Abstract
BACKGROUND: Longitudinal assessment of tumor burden using imaging helps to determine whether there has been a response to treatment both in trial and real-world settings. From a patient and clinical trial perspective alike, the time to develop disease progression, or progression-free survival, is an important endpoint. However, manual longitudinal response assessment is time-consuming and subject to interobserver variability. Automated response assessment techniques based on machine learning (ML) promise to enhance accuracy and reduce reliance on manual measurement. This paper evaluates the quality and performance accuracy of recently published studies. METHODS: Following PRISMA guidelines and the CLAIM checklist, we searched PUBMED, EMBASE, and Web of Science for articles (January 2010-November 2024). Our PROSPERO-registered study (CRD42024496126) focused on adult brain tumor automated treatment response assessment studies using ML methodologies. We determined the extent of development and validation of the tools and employed QUADAS-2 for study appraisal. RESULTS: Twenty (including 17 retrospective and 3 prospective) studies were included. Data extracted included information on the dataset, automated response assessment including pertinent steps within the pipeline (index tests), and reference standards. Only limited conclusions are appropriate given the high bias risk and applicability concerns (particularly regarding reference standards and patient selection), and the low-level evidence. There was insufficient homogenous data for meta-analysis. CONCLUSIONS: The study highlights the potential of ML to improve brain tumor longitudinal treatment response assessment. Interpretation is limited due to study bias and limited evidence of generalizability. Prospective studies with external datasets validating the latest neuro-oncology criteria are now required.