Modern paediatric cancer management needs improved methods for forecasting patient outcomes to enable tailored risk grouping. However, research evaluating the effectiveness, makeup, and shortcomings of current prognostic tools remains limited. This study sought to evaluate the predictive performance of conventional and more sophisticated prognostic approaches. This systematic review and meta-analysis (registered as CRTN42022370251) involved a comprehensive literature search across PubMed, Embase, Scopus, and the Cochrane Library, completed on 28 June 2024. Eligible investigations examined the precision of prognostic indicators or algorithms in paediatric blood cancers, central nervous system (CNS) tumours, or non-CNS solid tumours (NCNSST). Three main model types were examined: those based on (1) clinical variables, (2) genomic and transcriptomic information, and (3) artificial intelligence (AI) techniques. Key endpoints included the area under the receiver operating characteristic curve (AUROC) with 95% confidence intervals (CI) for different overall survival (OS) periods and event-free survival (EFS). Study selection and data gathering were carried out independently by two teams. Analyses drew on published results and open-access repositories [CRTN42022370251]. From 12,982 screened records, 358 studies contributed to the meta-analysis and 27 to the systematic review, though information on AI methods was scarce. The largest body of evidence concerned NCNSST patients at the 5-year OS mark. Here, Category-1 models achieved a pooled AUROC of 0.75 (CI: 0.72–0.79), which was significantly lower than Category-2 models at 0.85 (CI: 0.82–0.88) (p < 0.001). No significant difference emerged between Category-2 and Category-3 models (p = 0.2834; pooled AUROC 0.82, CI: 0.77–0.88). Investigations relying on internal validation demonstrated markedly higher performance than those with external validation, underscoring elevated risk of bias (RoB) associated with internal approaches. The most frequent RoB concerns appeared in the outcome measurement and statistical analysis domains, evaluated via PROBAST and QUIPS tools. Incorporating Category-2 and Category-3 models into routine practice is recommended, particularly for NCNSST cases to support better risk classification. Although AI-based forecasting in childhood cancer is still developing, its capabilities warrant continued investigation. Achieving this will demand systematic collection and responsible sharing of high-quality data from sufficient numbers of paediatric oncology patients.
