Page 23 - IJAMD-1-1

International Journal of AI
for Material and Design ML in 3D bioprinting of cultivated meat

SNV transformation was applied to standardize spectral compared to traditional approaches, including KNN,
data, and the optimal number of LVs was determined using MLP, and even LSTM. While these results are extremely
residual variances. The PLS-DA model achieved 88.1% promising, the use of DWT for denoising is only applicable
overall accuracy with 86.7%, 87.5%, and 88.9% accuracy for here due to their experimental setup of continuously
fresh, semi-fresh, and spoiled samples, respectively. While monitoring the beef samples over a period of 37 h. The
the PLS-DA model offers decent predictive accuracies, authors do note that DWT-LSTM with an E-nose system
there is a significant decrease in accuracy between the can be useful in storage monitoring purposes, but shorter
training and testing set, from 93.3% to 86.7% for the fresh durations where rapid determination of TVC is needed
samples and from 100.0% to 88.9% for the spoiled samples. would yield inaccurate predictions, and hence, alternative
These results may indicate some overfitting in the models. data preprocessing methods could be tested as well.
In addition, it is notable that the study builds models for
each of the selected species instead of relying on TVC as an 4.2. Detection of meat adulteration
indicator of microbial quality, as TVC is not representative Meat adulteration involves the deliberate or unintended
of microbial growth dynamics at different temperatures, introduction of substances into meat products, either
hence enabling more accurate models to be trained. through contamination or substitution. Such adulteration
encompasses the inclusion of lower-quality or cheaper
In a direct comparison between PLSR and SVM for ingredients, incorrect labeling, or the use of unauthorized
estimating TVC in raw pork using HSI spectral data, SVM additives. The impact of meat adulteration on human
demonstrated slightly superior performance (Table 4). consumption varies depending on the nature and extent of
HSI spectra were truncated to 400 – 1000 nm to eliminate the adulteration. Therefore, it is crucial to employ detection
noise and invalid information, and various preprocessing technologies that are rapid, effective, accurate, and reliable.
84
methods were applied. The best-performing PLSR model, The successful implementation of ML for detecting meat
using second derivative preprocessing, achieved an adulteration relies on access to high-quality, well-labeled
R = 0.94 and SEP = 0.7374. However, the top-performing training data, along with the continuous adaptation and
SVM model, with first derivative preprocessing, slightly enhancement of models as new data becomes available.
outperformed it, achieving R = 0.95 and a lower standard Collaboration among industry experts in the food sector,
error = 0.5264. The comparison of multiple data data scientists, and regulatory authorities is essential for
74
transformation methods also demonstrates, in this study, the development of dependable ML solutions for detecting
that the accuracy of SVM models is more sensitive to the meat adulteration (Figure 6).
preprocessing method than the PLSR models, as the SVM
SEP varies from 0.46 to 1.04 while the PLSR SEP only PCA and PLS-DA have been utilized for assessing the
varies from 0.74 to 0.98. Choosing an appropriate data presence of adulteration in beef samples, employing FTIR
-1
transformation method is, hence, especially important spectral data within the range of 4000 – 525 cm along with
when working with SVM. five physicochemical properties (protein, ash, sodium,
chloride, and phosphate content). FTIR data underwent
Several models were utilized with E-nose data from pre-processing through Savitzky-Golay smoothing (length
raw pork to estimate TVC and classify samples into four 15 and polynomial order 2), and a Bayesian threshold was
quality groups. The E-nose, consisting of 11 gas sensors, applied to the models’ predictions. Samples exceeding the
underwent denoising using discrete wavelet transform threshold in predicted values were considered adulterated.
(DWT) for all models except the long short-term memory The number of LVs was determined through Venetian
(LSTM) model. The performance metrics for different blinds cross-validation based on the smallest classification
models in estimating TVC in pork are as follows: k-nearest error. While the first two PCs in the PCA models
neighbor (KNN) achieved 77.73% accuracy for quality explained 90.81% and 79.93% of the variance using either
classification, with an R of 0.81 and RMSE of 0.46 for TVC; physicochemical or FTIR values, the models were unable
2
LDA demonstrated 76.82% accuracy, with an R of 0.83 and to differentiate between the samples. In contrast, the
2
RMSE of 0.42 for TVC; SVM attained 84.88% accuracy, PLS-DA model achieved 91% accuracy with low-level data
with an R of 0.91 with RMSE of 0.37 for TVC; MLP fusion of FTIR spectra and physicochemical properties
2
2
showed 56.19% accuracy, with an R of 0.41 and RMSE of using four LVs. These results may not be significant, as
86
10.64 for TVC; LSTM exhibited 85.14% accuracy, with an it was noted in the study that NaCl was one of the main
R of 0.96 and RMSE of 0.15 for TVC; and DWT-LSTM adulterants in the samples, and examination of the PLS
2
outperformed others with 94.83% accuracy, an R of 0.97, weights shows that chloride was the chemical variable
2
and a low RMSE (0.05) for TVC. The findings revealed with the highest discriminant. It is, therefore, possible that
83
that the DWT-LSTM exhibited superior performance the model would be especially effective at detecting NaCl
Volume 1 Issue 1 (2024) 17 https://doi.org/10.36922/ijamd.2279