🤖 Generated with AI assistance · Does not replace individualized medical advice.
AI in Neuroblastoma: A Systematic Review Highlights Proof-of-Concept but Significant Validation Gaps
AI Act Transparency Notice: This content was generated with the assistance of artificial intelligence tools and has been reviewed by a medical oncology expert. It is for informational purposes only and does not constitute medical advice.
Lead
A 2026 systematic review and meta-analysis of 53 studies evaluates AI’s role in neuroblastoma (NB) diagnosis, risk stratification, and prognosis. While machine learning models show promising diagnostic and prognostic performance, widespread lack of external validation and calibration renders clinical adoption premature.
Key Facts
| Metric | Value | Context |
|---|---|---|
| Included Studies | 53 | From 3,742 identified articles PMID: 42696227 |
| Diagnostic AUC (MLM) | 0.87 | vs. 0.83 for radiologists (NS) PMID: 42696227 |
| Risk Stratification AUC | 0.87 | Achieved by Hybrid Nomograms PMID: 42696227 |
| External Validation | 24.5% | Only 13 of 53 models validated externally PMID: 42696227 |
Why It Matters / Context
Neuroblastoma is the most common extracranial solid tumor in children, with outcomes heavily dependent on accurate risk stratification and early detection of high-risk features. Traditional prognostic markers, such as MYCN amplification and stage, are robust but do not capture the full heterogeneity of the disease. Artificial intelligence (AI), specifically Machine Learning Models (MLMs) and Hybrid Nomograms (HNs), offers a potential pathway to integrate complex genomic, imaging, and clinical data into unified predictive tools.
This meta-analysis is critical because it moves beyond single-center case reports to synthesize the current evidence base. The distinction between MLMs (pure algorithmic approaches) and HNs (integrating clinical variables into nomograms) is particularly relevant for oncologists, as HNs often align more closely with current clinical decision-making frameworks. The finding that AI-derived nomograms achieve an AUC of 0.9 for prognosis suggests a potential leap in predictive accuracy over conventional markers, provided the models are robustly validated. However, the field is currently characterized by a “validation gap,” where high internal performance metrics do not translate to reliable external utility.
Details
The study classified AI models into two primary categories: Machine Learning Models (MLMs) and Hybrid Nomograms (HNs). The performance across different clinical domains is summarized below. Note that “NS” indicates the difference was not statistically significant.
| Clinical Domain | Model Type | Performance Metric | Comparison / Notes |
|---|---|---|---|
| Differential Diagnosis | MLMs | AUC: 0.87 | Superior to radiologists (AUC: 0.83), but difference NS PMID: 42696227 |
| Risk Stratification | HNs | AUC: 0.87 | Stronger performance than other model types in this domain PMID: 42696227 |
| Prognosis | AI Nomograms | AUC: 0.90 | Descriptively outperformed conventional markers PMID: 42696227 |
| Prognosis | Gene Signatures | AUC: 0.80 | Descriptively outperformed conventional markers PMID: 42696227 |
| Chemo Response | All Types | Below Threshold | Failed to meet clinical utility standards across all models PMID: 42696227 |
The meta-analysis utilized random-effects models with logit-transformed AUCs and cluster-robust standard errors to account for heterogeneity among the included studies. The inclusion of 53 studies from five databases provides a comprehensive snapshot of the current literature, although the heterogeneity in model construction and dataset composition limits the precision of pooled estimates.
Limitations
Critical Limitation: Validation and Calibration Deficits
The most significant barrier to clinical translation identified in this review is the lack of rigorous methodological reporting. Only 33.9% of the included models reported calibration metrics, which are essential for assessing the agreement between predicted and observed probabilities PMID: 42696227. Furthermore, only 24.5% underwent external validation, meaning the majority of reported performance metrics are likely overfitted to their training datasets PMID: 42696227.
Additionally, the study notes persistent gaps in dataset size and the absence of pediatric-specific model development. Many models may have been trained on adult cancer data or small, single-institution pediatric cohorts, limiting their generalizability. The prediction of chemotherapy response, a key determinant of treatment intensity in NB, remained below clinical utility thresholds across all model types, indicating that AI has not yet solved the problem of identifying non-responders early in the treatment course PMID: 42696227.
Clinical Takeaway
Clinicians should view current AI tools for neuroblastoma as research instruments rather than clinical decision aids, prioritizing prospective multicenter validation through collaborative groups like COG or SIOPEN before integrating these models into routine risk stratification or diagnostic workflows.
Full Reference
Serbena DR, Cieslack ILF, Ratis RC, Chaves DVP, de Oliveira Filho SS, Neves H. Artificial Intelligence for Diagnosis, Risk Stratification, and Prognosis of Neuroblastoma – A Systematic Review and Meta-Analysis. Current Oncology Reports. 2026. PMID: 42696227 [DOI: 10.1007/s11912-026-01824-0]