From biosensing to perception: Collaborative few-shot learning for explainable digital biomarker identification in high-dimensional biomedical spectra

Abstract

The application of in vitro diagnostic biosensors for early cancer detection remains challenging due to the insufficient representation by a few molecular biomarkers. Digital biomarkers promise comprehensive disease phenotyping but face constraints of clinical data scarcity and obstacles of limited generalization. Here, we introduce Coupled Explainable Artificial Intelligence Recursive (CEAIR) learning, a computational framework that integrates computer vision and cooperative game theory for interpretable few-shot learning, enabling the extraction of domain-relevant digital biomarkers from high-dimensional surface-enhanced Raman spectroscopy biosensor data of limited serum samples. Applied to hepatocellular carcinoma detection, CEAIR-derived digital biomarkers significantly outperform circulating molecular biomarkers, achieving area under the curve values consistently exceeding 0.97 across multiple independent classifiers built with classic machine learning algorithms and demonstrating strong generalization upon external validation. Our findings underscore CEAIR's capacity to overcome fundamental limitations in generating clinically meaningful diagnostic knowledge from high-dimensional, small-sample biosensor data, learning reliable digital biomarkers for robust, non-invasive, and timely diagnosis of complex diseases.