Cellular morphometric biomarkers and large language model predict prognosis and treatment response in neuroblastoma patients: A retrospective and double-blind prospective single arm clinical study

Abstract

Background

The heterogeneity of Neuroblastoma (NB) leads to variation in response to treatment and outcomes. The aim of the current study is to discover AI-empowered cellular morphometric biomarkers (CMBs), to establish the corresponding CMB risk score (CMBRS), CMB risk group (CMBRG), large language model driven CMB risk score (CMB-LLM-RS), and large language model driven CMB risk group (CMB-LLM-RG), and to investigate and validate their prognostic and predictive power in NB.

Methods

In this study, the retrospective cohort enrolled 84 primary NBs between 1/2020 and 12/2021, followed up through 11/22/2024; the prospective cohort enrolled 67 primary NBs between 1/2022 and 7/2023, followed up through 11/22/2024.

Results

We identified 9 CMBs from a retrospective NB cohort, enabling the CMBRS, CMBRG, CMB-LLM-RS, and CMB-LLM-RG. Both CMBRG and CMB-LLM-RG are significantly associated with prognosis (p < 0.0001) and treatment response (p < 0.0001). Furthermore, we double-blindly validated the predictive power of CMBRG and CMB-LLM-RG in a prospective NB cohort, which confirms their potential value in real clinical settings. Importantly, CMBRG provides clinical value independent of the International Neuroblastoma Risk Group (INRG) classification system in both retrospective and prospective NB cohorts (p < 0.05); and the combination of CMBRG and INRG significantly increases prognostic and predictive performance for NB patients.

Conclusions

These findings suggest that CMBRG and CMB-LLM-RG have prognostic and predictive value for NB and warrants evaluation in larger multicenter cohorts.