The use of artificial intelligence and machine learning to predict tumor recurrence in high-grade gliomas: a systematic review

Abstract High-grade gliomas (HGGs) are aggressive tumors with a propensity for recurrence. Despite standardized therapies, definitive treatment is elusive. Advancements in artificial intelligence (AI) and machine learning (ML) can potentially identify recurrence probabilities and patterns facilitating individualized therapy. A systematic review was conducted in accordance with the PRISMA guidelines. PubMed, ScienceDirect, and Web of Science databases were queried for reports on the use of AI/ML for the prediction of disease recurrence in HGGs. Using a random-effects model and inverse variance weighting a pooled analysis of key performance metrics (sensitivity, specificity, and accuracy) was conducted on the top performing models from each manuscript. In total, 14 manuscripts encompassing 1,540 patients were selected for systematic review and analysis. Across the included studies, 13/14 (92.9%) were retrospective study designs, with 1/14 (7.1%) prospective study design. Among the 1,540 patients, 1,530 (99.3%) and 10 (0.7%) were histologically classified as WHO grade IV and III respectively. Nine studies (9/14, 64.3%) examined patients undergoing GTR following by adjuvant RT, and five studies (5/14, 35.7%) undergoing STR/NTR followed by adjuvant RT. The Random Forest (RF) model was most frequently utilized, with T2-FLAIR sequences most frequently incorporated into model training. The pooled sensitivity, specificity, and accuracy of the models were 81% (95% CI: 73–87; I² = 85.2%), 75% (95% CI: 65–85; I² = 91.9%), and 79% (95% CI: 64–92; I² = 87.8%), respectively. Following sensitivity analyses, the corresponding estimates were 81% (95% CI: 77–84; I² = 0.0%), 84% (95% CI: 79–88; I² = 38.8%), and 89% (95% CI: 83–95; I² = 0.0%), respectively. The development of an AI/ML model to predict tumor recurrence in HGGs has been an emerging area of research over the past decade. While ongoing validation in larger, prospective databases is needed, preliminary evidence suggests that existing models perform with reasonable sensitivity, specificity, and accuracy.