Predicting Dairy Milk Intake Recommendations Using Machine Learning Models: A Comparative Analysis

Predicting Dairy Milk Intake Recommendations Using Machine Learning Models: A Comparative Analysis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Predicting Dairy Milk Intake Recommendations Using Machine Learning Models: A Comparative Analysis Mian Muhammad Hamza, Ruhma Shahbaz This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8927461/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Background: Dairy milk occupies a central position in human nutrition, supplying indispensable micronutrients — including calcium, vitamins D and B12, and high-quality proteins — that sustain bone integrity, cardiovascular function, and metabolic homeostasis. Despite its nutritional value, the optimal level of dairy milk consumption remains subject to ongoing debate, owing to inter-individual variability in lactase persistence, cardiometabolic risk profiles, and the heterogeneous health effects of different dairy subtypes. Clinical guidelines have yet to converge on universally applicable intake thresholds, underscoring the need for personalized, data-driven approaches to dietary counselling. Methods: A publicly available synthetic dataset comprising 10,000 records and 17 clinico-demographic features was employed. Nominal categorical variables were encoded via one-hot encoding to preclude the imposition of spurious ordinal relationships/ Five supervised classification algorithms were trained and rigorously evaluated: Logistic Regression, Random Forest, Light-GBM, K-Nearest Neighbors (KNN), and a soft-voting Voting Classifier ensemble. Model assessment relied upon stratified 10-fold cross-validation as the primary evaluation paradigm, supplemented by a held-out 20% test set. Performance was quantified across accuracy, macro-averaged precision, recall, F1-score, one-vs-rest ROC-AUC, and per-class confusion matrices. SHAP (SHapley Additive exPlanations) analysis was subsequently applied to the two best-performing models to identify the most clinically influential predictive features. Results: Under stratified 10-fold cross-validation, Random Forest and the Voting Classifier attained mean accuracies of 0.982 ± 0.006 and 0.981 ± 0.007, respectively, while Light-GBM reached 0.978 ± 0.008. Logistic Regression and KNN achieved 0.964 ± 0.010 and 0.948 ± 0.012, respectively; all models substantially outperformed the majority-class dummy baseline (accuracy = 0.382). Perfect held-out test scores observed for Random Forest and the Voting Classifier are attributable to the deterministic, rule-based structure of the synthetic target variable and do not reflect real-world generalizability. SHAP analysis identified Age, Calcium Level, and Lactose Intolerance as the three most influential predictors, consistent with established clinical determinants of dairy intake guidance. Conclusion: Machine learning classifiers demonstrated strong predictive performance for physician-recommended dairy milk intake adjustments within a controlled synthetic setting. SHAP-based interpretability enhanced transparency by elucidating the clinical basis of model predictions. Nonetheless, the synthetic origin of the dataset constitutes a fundamental limitation; prospective validation on real-world clinical cohorts is a prerequisite before these models could be considered for integration into dietary decision-support systems. Machine Learning Dairy Milk Intake Physician Recommendation Prediction SHAP (SHapley Additive exPlanations) Random Forest Light-GBM Cross-Validation Clinical Decision Support Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Dairy milk and dairy products have constituted a cornerstone of human dietary practice across millennia, furnishing a concentrated array of essential nutrients — proteins, calcium, phosphorus, riboflavin, and vitamins D and B12 — that collectively sustain bone mineralization, neuromuscular function, and systemic metabolic processes [1,2]. Despite this established nutritional profile, the epidemiological evidence linking dairy consumption to long-term health outcomes presents a considerably more nuanced picture than historical dietary guidelines have acknowledged. Prospective studies and meta-analyses report that moderate dairy milk intake is inversely associated with the risk of stroke and certain manifestations of cardiovascular disease (CVD); however, the magnitude and direction of this association vary markedly according to dairy subtype, fat content, and the specific cardiovascular endpoint under investigation [3,4]. Total dairy consumption — encompassing both liquid milk and fermented products such as yogurt — has been correlated with a modestly reduced risk of overall CVD, while high-fat dairy products have been implicated in adverse cardiometabolic outcomes in selected cohorts [3,5]. Similarly, low-fat dairy variants have demonstrated favorable associations with blood pressure regulation, whereas full-fat products may confer neutral or marginally deleterious effects on cardiovascular mortality in certain populations [4]. In parallel, the precise roles of dairy milk in modulating the risk of type 2 diabetes mellitus and hormonally sensitive cancers remain subjects of active scientific inquiry, with published findings indicating inconsistent and population-specific associations [6]. These heterogeneous outcomes