Large-scale assessment of regression modeling practices in phytomedicine extraction process optimization Authors: Yuan Tao, Zhexuan Yu, Tianxue Chen, Wei Jin, Weifeng Jin, Li Yu Published: 2026 ABSTRACT SummaryPhytomedicine extraction-process optimization often relies on regression or machine learning models, but model choice is commonly judged by goodness of fit alone. Here, we assembled 1,148 structurally eligible extraction datasets from systematic literature screening and evaluated seven candidate models within a unified benchmark. The models included full quadratic regression, two subset-selected parsimonious regression models, quadratic ridge regression, support vector regression, partial least-squares regression, and Gaussian process regression. Performance was assessed across fitting, leave-one-out cross-validation, diagnostic behavior, and stability of model-predicted optima. Parsimonious regression models showed the most balanced profile, whereas some machine learning models reduced training errors without consistently improving cross-validated prediction or optimization stability. These results indicate that model selection can itself introduce optimization uncertainty and should be evaluated as a decision-oriented component of extraction-process modeling rather than by goodness of fit alone.Keywords: phytomedicine extraction, regression modeling, machine learning, optimization stability, RegressCheck Introduction Optimization of extraction processes is generally considered a foundation of phytomedicine.1,2,3 Although various extraction techniques have been applied in phytomedicine research,4 several fundamental methodological issues remain in the modeling and optimization of extraction processes.5 The first is the development of effective mathematical models capable of properly describing the relationships between extraction variables and target constituents. At present, both statistical methods6,7,8 and data-driven optimization techniques9,10 have been used in related studies, yet their robustness and interpretability vary across studies. Another important issue is the identification of appropriate extraction factors. This is commonly addressed using experimental design methods or heuristic optimization algorithms, including orthogonal design,11 genetic algorithms,9 and ant colony optimization.10 Although extensive experimental and data-driven studies have been conducted on the optimization of phytomedicine extraction processes, and a research pathway integrating multidisciplinary knowledge from chemistry, process engineering, data analysis, and pharmacology has gradually been established,9,10,12,13,14 model selection in practice still often relies on research conventions or a limited number of statistical indicators, rather than on a systematic and comprehensive evaluation framework.15,16,17,18,19,20 Because the same dataset can usually be fitted by multiple models, different models may produce different predicted results and even lead to inconsistent optimization conclusions.18 In this context, if model hypothesis testing and predictive validation are not adequately incorporated into the evaluation process, it is difficult to establish unified criteria for model comparison, and the generalizability of the resulting conclusions may also be affected.16,17,21,22,23 Therefore, a more rigorous and standardized examination of model selection and performance evaluation is needed.18,19,20,22,23,24,25,26 In addition, considering that the number of controllable factors and the sample size in laboratory extraction studies are usually limited, the introduction of complex models and intelligent optimization methods should be carefully assessed through strict validation procedures.16,17,21,23,27,28 To address this challenge, this study proposes and systematically examines the issue of “model-dependent optimization uncertainty” in phytomedicine extraction-process optimization; that is, the same extraction dataset may yield different predicted optimal responses and optimal process conditions when characterized using different candidate models. To our knowledge, independent comparative studies based on large-scale real-world extraction datasets that simultaneously assess predictive reliability, diagnostic adequacy, and optimization-conclusion stability remain lacking. Unlike previous studies that mainly focused on the fitting performance of individual models, this study evaluates model selection within a process-optimization decision-making framework. Specifically, this work examined whether different models, including three regression models and four machine learning models, could provide reliable prediction, reasonable diagnostic behavior, and stable judgments of optimal process conditions in small-sample extraction experiments. The overall study workflow is shown in Figure 1. The findings of this study provide systematic evidence for model selection in phytomedicine extraction-process optimization. They also suggest that modeling decisions should not be based solely on goodness of fit, but should also incorporate predictive validation, diagnostic evaluation, and the stability of optimization conclusions. Therefore, the main contribution of this study is not the development of a new modeling algorithm, but the reframing of model selection in phytomedicine extraction optimization from a simple comparison of goodness of fit into a decision-oriented evaluation problem that jointly considers predictive reliability, diagnostic adequacy, and optimization-conclusion stability. Study workflow for large-scale model comparison and optimization-stability evaluation See also Figure S1. Results Datasets and models After literature screening, data-integrity checking, and structural eligibility assessment, 1,148 datasets that met the requirements for unified quadratic regression modeling were included in the main M0–M6 model comparison. Another 251 datasets were not included in the main analysis because their sample sizes were insufficient to support the estimation of the full quadratic regression model. These datasets were retained for a machine-learning-only sensitivity analysis to further define the applicability boundary of the main findings. The complete process of literature screening, dataset exclusion, and analytical-set construction is shown in Figure S1. Seven candidate models were compared under a unified analytical framework, as summarized in Table 1. M0–M2 were traditional regression models constructed within the same quadratic candidate-term space, including the full quadratic response surface model and two parsimonious regression models based on global subset selection. M3–M6 were machine learning models suitable for small-sample, low-dimensional continuous-response prediction, including quadratic ridge regression, support vector regression, partial least-squares regression, and Gaussian process regression. The variable-processing procedures, candidate-term construction, and hyperparameter-tuning strategies for each model are described in STAR Methods. Table 1Summary of candidate modelsModelModel typeCore structureMain role in this studyMain limitationM0full quadratic regressionfull second-order polynomial with linear, squared, and two-way interaction termsbaseline response-surface reference modelmay retain redundant or unstable termsM1Max Rc2 selected regression submodelglobal subset of full quadratic candidate terms selected by maximum Rc2parsimonious model for prediction and optimizationmay favor explanatory fit over strict parsimonyM2Mallows’ Cr-selected regression submodelglobal subset of full quadratic candidate terms selected by Mallows’ Cr criterionparsimonious model emphasizing predictive error controlmay sacrifice some fit for predictive parsimonyM3quadratic ridge regressionsecond-order polynomial predictors were standardized before Ridge regression.regularized quadratic model for multicollinearity controlstill constrained by the quadratic feature spaceM4support vector regressionstandardized predictors with linear/RBF kernels.Kernel-based model for nonlinear predictionlower interpretability and hyperparameter sensitivityM5partial least-squares regressionstandardized predictors projected onto latent components.latent-variable model for collinear small-sample datamay underfit complex response surfacesM6Gaussian process regressionstandardized predictors with covariance kernels.probabilistic model for small-sample nonlinear predictionmay overfit small datasets and reduce CV stabilityNote: M0–M2 were constructed within the same full quadratic candidate-term space. Summary of candidate models Note: M0–M2 were constructed within the same full quadratic candidate-term space. Model evaluation was conducted across four complementary dimensions: overall fitting performance, leave-one-out cross-validation predictive ability, model diagnostics, and the stability of model-based optimization conclusions. Overall fitting performance and cross-validation were used to assess the explanatory ability of models for the training data and their predictive reliability for unseen samples, respectively. Model diagnostics were used to examine residual structure, sensitivity to abnormal observations, and error stability, whereas optimization-conclusion stability was used to determine whether different models produced consistent predicted optimal responses and extraction conditions. Because all candidate models were evaluated on the same set of datasets, between-model comparisons were performed under a paired analytical framework. The detailed procedures for literature screening, data extraction, structural eligibility assessment, model fitting, cross-validation, diagnostic analysis, optimization search, sensitivity analysis, and statistical testing are described in STAR Methods. Overall