The conversion of biomass from agro-wastes into animal feed through bioconversion is a potential method for generating sustainable ruminant feed. Additionally, to improve sustainability in livestock production, this method also ensures long-term feed security while creating an environmentally friendly method of feeding animals (Ponnampalam et al., 2025). Citronella residue (CR) represents a significant portion of agro-industrial wastes that are readily available within many tropical production systems. It is estimated that over 99.2% of all waste produced in the citronella essential oil extraction industry is CR (Manurung et al., 2015). CR is primarily wasted due to its high level of lignin, which limits the nutritional value of the material as ruminant feed. Although it was found that ruminants may have limited intake when fed CR alone, there are still opportunities to use this material as a source of energy for ruminants. Solid-state fermentation (SSF) using filamentous fungi has recently become increasingly popular as a means to biologically modify the chemical composition of these materials. As a result of biological modification, SSF can reduce lignin content, increase microbial protein concentration, and decrease the presence of anti-nutritional compounds present in the substrate(Wang et al., 2023); (Fan et al., 2024). As (Heidari et al., 2022) stated, SSF of canola meal using Pleurotus ostreatus resulted in significantly decreased phytate concentrations (up to 98.8%), as well as an increase in crude protein concentration of 11%-18% compared to unfermented canola meal.

In ideal circumstances, SSF will also increase the nutritional quality of the feed materials. In an SSF environment, organisms capable of producing exogenous enzyme production, i.e., cellulase and xylanase, may convert cellulose and hemicellulose-based oligo-saccharide polysaccharides into simpler oligo-saccharide chains for better accessibility by fiber-associated proteins (Chukwuma et al., 2020). The enzymatic activity in SSF processes occurs in a dynamic manner and is extremely sensitive to environmental variables, e.g., substrate variability, moisture distribution, temperature gradient differences, O2 diffusion limitations, etc.(Bamidele et al., 2025). Therefore, Jiang et al. (Jiang et al., 2015) asserted that to improve fermentation rate and prevent slow and/or poor fermentation rates, it is necessary to utilize effective online monitoring techniques, which will provide timely/real-time data for improving the consistency and quality of SSF processed products as value-added feedstuffs.

Near-infrared spectroscopy (NIRS) is a non-destructive analysis tool that has become increasingly popular across many different types of industries. Some examples include food (Jiang et al., 2023), agriculture (Munawar et al., 2024), pharmaceutical (Dégardin et al., 2016); (Massei et al., 2025), and animal feed (Samadi et al., 2023), (Samadi et al., 2024), (Samadi et al., 2025). One reason why NIRS is so popular is that it offers a relatively fast and accurate way to get chemically relevant information when compared to other methods of collecting this type of data for evaluating biological processes. Using NIRS theoretically enables one to identify changes in the overtone and combination bands related to vibrations caused by stretching of the covalent bonds found within molecules, i.e., O-H, C-H, N-H, and S-H. These molecular vibrations are occurring in the near-infrared (NIR) range (Cen & He, 2007). Therefore, chemical information about biological systems can be obtained in the NIR region. During SSF, the substrate undergoes a very rapid biochemical transformation as the microorganisms break down the cellulose and hemicellulose present on the substrate. This leads to a significant increase in the moisture content, as well as the amount of microbial biomass produced and the level of protein and structural carbohydrates in the substrate. With each of these biochemical transformations, new functional groups are formed while others may be destroyed. Each of these biochemical transformations will alter the absorbance spectrum of the substrate in the NIR wavelength (Jiang et al., 2015). Since there are differences in spectral characteristics for each of these biochemical transformations, spectral fingerprinting can be utilized to differentiate among the several levels of fermentation. Previous research has demonstrated that NIRS can be applied to measure moisture content and pH (Wahyudi et al., 2025), soluble protein and trypsin inhibitor content (Dai et al., 2023), and extent of fermentation (Jiang et al., 2015) during SSF. In summary, these studies demonstrate that NIRS is a rapidly applicable, reliable, and versatile analytical tool that can provide data on both physical/chemical attributes as well as changes in anti-nutritional compounds throughout the SSF process.

