Page 25 - IJAMD-1-1

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

A B

C

Figure 6. Schematic drawing illustrating the use of machine learning (ML) approach for detection of meat adulteration: (A) Sample preparation involved
mixing of minced pork with mutton essence at varying ratios (5 – 40% w/v); (B) a near-infrared hyperspectral imaging system was used for meat
adulteration analysis; and (C) image acquisition and processing to obtain the region of interest for ML. Figure reproduced from Fan et al. .
85
Abbreviation: ROI: Region of interest.
In a different study, SVM, Logistic Regression, DT, of 54.6%, 71.2%, and 65.2% with UV-Vis data, NIR data,
ANN, and an Ensemble Voting model were employed to and MIR data, respectively), while PLSR performed better,
classify chopped beef into seven bins of pork adulterant achieving R value = 0.81 and RMSEP = 8.61 using UV-Vis
2
2
percentages. An E-nose with nine gas sensors provided data, data, R value = 0.9184 and RMSEP = 5.79 using NIR
with six statistical parameters extracted from the sensor data, R value = 0.9, and RMSEP = 6.19 using MIR data,
2
data (average, standard deviation, kurtosis, skewness, and R value = 0.95 and RMSEP = 5.33 using fused data.
2
88
minimum, and maximum). SVM, Logistic Regression, DT, The findings indicated that NIR and MIR spectroscopy,
and ANN models were trained with optimal combinations recognized as non-selective analytical techniques, can
of the six statistical parameters, achieving accuracies of be regarded as rapid and dependable complements to
94.57%, 93.71%, 91.14%, and 92.85%, respectively. The traditional methods for species identification in minced
Ensemble Voting model outperformed others, achieving meat products especially when used in building ML models.
95.71% accuracy. The study is unique in its application In other studies, exclusive utilization of IR spectroscopy
87
of a meta-learning model to adulteration detection in beef, in conjunction with ML involved the application of PLSR
but it is yet to be seen whether the additional computational and soft independent model of class analogies (SIMCA).
requirement of running every model for the sake of the These methodologies, coupled with FTIR, were employed
Ensemble Voting model is worth the marginal improvement to predict and identify various adulterants such as horse
in accuracy. In addition, the finding that statistical feature meat, soy protein, and fat beef trimmings in minced beef.
extraction with smaller extraction chamber sizes and The models achieved extremely high accuracy, with the
hence higher gas concentrations gives data with stronger PLSR model exhibiting an R value exceeding 0.99 for
2
signals can be leveraged in future studies beyond adulterant all adulterants and the SIMCA model achieving 100%
detection, especially for situations where rapid prediction accuracy in recognizing and rejecting all adulterants.
89
of TVC from E-nose data is required. In separate studies, PLSR and FTIR were utilized to
For evaluating turkey adulteration concentration in predict the concentration of porcine adulteration in beef
minced beef, PCA, LDA, and PLSR were applied using meatballs. One study achieved an R value over 0.99 and
2
2
90
UV-visible (UV-Vis) data, NIR data, mid-infrared (MIR) RMSEP of 0.71, while the other achieved an R value
data, and a fused data matrix. Data preprocessing involved of 0.99 and RMSEP of 0.06, both using different FTIR
selecting relevant wavelengths, applying SNV, and Savitzky- fingerprint regions. The results showed that the detection
91
Golay smoothing. LDA struggled to classify adulterant of pork adulteration in beef meatballs is feasible through
percentages accurately (average prediction accuracies the application of FTIR spectroscopy using the attenuated
Volume 1 Issue 1 (2024) 19 https://doi.org/10.36922/ijamd.2279