performance Overall fitting performance was evaluated using three metrics: R2, root mean square error (RMSE), and mean absolute error (MAE). Each metric was summarized using the median and interquartile range, and its distributional characteristics are shown in Figures 2A–2C. The Shapiro-Wilk normality test29 indicated that most evaluation metrics did not satisfy the normality assumption (Table S1). Therefore, between-model differences were compared using paired Wilcoxon signed-rank tests, and the results are shown in Figures 2D–2F. In each heatmap, the upper triangle shows Holm-adjusted p values from paired Wilcoxon signed-rank tests, whereas the lower triangle shows the absolute Hodges-Lehmann paired difference estimate (|HL|). Triangle symbols denote the magnitude of |HL|: none, |HL| < 0.01; one, 0.01 ≤ |HL| < 0.03; two, 0.03 ≤ |HL| < 0.05; three, 0.05 ≤ |HL| < 0.07; and four, |HL| ≥ 0.07. Asterisks indicate statistical significance after Holm correction: ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001. The same visualization scheme was used for all subsequent model-comparison heatmaps. Overall fitting performance and pairwise model comparisons (A) Distribution of R2, (B) distribution of RMSE, (C) distribution of MAE, (D) pairwise comparison results for R2, (E) pairwise comparison results for RMSE, and (F) pairwise comparison results for MAE. For (A–C), data summarize n = 1,148 independent extraction-process datasets. Data are represented as violin plots with overlaid boxplots and jittered dataset-level points; boxplot center lines indicate medians, boxes indicate interquartile ranges (IQRs), and whiskers extend to 1.5 × IQR. Pairwise comparisons were performed using paired Wilcoxon signed-rank tests with Holm correction. In each heatmap, upper triangles show Holm-adjusted p values, and lower triangles show absolute Hodges-Lehmann paired difference estimates (|HL|). Triangle symbols indicate |HL| ranges: none, |HL| < 0.01; one, 0.01 ≤ |HL| < 0.03; two, 0.03 ≤ |HL| < 0.05; three, 0.05 ≤ |HL| < 0.07; four, |HL| ≥ 0.07. Asterisks indicate Holm-adjusted significance: ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001. See also Figure S2 and Table S1. Taken together, R2, RMSE, and MAE showed a relatively clear hierarchy in the overall performance of different models, although the advantages of each model were not fully consistent. In terms of explanatory ability, M6 had the highest median R2 value of 1.0000, which was higher than those of M0(0.9610) and M1 (0.9558), indicating that M6 achieved the best fitting performance for most datasets. Meanwhile, the R2 distributions of M0 and M1 also remained at a relatively high level, suggesting that both models also showed good explanatory ability. Further comparison showed that although the Wilcoxon paired test for R2 between M0 and M1 remained statistically significant after Holm correction, the Hodges-Lehmann estimate of the difference was only 0.0075 (95% CI: (0.0064–0.0083)). If ±0.01 was defined as the range of practically negligible differences, this confidence interval fell entirely within the prespecified range, indicating that the difference in explanatory ability between the two models was very small and that their performance could be considered approximately equivalent in practical terms. In terms of error control, M6 performed best for both RMSE and MAE, with the lowest median RMSE of 0.0235 and the lowest median MAE of 0.0208. M4 ranked second, with median RMSE and MAE values of 0.2040 and 0.1748, respectively. Further comparison showed that M4 and M6 still differed significantly in RMSE, whereas their difference in MAE was not statistically significant. Moreover, the HL estimate of the difference was only 0.0043, and its 95% confidence interval crossed 0, suggesting that the practical difference between the two models in terms of absolute error was already limited. Overall, among the machine learning models, M6 showed the best performance, followed by M4. Among the traditional regression models, M0 and M1 showed relatively balanced performance, with both high goodness of fit and low error levels, whereas M2 performed slightly less well. In addition, supplementary analyses were conducted for the traditional regression models M0–M2 to compare model complexity, Akaike information criterion (AIC), overall model significance, and the significance of retained terms.30,31 The results are shown in Figure S2. Although M0 achieved a high goodness of fit, it tended to retain a larger number of terms with limited statistical contribution. In contrast, M1 and M2 reduced redundant terms through global subset selection and achieved a better balance between fitting performance and model parsimony. Among them, M1 showed a more balanced performance in terms of AIC, overall model significance, and retained-term significance, indicating stronger overall applicability among the traditional regression models. Model predictive ability Predictive ability was further evaluated using three leave-one-out cross-validation metrics: Q2, root mean square error of cross-validation (RMSECV), and mean absolute error of cross-validation (MAECV). Each metric was summarized using the median and interquartile range, and its distributional characteristics are shown in Figures 3A–3C. The Shapiro-Wilk normality test indicated that most cross-validation metrics did not satisfy the normality assumption (Table S2). Therefore, between-model differences were compared using paired Wilcoxon signed-rank tests, and the results are shown in Figures 3D–3F. Distribution characteristics of the three cross-validation metrics and heatmaps of pairwise model comparisons (A) Distribution of Q2, (B) distribution of RMSECV, (C) distribution of MAECV, (D) pairwise model comparison results for Q2, (E) pairwise model comparison results for RMSECV, and (F) pairwise model comparison results for MAECV. For A–C, data summarize n = 1,148 independent extraction-process datasets. Data are represented as violin plots with overlaid boxplots and jittered dataset-level points; boxplot center lines indicate medians, boxes indicate interquartile ranges (IQRs), and whiskers extend to 1.5 × IQR. Pairwise comparisons were performed using paired Wilcoxon signed-rank tests with Holm correction. In each heatmap, upper triangles show Holm-adjusted p values, and lower triangles show absolute Hodges-Lehmann paired difference estimates (|HL|). Triangle symbols indicate |HL| ranges: none, |HL| < 0.01; one, 0.01 ≤ |HL| < 0.03; two, 0.03 ≤ |HL| < 0.05; three, 0.05 ≤ |HL| < 0.07; four, |HL| ≥ 0.07. Asterisks indicate Holm-adjusted significance: ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001. See also Table S2. Collectively, these three cross-validation metrics showed a relatively clear hierarchy in the cross-validated predictive ability of different models. In terms of predictive explanatory ability, M1 had the highest median Q2 value of 0.6436, which was higher than those of M2 (0.5579) and M4 (0.4981). In contrast, M6 had the lowest median Q2 value of −24.8911, indicating markedly insufficient predictive stability. In terms of predictive error control, M1 also performed best for both RMSECV and MAECV, with median values of 1.0105 and 0.7881, respectively. M2 and M4 ranked next, and their error levels were relatively close. Further comparison showed that the advantage of M1 over M2 and M4 was supported by the paired nonparametric tests, whereas the HL estimate of the difference between M2 and M4 was close to 0 and its 95% confidence interval crossed 0, suggesting that the practical difference between these two models was limited. Overall, M1 performed best across all three cross-validation metrics, showing the strongest overall generalization ability. M2 and M4 could be regarded as a second tier with comparable performance. Although M0 still retained a certain level of predictive ability, its overall performance was inferior to that of M1, M2, and M4. M3 and M5 showed moderate performance, whereas M6 was clearly at a disadvantage. Model diagnostic analysis On the basis of the evaluations of fitting performance and predictive ability, model diagnostics were further conducted from three aspects: residual structure, sensitivity to abnormal observations, and error stability. The main diagnostic metrics included the p value of the residual normality test (pn), the p value of the residual mean-deviation test (pm), the maximum absolute residual, and the interquartile range of absolute errors (AE-IQRs). To more directly reflect the degree of residual mean deviation, pm is displayed as 1—pm in the figure. The distributional characteristics of the main diagnostic metrics and the between-model comparison results are shown in Figures 4A–4H. The corresponding Shapiro-Wilk normality test results for the diagnostic metrics are reported in Table S3. For the traditional regression models, Cook’s distance and the Breusch-Pagan test were further used to assess sensitivity to highly influential observations and residual variance stability, respectively.32,33 The corresponding results are shown in Figure S3. Distribution characteristics of diagnostic metrics and heatmaps of pairwise model comparisons (A) Distribution of the residual normality test p value (pn), (B) distribution of 1—pm, (C) distribution of the maximum absolute residual, (D) distribution of the interquartile range of absolute errors (AE-IQR), (E) pairwise model comparison results for pn, (F) pairwise model comparison results for pm, (G) pairwise model comparison results for the maximum absolute residual, (H) pairwise model comparison results for AE-IQR. For A–D, data summarize n = 1,148 independent extraction-process datasets. Data are represented as violin plots with overlaid boxplots and jittered dataset-level points; boxplot center