Main Session
Sep
27
PQA 02 - Pediatric Cancer, Sarcoma and Cutaneous Tumors, Medical Education & Professional Development, and Health Services Research
2373 - Integrating Gene Expression and Clinical Variables Improves Prediction of Disease Progression in Neuroblastoma: A Secondary Analysis of the SEQC Cohort
Presenter(s)
Soumyajit Roy, MBBS, MSc. - University Hospitals Seidman Cancer Center, Cleveland, OH
S. Roy1, S. Lichtman-Mikol2, A. Y. Jia2, D. E. Spratt3, and D. Mansur4; 1Rush University Medical Centre, Chicago, IL, 2Department of Radiation Oncology, University Hospitals Cleveland Medical Center/ Seidman Cancer Center, Cleveland, OH, 3University Hospitals Seidman Cancer Center, Case Western Reserve University, Cleveland, OH, 4Department of Radiation Oncolog, University Hospitals Cleveland Medical Center, Cleveland, OH
Purpose/Objective(s):
Neuroblastoma risk stratification relies primarily on clinical factors, yet molecular heterogeneity suggests that gene expression may provide additional prognostic information. The incremental value of integrated clinico-genomic models over gene-expression–only approaches and established clinico-genetic subgrouping remains uncertain. Using a large, publicly available neuroblastoma cohort, we evaluated whether combining gene expression with clinical variables improves prediction of disease progression.Materials/Methods:
Microarray gene-expression profiles from the MAQC-III/SEQC study were merged with curated clinical data. Of 44,708 probe sets, 8,819 mapped to gene symbols and were collapsed to 7,005 unique genes. After data cleaning, 373 patients had complete clinico-genomic features, including 147 progression events. Data were split into training (70%) and testing (30%) cohorts stratified by progression status. Gradient-boosted classifiers were trained using (1) PCA-reduced gene expression (95% variance retained) and (2) combined clinico-genomic features including age, sex, INSS stage, and MYCN amplification. Models were optimized using 5-fold cross-validation and evaluated once on the held-out test set (n=112). Discrimination was assessed using ROC AUC, PR-AUC, and Brier score. Model performance was compared with the established clinico-genetic subgroup classification (ST1, ST4S, ST4, MNA). Differences in AUC were estimated using nonparametric bootstrap resampling (2,000 iterations) to derive 95% confidence intervals (CI).Results:
In the test cohort (n=112), the combined clinico-genomic model achieved a ROC AUC of 0.81, PR-AUC of 0.64, and Brier score of 0.19, compared with 0.76, 0.63, and 0.21 for the gene-expression–only model. The difference in AUC was statistically significant (?AUC 0.05; 95% CI 0.01 to 0.09; two-sided p=0.021). Compared with existing clinico-genetic subgrouping (ST1, ST4S, ST4, MNA), the combined model demonstrated non-significantly higher discrimination for progression (?AUC 0.02; 95% CI -0.04 to 0.10) and adisease-specific deaths (?AUC 0.05; 95% CI -0.04 to 0.14), respectively. The lack of significance with wide 95% CI could be attributed to small sample size.Conclusion:
Integrated clinico-genomic modeling provided statistically significant but clinically modest improvements over gene-expression–only models and demonstrated non-significantly improved discrimination compared with established clinico-genetic subgrouping. These findings suggest that continuous, model-derived risk estimates may offer greater prognostic resolution than categorical subgrouping, although external validation and decision analytics in a large dataset is needed to define clinical utility.