Toward Real-Time, Scalable Vis–SWIR Diagnostics: Evaluating Machine-Learning Classification Performance with Reduced-Spectra Acquisition Protocols

Near-infrared spectroscopy (NIRS) is increasingly studied as a non-invasive optical investigation tool for in vivo tissue characterization, including applications to skeletal muscle and brain regions. In this context, previous studies have demonstrated reliability in differentiating muscle sites, typically relying on dense acquisition schemes (≥50 spectra acquired per site) to ensure signal stability. However, this requirement may limit throughput and hinder real-world clinical translation. Optimizing the trade-off between acquisition burden and classification performance represents a key design problem for device scalability and feasibility of bedside deployment. In this study, we explored the impact of spectral sampling density on machine learning-based muscle discrimination. Thirty healthy adults provided 50 Vis–SWIR (Visible–Short-Wave Infrared; 350–2500 nm) reflectance spectra per biceps and triceps muscle sites (3000 spectra). Seven datasets were generated by random subsampling, progressively reducing the number of spectra (from 50 to 1 spectra/muscle/subject). All datasets underwent an identical preprocessing pipeline and were subjected to Partial Least-Squares Discriminant Analysis (PLS-DA) classification. PLS-DA achieved near-perfect discrimination from 50 to 5 spectra per muscle with a mean cross-validation (CV) accuracy ≥ 99.5%, whereas performance collapsed abruptly at three spectra (CV accuracy ~39%) and one spectrum (CV accuracy ~15%). Therefore, high machine learning classification performance is retained even when the number of acquired spectra is substantially reduced. These findings support the feasibility of acquisition-efficient protocols that may enhance device portability and reduce measurement time, thus enabling NIRS integration into clinical workflows. From a biomedical engineering standpoint, spectra number reduction without loss of predictive performance represents a key step toward scalable, real-time, and patient-centered Vis–SWIR diagnostic platforms.