lines indicate medians, boxes indicate interquartile ranges (IQRs), and whiskers extend to 1.5 × IQR. Pairwise comparisons were performed using paired Wilcoxon signed-rank tests with Holm correction. In each heatmap, upper triangles show Holm-adjusted p values, and lower triangles show absolute Hodges-Lehmann paired difference estimates (|HL|). Triangle symbols indicate |HL| ranges: none, |HL| < 0.01; one, 0.01 ≤ |HL| < 0.03; two, 0.03 ≤ |HL| < 0.05; three, 0.05 ≤ |HL| < 0.07; four, |HL| ≥ 0.07. Asterisks indicate Holm-adjusted significance: ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001. See also Figure S3 and Tables S3 and S4. Overall, different models showed distinct patterns in residual structure and training-set residual dispersion. With respect to residual structure, the traditional regression models generally performed better (Figure 4A). The median p values of the residual normality test were 0.6302, 0.5285, and 0.4967 for M0, M1, and M2, respectively, which were markedly higher than that of M6. This suggests that the residual distributions of the traditional regression models were generally closer to the normality assumption. Pairwise model comparisons further showed clear differences in the distribution of pn among models (Figure 4E). At the same time, the p values for the residual mean-deviation test were generally close to 1 across models. Because Figure 4B displays this metric as 1 — pm, values closer to 0 indicate the absence of an obvious systematic deviation in the mean residual. The corresponding between-model comparison results are shown in Figure 4F. Overall, these findings suggest that none of the models showed a clear tendency toward systematic overestimation or underestimation. In terms of training-set residual dispersion, M6 showed the smallest residual variation (Figures 4C and 4D). Its median maximum absolute residual was 0.0035, and its median AE-IQR was only 0.0011, both of which were the lowest among all models. Between-model comparisons also showed that M6 differed clearly from most other models in terms of both maximum absolute residual and AE-IQR (Figures 4G and 4H). However, the apparent advantage of M6 in training-set residual metrics should be interpreted cautiously, because both the maximum absolute residual and AE-IQR were calculated from training-set fitted residuals. Taken together with the cross-validation results described above, the extremely low training-set residual dispersion of M6 mainly indicates that Gaussian process regression could closely reproduce the response values at the observed experimental points, but does not directly demonstrate stable cross-validated predictive ability. This was also reflected in the marked contrast between the training-set fitting performance and leave-one-out cross-validation (LOOCV) predictive performance of M6. For M6, the median training-set R2 was 0.9997, indicating that Gaussian process regression (GPR) achieved nearly perfect in-sample fitting for most datasets. However, its median Q2 was only −24.8911. When calculated dataset by dataset, the median difference between training-set R2 and Q2 reached 22.4614. Further statistics showed that, under the GPR model, 97.74% of the datasets yielded negative Q2 values, and 64.37% had Q2 values below −10 (Table S4). These results indicate that the excellent in-sample fitting performance of GPR did not translate into stable leave-one-out cross-validation predictive ability. Instead, they suggest that, for small-sample phytomedicine extraction datasets, GPR may tend to closely fit the observed experimental points, while producing unstable predictions for held-out observations. To further determine whether this phenomenon was mainly related to overfitting, kernel selection, or feature scaling, a dedicated GPR analysis was conducted (Table S4). This table summarizes the gap between GPR training-set fitting performance and LOOCV predictive performance, the proportion of negative Q2 values, the selected kernel types, the use of WhiteKernel, and model performance stratified by sample size. The results showed that the poor cross-validation performance of GPR was not confined to a specific kernel structure and was unlikely to be explained solely by differences in feature scaling. Rather, it was more likely to reflect the general instability of highly flexible nonparametric models under sparse small-sample experimental designs. The sample-size-stratified results also showed that the gap between training-set R2 and Q2 for GPR was particularly pronounced in small-sample datasets. Although the degree of instability varied across sample-size strata, the overall pattern remained consistent: The advantage in training-set fitting did not reliably translate into improved cross-validation predictive performance. For the traditional regression models, the supplementary diagnostic results showed that M1 achieved a more balanced performance in terms of sensitivity to highly influential observations and variance stability (Figure S3). The median proportion of observations exceeding the Cook’s distance threshold was 0.2083 for M1, which was lower than that for M0 (0.2308) and M2 (0.3333), indicating relatively lower sensitivity to highly influential observations. Meanwhile, the median p value of the Breusch-Pagan test for M1 was 0.3268, higher than those for M0 (0.2792) and M2 (0.1626), suggesting relatively better residual variance stability. These results indicate that, among the traditional regression models, M1 showed a more balanced diagnostic profile in terms of abnormal-observation sensitivity and variance stability. Taken together, the model diagnostic results indicate that no single model performed best across all diagnostic dimensions. The traditional regression models showed advantages in terms of residual-structure adequacy (Figures 4A and 4B), whereas GPR exhibited strong training-set residual control (Figures 4C and 4D) but clearly insufficient cross-validation stability (Table S4). These findings further support the core judgment of this study: In small-sample phytomedicine extraction-process optimization, high in-sample goodness of fit and low training-set residuals should not be directly interpreted as evidence of reliable predictive ability or robust optimization decisions unless they are also supported by cross-validation performance and optimization-stability analysis. Consistency analysis of model-based optimization conclusions In addition to the fitting, prediction, and diagnostic metrics, this study further compared the consistency of different models in terms of optimal extraction conditions and predicted optimal responses, so as to evaluate the influence of model selection on the stability of optimization conclusions. The relevant distributional characteristics are shown in Figure 5. The analysis of the seven-model comparison followed the same approach as described above, and the distributional characteristics of coefficient of variation of the response (response CV), average coefficient of variation of the factors (average factor CV), 95% confidence interval (CI) width, and 95% prediction interval (PI) width were also reported. Comparison of the distributions of consistency metrics for optimization conclusions under the three-model and seven-model settings (A) Distribution of response CV, (B) distribution of average factor CV, (C) distribution of 95% CI width, and (D) distribution of 95% PI width. For each model-comparison setting, data summarize n = 1,148 independent extraction-process datasets. Data are represented as right-half violin plots with overlaid boxplots; boxplot center lines indicate medians, boxes indicate interquartile ranges (IQRs), and whiskers extend to 1.5 × IQR. The 95% CI width and 95% PI width are shown on log10 scales. See also Table S7. For the comparison among M0, M1, and M2, the model-based optimization conclusions were generally consistent. For the predicted optimal response, the median response CV was 0.0377, with an interquartile range of 0.0076–0.1051. Among the 1,148 datasets, 73.95% had CV values below 10%, and only 6.79% exceeded 30%. Meanwhile, the median 95% CI width and 95% PI width were 0.2412 and 0.4823, respectively, with interquartile ranges of 0.0229–1.7877 and 0.0457–3.5756. These results suggest that the three models produced largely similar estimates of the predicted optimal response, with relatively low overall uncertainty. For the optimal extraction conditions, the consistency among the three models was also high. The median average factor CV was only 0.0106, with an interquartile range of 0.0007–0.0601, and 81.88% of the datasets had values below 10%. This indicates that the traditional regression models generally produced similar judgments regarding the optimal combination of process parameters. In contrast, when all seven models were compared, the discrepancies among models increased markedly. For the predicted optimal response, the median response CV increased to 0.2503, with an interquartile range of 0.1353–0.4801. Only 14.46% of the datasets had CV values below 10%, whereas 42.94% exceeded 30%. Meanwhile, the median 95% CI width and 95% PI width increased to 1.0753 and 3.0414, respectively, with interquartile ranges of 0.1848–5.2652 and 0.5225–14.8924. The PI was also markedly wider than the CI, indicating that the dispersion of the predicted optimal responses increased substantially under the seven-model comparison. For the optimal extraction conditions, the median average factor CV was 0.1215, with an interquartile range of 0.0587–0.2017. Only 24.30% of the datasets had values below 10%, whereas 41.47% exceeded 30%. This indicates that, under the seven-model comparison setting, discrepancies in the optimal process-parameter combinations obtained from different models also increased