Use of artificial intelligence models for prognosis and survival prediction in brain tumor patients: a systematic review and meta-analysis

Abstract

Brain tumors are associated with significant morbidity and mortality. Early prognostication of preoperative cases via AI models may improve treatment planning and clinical outcome. This systematic review and meta-analysis followed PRISMA 2020 guidelines. Literature review was conducted across various databases including PubMed, Web of Science, Cochrane etc. Out of the 433 papers obtained, 17 retrospective observational studies met the inclusion criteria. Risk of bias assessment was carried out using PROBAST+AI tool. Meta-analysis was carried out for the reporting area under the curve (AUC) or C-index for survival prediction models. With a total of 98,464 observations across 17 studies, machine learning (ML) and deep learning (DL) models were used to predict survival and prognosis in brain tumor patients. The risk of bias was low in 6% studies, moderate in 59% studies and high in 35% of the studies. The pooled AUC was 0.87 (SE: 0.04) with a 95% prediction interval between studies ranging between 0.61 and 1.13. Cochran’s Q statistic for heterogeneity was 48.81 (p<0.001). Subgroup analyses showed pooled AUCs of 0.92 for ML models and 0.81 for DL models. Significant publication bias was demonstrated by Funnel plot and Egger’s test. This systematic review and meta-analysis are the first to include multiple brain tumor types for predicting prognosis via AI models, with ML models showing slightly higher pooled performance than DL models. However, variability in datasets, limited external validation, and high heterogeneity among studies highlight the need for standardization and further research.

Ähnliche Arbeiten