Alternative approaches for accelerated shelf life testing (ASLT) of low-moisture food products: A case study of dried shrimp

Abstract

Alternative approaches for shelf life evaluation of food based on accelerated shelf life testing (ASLT) assisted with principal component analysis-based machine learning (ML-ASLT), and hyperspectral imaging integrated with ML-ASLT (HSI-ML-ASLT) were developed and compared with the conventional ASLT (C-ASLT). Dried shrimp was stored at 15, 30, and 40 °C and monitored for changes in sensory characteristics including odor, color and overall acceptability during the storage. Physicochemical properties of the dried shrimp were monitored to describe the changes in sensory attributes. Volatile compounds of initial and unacceptable dried shrimp were analyzed using gas chromatograph-mass spectrometer (GC–MS). The samples were also analyzed using HSI in the wavelength range of 400–1000 nm. PCA-based machine learning was applied with ASLT and HSI-ASLT to determine quality degradation rate constant and shelf life. The ML-ASLT gave the predicted shelf life of 9.87, 6.13, and 3.34 wk., while HSI-ML-ASLT gave the shelf life of 9.71, 6.51 and 3.31 wk. for dried shrimp stored at 15, 30, and 40 °C, respectively, which were more accurate than C-ASLT (10.29, 6.80 and 3.49 wk), when compared to the actual shelf life (10, 6 and 3 wk). By integrating multivariate quality attributes, the PCA-based machine learning model enhances the ASLT approach to yield precise, practical shelf life predictions for dried shrimp. Furthermore, HSI-ML-ASLT provides a rapid, simple, and nondestructive method for evaluating food shelf life without the need for chemical reagents.

Highlights

• •Alternative approaches for accelerated shelf life testing (ASLT) were developed.
• •PCA-based machine learning (ML) improved the accuracy of shelf life prediction.
• •Hyperspectral imaging (HSI) integrated with ML enhanced the ASLT approach.
• •Rapid and chemical-free shelf life prediction can be done using HSI-ML-ASLT.

1. Introduction

Shelf life evaluation is crucial to ensure product quality and minimize food waste caused by unnecessary rejections (Lee, 2024). Conventional shelf life evaluation of shelf-stable or low-moisture foods under ambient conditions usually requires long periods of time. To shorten the testing time, shelf life of food can be evaluated using accelerated shelf life testing (ASLT) by measuring the changes in quality attributes of food stored under extreme stress conditions in order to expedite the degradation process followed by determining a kinetic model of food degradation, and projecting to predict the product shelf life (Hough et al., 2006; Saavedra et al., 2013). ASLT is often used in the development of food products in the industry to provide an early indication of product shelf life and thereby shorten the product development time and product launching schedule (Reid et al., 2003; Siripatrawan & Jantawat, 2009). ASLT also helps prevent food industry from distributing low-quality products and dealing with consumers complaints or product recalls due to unacceptable product quality.

However, there are limitations of conventional ASLT. Although a suitable estimation of shelf life can be achieved with the ASLT models, they do not always represent the nature of the food aging process. In all instances, ASLT using the empirical models usually focuses on only one quality attribute to determine the shelf life termination. Whereas, food compositions are complex and quality degradation of food is multivariate processes and thus, shelf life evaluation may become complicated (Derossi et al., 2016; Siripatrawan, 2018). Therefore, monitoring only one critical quality attribute may not be able to practically reflect the changes in food quality. When there are various quality attributes involved in food deterioration, those attributes need to be taken into account to decide the end of shelf life (Chaudhry et al., 2018; Saavedra et al., 2013;).

Because food deterioration involves multivariate phenomena, comprehensive shelf life studies require advanced, simultaneous multi-attribute modeling to ensure precise and practical quality evaluations (Cui et al., 2024; Derossi et al., 2016). Recent advancements in multivariate accelerated shelf life testing have facilitated the development of accurate shelf life prediction models. The concept of multivariate kinetics has been applied to various food products such as low fat UHT milk (Richards et al., 2014), fresh-cut lettuce (Derossi et al., 2016), cape gooseberry (Figueroa-Avalos et al., 2023), and extra virgin olive oil (Martín-Torres et al., 2023). Although multivariate ASLT is more practical and can improve the accuracy of the shelf life prediction (Chaudhry et al., 2018; Derossi et al., 2016; Saavedra et al., 2013), it is still laborious and time consuming in that it requires a number of sample preparations, physicochemical and/or microbiological analyses, and chemicals. Hyperspectral Imaging (HSI) is a rapid and nondestructive technique that combines spectroscopy and imaging. It requires no complex sample preparation or chemical reagents. The HSI technology has been widely used for agriculture, pharmaceutical and food applications. For instance, monitoring of surimi quality during thermal processing (Xia et al., 2025), detection of latent stage of anthracnose infection in mango (Siripatrawan & Makino, 2024a), classification of fungal infection in rice grain (Siripatrawan & Makino, 2024b).

Artificial intelligence (AI) is computational technologies that enable machines to analyze complex data and resolve multivariate computational problems. AI facilitates the processing of high-volume, multidimensional datasets, enhancing advanced analytical applications, such as pattern recognition, prediction or decision making, and error correction (Siripatrawan & Makino, 2025; Sun et al., 2026). With the developments of AI in the area of agricultural and food chemistry, innovative solutions to the food industry have been proposed. Machine learning, a category of AI which uses algorithms that learn from data to make predictions by incorporating several factors simultaneously, is considered useful and can be applied to determine the shelf life of food (Chaudhry et al., 2018; Derossi et al., 2016; Saavedra et al., 2013). Novel applications based on ML has gained increasing interest in food processing, packaging and shelf life due to their simplicity, flexibility, and ability to analyze large and complicated data (Bonneu et al., 2024; Siripatrawan & Makino, 2025). Recently, Cui et al. (2024) have developed a comprehensive machine learning models with real-time shelf life information to facilitate food safety and quality control. Numerous machine learning algorithms (e.g. support vector machine, k-nearest neighbor and artificial neural network) have been also applied to complex food analytical data related to agriculture and food products (Kim et al., 2025; Siripatrawan & Makino, 2024b). When integrated to HSI, machine learning can help process multidimensional complicated hypercube data and give rapid results with more precision (Siripatrawan & Makino, 2024).