Gross energy (GE) and crude protein (CP) and crude fat (ether extract or EE) are generally considered among the three most important of all of the nutrient parameters in evaluating ingredients of animal feeds; they provide the best general indication of the potential energy content of an ingredient to animals (Kil et al., 2013); indicate the bioavailability of amino acids and microbial nitrogen availability (Dias et al., 2025); and serve as the principal measure of the biochemical stability of fats in animal diets (Rolls, 2009). In recent years, there have been significant advancements in the use of NIRS technology, along with multivariate statistical techniques such as chemometrics and machine learning (ML), primarily in the field of animal feed and food science. For instance, (Cavallini et al., 2025) indicated that a number of researchers have employed technologies such as those mentioned previously to develop “precision feeding” programs using a variety of fibrous agro-industrial by-product substrates, and consequently reducing the environmental impact associated with livestock production systems. (Gasparini et al., 2026) argue that, while NIRS-based methodologies have been increasingly accepted due to their rapid determination of the chemical composition of food and feed ingredients at an industrial level, there is increasing acknowledgment of the relevance of NIRS-based methodologies for rapid assessment of the existence of contaminants and for identification of the origin and/or identity of individual feed ingredients within global supply chains. Therefore, although it is well established that NIRS-machine learning (NIRS-ML) can be successfully applied in a wide variety of applications beyond just predicting nutrients, this study aims to explore, for the first time, to the authors’ knowledge, the capability of this methodology for the dynamic assessment of proximate nutritional component characteristics (i.e., GE, CP, and EE) during SSF of CR. Therefore, while the present study does not propose the introduction of NIRS-ML as a new paradigm for predicting proximate composition, we do intend to provide evidence (to our knowledge, for the first time) of its utility for dynamically measuring key proximate nutrient components (GE, CP, and EE) specifically in CR undergoing SSF and in a substrate/application context that has never previously been examined.

Although there is a potential for high levels of variability in the proximate nutritional values produced from SSF, this study aimed at developing and testing an NIRS-based predictive model that could predict changes in three key proximate nutritional components (gross energy [GE], crude protein [CP], and ether extract [EE]) as they occur during the SSF process. As the SSF process involves dynamic interactions among microbial populations, substrate materials, temperature, moisture, etc., it also introduces significant complexity into the data, potentially leading to redundant information in multidimensional NIR spectral datasets. Due to this level of complexity, variable selection methods are required to isolate wavelengths that contain the most useful information for predicting proximate nutrient content while minimizing interference with prediction performance.

To address these challenges, a successive projection algorithm (SPA) was applied to identify the appropriate wavelengths with the least multicollinearity prior to modeling. The three machine learning algorithms used to build the models are: light gradient boosting machine (LightGBM), partial least squares regression (PLSR), and support vector regression (SVR). These algorithms are used together because they have different advantages and disadvantages and have been demonstrated to be successful in chemometrics with small to medium sample sizes (n = 45 for calibration). PLSR is considered to be the traditional standard in NIRS analysis, as it addresses the problem of collinearity in the spectral data by reducing the dimensionality of the data into fewer principal components, which can perform adequately if there is a relatively simple linear relationship between the spectra and the measured property. SVR was used to address potential nonlinear relationships between spectra and the analyte, given the complexities of biochemical reactions during SSF. SVR has particular appeal for the large number of variables with few samples, since it uses both regularization and the “kernel trick” to reduce the risk of overfitting(Wahyudi et al., 2025). A further advantage of LightGBM compared to many other tree-based methods is both faster training times and lower memory usage. These advantages should be especially beneficial when you have multiple iterations of cross-validation and/or hyperparameter tuning. Due to the small sample sizes, alternatives such as deep neural networks were deliberately excluded from this comparison, as they would likely exhibit severe overfitting and poor generalization. This research is designed to test the effectiveness of three techniques (a linear latent variables technique (PLSR), a nonlinear regression technique (SVR), and a decision tree ensemble method (LightGBM) to identify which one can be most effective to follow the dynamic changes of substrate concentration during SSF fermentation by analyzing nutritional changes.

The results section provides information on gross energy (GE), crude protein (CP), and crude fat (EE) using a proposed workflow based on NIRS. Initially, the subsets used to calibrate the model and predict the data are described (Table 1). Following this description of subsets, the wavelength selected from MSC pre-processed data after selecting principal variables by SPA (Tables 2-3; Figures 2-4) and subsequently the predictive capabilities of models derived with PLSR, SVR and LightGBM using R2, root mean square error (RMSE), and mean absolute error (MAE) on an independent test dataset (Tables 3; Figures 5-8).

The higher-order value of creating fast, reliable monitoring systems for SSF-produced feed ingredients is illustrated by the increased reliance on agro-industrial residues in animal feed production. Such materials exhibit natural variability in composition and, as seen in the recent cases cited in (Gasparini et al., 2024), this variability may pose feed safety concerns due to the diversity of supply chains they originate from. As such, NIRS-based screening methods that track critical nutritional components during fermentation timelines at reasonable speeds (as shown to be possible with moderate reliability for CP and GE in this study) would provide an important first level of monitoring capability. The collected data collectively support a feasible analytical determination; however, this depends on the parameter analyzed. CP provides the highest degree of reliable analytical prediction (PLSR analysis, Tables 1-3,Figures 5-8); GE demonstrates some ability to predict moderately well using nonlinear models (LightGBM/SVR, Table 3, Figure 5); and EE does not demonstrate sufficient ability to predict accurately (Table 3, Figure 7). The grouped MAE comparison in Figure 8 succinctly summarizes this hierarchy performance.

