Answers are grounded in this paper and cite their sources.
Background Deep learning is reshaping non-invasive biomarker assessment in non-small cell lung cancer (NSCLC), offering opportunities to refine diagnosis, treatment selection, and prognosis. Recent systematic reviews and meta-analyses have catalogued the diagnostic accuracy of these models [1–4], yet none has formally appraised the methodological quality of the underlying prediction-model studies using a dedicated risk-of-bias tool. Objectives To characterise the deep-learning and machine-learning models reported between 2018 and 2026 for NSCLC biomarker classification, and—uniquely among current reviews—to appraise their methodological quality with the PROBAST risk-of-bias tool, thereby assessing whether the reported performance is trustworthy and what would still be required to establish clinical utility. Methods A systematic literature review was conducted following the PRISMA 2020 statement. IEEE Xplore, Scopus, Springer Link, PubMed, Sage Journals, ScienceDirect, and BioMed Central were searched for peer-reviewed primary studies published between 1 January 2018 and 31 March 2024 that applied deep-learning or machine-learning models to NSCLC biomarker classification; a systematic update search extended coverage to 31 May 2026. Eligibility criteria, search strings, and screening procedures are reported in full in the Methodology section. Two reviewers independently screened titles, abstracts, and full texts, with disagreements resolved by consensus. Risk of bias and methodological quality of the included prediction-model studies were appraised with the PROBAST tool. Results Eighty-two primary studies met all inclusion criteria, spanning five thematic categories: survival prediction, treatment-response prediction, biomarker identification, diagnostic/staging imaging analysis, and radiogenomics/transcriptomics. Convolutional neural networks were the single most common architecture (28 of 82 studies, 34%), followed by hybrid and multi-modal architectures (20 of 82, 24%) that combined imaging with genomic, transcriptomic, or clinical inputs. PD-L1 and EGFR were the most frequently modelled biomarkers, with several studies reporting concordance with pathologist scoring above 0.90 in internally validated cohorts. Datasets were concentrated in China and the United States, with limited representation of African, Latin-American, and South-Asian cohorts. Only 13 of the 80 primary predictive-model studies (16%) reported validation on an explicitly independent external cohort; a further 22 (28%) described multi-centre or multi-cohort data without stating whether the test set was independent of training. PROBAST appraisal indicated a high or unclear risk of bias in the majority, driven primarily by the analysis and participants domains. Conclusions Deep learning has materially advanced NSCLC biomarker classification, particularly for PD-L1 and EGFR. Translation to routine clinical use remains constrained by data heterogeneity, single-centre cohort dominance, sparse external validation, and limited reporting of uncertainty. Future work should prioritise prospective multi-institutional validation, harmonised acquisition protocols, interpretable modelling, and consistent reporting against the PROBAST [5] and TRIPOD+AI [6] frameworks. Registration The protocol was not pre-registered on PROSPERO; methodological details are available from the corresponding author on reasonable request.
No takes yet. Share an insight, caveat, or question.
Karim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: