Explainable and Privacy-Preserving Multimodal Machine Learning for Scalable Cancer Diagnosis: A Review

Abstract

Cancer struck an estimated 20.6 million people and killed 9.8 million worldwide in 2024, a burden projected to grow to 34 million new cases and 17.5 million deaths annually by 2050, yet only one in three countries currently reports high-quality cancer incidence data and only one in four reports reliable mortality data. Machine learning promises to help close this gap, but only if three requirements are met at once: clinicians need explanations they can act on and defend, patient data protected by regulations such as HIPAA and the GDPR cannot simply be pooled across institutions or borders, and the diagnostic signal that actually distinguishes malignant from benign disease is often split across multiple data modalities histopathology, radiological imaging, and structured clinical variables that no single-modality model can fully exploit. This review synthesizes five studies addressing exactly these requirements. On explainability, two independent studies applied to the same Wisconsin breast-cancer cytological benchmark two years apart reach 96% and 97.07% accuracy respectively, the more recent adding SHAP-based feature attribution and systematic class-imbalance handling within a predictive-preventive-personalized-medicine framework. On privacy-preserving multimodal fusion, a federated framework combining skin-lesion images with tabular metadata reaches 92.32% accuracy while losing only 2 percentage points under substantial missing-modality conditions, and a tri-modal federated framework fusing ultrasound, MRI, and mammography reaches 93.10% fused accuracy while cutting communication bandwidth by 84.9% through reduced-precision weight transmission. A sixth study, whose framing a multimodal, privacy-preserving, scalable diagnostic framework across healthcare systems matches this review's scope unusually closely, could not be independently verified on its quantitative results and is discussed with that limitation stated explicitly. Read together, these studies show genuine, measurable progress on explainability, privacy preservation, and multimodal fusion individually, but no study reviewed here demonstrates all three simultaneously at the scale the global cancer-data gap actually requires.