Including pH and temperature data improves NIRS prediction models for drip loss in pork slices

In the last decades NIRS has proven to be an efficient tool for estimating the nutritional content of meat and meat products such as moisture, fat and protein content (AOAC approved method, Anderson 2007). However, physical parameters of meat quality remain difficult to predict by NIR spectroscopy (R2 for drip loss (DL) mostly below 0.6). The aim of this work was to determine the variability of DL between replicates and among DL reference methods and to evaluate the possibility to improve NIRS predictive models for DL by including further physical and physico-chemical data (temperature and pH of meat early post mortem) in the NIR calibration models.
Meat samples of 86 pigs from the pig performance testing station (Suisag, in Sempach, Switzerland) of different breeds (Duroc, Pietrain, Large White, Swiss Landrace and crossbreeds) were included in this work. Temperature and pH-measurements were taken at 90 min and 24 h post mortem, in the Longissimus dorsi muscle (LD) at the 13th rib level. Samples of the LD (10th to 13th rib) were collected at 24 h post mortem. Subsequently, with 3 cm thick chops of ~80 g weight sealed in bags, mass-standardized drip loss (DLsm) measurements were performed for 48 h at 2°C. In parallel, bag drip loss (DLb) measurements were performed in 3 to 4 whole chops of 2 cm thickness for 48 h at 4°C. Prior to DLb measurement the NIR spectra were taken on 2 chops per sample. A NIRFlex N-500 (Büchi, Flawil, Switzerland) equipped with a rotary flat accessory designed for meat chops was used. Per chop, 3 to 5 replicates were taken with 21 scans per replicate. Each replicate was taken along an arch of ~180° in the range of 4000 to 10000 cm-1. The NIR models were built using NIRCal (Büchi, Flawil, Switzerland) with NIR spectra only, and R software (R Core Team, 2013) and Unscrambler (Camo, Oslo, Norway) for models with combined parameters.
The DL reference measurements ranged from 2.59 to 11.12% for DLb and from 1.52 to 9.70% for DLsm. The respective average DL and standard deviation were 5.90 ± 1.52% (DLb) and 3.27 ± 1.35% (DLsm), with a low correlation between DLb (including all replicates) and DLsm (R2 = 0.42).
Standard error of prediction (SEP), for models based on NIR spectra only, oscillated around 1.3%. However, when scaled data of pH and temperature at 90 min and 24 h post mortem were included, SEP values for the prediction of DLb decreased to a range of 1.10 to 1.05%. Most importantly, by including in the model the pH and temperature the number of components necessary to describe the data drastically reduced from > 13 to 2 - 4. In conclusion, the large variability between replicates, especially when compared to the relatively short range of DL eventually explains the poor predictive power of NIRS for DL. However, the present results show a clear improvement of the predictive power of NIRS for DL by the inclusion of pH and temperature data in the NIRS calibration models.