Journal of Innovative Science

Abstract: The earliest possible diagnosis is currently the most crucial element affecting the survival rate in cancer patients, but the existing screening and diagnostic methods are unable to provide early clues about malignancies. However, the last decade has witnessed the advent of artificial intelligence, or AI, in predictive oncology, which combines radiologic, molecular, and clinical variables for the improved specificity of diagnosis. The existing literature has revealed the efficacy of predictive models based on deep learning with the help of AI in the early diagnosis of malignancies, even beyond the capabilities of human interpretation. This narrative review synthesizes findings from recent studies that developed or validated AI models for early cancer diagnostics. The data show the tremendous success achieved in imaging radiomics, cfDNA fragmentomics, proteomics, or EHR-based model approaches with highly discriminative abilities, with AUC often well beyond 0.85, mostly for lung, colon, or breast cancers. However, their quality in terms of statistical validation or clinical transferability is often questionable, with many publications predominantly focusing on validation testing, without much care about the aspects of calibration, multicentric validation, or transferability to other populations. The article showcases the prominent paradigm shifts in the methods being employed, such as the need for explainable AI, the integration of deep learning with traditional machine learning, or the novel application of real-time or edge computing in the context of poor resources. Overfitting, imbalance, and lack of translucency in reporting are some of the challenges yet to be fully remedied in the pursuit of clinical-grade translatability. The future will require validation, benchmarking, or sound application practices to ensure these aspects. Although the predictive capabilities in cancer, utilizing AI, are revolutionizing the future of early cancer diagnosis, their successful implementation awaits improved statistical grounding, validation, and fair global application.