Although ASLT integrated with multivariate data analysis has been applied to indicate shelf life of various food products, ASLT assisted with principal component analysis (PCA)-based machine learning or HSI integrated with ML-ASLT for shelf life estimation of shelf stable food product like dried shrimp has never been reported. Therefore, the aim of this study was to develop alternative methods for prediction of food shelf life by including various quality attributes simultaneously, instead of using only one quality, into the model using (1) ASLT assisted with PCA-based machine learning (ML-ASLT) and (2) HSI integrated with ML-ASLT (HSI-ML-ASLT), and then compared with the conventional ASLT (C-ASLT) and actual shelf life.

2. Materials and methods

2.1. Sample preparation

Dried shrimp was prepared according to Siripatrawan et al. (2022). Shrimp (Macrobrachium equidens) received from Phatthalung Province was washed with water, soaked with salt solution, and then dried in a hot air oven at 120 °C. Fifty grams of dried shrimp, which contained 52.4% crude protein and 3.6% crude fat on a dry weight basis, were packaged in a commercial oriented polypropylene (PP), heat sealed, and stored at 15, 30, and 40 °C for 12 weeks. The PP film has thickness, water vapor transmission rate, oxygen transmission rate at 25 °C and 75% relative humidity of 0.077 ± 0.002 mm, 57.25 ± 2.63 g. m⁻² day⁻¹, and 635.52 ± 24.41 cc. m⁻² day⁻¹, respectively. The packaged samples were evaluated for color, astaxanthin content, thiobarbituric acid reactive substance (TBARS), total volatile basic nitrogen (TVB-N), volatile compounds, sensory characteristics (odor, color and overall acceptability), and hyperspectral imaging.

2.2. Physicochemical analyses

CIELAB color values of L*, a* and b* of dried shrimp were assessed using a chroma meter (Minolta CR-300 Series, Minolta, Osaka, Japan).

Astaxanthin content was analyzed using a spectrophotometric technique according to the method of Tolasa et al. (2005) and Jitpasutham et al. (2024) with modifications. Five grams of dried shrimp were extracted with 20 mL of hexane/ethanol/acetone solvent (2:1:1 v/v). The samples were homogenized and centrifuged at 3000 rpm. After the phase separation completed, the absorbance of the supernatant was assessed at the wavelength of 473 nm using a UV–Vis spectrophotometer (XENON U-5100, Hitachi, Ltd., Japan). The astaxanthin content (μg/g) of dried shrimp was determined from a calibration curve of absorbance versus concentration.

Thiobarbituric acid reactive substance assay was carried out following the method described by Siripatrawan and Makino (2018). In brief, ten grams of dried shrimp sample were mixed with 97.5 mL of distilled water and 2.5 mL of 4 N HCl before distillation. Five milliliters of the obtained distillate was mixed with 5 mL of thiobarbituric acid reagent. The solution test tube was then heated in boiling water for 35 min and cooled down instantly. The absorbance of the distillate (Abs) was assessed at 538 nm using a UV-VIS spectrophotometer (XENON U-5100, Hitachi, Ltd., Japan). Distilled water was used as a blank solution. The TBARS, reported as milligram of malondialdehyde (MDA) equivalents per kg of sample (W_S), was calculated according to Eq. (1).

$$\text{TBARS}=\frac{7.8\times 10\times \mathit{Abs}}{W_s}$$ (1)

Total volatile basic nitrogen of dried shrimp was analyzed using the convey micro-diffusion method (Conway and Byrne, 1933). The inner ring of the Conway dish contained 1% boric acid solution. Three grams of ground dried shrimp were mixed with 17 mL of 4% trichloroacetic acid solution at 25 °C for 30 min. The mixture was then filtered through filter paper (Whatman No.41) and 1 mL of the filtered mixture was transferred to the outer ring of the Conway dish and added with potassium carbonate solution. The dish was incubated at 37 °C for 1 h. The inner solution was then titrated with 0.02 N HCl until the color turned from green to pink (reaching the endpoint). Trichloroacetic acid solution (4%) was used as a blank. The TVB-N value was calculated using Eq. (2) and reported as mg N per 100 g of sample.

$$\mathrm{TVB}-\mathrm{N}=\frac{N\times 14\times \left(V_s-V_b\right)\times V\times 100}{W_s}$$ (2)

where N is the HCl concentration (0.02 N), V_S is the HCl titration volume of dried shrimp sample (mL), V_b is the HCl titration volume of blank (mL), V is the total volume of sample and TCA solution (mL), and W_S is the sample weight (g).

2.3. Volatile compound analysis

Volatile compounds of dried shrimp were collected using the headspace solid-phase microextraction (SPME) and analyzed using gas chromatography–mass spectrometry (GC–MS) (Agilent 7890 A GC-7000 Mass Triple Quad, CA, USA) following the method described by Chuesiang et al. (2023) and Ruengdech and Siripatrawan (2020). Dried shrimp (3 g) was placed in a 20 mL glass vial capped with 20 mm aluminum crimp cap attached with a PTFE silicone septum. Headspace volatiles were collected using a 50/30 μm Divinylbenzene/Carboxen/Polydimethylsiloxane SPME fiber (Supelco, PA, USA) at 30 °C for 20 min. The GC–MS with a DB-WAX, 60 m × 0.25 mm × 0.25 μm capillary column (J&W Scientific, Folsom, CA, USA) was used. The detector was a quadrupole mass analyzer. The carrier gas was helium. The mass spectrometer was used in the electron ionization mode. The ion source temperature was 230 °C, and the ionization energy was 70 eV. The scan mode was in the range of 25–400 m/z. Compounds were tentatively identified using the National Institute of Standards and Technology (NIST) library. Data was analyzed using MassHunter Qualitative Analysis B.04.00 software (Agilent Technologies, Inc., CA, USA).

2.4. Sensory evaluation

Sensory evaluation was conducted by ten trained-experienced members (aged 22–47 years) from the Department of Food Technology and the Center of Excellence in Food Processing Pilot plant, Chulalongkorn University. An initial assembly was carried out prior to testing in order that each member could elucidate each attribute (Chuesiang et al., 2023). Sensory assessment of the samples was performed in a sensory laboratory free from odor and noise. There were two assemblages of sensory assessment at each storage time interval. Sensory attributes including odor, color and overall acceptability were evaluated using seven-point scoring. The sensory score was classified into: 7 = highly desirable; 1 = highly unacceptable. End of the sample shelf life was reached when the score was ≤4.

2.5. Hyperspectral imaging

A hyperspectral camera (JFE, Techno-Research Corporation, Tokyo, Japan) attached with a stage automation control system (Model SGSP 26–200, Sigma–Kaki Co., Ltd., Tokyo, Japan) was used to acquire full spectral and spatial information. The HSI system comprised a Pentax C33500 having a 16 mm focal length for 2/3-in. charge-coupled device camera with aperture range of f/1.6–f/16 lens. The light sources include 150 W Tungsten halogen lamp (ColdSpot PCS-UHX, NPI, Tokyo, Japan) and 150 W Xe lamp (Super Bright 152S, SAN-EI Electric, Osaka, Japan). Data acquisition was performed following the method described by Siripatrawan and Makino (2025). A hyperspectral image cube was obtained by scanning in the direction perpendicular to the spatial stage. The hyperspectral images were attained using ImSpector (Model V10, Spectral Imaging Ltd., Oulu, Finland) which creates a spectrum ranging from 400 to 1000 nm with a spectral resolution of 5 nm. The hyperspectral images were assembled using SpectrumAnalyzer software (Version 1.8.5, JFE, Techno-Research Corporation, Tokyo, Japan).

2.6. Conventional ASLT

The criterion for the end-point of dried shrimp shelf life was set as the time when the sensory score was equal or lower than 4 to grade the samples as being slightly unacceptable. Kinetics of quality degradation of each quality attribute (Eq. (3)) was determined by individually plotting odor, color, and overall acceptability sensory scores versus time at 15, 30, and 40 °C. A plot of C-C₀ or Ln (C/C₀) vs. time was constructed to estimate the value of k by linear regression (Tsironi et al., 2009). Degradation of quality was fit with zero-order (n = 0) or first-order reactions (n = 1) as shown in Eq. (4) and Eq. (5), respectively. The coefficient of determination (R²) was calculated to find the optimal reaction order of the quality attributes of the dried shrimp.

$$\frac{\mathit{dC}}{\mathit{dt}}=k{C}^n$$ (3)

$$C-C_0=-\mathit{kt}$$ (4)

$$\mathit{ln}\frac{C}{C_0}=-\mathit{kt}$$ (5)

where C₀ is the initial quality score; C is the quality score at any time t; k is the rate constant (wk⁻¹); n is the reaction order; and t is time.

2.7. PCA-based machine learning for analysis of ASLT and HSI-ASLT

PCA-based machine learning was used to assist the ASLT determined by sensory evaluation and the ASLT determined by hyperspectral imaging. Unsupervised machine learning, namely PCA which is commonly used for data exploration and reduction in statistical pattern recognition and signal processing (Siripatrawan & Makino, 2024a), was used to analyze the sensory data and hyperspectral data by transforming a large set of variables into a smaller one that still contains most of the information in the large set (Jiarpinijnun et al., 2020). For ML-ASLT, PCA was used to explore the time-associated sensory quality degradation of the dried shrimp at different storage temperatures (15, 30, and 40 °C). For HSI-ML-ASLT, PCA was used to explore the time-associated hyperspectral data corresponding to the quality of dried shrimp. The kinetics of the quality degradation reactions was described based on the PC scores obtained from the PCA and the shelf life was determined (Chaudhry et al., 2018). Data analysis was performed using MATLAB software (MathWorks, Inc., Natick, Massachusetts, USA).

2.8. Statistical analysis

The quality measurements of dried shrimp were performed using a completely randomized design (CRD). All measurements were conducted in triplicate, and the data are presented as mean values ± standard deviation (S.D.) from triplicate measurements. All experimental results of dried shrimp physicochemical quality were analyzed using one way analysis of variance (ANOVA). Multiple comparison test was used to compare the means for statistically significant differences among the individual means established at p < 0.05 using Duncan's Multiple Range Test. Statistical analyses were performed using IBM SPSS Statistics for Windows, Version 30.0 (IBM Corp., NY, USA).

3. Results & discussion

3.1. Changes in physicochemical quality of dried shrimp

Fig. 1 shows the physicochemical properties including L*, a*, b* color values (Fig. 1 (A)), astaxanthin content, TBARS, and TVB-N (Fig. 1B) of the dried shrimp during storage at 15, 30 and 40 °C. Changes of these indices showed a time-associated increase with storage temperatures. L* and b* tended to increase whereas a* value tended to decrease but not significantly different (p > 0.05) with storage temperature and time, suggesting that the color of dried shrimp changed from vermillion-orange (red-orange) color to pale orange color. Li et al. (2019) also found that L* and a* values of hot air dried shrimp decreased, whereas b* value increased during storage. The color changes in dried shrimp may be attributed to astaxanthin degradation caused by oxidation when exposed to oxygen and light and the degradation reaction was accelerated by higher temperature (Gómez-Estaca et al., 2017; Inthamat & Siripatrawan, 2025). These results agree with those of Ahmed et al. (2015) and Hernández Becerra et al. (2014) who found that degradation of astaxanthin increased with temperature.

Fig. 1.

Color values (A), physicochemical properties (B), and sensory quality (odor, color and overall acceptability) (C) of dried shrimp during storage at 15, 30 and 40 °C. All measurements are presented with their respective standard deviation (Mean ± SD), L* = lightness, a* = red (+) and green (−), b* = yellow (+) and blue (−), TBARS = thiobarbituric acid reactive substances, TVB-N = total volatile base nitrogen. Pink dash line indicates the cut-off criterion of the sensory score. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

As shown in Fig. 1B, TBARS values of dried shrimp increased significantly (p ≤ 0.05) with storage time and temperature because dried shrimp contains cholesterol, monounsaturated and polyunsaturated fatty acids, which are susceptible to the oxidation reaction (Jitpasutham et al., 2024; Li et al., 2019). The increase in TBARS values infers the incident of lipid oxidation causing the degradation of dried shrimp quality in terms of discoloration, off-odor, off-flavor, or nutritional loss (Kongboonkird et al., 2026; Siripatrawan et al., 2022). These results concur with those of Li et al. (2019) who found oxidative indices (peroxide value and TBARS) of dried shrimp increased with storage temperature. Wang et al. (2024) also found the changes in lipid profiles and an increase in TBARS due to lipid oxidation of phosphatidylcholine, phosphatidylethanolamine, triglyceride, and diglyceride in dried shrimp during storage.

An increase in TVB-N of dried shrimp with storage time and temperature (Fig. 1B) was possibly caused by the protein break down into amino acid and peptide (Jitpasutham et al., 2024). Moreover, it is also possible that microbial growth initiates the degradation of proteins and nitrogen-containing compounds, leading to an accumulation of volatile organic amines such as dimethylamine (DMA), trimethylamine (TMA) and ammonia (Sun et al., 2022). TVB-N and TMA are generally used as chemical indicators of seafood freshness. In this study, TVB-N was used as a primary indicator for dried shrimp quality. This is because while TMA, produced by bacterial reduction of trimethylamine oxide (TMAO), is a good indicator for freshness or early spoilage in fresh or raw seafood, TVB-N, measuring the total amount of volatile nitrogenous compounds, presents a more comprehensive indicator of protein degradation in dried seafood (Fu et al., 2021; Huss, 1995). Moreover, TMA may deplete or become less active than the ammonia formed during degradation of dried seafood, making TVB-N a reliable indicator for predicting shelf life of dried shrimp (Huss, 1995; Siripatrawan et al., 2022).

3.2. Volatile compounds of dried shrimp

Volatile compounds including alcohols, pyrazines, aldehydes, ketones, and N-containing compounds found in the dried shrimp before and after storage for 6 wk. are listed in Table 1. These volatile compounds are also found in hot air-dried shrimp in the studies of Zhang et al. (2020) and Sun et al. (2022). Zhang et al. (2020) suggested that desirable flavors of dried shrimp are generated by volatile compounds such as alcohols, pyrazines, aldehydes, and ketones produced by Maillard reaction during thermal processing. Pyrazines (e.g. 2,5-dimethyl-Pyrazine, trimethyl-Pyrazine) and other nitrogen-containing heterocyclic compounds are associated with distinctive aromas of hot-air dried shrimps (Zheng et al., 2025). Hao et al. (2025) also documented that volatile compounds, such as trimethyl pyrazine, 3-ethyl-2,5-dimethylpyrazine, 2-methylpyrazine, and 2-heptanone, are the most plentiful Maillard reaction products and are correlated with umami and richness.

Table 1.

Volatile compounds in dried shrimp after storage at 30 °C for 0 wk. and 6 wk.

Compound	Intensity (unit)*
	At 0 wk.	At 6 wk
Methylamine, N,N-dimethyl-	9.63E+06	2.60E+07
n-Hexane	–	1.62E+07
Cyclopentane, methyl-	7.92E+05	5.56E+06
Heptane, 2,4-dimethyl-	5.42E+05	–
Octane, 2,6,6-trimethyl-	–	1.09E+06
Octane, 6-ethyl-2-methyl-	7.22E+05	1.21E+06
Toluene	3.51E+05	–
Undecane	4.18E+05	9.43E+06
Undecane, 4,7-dimethyl-	–	3.84E+06
Hexanal	1.26E+06	–
Nonane, 3,7-dimethyl-	8.66E+05	1.85E+06
Octane, 2,3,6,7-tetramethyl-	4.57E+05	1.36E+06
Nonane, 4,5-dimethyl-	5.09E+05	5.74E+06
Undecane, 4,7-dimethyl-	–	3.24E+06
Octane, 2,3,6,7-tetramethyl-	–	1.25E+06
Cyclohexane, 1-ethyl-2-propyl-	5.26E+05	1.51E+06
1-Penten-3-ol	1.27E+06	3.61E+06
2-Heptanone	–	2.87E+06
Furan, 2-pentyl-	–	9.66E+05
Decane, 3,8-dimethyl-	–	5.45E+05
Dodecane, 2,7,10-trimethyl	2.15E+05	3.42E+06
Pyrazine, methyl-	–	8.32E+05
Pentadecane	–	6.87E+05
Hexadecane	–	5.17E+05
Dodecane, 2,6,11-trimethyl-	2.40E+05	6.78E+05
2-Octanone	–	6.67E+05
Tetradecane, 4-methyl-	–	6.57E+05
Dodecane, 2,6,11-trimethyl-	–	2.36E+06
Dodecane, 2,7,10-trimethyl-	2.40E+05	5.60E+05
Pyrazine, 2,5-dimethyl-	4.83E+05	1.49E+06
1-Decanol, 2-hexyl-	–	1.31E+06
Octadecane	–	5.90E+05
4,4-Dimethyl-cyclohex-2-en-1-ol	–	1.04E+06
1-Hepten-3-ol	–	5.59E+05
Pyrazine, trimethyl-	4.82E+05	9.95E+05
Pyrazine, 2-ethyl-3,5-dimethyl-	3.33E+05	1.24E+06
1-Octen-3-ol	1.48E+06	7.71E+06
2-Hexene, 3,5,5-trimethyl-	8.92E+05	4.03E+06
Benzaldehyde	5.44E+05	2.44E+06
5-Octen-1-ol, (Z)-	–	5.00E+05
Benzyl alcohol	–	6.55E+05

