Page 20 - IJAMD-1-1

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

data characterized by high variability and noise due to the respectively, with an RMSE of 0.85 and a standard error of
diverse composition of food matrices, the simultaneous prediction (SEP) of 12.94% for TVC. In addition, an MLP
75
presence of multiple microbes at varying concentrations, with an E-nose consisting of 11 gas sensors was used to sort
and inherent variations in the chemical composition of raw beef into four quality categories. The MLP, with two
food samples. 69 hidden layers, achieved an average classification accuracy
76
ML algorithms become essential for ensuring the of 93.73%. This study had a greater focus on lowering
robustness and accuracy of predictions derived from such the computational requirement of the MLP by reducing
highly variable data. Studies have demonstrated their the model to run on a low-cost, low-power consumption
capabilities to achieve precise detection and quantification field-programmable gate array. While it was expected that
of contaminants. For instance, a PLSR model achieved a simplified model would see a decrease in accuracy, it
an R of 0.987 in predicting the level of adulteration in still performs with 92.72% accuracy, demonstrating the
2
2
70
butter, while a BPNN model achieved an R of 0.978. robustness of the MLP model.
In terms of classification accuracy, reports indicate 100% In another study, a dataset encompassing an averaged
detection rates for identifying adulteration in honey. Raman shift ranges from 500 to 1800 cm , TVC, and the
71
-1
Raman hyperspectral imaging (HSI), without the need for lactic acid bacteria (LAB) count for the Chinese crossbred
sample preprocessing, achieved a reasonably high 75.55% yellow cattle beef sample was utilized. PLSR analysis,
accuracy in detecting the presence of Aspergillus flavus by coupled with leave-one-out cross-validation, was applied.
directly imaging whole corn kernels. The development The optimal PLSR model, selected based on achieving
72
of ML-augmented HSI holds promise for enhancing the minimum root mean square error of prediction
accuracy, potentially enabling online and near-real-time (RMSEP) and the lowest number of latent variables (LVs),
monitoring of food quality. In addition, ML models have demonstrated an R of 0.90 and an RMSEP of 0.38 for
2
demonstrated the ability to accurately identify the species TVC in modified air packaging after a 21-day storage
of individual cells. A Kernel PCA-DT model achieves 88.6 period. The model also achieved an R of 0.75 and RMSEP
2
– 95.8% accuracy in identifying single bacterial cells at the of 0.60 for LAB count in the same conditions and time
genus level, though its direct application to food matrices frame. It must, however, be noted that the correlation for
73
77
is yet to be tested. the 3-day and 7-day storage times was generally relatively
4.1. Detection of TVC in meat weak, indicating that the microbial exponential growth
phase likely introduces noise into the data that PLSR could
TVC is a microbiological measurement that assesses the not account for in its models. Another aspect involved
total number of viable microorganisms present in each predicting TVC, Pseudomonas spp., and Brochothrix
food sample. It is an important parameter used to evaluate thermosphacta counts using multispectral imaging (MSI)
the overall microbiological quality and hygiene of a food spectroscopic data. An 18-wavelength MSI and plate
product. The TVC is expressed as the number of colony- counts of respective microbes were used in its dataset.
forming units per unit of sample. High TVC may indicate A partial least squares-discriminant analysis (PLS-DA)
poor hygiene during food processing, storage, or handling, classifier categorized beef samples into three different
and they can be associated with a higher risk of spoilage quality classes with an 80.0% accuracy. In contrast, the
or the presence of harmful microorganisms. Monitoring PLSR model, optimized with five LVs, achieved correlation
TVC is a common practice in the food industry to ensure coefficients of prediction (r ) and RMSEP of 0.78 and 1.29,
compliance with food safety standards and regulations. respectively, for TVC. In addition, it yielded values of
p
Elevated TVC levels might prompt further investigation 0.84 and 1.12 for Pseudomonas spp., and 0.86 and 1.00 for
and corrective actions to maintain the quality and safety B. thermosphacta. In contrast to the previous study, this
78
of food products. In the following sections, the utilization
of ML approaches for the identification of TVC in meat is study did not separate its data based on storage duration
discussed (Figure 5). to train different models but was still able to achieve
reasonably accurate prediction results, making it more
4.1.1. Detection of TVC in beef broadly generalizable to predicting microbial counts.
A multilayer perceptron (MLP) was employed to classify An SVR model was applied to MSI data with a pre-
TVC in raw beef into three bins using FTIR data. The clustering step for TVC prediction in raw beef. The SVR,
2
model, trained with the first five PCs, storage temperature, using a radial basis function kernel, achieved an R of
and duration, achieved an overall classification accuracy 0.98, RMSEP of 0.14, and an 80.8% accuracy in classifying
of 90.5%. It demonstrated 91.7%, 81.2%, and 94.1% TVC into two bins. In the study, three training sets at
79
sensitivities for fresh, semi-fresh, and spoiled samples, temperatures of 2°C, 8°C, and 15°C were made, and it was

Volume 1 Issue 1 (2024) 14 https://doi.org/10.36922/ijamd.2279