Label-free Raman spectroscopy combined with artificial intelligence for functional subtyping of human sperm and prediction of embryo development outcomes.

Raman spectroscopy delivers label-free, non-destructive, single-cell measurements and directly senses intrinsic biochemical signals of nucleic acids, proteins, and lipids. Conventional semen analysis offers limited prognostic power for assisted reproductive technology (ART). We developed a single-cell, label-free Raman workflow integrated with artificial intelligence (AI) to assess sperm functional status and predict outcomes of ART. Spectra from sperm of thirty-one ART patients were processed to extract diagnostically informative bands spanning nucleic acids, proteins, lipids, and glycogen. Spectral analysis resolved two robust spectral subpopulations. The subpopulation exhibiting higher chromatin integrity, more favorable protein conformation, and better membrane lipid dynamics molecular features was associated with better outcomes, higher two-pronuclear (2PN) fertilization rate (76.08% vs 50.17%, P = 0.044), high-quality embryo rate (40.53% vs 17.57%, P = 0.010), and blastocyst formation (59% vs 9%, P < 0.001). Among evaluated classifiers, a gated recurrent unit (GRU) model showed the best predictive performance (accuracy 94%, sensitivity 94%, AUC 0.98). This Raman-AI assay enables objective, non-invasive stratification of sperm quality and clinically relevant prognosis, offering an analytically rigorous tool to guide precise sperm selection in ART.