The trimethylamine found in the initial dried shrimp was possibly caused by the thermal degradation of choline, betaine, methionine, or trimethylamine oxide during heating (Sun et al., 2022). The results also showed that methylamine and other volatile nitrogen of stored dried shrimp had higher intensities than those of dried shrimp at 0 wk. An increase of pyrazines in dried shrimp after 6 wk. storage at 30 °C was probably formed by interactions of lipid degradation products (Bai et al., 2022; Zhang et al., 2020). Stored dried shrimp was found to contain higher 1-Octen-3-ol than the initial dried shrimp. Mall and Schieberle (2017) described that 1-Octen-3-ol which is belonged to aliphatic unsaturated alcohols contributes to the mushroom and fishy smells of dried shrimps.

Since dried shrimp contains crude protein (52.4% dry basis), crude fat (3.6% dry basis), cholesterol and polyunsaturated fatty acids, higher concentration of alcohols in stored dries shrimp may be due to the peroxidation of polyunsaturated fatty acids or a reduction of carbonyl compounds (Mall & Schieberle, 2017; Zhang et al., 2020). Aldehydes may be produced by the oxidation of fat (Zhang et al., 2020). Since aldehydes has low threshold, even low intensity may affect flavor attribute of dried shrimp. Ketones, such as 2-Heptanone and 2-Octanone was found only in stored dried shrimp. This was probably produced by the oxidation of monounsaturated and polyunsaturated fatty acids or the degradation of amino acids. Sun et al. (2022) also suggested that ketones like 2-Nonanone may be caused by polyunsaturated fatty acids or amino acid degradation and give a distinctive aroma to food. The volatile compounds found in dried shrimp before and after storage are in accordance with the results of TBARS and TVB-N.

3.3. Changes in sensory quality of dried shrimp

All sensory attributes (odor, color, and overall acceptability) showed declining scores with storage time and temperature (Fig. 1C). The storage temperature and time induced various changes in physicochemical properties that influence the sensory attributes of the product. Odor, color and overall acceptability were considered as good quality indices, as they were well correlated with physicochemical indices and consumer acceptance. The eventual unacceptance of dried shrimp is mainly caused by changes in color and odor which are related to astaxanthin degradation, lipid oxidation, TVB-N accumulation and some volatile compounds generated during the storage. For instance, the degradation of astaxanthin causes color changes and consequently affects sensory scores of the dried shrimp (Hernández Becerra et al., 2014). In addition, as suggested by the GC–MS results, aldehydes and ketones, such as 2-Heptanone and 2-Octanone, initiated by the oxidation of fatty acids or degradation of amino acids were found only in stored dried shrimp causing a decrease in odor scores. These compounds even at low intensity can influence the odor of dried shrimp because of their low threshold (Zhang et al., 2020).

It was evident that all sensory indices correlated with the changes in physicochemical properties (astaxanthin, TBARS, TVB-N and volatile compounds) during the storage of dried shrimp. Therefore, dried shrimp shelf life was determined based on sensory evaluation. In this study, the sensory score of ≤4 was established to categorize samples as unacceptable and used to define the shelf life end-point criterion for dried shrimp.

3.4. Actual shelf life

Although instrumental analyses give useful information about food quality and can accompany data from sensory analysis, for many food products what constitutes their shelf life is often difficult to define (Lu et al., 2011). In this respect, sensory assessment of foods which is a direct measure of consumer response to products can practically describe the complexity of food quality than instrumental procedures (Ruengdech & Siripatrawan, 2020; Siripatrawan, 2026). Therefore, in this study, organoleptic quality was used as a criterion to determine the product shelf life. For dried shrimp, there is no standard reference of the instrumental colors, astaxanthin content, TBARS and TVB-N values to judge the end of shelf life or marketability limits. Therefore, instead of using instrumental measurements, the cut-off criterion was well represented by sensory attributes. Derossi et al. (2016) also suggested that for various food products, when their shelf life is difficult to define, it can be based on consumer acceptance instead of objective criteria, the reference limits for sensorial attributes should be considered as cut-off criterion.

In this study, a cut-off criterion for indicating the end of shelf life of dried shrimp was determined based on odor, color and overall acceptability. The cut-off criterion was set at the sensory score of ≤4 to grade the samples as being unacceptable. The results in Table 2 shows that actual shelf life of dried shrimp based on odor, color, and overall acceptability was 12, 10 and 10 wk. at 15 °C; 6, 6 and 6 wk. at 30 °C; and 3, 3, and 3 wk., respectively at 40 °C. It should be noted that dried shrimp stored at 15 °C was rejected due to unacceptable color and overall acceptability after storage for 10 wk. even though its odor was still acceptable. In such case, basing on the least time point the samples to reach the unacceptable quality, the shelf life was thus considered to be 10 wk. Freitas et al. (2003) also suggested the usual procedure for indicating the shelf life is to approve of the smallest estimated value. Therefore, the shelf life values of dried shrimp were 10, 6 and 3 wk. at 15, 30 and 40 °C, respectively.

