Machine learning across label-free optical measurement platforms for cellular dynamics and biomechanics

Label-free optical biosensing combined with machine learning enables live-cell analysis with high spatial and temporal resolution and can improve cell-state classification, response profiling, and estimation of biomechanically relevant variables. This review organizes works into single-modal and multimodal workflows. In single-modal analysis, representation-oriented approaches improve signal quality and provide data reconstruction or calibration before modeling, whereas inference-oriented approaches map optical data to phenotypes, adhesion behavior, or other biologically relevant variables. These roles are examined across surface-enhanced Raman spectroscopy (SERS), surface plasmon resonance/resonant waveguide grating (SPR/RWG), and digital holographic microscopy (DHM). Multimodal workflows are grouped into reference-based calibration, in which an auxiliary modality supervises a primary platform, and joint multimodal inference, in which complementary readouts are fused to estimate cell state robustly.