markedly. From a practical perspective, the increased discrepancies in optimization conclusions under model comparison do not necessarily mean that the optimal conditions predicted by a given model are incorrect. Rather, they indicate that the current data may provide insufficient support for a single optimal point. For phytomedicine extraction processes, optimal conditions usually need to account for multiple practical considerations, including target-compound yield, quality stability, solvent consumption, energy use, equipment compatibility, and safety. Therefore, when different models yield similar predicted optimal responses and comparable factor combinations, the corresponding region can be regarded as a relatively reliable candidate process window. Conversely, when the model-predicted optima are dispersed, the predicted responses differ substantially, or the prediction intervals are wide, the model results are more appropriately interpreted as exploratory evidence that requires further confirmation through additional experiments or local validation. Sensitivity and stratified robustness analyses This study first compared the 1,148 datasets included in the main analysis with the 251 datasets that did not meet the structural inclusion criterion (Table S5). The results showed that the latter had markedly lower effective sample sizes and s–t values than the former, indicating that their exclusion was mainly due to insufficient structural support for estimating the full quadratic regression model, rather than selection based on model performance. A further machine-learning-only sensitivity analysis of these 251 datasets using M3–M6 (Table S6) showed that support vector regression performed relatively well in terms of Q2 and MAECV, suggesting that datasets with insufficient structural support may exhibit a different pattern of model applicability. Therefore, the conclusions from the complete M0–M6 comparison should be restricted to datasets that meet the requirements for unified regression modeling. Stratified robustness analyses were further conducted according to experimental design type, effective sample size, and number of factors to assess whether differences in data structure affected the main conclusions (Table S7). Overall, although the magnitude of model differences varied across strata, the main patterns remained consistent. Regression models based on global subset selection, represented by M1, maintained good cross-validation predictive ability in most strata. Gaussian process regression showed some advantage in controlling training-set residual dispersion, but this advantage did not consistently translate into better cross-validated predictive performance. When the comparison was expanded from traditional regression models to all candidate models, discrepancies in the predicted optimal responses and optimal factor combinations increased markedly across all strata. These results indicate that the main findings regarding model predictive performance and model-dependent optimization uncertainty were not driven by any single experimental design, sample size, or number of factors, but were generally robust across different data structures. RegressCheck platform and worked example RegressCheck was designed as a standardized web-based platform for regression analysis under various experimental settings. The platform integrates automatic model fitting, hypothesis testing, statistical evaluation, and graphical visualization, thereby providing a systematic workflow for model comparison and diagnostic interpretation. Data are imported in excel format, with the independent variables arranged in the first n−1 columns and the response variable placed in the final column. As shown in Figure 6, the platform interface contains several functional modules that guide users through the complete analytical process, including data import, results overview, candidate model comparison, single-model result inspection, and diagnostic plot analysis. Main functional interfaces of the RegressCheck platform (A) Data import module for uploading data files and specifying the response variable, (B) results overview page for displaying the analysis results of candidate models M0–M6, (C) candidate model page for listing candidate models under the selected evaluation criterion and showing their R2, Rc2 and variable composition, (D) detailed model results page for presenting the regression equation, variable combination, key evaluation metrics, and parameter estimates of the selected model, including R2, Rc2, Q2, regression coefficients, and corresponding p values, to support model comparison, parameter interpretation, and diagnostic analysis, (E) diagnostic plots page, including the predicted-versus-observed plot, residuals-versus-fitted plot, residual Q-Q plot, and p value distribution plot for regression coefficients. See also Table S8. Platform design and analytical workflow RegressCheck was developed as a standardized platform for model comparison and diagnostic evaluation of extraction-process experimental data. Unlike software such as design-expert, which focuses mainly on experimental design and response surface optimization, RegressCheck is not intended to replace design of experiments (DOE) or process-optimization software. Instead, it focuses on transparent candidate-model comparison, predictive validation, diagnostic checking, and optimization-stability assessment after experimental data have been obtained. Therefore, the platform is better positioned as a complementary tool to conventional experimental-design and optimization software. The main value of the platform lies in integrating model fitting, cross-validation, parameter significance testing, diagnostic visualization, and candidate-model comparison into a single analytical workflow. This allows researchers to avoid selecting models solely on the basis of training-set R2 or a single error metric. Its core analytical implementation is consistent with the local analysis scripts used in this study, and the corresponding reproducible code is available through the public repository listed in the Data and code availability section. Therefore, RegressCheck is not intended to automatically identify a single “best” model. Rather, it helps researchers detect potential overfitting risks, diagnose abnormal results, and determine whether model selection may affect optimization conclusions.18,24,25,34,35,36,37,38,39 Platform demonstration Dataset 178040 was selected as the worked example in this section. After the dataset was uploaded to RegressCheck, the platform automatically performed model fitting, LOOCV, diagnostic evaluation, and model search. The main results are shown in Table 2, and the complete results are provided in Table S8. If only training-set goodness of fit was considered, M6 might be regarded as the best-performing model because its training-set R2 reached 1.0000. However, its LOOCV performance was not superior, indicating that its apparent fitting advantage did not translate into stable predictive ability. In contrast, M1 maintained a high training-set fit while achieving better cross-validation performance, suggesting higher predictive reliability. Although M2 showed a slightly lower training-set fit, its error distribution was relatively stable and could serve as a supplementary comparison from the perspective of model diagnostics. Table 2Model comparison and model-predicted optimum for dataset 1780ModelR2Q2RMSECVpnMaxAbsResAE-IQRPredicted response (mg/g)A: time (h)B: temperature (°C)C: liquid/material ratio (mL/g)M00.9353−0.1694.3310.6581.5830.873123.6752.38351.00050.000M10.93010.4093.0790.3232.0000.886123.4312.35050.50050.000M20.87120.2963.3610.8903.1610.722122.7262.23352.50050.000M61.0000−0.1364.269–––121.9272.50050.00050.000M0 (reported)0.9388–––––120.9602.20053.40045.700Note: “—” indicates that the value was unavailable or not applicable. The original reported results were taken from the source article, where R2 refers to the model fit reported there. Model comparison and model-predicted optimum for dataset 1780 Note: “—” indicates that the value was unavailable or not applicable. The original reported results were taken from the source article, where R2 refers to the model fit reported there. Table 2 further compares the worked-example results with the optimum reported in the original study. Overall, M0 was closest in model form to the full quadratic RSM model commonly used in the source study, and its model-predicted response and recommended conditions were relatively close to the original results. This suggests that RegressCheck could reasonably reproduce the original analytical trend at the level of conventional RSM fitting and point optimization. However, the LOOCV results led to a different interpretation: Although M0 was close to the original results, its predictive stability was insufficient; M1 did not have the highest training-set R2 but showed better predictive reliability; and M6 achieved the highest training-set fit and produced a point optimum with a predicted response close to the original report, but its LOOCV performance was not superior. This contrast indicates that a high training-set fit or a seemingly plausible point optimum does not necessarily demonstrate stable predictive reliability. This worked example shows that proximity to the original RSM results or a high training-set fit does not necessarily indicate better predictive reliability or more stable optimization recommendations. Therefore, model selection should jointly consider fitting performance, predictive reliability, model diagnostics, and the consistency of recommended optimal conditions, rather than relying solely on training-set R2 or similarity in model form. Discussion This study was not intended to develop a new modeling