Table 2.

Actual shelf life of dried shrimp stored at 15, 30 and 40 °C.

Storage condition	Time to reach the rejected sensory score (wk)			Actual Shelf Life(wk)
	Odor	Color	Overall acceptability
15 °C	12.00	10.00	10.00	10
30 °C	6.00	6.00	6.00	6
40 °C	3.00	3.00	3.00	3

3.5. Conventional ASLT-calculated shelf life

In this study, the shelf life of dried shrimp was determined basing on the sensory evaluation. The sensory quality attributes were applied with zero order reaction that represents a linear relationship of the quality loss and first-order reaction that represents an exponential evolution of the quality. This method of fitting the degradation data with linear and non-linear regression has been widely used and reported to be practical for shelf life studies (Boonmeesuk et al., 2024; Palazón et al., 2009; Reid et al., 2003). It was found that zero-order kinetics better explained the quality degradation with time since coefficient of determination (R²) was higher than those obtained from first-order reaction (data not shown). The loss of quality attributes has been found to follow the zero-order kinetic reaction. The kinetic parameters of the sensory quality degradation were then measured by fitting the sensory scores with time using a linear regression method (Fig. 1C).

Odor, color, and overall degradation during storage at 15, 30 and 40 °C followed zero-order kinetics and the rate constant (k) of 0.255–0.451 wk.⁻¹ with a R² of 0.763–0.906 were obtained as summarized in Table 3. Coefficient of determination values of the conventional ASLT (C-ASLT)-calculated shelf life at 15, 30 and 40 °C were in the range of 0.763–0.797, 0.904–0.919, 0.825–0.890, respectively. The results suggested that sensory scores of odor, color, and overall acceptability were modeled by apparent zero-order reactions and the results for temperature dependence were in agreement with those for chemical indices.

Table 3.

Kinetic rate constant and shelf life of dried shrimp stored at 15, 30, 40 °C using C-ASLT.

Storage condition	Quality	Linear equation	K-rate constant⁎ (wk−1)	R2	Shelf life (wk)
15 °C	Odor	y = −0.255 x + 6.769	0.255	0.763	10.47
	Color	y = −0.270 x + 6.935	0.276	0.797	10.29
	Overall acceptability	y = −0.246 x + 6.795	0.246	0.782	10.97
30 °C	odor	y = −0.382 x + 6.771	0.382	0.906	7.25
	Color	y = −0.404 x + 6.779	0.404	0.919	6.88
	Overall acceptability	y = −0.399 x + 6.714	0.399	0.904	6.80
40 °C	Odor	y = −0.452 x + 5.998	0.451	0.890	4.20
	Color	y = −0.449 x + 5.667	0.449	0.825	3.49
	Overall acceptability	y = −0.444 x + 5.845	0.444	0.847	3.93

(^⁎) Kinetic rate constants of the sensory quality (odor, color and overall acceptability): zero-order kinetics expressed as wk.⁻¹, R² = Coefficient of determination.

The results in Table 3 show that the conventional ASLT-calculated shelf life values of dried shrimp based on odor, color, and overall acceptability, respectively, were 10.47, 10.28 and 10.97 wk. when stored at 15 °C; 7.25, 6.88 and 6.80 wk. at 30 °C; and 4.20, 3.49, and 3.93 wk. at 40 °C. The fact that when one quality attribute was unacceptable the consumer would reject the product, notwithstanding other quality attributes are still acceptable. Moreover, the procedure for indicating the shelf life by accepting the least estimated value has also been suggested by Freitas et al. (2003). Therefore, the end of shelf life is the least time point the samples reach the unacceptable quality. The conventional ASLT-calculated shelf life was 10.29, 6.80 and 3.49 wk. at 15, 30 and 40 °C, respectively.

3.6. ML-ASLT-calculated shelf life

It has been suggested that the degradation process of dried shrimp is a multivariate phenomenon depending on several factors and causing changes in various quality attributes (Cui et al., 2024; Derossi et al., 2016; Siripatrawan & Makino, 2018). Therefore, an unsupervised machine learning technique, namely principal component analysis (PCA), was integrated with ASLT to simultaneously include several quality attributes as criteria for shelf life prediction (Chaudhry et al., 2018). PCA finds a low dimensional representation of a dataset such that the selected features contain as much variation in the dataset as possible (i.e. it aims to maximize the variance over a set of linear combinations). The multidimensional data can be described with smaller numbers of principal components (PC).

In this study, the data matrix (X) of i × j was arranged, where i is the number of samples collected from all time points throughout the storage (i = 1,2, …,m) and j is the number of quality attributes (variables, j = 1,2, …,p) used in the shelf life study. To integrate PCA with ASLT, few steps are involved and the mathematical formulations are given in Eq. (6)- Eq. (10). First, the original data matrix (X) is standardized using autoscaling (Eq. (6)) to ensure that all data patterns have a comparable scale. Covariance matrix (ϑ), determined using Eq. (7), is then decomposed to eigenvectors (principal components, q), which are unit vectors indicating directions of the PC, and eigenvalues (λ) which represent the amount of variance explained by each PC, as shown in Eq. (8). These eigenvectors form a new orthogonal coordinate and the original data is projected onto these PC. The result is a lower-dimensional representation of the original data that retains as much variance as possible. The eigenvectors are ranked based on their corresponding eigenvalues in descending order. The eigenvector with the highest eigenvalue corresponds to the first principal component, the one with the second-highest eigenvalue corresponds to the second principal component, and so forth.

$$x_{i,j_a}=\frac{x_{i,j}-{\overline{x}}_j}{{\sigma }_j}$$ (6)

$$\vartheta =\frac{1}{m}\sum _{i=1}^m\mathbf{x}\left(i\right){\mathbf{x}}^T\left(i\right)$$ (7)

$$\vartheta q_j={\lambda }_j{q}_j,j=\mathit{1},\mathit{2},\dots ,p$$ (8)

where q is the eigenvector, λ is the eigenvalues, T is the matrix transposition.