The use of four 7-day cycles (over 28 days) using SSF conditions in addition to the presence of replicates suggests that NIRS, coupled with chemometric/machine learning analysis, has potential as a fast-screening method to help monitor production processes specifically when nutritional enhancements are related to proteins (where predictions have been shown to be most reliable) (Jiang et al., 2015); (Dai et al., 2023). As the study did not provide empirical data to show generalization of the results to previously unseen periods of fermentation or novel uses of fungal treatments (i.e., time-hold-out validation or treatment holdout validation); the term “dynamic monitoring” should be defined as predictive models based on sampling at various points over time and/or treatment conditions as opposed to validated deployment under completely new periods/treatment (Ezenarro & Schorn‐García, 2025).

Results from this study need to be understood within the context of several methodological limitations that reduce both the generalizability and reliability of the findings. As a limitation, the predictive accuracy of the models developed was based on a single holdout methodology (the 3:1 split shown in Table 1). Although this provides a starting point for estimating the potential predictive value of these models with respect to an independent test data set, this approach also does not provide estimates of predictive accuracy variability associated with the use of different splits or predictability when using multiple independent fermentation batches, which has become recognized as best practice in NIRS chemometric validation (Ezenarro & Schorn‐García, 2025). Second, samples were generated under a structured SSF design (fermentation period × fungal treatment). Consequently, the reported prediction performance should be viewed as feasible within the sampled domain rather than evidence of universal generalization to unseen operating conditions, time periods, or fungal treatments. Third, the risk of overfitting was a specific concern for LightGBM given the small calibration set (n = 45). Several measures were taken to mitigate this: (i) a conservative hyperparameter grid was used, favouring shallow trees (num_leaves: 15-63, max_depth: -1 to 10) and a minimum child sample constraint (min_child_samples: 10-30) to limit model complexity; (ii) hyperparameter selection was based on 10-fold cross-validation RMSE rather than training error; and (iii) the num_leaves and min_child_samples parameters were explicitly constrained to prevent leaf-level memorisation of noise. Despite these measures, the calibration-prediction gap for GE (R²c = 0.82 vs. R²p = 0.55) indicates that overfitting was not fully eliminated. Accordingly, GE prediction should be framed as moderately informative for screening/process support within the present dataset rather than high-precision routine quantification. Finally, crude fat (EE) could not be predicted reliably in the current configuration, as indicated by low or negative validation R2 values and prediction errors that remain high relative to the natural dispersion of EE in the dataset (Tables 1 and 3, Figures 7-8).

The limitations should be identified and explicitly communicated for each specific analytical component, with the understanding that they represent analyte-specific constraints associated with the current dataset and modeling configuration. Moreover, the structured nature of the data collected (i.e., samples were non-independent; rather, samples were nested within a defined experimental configuration) will limit the ability to generalize findings based upon sample number; therefore, it is likely that the effective size of the calibration set will be less than the total number of samples analyzed. Furthermore, hyperparameter optimization was performed with a single 10-fold cross-validation run using the full calibration set. Optimization via repeated or nested cross-validation runs could yield more stable estimates of model performance and reduced sensitivity to partitioned subsets of the calibration data (Jalal et al., 2025).

While there are certainly some constraints associated with the research that was conducted here, this work is consistent with other reports concerning the use of spectroscopy as a monitoring tool for SSF processes by applying chemometrics. More specifically, FT-NIR has been demonstrated as an effective method to track key SSF process parameters, including the alcohol content and residual glucose (Jiang et al., 2018), which provides considerable justification for the use of NIRS as a monitoring tool during the SSF process independent of differing targeted analytical parameters. Therefore, this project may be viewed as an expansion of the monitoring process variables to include proximate nutrients. There was strong evidence for predicting crude protein content and moderate evidence for predicting total energy content, given the experimental parameters. The superior performance of nonlinear statistical methods demonstrates that the relationship between NIR spectral information and GE content in the complex SSF environment studied herein does not follow a strictly linear model. This outcome is also consistent with biological expectations, since GE content reflects the cumulative contributions of multiple biochemically distinct compounds (e.g., protein, fat, dietary fiber, residual carbohydrate) that vary independently in their rates and patterns of degradation throughout fermentation. Moreover, it is likely that overlapping absorption bands and matrix effects lead to nonlinear relationships among NIR spectral features that cannot be captured by linear latent variable models such as PLSR. Nevertheless, the calibration-prediction gap observed for LightGBM (R2c = 0.82 vs. R2p = 0.55) indicates that the calibration set was overly fit to the calibration data; a problem that occurs when flexible tree-based models (e.g., LightGBM) are employed with limited sample sizes (i.e., n = 45). Hence, while GE content prediction could yield reasonable results for screening (RMSEp ≈ 0.65 × SD), the developed models should not be used for quantitative accuracy assessment without additional independent validation.

