Support Vector Machine Models for Cancer Detection and Related Clinical Applications Using Omics and Omics-Adjacent Data: A Systematic Review Publisher Pubmed

Summary: How reliable are cancer AI models? A review of 75 SVM studies found widespread bias risks and limited external validation, making a pooled accuracy claim misleading. #CancerAI #ModelValidation

Mohamadi Z ; Mohammadi R ; Abtahi E ; Shayegh Z S ; Kachouei M A ; Fakhar A ; Shirani M M ; Malekian M ; Zinatshoar A ; Biglari M ; Rezaei F ; Zarinkhat A
Source: BioMed Research International Published:2026

Abstract

Background: Support vector machines (SVMs) have been widely applied to high-dimensional molecular and related biomedical data for cancer detection and other clinical classification tasks. The evidence spans diverse cancer types, specimens, platforms, objectives, and validation designs, limiting direct comparison. Methods: We systematically reviewed studies evaluating SVM-based models using molecular omics or high-dimensional omics-adjacent data for cancer-related clinical applications. Article-level characteristics, SVM implementation, feature selection, validation, comparator models, and reported performance measures were extracted. Risk of bias was evaluated with the Prediction Model Risk Of Bias Assessment Tool (PROBAST). Because the studies did not estimate a common clinically interpretable quantity, no meta-analysis, pooled estimate, median performance comparison, or statistical subgroup comparison was performed. Findings were synthesized descriptively by data modality, clinical task, validation approach, comparative modeling, and methodological quality. Results: Seventy-five studies met the revised eligibility criteria. Sixty used molecular omics data; the other 15 used omics-adjacent inputs (10 spectroscopy-derived, two sensor derived, two mixed-data, and one imaging-derived). Cancer detection was the leading application. Internal validation alone was reported in 60 studies; eight combined internal and external validation, four relied on external validation only, two reported no validation, and one remained unclear. Seventy-two studies contributed 544 article-level SVM performance entries. In a separate PROBAST assessment, 72 studies were rated at high overall risk of bias and three at unclear risk. Conclusions: SVM-based models have been studied across a broad range of cancer-related applications, but the evidence is heterogeneous, predominantly internally validated, and frequently at high risk of bias. A single summary performance estimate would be misleading. This review maps the evidence and identifies cancer-specific, modality-specific, and externally validated questions suitable for future focused studies. Copyright © 2026 Zhina Mohamadi et al. BioMed Research International published by John Wiley & Sons Ltd.