The original vector x may be reconstructed as in Eq. (9). Where a_j are the principal components which are the projections of x onto the principal directions. A set of principal components retained ($\widehat{\mathbf{x}}$) that are needed for effective data representation is computed by discarding those linear combinations as shown in Eq. (10).

$$\mathbf{x}=\sum _{j=1}^p{a}_j{q}_j$$ (9)

$$\widehat{\mathbf{x}}=\sum _{j=1}^l{a}_j{q}_j,l\le p$$ (10)

To calculate the shelf life using ML-ASLT, individual reference limits (x_s) of each quality attribute (color, odor and overall acceptability) are chosen and then multiplied by the loadings (L) of the PCA to obtain a vector of acceptable values (τ) using Eq. (11). The cut-off criterion (τ_c), as shown in Eq. (12), is then calculated as maximum acceptable value of the vector τ. The shelf life is then determined using the cut-off criterion as shown on Fig. 2 and summarized in Table 4.

$$\mathbf{\tau }={\mathbf{x}}_{s}L$$ (11)

$${\mathbf{\tau }}_{\mathbf{c}}=\mathit{Max}\left(\mathbf{\tau }\right)$$ (12)

where ${\mathbf{x}}_s$ is the row vector of the auto-scaled reference (critical) values of each quality attribute. L is the loading matrix of the time-related PC (Derossi et al., 2018).

Fig. 2.

ASLT assisted with PCA-base machine learning (ML-ASLT): PC1 scores of dried shrimp as a function of time during storage at 15, 30 and 40 °C. Pink dash line indicates the cut-off criterion. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 4.

Kinetic rate constant and shelf life of dried shrimp stored at 15, 30, 40 °C using the ML-ASLT and HSI-ML-ASLT.

Shelf life method	Storage condition	Linear equation	Kinetic rate constant⁎(wk−1)	R2	Shelf life (wk)
ML-ASLT	15 °C	y = −0.449× + 4.446	0.449	0.991	9.87
	30 °C	y = −0.684× + 4.309	0.684	0.968	6.13
	40 °C	y = −0.779× + 2.720	0.779	0.895	3.34
HSI-ML-ASLT	15 °C	y = 0.345× - 1.163	0.345	0.964	9.71
	30 °C	Y = 0.562× - 1.415	0.562	0.996	6.51
	40 °C	y = 0.656× - 1.372	0.656	0.915	3.31

(^⁎) Kinetic rate constants of the sensory quality (odor, color and overall acceptability): zero-order kinetics expressed as wk.⁻¹, R² = Coefficient of determination.

PCA facilitates data exploration and interpretation of multivariate datasets by transforming the measured data into new uncorrelated variables called principal components (PC) (Jiarpinijnun et al., 2020; Ruengdech et al., 2019). Each of these principal components explains a percentage of the total variation in the dataset. The principal components (e.g. PC1, PC2, …) are arranged in order of decreasing variance. In this study, two principal components (PC1 and PC2) were obtained. These are linear combinations of the original variables, with the first principal component (PC1) having the largest variance, and the second principal component (PC2) having the second largest variance. As a result, the data dimension is reduced by preserving the systematic variation in the data retained in the selected components. PC1 could explain 96.86% and PC2 could explain 1.85% of the total variance, and these 2 components explained 98.71% of the total variance. The PCA results suggested that as high as 96.86% of the original information in the dataset (having 3 variables: odor, color and overall acceptability) can be described by the first PC.

In this case, PC1 was hence selected as it possesses highest variances. The structure of the multidimensional data (odor, color and overall acceptability) can be examined graphically with this new coordinate system. The scores of the first PC1 which conveys most information regarding the original data were then plotted against storage time as shown in Fig. 2. The reaction order of the obtained PC scores versus time was then described by fitting with the zero-order or first-order kinetics. The fitting of PC1 scores as a function of time was optimal with the zero-order model. Fig. 2 shows the PC1 scores as a function of time, the fittings obtained with the zero-order model, and the cut-off criterion. The results showed that the fits matched the PC1 scores well for all three storage temperatures, demonstrating that the zero-order kinetics explains the multivariate criteria of sample well. The reaction order and multivariate kinetic parameters (rate constants) were obtained and the results are displayed in Table 4.

According to Eq. (11)-Eq. (12), for the selection of cut-off criterion (τ_c) for the quality attributes under study (color, odor and overall acceptability), individual reference limits were identified for each quality characteristic (when the score ≤ 4). Then the reference limit value for each quality characteristic was auto-scaled as described by Pedro and Ferreira (2006) and then multiplied with the PC1 loading of the corresponding dependent variables as shown in Eq. (11). The cut-off criterion (τ_c = 0.107) was selected from the calculated maximum acceptable value of the vector τ (Eq. (12)). The shelf life was then predicted by using the loadings of the PCA to calculate the cut-off criterion (τ_c) for the time-related PC using Eq. (12). The results in Fig. 2 and Table 4 show that the ML-ASLT-calculated shelf life of dried shrimp stored at 15, 30 and 40 °C was 9.87, 6.13, and 3.34 wk., respectively, and the coefficient of determination was 0.991, 0.968, 0.895, respectively.