algorithm, but to re-examine model selection in phytomedicine extraction-process optimization within a decision-oriented evaluation framework. Existing studies on phytomedicine extraction optimization have widely used RSM, regression models, and some machine learning methods to fit the relationships between process factors and responses. However, model evaluation has often focused on training-set R2, regression significance, or training error, with less simultaneous attention to cross-validation predictive ability, diagnostic adequacy, and the influence of model selection on optimal process conditions. The findings of this study support the view emphasized in previous work that model validation and predictive evaluation are essential: In small-sample experimental designs, high in-sample goodness of fit does not necessarily imply reliable cross-validated predictive ability.16,17,19,20,24,26,41,42,43,44 This study proposed and systematically examined the issue of model-dependent optimization uncertainty; that is, optimal conditions are not directly observed outcomes, but are predicted by models within the factor space. Therefore, when different models generate different response surfaces from the same data, model differences may propagate from statistical evaluation to optimization conclusions. We found that, when the comparison was expanded from traditional regression models to all candidate models, the discrepancies in predicted optimal responses and optimal factor combinations increased markedly. This indicates that, when reporting optimal extraction conditions, providing only the fitting metrics or regression equation of a single model is insufficient. Researchers should also clarify whether model selection may alter the optimization judgment. From the perspective of traditional regression models, the findings of this study partly support the continued applicability of RSM in small-sample phytomedicine extraction-process optimization,6,7,8 while also refining the conventional default use of full quadratic models. Existing studies often emphasize the ability of full quadratic response surface models to represent main effects, quadratic effects, and interaction effects, and this advantage was still supported by the present results. However, when sample size is limited, retaining all candidate terms may also introduce redundant complexity. In contrast, parsimonious regression models based on global subset selection reduced unnecessary regression terms while preserving the basic structure of the quadratic response surface. As a result, they achieved a more balanced performance among predictive ability, model parsimony, and optimization stability. This finding suggests that, in small-sample, low-dimensional extraction experiments with continuous responses, controlling model complexity may be more practically meaningful than simply pursuing training-set goodness of fit.16,17,18,25,26,42,43,44 The results for the machine learning models also temper the optimistic expectations for highly flexible models in some previous studies. In recent years, machine learning and intelligent optimization methods have increasingly been introduced into phytomedicine extraction-process research, often with high goodness of fit used as the main evidence of model effectiveness. The present results show that, at least for the evaluated M3–M6 models and the data structure of the main analysis, reductions in training-set error did not consistently translate into better LOOCV predictive performance, nor did they consistently support more concordant model-recommended optima. This contrast was particularly evident for GPR, whose high in-sample fitting performance was accompanied by poor cross-validation performance. This suggests that highly flexible models may more readily fit the observed experimental points in small-sample and sparse factor spaces, but do not necessarily provide stable cross-validated predictive ability.16,17,22,23,24,27,28,41,42 Therefore, this study does not dismiss the potential value of machine learning methods in complex extraction systems. Rather, it emphasizes that their application to small-sample process optimization should be supported by predictive validation, diagnostic checking, and evaluation of optimization-conclusion stability. The conclusions of this study should still be interpreted in relation to data structure. The complete M0–M6 comparison mainly applies to datasets that meet the requirements for unified regression modeling. For datasets with fewer experimental runs that cannot support full quadratic regression comparison, the pattern of model applicability may differ. In the sensitivity analysis, support vector regression (SVR) showed a relative advantage in the structurally excluded datasets, indicating that the present findings should not be generalized as “regression models outperform machine learning models.” Rather, they should be understood as follows: among the small-sample, low-dimensional, continuous-response datasets examined in this study that met the requirements for unified regression modeling, parsimonious regression models exhibited a more balanced overall profile. RegressCheck and the worked example further demonstrate that integrating fitting performance, cross-validation, diagnostics, and model-recommended optima can help identify potential overfitting risks and model-dependent optimization uncertainty. Limitations of the study This study has several limitations. First, the complete multi-model comparison was based only on datasets that met the requirements for unified regression modeling; therefore, the conclusions should not be unconditionally generalized to all phytomedicine extraction datasets. Second, the machine learning models included in this study were limited to M3–M6 and did not cover all possible approaches, such as random forest, gradient boosting, and neural networks. Third, this study was mainly a retrospective benchmark evaluation based on literature-extracted data, and the practical process value of the model-predicted optimal conditions still requires validation through further prospective confirmation experiments. Overall, this study shows that model selection in phytomedicine extraction-process optimization should not rely solely on goodness of fit, but should also consider predictive reliability, diagnostic adequacy, and optimization-conclusion stability. The model-predicted optimum should also be interpreted as a candidate operating region for subsequent confirmation experiments, rather than as a definitive production condition that requires no further validation. Resource availability Lead contact Further information and requests for resources should be directed to and will be fulfilled by the lead contact: Li Yu (yuli9119@126.com). Materials availability This study did not generate new unique reagents or physical materials. Data and code availability •The curated datasets and dataset-level summary files supporting the findings of this study are available at Mendeley data: https://doi.org/10.17632/nptwtyrc8w.1. The analysis scripts, modeling code, optimization procedures, result summary files, and archived reproducibility package are available at Zenodo: https://doi.org/10.5281/zenodo.20270109. The public code repository is also available at GitHub: https://github.com/yuan1204t-hub/Large-scale-assessment. The curated datasets and dataset-level summary files supporting the findings of this study are available at Mendeley data: https://doi.org/10.17632/nptwtyrc8w.1. The analysis scripts, modeling code, optimization procedures, result summary files, and archived reproducibility package are available at Zenodo: https://doi.org/10.5281/zenodo.20270109. The public code repository is also available at GitHub: https://github.com/yuan1204t-hub/Large-scale-assessment. •The archived reproducibility package includes the scripts and documentation required for data preprocessing, M0–M6 model fitting, leave-one-out cross-validation, model diagnostic analysis, model-based optimization, sensitivity analysis related to structural inclusion criteria, and stratified robustness analysis. Further information required to reanalyze the data reported in this article is available from the lead contact upon reasonable request. The archived reproducibility package includes the scripts and documentation required for data preprocessing, M0–M6 model fitting, leave-one-out cross-validation, model diagnostic analysis, model-based optimization, sensitivity analysis related to structural inclusion criteria, and stratified robustness analysis. Further information required to reanalyze the data reported in this article is available from the lead contact upon reasonable request. Acknowledgments This research was funded by the Zhejiang Provincial Natural Science Foundation (grant nos. LZYQ25H270001 and LY24H270007). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Author contributions Y.T. writing—original draft, formal analysis, software, data curation, and web-based tool development. Z.Y. writing—original draft, formal analysis, and methodology. T.C. formal analysis and writing—review and editing. W.J. conceptualization, model construction, and writing—review and editing. W.J. software design, methodology, funding acquisition, and writing—review and editing. L.Y. validation, funding acquisition, and writing—review and editing. All authors have read and approved the final manuscript and agree to be accountable for all aspects of the work. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataCurated phytomedicine extraction datasetsThis paper; Mendeley Datahttps://doi.org/10.17632/nptwtyrc8w.1Dataset-level information and summary tablesThis paper; Mendeley Datahttps://doi.org/10.17632/nptwtyrc8w.1Software and algorithmsAnalysis scripts and modeling codeThis paper; GitHubhttps://github.com/yuan1204t-hub/Large-scale-assessmentArchived code repositoryThis paper; Zenodohttps://doi.org/10.5281/zenodo.20270109RegressCheck analytical implementationThis paper; GitHub and Zenodohttps://github.com/yuan1204t-hub/Large-scale-assessment; https://doi.org/10.5281/zenodo.20270109PythonPython Software Foundationhttps://www.python.org/RR Foundationhttps://www.r-project.org/NumPyNumPy Developershttps://numpy.org/Pandaspandas Developershttps://pandas.pydata.org/SciPySciPy Developershttps://scipy.org/scikit-learnscikit-learn Developershttps://scikit-learn.org/Statsmodelsstatsmodels Developershttps://www.statsmodels.org/ggplot2R packagehttps://ggplot2.tidyverse.org/ComplexHeatmapR/Bioconductor packagehttps://bioconductor.org/packages/ComplexHeatmap/OtherRegressCheck web platformThis paperhttps://tyxone.online/ Experimental model and study participant details This study did not involve human participants, animals, human samples, cell lines, primary cultures, microorganisms, or newly collected plant specimens. All analyses were based exclusively on secondary numerical datasets extracted from published phytomedicine extraction-process studies. Therefore, ethical approval, informed consent, animal-use approval, and sex- or gender-based analyses were not applicable. Method details Literature search and screening Literature search and screening were organized with reference to systematic-review reporting principles.45 A systematic literature search was conducted in CNKI, Wanfang, VIP, SinoMed, PubMed, Springer Nature Link, ScienceDirect, and Scopus. The search covered studies published during the most recent five-year period up to March 27, 2026. The core search strategy was as follows: (“TCM” OR “Herbal” OR “Phytomedicine”) AND “extraction” AND (“Response surface” OR “Box-Behnken” OR “BBD” OR “CCD”). Equivalent Chinese search terms and database-specific topic, title, abstract, and free-text fields were used when applicable to improve retrieval coverage. No language restriction was applied. The initial search yielded 3,241 records. Duplicate records were removed primarily on the basis of DOI information. For records without available DOI information, potential duplicates were further checked by title, authors, publication year, and journal information when necessary. The remaining records were screened by title and abstract, followed by full-text assessment. Two reviewers independently screened the literature and cross-checked the results. Disagreements were resolved through further review and discussion. Forward and backward citation tracking was used when necessary to reduce the risk of missing relevant studies. Studies were included if they investigated phytomedicine extraction-process optimization, reported original experimental data or recoverable numerical values, used a multifactor experimental design suitable for subsequent modeling, and provided sufficient information on factor settings, factor levels, and response values. Studies were excluded if they lacked original data, had incomplete key information, did not involve mathematical modeling, or focused on mineral or animal medicines rather than phytomedicine extraction. Study datasets The unit of analysis was an extraction-process dataset comprising multiple experimental runs, process factors, and one continuous response variable. Systematic literature screening identified 704 eligible studies. After data extraction, integrity checking, and preprocessing, 1,399 datasets were available for structural assessment. Of these, 1,148 met the structural inclusion criterion and were included in the main M0–M6 model comparison, whereas 251 did not meet this criterion and were retained for a separate machine-learning-only sensitivity analysis. Thus, the main analytical dataset comprised 1,148 phytomedicine extraction datasets suitable for unified quadratic regression modeling. Data extraction and preprocessing For each eligible study, extraction-process data were extracted from tables, text, or other recoverable numerical information. Each dataset was organized in a standard tabular format, with process factors as predictors and the target extraction outcome as the response. For modeling, the first n−1 columns were treated as independent variables, and the final column as the response variable. When multiple responses were reported in a study, each recoverable response was treated as a separate single-response modeling dataset. Data were checked for completeness, numerical consistency, sample size, number of factors, factor-level structure, and experimental design type. Datasets with missing key factor or response values were excluded from model comparison. Experimental designs were classified as Box-Behnken design, central composite design, full factorial design, or other/undetermined according to the reported design structure and factor-level pattern. In the main analysis, 851 datasets were classified as Box-Behnken designs, 251 as central composite designs, 41 as full factorial designs, and 5 as other or undetermined because of insufficient information in the original reports. Structural eligibility To ensure that the full quadratic regression model was estimable, a structural inclusion criterion was applied prior to the main model comparison.6,7,8,34,35 Let s denote the effective sample size of a dataset and t denote the number of non-intercept candidate terms in the full quadratic model. The structural criterion was defined as follows: For a dataset with k extraction factors, the full quadratic candidate space comprised k linear terms, k squared terms, and k(k–1)/2 two-way interaction terms. Therefore: Datasets meeting this criterion were included in the main M0–M6 model comparison. Those that did not meet the criterion were not forced into the full regression comparison, but were analyzed separately in a machine-learning-only sensitivity analysis. Therefore, the full M0–M6 comparison should be interpreted as applying primarily to datasets with sufficient structural support for unified quadratic regression modeling. Candidate model framework Seven candidate models were compared under a unified analytical framework. M0–M2 were traditional regression models constructed within the same full quadratic candidate-term space.6,7,8,34,35 M0 was the full quadratic response surface model, retaining all linear, squared, and two-way interaction terms. M1 was selected by maximizing Rc2, whereas M2 was selected using a Mallows’ Cr-type criterion, as defined in the model summary table in the main text.31 M3–M6 were machine learning models selected for small-sample, low-dimensional, continuous-response extraction modeling.17,22,23,27,28,36 Specifically, M3 was quadratic Ridge regression based on the full quadratic feature matrix,46 M4 was support vector regression,47 M5 was partial least-squares regression,48 and M6 was Gaussian process regression.49 These models were included because they generate continuous prediction functions and can therefore be used for both response prediction and model-based searches for optimal extraction conditions. The machine learning models were included to assess whether flexible data-driven methods could improve prediction and optimization stability relative to traditional regression models in small-sample phytomedicine extraction datasets. They were not intended to represent the full spectrum of machine learning algorithms. Regression model fitting For M0–M2, regression models were constructed within the same second-order polynomial candidate-term space.6,7,8,34,35 For a dataset with p extraction factors, the general quadratic response surface model was expressed as follows: where y is the response variable, xi and xj are extraction-process factors, β0 is the intercept, βi, βii, and βij are the coefficients of the linear, quadratic, and interaction terms, respectively; and ε is the residual error. M0 was defined as the full quadratic response surface model and retained all linear, quadratic, and two-way interaction terms.6,7,8 M1 and M2 were obtained from the same candidate-term space through global subset enumeration. For M1, the optimal submodel was identified according to the corrected coefficient of determination: where Rc2 denotes the corrected coefficient of determination, n is the sample size, and k is the number of retained explanatory terms in the candidate submodel. The submodel with the highest Rc2 value was selected as M1. Furthermore, M2 was selected using a Mallows’ Cr-based criterion, with the optimal submodel defined as the one whose Cr was closest to r + 1, where r denotes the number of model terms excluding the intercept.31 Cr was calculated as follows: where SSE denotes the residual sum of squares of the candidate submodel, σˆ2 denotes the error variance estimated from the full quadratic model, and n denotes the sample size. In general, when Cr is close to r + 1, the model is considered to achieve a reasonable balance between goodness of fit and model complexity.31 Therefore, the absolute deviation between Cr and r + 1 was used as the optimization objective: The candidate submodel minimizing |Cr – (r + 1)| was defined as M2. Compared with M1, which emphasizes explanatory ability, M2 places greater emphasis on balancing model fit and structural parsimony. Machine learning fitting Considering the generally small sample sizes of the included datasets, four machine learning models suitable for small-sample continuous-response prediction were evaluated: quadratic Ridge regression, support vector regression, partial least-squares regression, and Gaussian process regression, denoted as M3, M4, M5, and M6, respectively.17,22,23,27,28,36,46,47,48,49 These models were included to assess whether flexible data-driven methods could improve prediction and optimization