The practical benefit of the successful prediction of CP (R2p = 0.72; RMSE = 0.29%) will be very useful in practice. CP is one of the primary specifications used by commercial feed manufacturers to determine how much to pay for their ingredients and to develop rations. Traditional Kjeldahl chemical methods require approximately 4-6 hours to obtain results, plus additional time for sample preparation. However, using NIRS to predict crude protein content, with the developed PLSR model, estimates can be obtained immediately after measurement. Therefore, this rapidity allows the operator(s) to make adjustments to fermentation time, aeration rate, or inoculum quantity during processing that would otherwise occur at the end of processing, prior to product completion, thereby reducing variability from batch to batch. Additionally, NIRS provides small-scale or resource-limited feed manufacturers (including many feed producers in tropical areas where large quantities of citronella residue are readily available) with a cost-effective alternative to expensive, laboratory-based wet-chemical analytical techniques. Although the prediction accuracy for GE (R2p = 0.55; RMSEP = 192.34 J/g) does not meet the quality control thresholds (R2p ≥ 0.90), it may still serve as a screening tool. For instance, if there are large deviations from predicted GE values during fermentation monitoring, this would indicate abnormal substrate degradation and/or incomplete fermentation. As such, targeted chemical analyses could then be performed on these suspect lots. It appears that the better-performing nonlinear models (LightGBM, SVR) capture complex relationships among the components (carbohydrate, protein, and fat) of biomass. However, the calibration prediction gap observed with LightGBM (R2c = 0.82; R2p = 0.55) suggests that current model complexity has yet to achieve full generality; therefore, GE predictions should be viewed cautiously and considered only as trending indicators.

In addition to being theoretically important, the inability to predict EE (negative R2p for SVR and LightGBM) provides practical information about how far NIRS can go with this matrix and compound. The inability to predict EE can be explained by several interacting chemical and spectroscopic variables. First, in the present data set, EE concentrations were very low and highly variable (range = 0.20%-4.39% in calibration; see Table 1), reducing the SVR available for model development. Lipid bands associated with CH-stretching overtone transitions from 1704-1780 nm, and CH-combination band transitions from 2300-2370 nm (Liu et al., 2015) are responsible for most of the absorption due to lipids in NIR spectra. However, in the more complex matrices of fermented substrates like those used here, these lipid bands are much weaker than the strong absorptions of water (OH, ~1450 & ~1940 nm), protein (NH, ~2050 & ~2180 nm), and structural carbohydrate molecules. These relatively large absorptivities will have masked the fat-related spectral signals. Additionally, the biochemical heterogeneity resulting from SSF processes, including partial breakdown of cell walls and redistribution of lipid fractions, is expected to increase spectral complexity and reduce the consistency of fat-related absorption patterns in samples. Manufacturers interested in using NIRS to measure fat levels in fermented citronella residue would need to supplement the substrate with sufficient fat (>5% EE) to obtain a measurable response or use an alternate method to rapidly measure fat content in their products, such as mid-infrared (MIR) spectroscopy. MIR spectroscopy provides significantly stronger, more selective lipid absorption bands than NIRS and has enabled improved predictions of fat content in more complex biological matrices (Lozano et al., 2017). From a fermentation monitoring perspective, EE is not a reliable process indicator under current conditions; moisture, pH, or CP are more appropriate targets.

This study confirmed that NIRS coupled with both chemometrics and machine learning can predict certain nutritional characteristics in citronella residue undergoing solid-state fermentation. The most reliable results were obtained by estimating CP utilizing PLSR. Gross energy (GE) was estimated at low reliability but with good enough values to provide only approximate trending estimates. LightGBM had the best predictive performance for all models tested on GE. Crude fat (EE), however, could not be accurately predicted using this technique under these data sets and model configurations. While there is obvious value in NIRS technology as a quick screening method for tracking changes in crude protein levels throughout solid-state fermentation, it is also important to understand that the utility of NIRS has limitations; i.e., predictions for GE are likely only useful for general trending purposes and should not be taken as exact quantitative estimates. Additionally, the lack of reliable EE estimation means that no recommendations for use can be made without either advancements in methodology or additional calibration samples.

The authors declare that there are no known financial or personal conflicts of interest that could have influenced the findings presented in this study.

The authors express their gratitude to Universitas Syiah Kuala for supporting this research through the H-Index Scheme in 2025, under contract number 327/UN11.L1/PG.01.03/14641-PTNBH/2025.

The authors state that no generative AI or AI-assisted technologies were used in the writing of this manuscript.