3.7. HSI-ML-ASLT calculated shelf life

The HSI data were collected from dries shrimp stored at 15, 30 and 40 °C for 0, 4, 8, and 12 wk. and there were 20 samples at each storage time. The detector signal intensity counts were transformed into reflectance units and mathematically pretreated using Sawitzky–Golay smoothing. Fig. 3 shows the time-related reflectance spectra of samples at 15, 30 and 40 °C, respectively. The patterns of the spectra at different storage time and temperature were almost similar but there are different intensities, probably due to the changes in physicochemical properties and compositions such as colors, astaxanthin content, TBARS, TVB-N, and volatile compounds occurred during storage as described earlier. Therefore, the HSI results enable a comprehensive evaluation of the changes in dried shrimp quality.

Fig. 3.

HSI spectra of dried shrimp during storage at 15, 30 and 40 °C.

HSI-ML-ASLT shelf life calculation was performed in the same manner as ML-ASLT, but the HSI-ML-ASLT used reflectance spectra instead of the sensory attribute data. The PCA-based machine learning was used to explore and reduce the hyperspectral data and gave PC score and loading matrices. Two principal components were obtained. PC1 and PC2 could explain 99.10% and 0.58%, respectively, of the total variance. In this case, PC1 was hence selected as it possessed highest variances. To determine the shelf life of dried shrimp using HSI-ML-ASLT, the average PC1 scores of time-related HSI data at each storage temperature were then plotted against the storage time as shown in Fig. 4. The fitting of PC1 scores versus time was optimal with the zero-order model for all three storage temperatures. The reaction order and multivariate kinetic parameters (rate constants) were obtained and the results are displayed in Table 4. The cut-off criterion (τ_c) for judging the end of shelf life was then calculated according to Eq. (11)-Eq. (12). The results in Fig. 4 and Table 4 show that the HSI-ML-ASLT-calculated shelf life of dried shrimp stored at 15, 30 and 40 °C was 9.71, 6.51, and 3.31 wk., respectively, with the coefficient of determination of 0.964, 0.996, 0.915, respectively. This high coefficient of determination indicates a goodness of fit.

Fig. 4.

HSI-ASLT assisted with PCA-based machine learning (HSI-ML-ASLT): PC1 scores of HSI spectra as a function of time during storage at 15, 30 and 40 °C. Pink dash line indicates the cut-off criterion. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 5 shows the comparison among the actual shelf life, and the shelf life calculated using C-ASLT, ML-ASLT, and HSI-ML-ASLT. The differences between C-ASLT and ML-ASLT results potentially because both techniques gave different kinetic rate constants, indicating different paces of quality degradations and thus different cut-off criteria. The C-ASLT seems to give overestimated shelf life values when compared to the actual shelf life. On the contrary, the ML-ASLT and HSI-ML-ASLT gave more accurate shelf life estimation because the ML-kinetics took all quality attributes into account simultaneously. Thus, the calculated shelf life was based on overall degradations, instead of focusing only on one quality. The HSI-ML-ASLT gave shelf life values close to ML-ASLT. In addition, the HSI technique provides a rapid, simple, non-destructive analysis without requiring sample preparation or chemical reagents.

Fig. 5.

Comparison among actual shelf life, C-ASLT calculated-, ML-ASLT calculated-, and HSI-ML-ASLT calculated shelf life.

It is worth noting that although the ASLT assisted with PCA-based ML algorithm used in this study provides more accurate shelf life results, as more quality parameters are considered simultaneously, than the conventional ASLT and offers a reduction in time and cost required to estimate shelf-life, its application may be limited by its inherent mathematical constraints. PCA assumes linear relationships among variables, whereas the degradation of complex food matrices often involves non-linear kinetic reactions. This potentially leads to imprecise shelf-life projections. However, this constraint may be eliminated by implementing advanced machine learning or deep learning algorithms to enable adaptive shelf-life prediction. This study also demonstrated that the HSI integrated with PCA-based machine learning approach is a rapid, nondestructive and chemical-free technique that can provide conclusive, accurate, and more complete results. Consequently, the practical implementation of this promising HSI-ASLT approach in food shelf-life prediction seems to be unavoidably increased in the near future. However, its adoption may be limited by technical and operational factors, primarily the high cost of HSI instrument, massive data processing requirements, and complex calibration processes. Nevertheless, future trends point to the broader adaptation of cost-effective HSI equipment combined with practical implementation of advanced ML computational techniques and sophisticated software.

4. Conclusion

The results indicate that sensory attributes (color, odor and overall acceptability) of dried shrimp correlate with the changes in physicochemical properties (astaxanthin, TBARS, TVB-N and volatile compounds) during the storage at 15, 30 and 40 °C. Dried shrimp shelf life was determined based on sensory evaluation, with a sensory score of ≤4 established as the end-point criterion. By integrating multivariate quality attributes, the PCA-based machine learning model enhances the ASLT approach to yield precise, practical shelf life predictions for dried shrimp. Compared to the conventional ASLT, the ML-ASLT and HSI-ML-ASLT approaches demonstrate improved accuracy in quantifying the shelf life of dried shrimp under the tested storage conditions at 15, 30 and 40 °C. Furthermore, the HSI-ML-ASLT technique is rapid, convenient, nondestructive and not labor-intensive. Reliable shelf life predictions help minimize food waste and support food safety by preventing consumers from receiving low quality or expiry food products.

Sensory evaluation ethics statement

Ethical permission was not required for the sensory evaluation in this research. The sensory evaluation conducted in this study were voluntarily participated. The participants were fully informed of the objectives of the study. No personal identifying information was collected, and all responses were treated confidentially and anonymously. The sensory evaluation was conducted in accordance with appropriate ethical principles for research involving human participants, ensuring the protection of participants' rights and privacy throughout the study. The participants agreed to participate in the sensory evaluation and consented to use their information.

CRediT authorship contribution statement

Ubonrat Siripatrawan: Writing – original draft, Visualization, Validation, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Weeranuch Hwansena: Investigation, Formal analysis. Marisa Kongboonkird: Writing – review & editing, Validation. Yoshio Makino: Writing – review & editing, Visualization, Validation, Resources.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.