stability relative to traditional regression models in small-sample phytomedicine extraction datasets, rather than to represent the full spectrum of machine learning algorithms. For all machine learning models, hyperparameters were optimized using cross-validation combined with grid search.17,21,22,23,27,36 The parameter combination yielding the lowest cross-validation error was retained as the optimal setting. Feature construction, scaling, model fitting, and hyperparameter optimization were performed within each model-specific fitting procedure. For M3, the independent variables were first expanded by second-order polynomial transformation to generate linear, squared, and two-way interaction terms.6,7,8,46 The expanded feature matrix was then standardized before Ridge regression. The optimized hyperparameters included the penalty parameter α and the intercept option. Candidate α values were 0.001, 0.01, 0.1, 1, 10, 100, and 1000, and the intercept option was set to either True or False. For M4, input features were standardized before support vector regression fitting.47 Two kernels were evaluated: linear and radial basis function (RBF). For the linear kernel, the optimized hyperparameters were C and ε; for the RBF kernel, γ was additionally optimized. Candidate values were 0.1, 1, 10, 50, 100, and 500 for C; 0.001, 0.01, 0.05, 0.1, and 0.2 for ε; and scale, auto, 0.001, 0.01, and 0.1 for γ. For M5, independent variables were standardized before partial least-squares regression fitting.48 The main hyperparameter was the number of latent variables, n_components. The search range was determined dynamically according to the sample size and number of predictors in each dataset. The upper limit was set to the smaller of the number of independent variables and n – 1, and candidate values were searched sequentially from 1 to this upper limit. For M6, input features were standardized before Gaussian process regression fitting.49 The fmin_l_bfgs_b algorithm was used for kernel-parameter optimization, and n_restarts_optimizer was set to 2. Hyperparameter optimization focused on the kernel structure, the noise parameter α, and the normalize_y option. Candidate kernels included a constant kernel combined with a radial basis function kernel, a Matérn kernel, or a Rational Quadratic kernel, as well as corresponding extended forms further combined with a WhiteKernel. Candidate α values were 10−10, 10−6, 10−4, and 10−2, and normalize_y was set to either True or False. Cross-validation Leave-one-out cross-validation was used to assess predictive ability.21,22,23,27,42,43,44 For each dataset, one observation was held out at a time, and the model was trained on the remaining n−1 observations. The held-out observation was then predicted using the model fitted on the corresponding training subset. This procedure was repeated until each observation had been held out once. The resulting leave-one-out predictions were used to calculate Q2, RMSECV, and MAECV. For models requiring feature construction, scaling, or hyperparameter optimization, all corresponding procedures were repeated within each leave-one-out training fold.21,22,23,27 Specifically, polynomial feature construction, standardization, grid search, hyperparameter optimization, and model fitting were performed using only the training subset in each fold. The held-out sample was used only for prediction and was not involved in feature scaling, hyperparameter tuning, or model fitting. This nested implementation was used to reduce information leakage and provide a more rigorous estimate of predictive performance. Model diagnostics Model diagnostics were conducted to evaluate residual structure, sensitivity to abnormal observations, and error stability.24,29,32,33,34,35 For each fitted model, residuals were calculated as observed values minus fitted values. The Shapiro–Wilk test was used to assess residual normality, and the corresponding p value was denoted as pn.29 A one-sample t test was used to determine whether the mean residual significantly deviated from zero, and the resulting p value was denoted as pm. The maximum absolute residual was recorded to characterize the largest single-sample deviation, whereas error stability was summarized using the interquartile range of absolute errors. For traditional regression models, additional diagnostic analyses were performed. Cook’s distance was used to identify highly influential observations, and both the number and proportion of observations exceeding the empirical threshold of 4/n were recorded.32 The Breusch–Pagan test was used as a supplementary assessment of homoscedasticity.33 These regression-specific diagnostics were not applied to machine learning models because they rely on assumptions specific to parametric regression. Optimization search For each dataset and model, model-based optimal extraction conditions were identified through grid search.6,7,8,18,25 For each factor, the search space was defined by its observed lower and upper bounds in the original dataset and discretized into 100 evenly spaced grid points. Candidate factor combinations within the grid were evaluated using the fitted model, and the combination yielding the highest predicted response was defined as the model-predicted optimum. The same search space and grid resolution were applied across models for each dataset to ensure comparability. The predicted optimum was interpreted as a candidate operating region requiring experimental confirmation, rather than as a definitive production condition.6,7,8,15,19,20,25 Optimization stability Optimization-stability analysis was conducted to evaluate whether different models produced consistent optimal extraction conclusions.18,25 For each dataset, the model-predicted optimal responses and corresponding optimal factor combinations were compared across models. The coefficient of variation of the predicted optimal responses was calculated to quantify their relative dispersion across models. Values below 10%, between 10% and 30%, and above 30% were interpreted as indicating high consistency, moderate discrepancy, and large discrepancy, respectively. These thresholds were used only for relative comparisons among model-based optimization conclusions and were not intended to represent absolute industrial tolerance limits. To evaluate the dispersion of predicted optimal factor settings, the coefficient of variation was calculated for each factor across models and then averaged across factors. In addition, the widths of the 95% confidence interval and 95% prediction interval were calculated to describe uncertainty in the predicted optimal responses. These metrics were used to assess whether model differences propagated from statistical evaluation to optimization-related decision making. Sensitivity analyses Sensitivity analysis was performed on the 251 structurally excluded datasets that did not meet the inclusion criterion for full quadratic regression.18,22,23 Because these datasets could not support the complete M0–M6 comparison, only machine learning models were fitted and evaluated. This analysis was used to assess whether the main conclusions depended on the structural inclusion criterion and to clarify the applicability boundary of the full model-comparison framework. Stratified robustness analyses were conducted according to experimental design type, number of factors, and sample-size category.18,22,23,27,28 These descriptive analyses were used to determine whether the main patterns in predictive performance, diagnostic behavior, and model-dependent optimization uncertainty were consistent across dataset structures, rather than being driven by a specific design type, factor number, or sample-size group. RegressCheck platform RegressCheck was developed as a web-based platform for standardized regression model checking, candidate-model comparison, diagnostic evaluation, and visualization. Users upload Excel-formatted datasets in which independent variables are arranged before the response variable. The platform then fits candidate models, summarizes evaluation metrics, reports model-comparison results, and generates diagnostic visualizations. The platform includes modules for data import, results overview, candidate-model comparison, single-model inspection, and diagnostic plot analysis. It reports model equations, retained variables, fitting metrics, cross-validation metrics, regression coefficients, p values, and diagnostic plots, including predicted-versus-observed plots, residual plots, Q–Q plots, and coefficient-significance visualizations. The analytical logic implemented in RegressCheck is consistent with the local analysis scripts used for the empirical analyses reported in this paper. The full web deployment environment is not required to reproduce the model-comparison results. The corresponding analytical implementation is available through the public and archived code repositories listed in the key resources table. Quantification and statistical analysis Software implementation All analyses were implemented in Python and R. NumPy, pandas, SciPy, scikit-learn, and statsmodels were used for data preprocessing, candidate-model fitting, hyperparameter tuning, cross-validation, diagnostic analyses, optimization search, and statistical testing. ggplot2 and ComplexHeatmap were used for figure generation and visualization of model-comparison heatmaps. Fitting metrics For each dataset, let yi denote the observed value of the i-th sample, yˆi denote the model-predicted value, y¯ denote the sample mean of the response variable, and n denote the sample size. The coefficient of determination was calculated as: The root-mean-square error was calculated as The mean absolute error was calculated as Because R2 uses the response mean as the baseline prediction, it can be negative when