From simple factors to artificial intelligence: evolution of prognosis prediction in childhood cancer: a systematic...

Current paediatric cancer care requires innovative approaches to predict prognosis that facilitates personalised stratification, yet studies on the performance, composition and limitations of contemporary prognostic models are lacking. We aimed to compare the accuracy of traditional and advanced prognostic models.

A systematic search for this systematic review and meta-analysis (CRTN42022370251) was conducted in PubMed, Embase, Scopus, and the Cochrane Library databases on 28 June 2024. Studies on the accuracy of prognostic markers or models used in paediatric haematological malignancies, central nervous system (CNS), or non-CNS solid tumours (NCNSST) were included. Three model categories were defined using: 1-clinical parameters, 2-genomic-transcriptomic data, and 3-artificial intelligence (AI). Primary outcomes were area under the receiver operating characteristic curve with a 95% confidence interval (CI) for various overall survival intervals and event-free survival. Two independent groups performed selection and data extraction. We used data published by the authors and publicly available databases.

Of 12,982 studies, 358 were included in the meta-analysis and 27 in the systematic review, with limited data on AI-approaches. Most data were reported on NCNSST at 5-year OS, where a statistically significant difference was observed between Category-1 (0.75 CI: 0.72-0.79) and Category-2 (0.85 CI: 0.82-0.88) (p < 0.001), but not between Categories-2 and -3 (p = 0.2834) (0.82 CI: 0.77-0.88). Internal validation studies showed significantly better performance compared to those using external validation, highlighting the high risk of bias (ROB) inherent in internal validation. High ROB was most commonly experienced in the outcomes and statistical analysis domains, assessed using PROBAST and QUIPS.

It is advisable to introduce Category-2 and -3 models in a clinical setting, especially for NCNSST prognostic for aiding risk-stratification. Although AI-supported predictions in paediatric oncology are at an early stage of development, it is imperative to further explore their potential. This requires structured data collection and ethical sharing from paediatric oncology patients in sufficient quantity and quality.

None.