mentions none. We therefore do not know whether ScoreNet is compared with a Cox model (the reference statistical model for survival data), an existing score or an unconstrained network.
The results
ML metrics. In the version we could read, no concordance index (C-index, the probability that the model correctly ranks two patients by survival), hazard ratio, confidence interval or calibration is accessible. We therefore do not quote them, and we refuse to guess. What is established: the score is a weighted sum of the eight variables, it separates three risk groups for OS and PFS, and the stratification applies, according to the authors, to patients treated with PD-1/PD-L1 inhibitors, chemotherapy, CTLA-4 inhibitors, BRAF/MEK inhibitors and their combinations.
Clinical translation. Without survival figures per group, it is impossible to say how many patients out of 1,000 would be wrongly reclassified. One can, however, ask the question that matters: is the survival difference between the high-risk and the low-risk group large enough to change a decision (treatment, surveillance, trial enrolment)? That is the value of the score, and it is what the full text must document.
What is good
A readable eight-variable score. All the elements are routine oncology measurements (LDH, haemoglobin, pulse, ECOG, imaging, PD-L1). A clinician can check the score by hand and challenge a value, which is impossible with a black box.
Regulatory-grade trial data. The five training trials supported drug approvals: measurements are collected according to a protocol, with systematic follow-up of events. This is a cleaner base than retrospective hospital records.
Validation on two trials distinct from training. Validating on trials not used for learning is a minimum requirement that many prognostic studies fail to meet. The authors also declare no competing interests and no specific funding, and the article is open access under a CC BY 4.0 licence.
What is less good
A trial-related population bias. Trial patients are selected (preserved general condition, acceptable organ function, no heavy comorbidity). A score trained there may behave poorly on real-world patients, older and more fragile, who are precisely those for whom decision support is most needed. This is the classic limit of moving from trial to practice.
Heterogeneous treatments and a changing era. The score is presented as applicable to patients treated with chemotherapy, anti-CTLA-4, anti-PD-1 and targeted therapies, in trials run over several years. But care has changed: a prognostic score easily mixes treatment effect with the effect of the era. Validation on patients treated with today's immunotherapy regimens, including recent combinations, is essential, and we do not know what the two validation trials cover.
A comparator and metrics we could not see. Without a comparison to a Cox model using the same eight variables, we do not know what the neural network itself adds over classical regression. If the gain is small, the interpretability argument favours regression. The metric to demand is twofold: discrimination (C-index) and calibration, that is, agreement between predicted and observed risk, often forgotten in points-based scores.
What it changes
For the research community. The paper illustrates a useful route: use deep learning to learn the weights, then make the result transparent. The logical next steps are systematic comparison with a penalised Cox model and release of the code and weights, which we did not find mentioned.
For clinicians. Nothing changes today: this is a retrospective prognostic score, with no study of its impact on decisions. It could, however, be used to stratify patients when designing future trials, which probably explains the presence of several FDA offices among the affiliations.
For patients and the public. A prognostic score is not an individual verdict: it describes groups, with a margin of uncertainty. Its value is to help talk about risk more clearly, not to decide in place of the patient and their oncologist.
Going further
Paper published in npj Digital Medicine: doi.org/10.1038/s41746-026-03379-7 (received 15 November 2025, accepted 27 September 2026, published online 13 October 2026, early-access version, subject to further edits). Corresponding author: Qi Liu (FDA, Center for Drug Evaluation and Research); lead authors: Ruihao Huang, Hao Zhu, Yue Huang, Hanrui Zhang and Yuanfang Guan. Affiliations: several FDA centres, University of Michigan Medical School and Stanford University. No specific funding or competing interests declared.
Editorial transparency: this article was produced with the help of artificial intelligence tools from the early-access article page (abstract and metadata); the Methods and Results sections could not be consulted, so no numerical result is quoted and the journal version may differ. The English, Spanish and Chinese versions are machine translations.