Ground beef lymph-node adulteration moves past visual inspection: four-excitation fluorescence plus LSTM holds RPD 9.3 at unseen slaughterhouses

Screening ground beef for lymph-node tissue shifts from single-wavelength fluorescence to front-face fluorescence at four excitation wavelengths, paired with an explainable LSTM regressor that outputs a lymph-node mass fraction directly. Earlier fluorescence work mostly used a single excitation near 280 nm, which tracks total aromatic amino acid fluorescence and falls with any dilution of muscle protein, so it could not separate lymphatic dilution from added fat, lipid oxidation or muscle variability. On two external slaughterhouses held out from training, the model reached R2P 0.993 (Shandong, n=90) and 0.988 (Sichuan, n=90), with RMSEP 2.60 and 3.35% w/w and RPD 12.0 and 9.3, both in the excellent RPD≥5.0 band; the internal test set gave RPD 19.5.

Each excitation slice enters as a sequence ordered along its own emission axis, and the four slices share one two-layer encoder whose four hidden states are concatenated before regression. Shared weights cut trainable parameters by roughly three-quarters, which matters at the n=370 calibration set size, and force one spectral-to-latent mapping across slices. Modelling used 824 ground beef samples from five commercial slaughterhouses, covering three breeds, two slaughter-age groups and three muscle cuts, with adulteration prepared gravimetrically at 0, 2, 5, 10, 15, 20, 30, 60 and 100% w/w.

The specificity is attributed to multi-excitation sampling itself rather than to the deep-learning architecture. The same shared-weight architecture fed only one excitation slice produced clearly inflated false-positive predictions under adipose, saline and accelerated-oxidation challenge. The 280 nm slice alone dropped to R2P 0.919 and RMSEP 8.90% w/w and predicted 3.9–6.8% lymphatic content under interferents, while the four-excitation model returned 0.11–0.48% on the same samples and passed equivalence testing against a pre-specified ±1.0% w/w margin.

The method targets homogeneous comminuted beef, and a single-point front-face fluorescence measurement returns one bulk estimate that cannot localize a small lymph-node inclusion, so spatial mapping remains the role of hyperspectral imaging. The study covers five commercial slaughterhouses, three breeds, two slaughter-age groups and three muscle cuts, with external validation from two unseen slaughterhouses in Shandong and Sichuan. The method is positioned as a rapid low-cost screen, with important or positive results requiring confirmation by an established analytical method. Next steps named are instrument-transfer studies, testing of processed beef products, broader multi-slaughterhouse and multi-laboratory validation, and reproducibility studies; a preliminary cost comparison suggests benchtop spectrofluorometers may have a lower initial cost than turnkey NIR-HSI systems, though a formal economic analysis is still needed.

Interferents were tested one at a time, so combined effects such as elevated fat alongside advanced oxidation remain untested, and at n=8 per condition the test detects differences of about 0.15% w/w, unable to exclude a smaller systematic bias. Breed is nested within farm, so breed and farm effects cannot be separated. Internal-to-external RPD falls from 19.5 to 9.3–12.0, which the authors attribute to genuine inter-farm biological variability, but two farms are not the supply chain, and breeds, feed regimes and practices beyond these five are untested. The 1.8% w/w detection limit sits below the roughly 2% indicative threshold while the 6.0% w/w quantification limit does not, so near 2% the method detects but cannot reliably quantify; that threshold is a literature-derived benchmark, not a compliance limit.