the residual sum of squares exceeds the total sum of squares. For traditional regression models, the Akaike information criterion was additionally calculated as follows30: where RSS denotes the residual sum of squares and p denotes the number of estimated parameters, including the intercept. The overall model-significance p value and the maximum p value among retained regression coefficients were also recorded. Predictive metrics Let yˆi,-i denote the predicted value for the i-th sample obtained from the model trained after excluding that sample. The leave-one-out cross-validated coefficient of determination was calculated as follows42,43,44: The root-mean-square error of cross-validation was calculated as The mean absolute error of cross-validation was calculated as These metrics were used to evaluate whether apparent training-set performance translated into reliable prediction for unseen observations.15,16,17,21,22,23,26,27,28,41,42,43,44 Diagnostic metrics For each dataset-model combination, residuals were defined as: Residual normality was assessed using the Shapiro–Wilk test, and the resulting p value was denoted as pn. A one-sample t test was used to determine whether the mean residual significantly deviated from zero, and the corresponding p value was denoted as pm. The maximum absolute residual was calculated as: The interquartile range of absolute errors was calculated as where Q75 and Q25 denote the 75th and 25th percentiles, respectively. For traditional regression models, Cook’s distance was used to evaluate the influence of individual observations. Observations with Cook’s distance exceeding 4/n were counted as potentially influential, and both their number and proportion were recorded. The Breusch–Pagan test was used as a supplementary test of homoscedasticity, and its p value was recorded. Optimization metrics For a given dataset, let yˆm∗ denote the predicted optimal response produced by model m, and let K denote the number of models included in the comparison. For the three-model comparison, K = 3, corresponding to M0–M2. For the seven-model comparison, K = 7, corresponding to M0–M6. The coefficient of variation of the predicted optimal responses was calculated as: For optimal extraction conditions, let xjm∗ denote the optimal value of factor j predicted by model m. The coefficient of variation for factor j was calculated as: The average factor coefficient of variation was then calculated as where k denotes the number of extraction factors in the dataset. For each dataset, the width of the 95% confidence interval for the predicted optimal responses across models was calculated as: where s∗ denotes the standard deviation of the model-predicted optimal responses, and t0.975,K-1 denotes the 97.5th percentile of the t distribution with K−1 degrees of freedom. The width of the 95% prediction interval was calculated as The confidence interval was used to describe uncertainty around the mean predicted optimal response across models, whereas the prediction interval characterized the expected dispersion of model-specific optimal response predictions. Statistical comparison All statistical analyses were conducted under a unified paired-comparison framework. The primary sample size for the main analysis was n = 1,148, where n represents the number of independent extraction-process datasets satisfying the structural inclusion criterion. The sensitivity analysis included n = 251 structurally excluded datasets. Continuous evaluation metrics were summarized using medians and interquartile ranges because many metrics were skewed or sensitive to extreme values. Normality was assessed using the Shapiro–Wilk test. Because all candidate models were evaluated on the same datasets, between-model comparisons were treated as paired comparisons. Overall differences among candidate models were first evaluated using the Friedman test. When the Friedman test indicated a significant overall difference, pairwise comparisons were performed using the Wilcoxon signed-rank test. p values were adjusted using the Holm method to control the family-wise type I error rate.50 Statistical significance was defined using a two-sided α level of 0.05. Where statistical significance is indicated by asterisks in the figure legends, the significance coding is ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001 after Holm correction. To avoid relying solely on p values, Hodges–Lehmann paired difference estimates were calculated to quantify the magnitude of between-model differences.51 When applicable, 95% confidence intervals were reported for these estimates. For metrics with predefined ranges of practically negligible differences, two models were considered approximately equivalent in practical terms if the 95% confidence interval of the Hodges–Lehmann estimate fell entirely within the predefined negligible range. For comparisons between structurally included and excluded datasets, Mann–Whitney U tests were used to evaluate differences in dataset-level structural characteristics. Stratified robustness analyses were summarized descriptively rather than treated as independent confirmatory tests. All statistical details, including exact n values, definitions of center and dispersion, statistical tests, and significance coding, are reported in the figure legends, Results section, and supplementary tables. Normality test results for the main fitting, predictive, and diagnostic metrics are reported in Tables S1–S3. These tests were used to guide the choice of nonparametric paired comparisons and were not interpreted as primary outcome analyses. Visualization of pairwise comparisons Pairwise model-comparison results were visualized using triangular heatmaps. In each heatmap, the upper triangular region shows Holm-adjusted p values from paired Wilcoxon signed-rank tests, whereas the lower triangular region shows the absolute Hodges-Lehmann paired difference estimates (|HL|).50,51 Smaller adjusted p values indicate stronger statistical evidence for between-model differences, whereas larger |HL| values indicate greater difference magnitude. To facilitate visual interpretation of practical difference magnitude, triangle symbols were used to denote predefined |HL| intervals. Zero, one, two, three, and four triangles indicate |HL| < 0.01, 0.01 ≤ |HL| < 0.03, 0.03 ≤ |HL| < 0.05, 0.05 ≤ |HL| < 0.07, and |HL| ≥ 0.07, respectively. These thresholds were used only as descriptive visual aids for comparing the relative magnitude of between-model differences and were not treated as independent statistical tests. Additional resources RegressCheck web platform: https://tyxone.online/. RegressCheck is a web-based platform developed in this study for standardized regression model checking, candidate-model comparison, diagnostic evaluation, and visualization of extraction-process datasets. The platform supports Excel-format data upload and provides model-fitting results, cross-validation metrics, regression diagnostics, and graphical outputs to assist users in evaluating model reliability and optimization-related decisions. Public code repository: https://github.com/yuan1204t-hub/Large-scale-assessment. Archived code repository: https://doi.org/10.5281/zenodo.20270109. Curated datasets and summary files: https://doi.org/10.17632/nptwtyrc8w.1. Published: July 22, 2026 Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.isci.2026.116883. Contributor Information Weifeng Jin, Email: jin_weifeng@126.com. Li Yu, Email: yuli9119@126.com. Supplemental information Document S1. Figures S1–S3 and Tables S1–S6, and S8 Table S7. Stratified robustness analyses according to experimental design type, effective sample size, and number of factors, related to Figures 2–5 References Associated Data Supplementary Materials Document S1. Figures S1–S3 and Tables S1–S6, and S8 Table S7. Stratified robustness analyses according to experimental design type, effective sample size, and number of factors, related to Figures 2–5 Data Availability Statement •The curated datasets and dataset-level summary files supporting the findings of this study are available at Mendeley data: https://doi.org/10.17632/nptwtyrc8w.1. The analysis scripts, modeling code, optimization procedures, result summary files, and archived reproducibility package are available at Zenodo: https://doi.org/10.5281/zenodo.20270109. The public code repository is also available at GitHub: https://github.com/yuan1204t-hub/Large-scale-assessment. The curated datasets and dataset-level summary files supporting the findings of this study are available at Mendeley data: https://doi.org/10.17632/nptwtyrc8w.1. The analysis scripts, modeling code, optimization procedures, result summary files, and archived reproducibility package are available at Zenodo: https://doi.org/10.5281/zenodo.20270109. The public code repository is also available at GitHub: https://github.com/yuan1204t-hub/Large-scale-assessment. •The archived reproducibility package includes the scripts and documentation required for data preprocessing, M0–M6 model fitting, leave-one-out cross-validation, model diagnostic analysis, model-based optimization, sensitivity analysis related to structural inclusion criteria, and stratified robustness analysis. Further information required to reanalyze the data reported in this article is available from the lead contact upon reasonable request. The archived reproducibility package includes the scripts and documentation required for data preprocessing, M0–M6 model fitting, leave-one-out cross-validation, model diagnostic analysis, model-based optimization, sensitivity analysis related to structural inclusion criteria, and stratified robustness analysis. Further information required to reanalyze the data reported in this article is available from the lead contact upon reasonable request.