underscore the inadequacy of uniform dietary recommendations and highlight an urgent clinical need for individualized guidance grounded in patient-specific clinical and lifestyle profiles. The convergence of high-dimensional electronic health records, clinical biomarker data, and advanced machine learning (ML) methodologies offers a promising avenue for meeting this need [4,5]. Supervised classification algorithms, particularly ensemble-based methods such as Random Forest and gradient-boosted frameworks, have demonstrated robust performance in predicting cardiometabolic risk and treatment outcomes from structured clinical data, frequently surpassing traditional statistical models in both accuracy and generalizability [5,13]. In the domain of ML interpretability, SHAP (SHapley Additive exPlanations) has emerged as the de facto standard for attributing feature-level contributions in complex, non-linear models [13]. Applications of SHAP to clinical prediction tools have been shown to enhance clinician trust, facilitate regulatory scrutiny, and support the responsible deployment of AI-driven decision-support systems. The utility of SHAP is particularly pronounced in the context of tree-based ensembles, where non-linear feature interactions and higher-order dependencies cannot be adequately characterised through coefficient-based approaches alone. Stratified k-fold cross-validation represents the methodological standard for evaluating ML classifiers on datasets of moderate size, providing substantially more stable and conservative performance estimates than single train-test partitions, particularly in the presence of class imbalance [12]. One-hot encoding is the established preprocessing strategy for nominal categorical variables — such as country of origin and milk type — as it preserves their nominal character without introducing artificial ordinal constraints that may systematically bias distance-based or linear classification algorithms [8]. The present study addresses the aforementioned gap by constructing an ML-based classification framework for the prediction of physician-recommended dairy milk intake adjustments. In contrast to prior work focused on agricultural yield forecasting or pathological disease detection, the clinical focus of this study — determining whether a clinician would advise a patient to maintain, increase, or reduce dairy milk consumption — positions it squarely within the domain of personalized nutritional decision support. This focus is informed by recent umbrella reviews that emphasize the heterogeneity of dairy-specific health effects across subtypes, age groups, and geographic populations [3,11]. 1.1 Study Objectives (i) To develop and comparatively evaluate ML classifiers — encompassing Random Forest, Logistic Regression, Light-GBM, KNN, and a Voting Classifier ensemble — for predicting physician-recommended dairy milk intake adjustments. (ii) To apply stratified k-fold cross-validation as the primary evaluation framework, thereby ensuring reliable and unbiased performance estimates. (iii) To employ SHAP analysis on the two best-performing models to identify and clinically interpret the most influential predictive features. (iv) To critically assess the limitations inherent to synthetic data and to outline a prospective validation roadmap using real-world clinical datasets. 1.2 Research Questions “How can be the multiple ML models be developed for prediction of possible milk effect on humans, belongs to multiple regions.?” “How can SHAP interpretability be integrated with predictive methodology in other to ensure practical clinical interpretabilities.?” 2. Materials and Methods 2.1 Dataset Characteristics and Synthetic Nature The dataset employed in this study is publicly accessible via the Kaggle repository (https://www.kaggle.com/datasets/hammadansari7/milks-effect-on-human-health) and comprises 10,000 synthetic records [1]. The dataset was computationally generated to simulate a diverse clinical cohort by sampling feature values from realistic marginal distributions. The target variable (Doctor_Recommendation) was assigned through a rule-based decision logic intended to approximate clinical guidelines governing dairy intake recommendations. The dataset encompasses 17 features spanning demographic, clinical, and lifestyle domains. Table 1 provides a comprehensive summary of all features, their measurement scales, value ranges, and descriptive statistics. Numerical features include: Age (18–80 years; mean = 47.3, SD = 17.1), BMI (kg/m²; mean = 26.4, SD = 4.8), Daily_Milk_Intake_ml (mean = 310 mL, SD = 145), Cholesterol_mg_dl (mean = 195 mg/dL, SD = 38), Calcium_Level_mg_dl (mean = 9.4 mg/dL, SD = 0.9), and Physical_Activity_Level (hours/week; mean = 3.2, SD = 2.1). Categorical features encompass Gender, Country (10 categories), Milk_Type (Whole, Low-fat, Skim, Plant-based), Lactose_Intolerance, Smoking, and Diabetes. It is explicitly acknowledged that synthetic datasets generated via statistical sampling and rule-based label assignment cannot fully replicate the distributional complexity, measurement noise, confounding structure, and inter-individual variability inherent in real-world clinical data. Accordingly, all reported performance metrics should be interpreted as upper-bound estimates that are unlikely to be sustained under external real-world validation. Table 1. Summary of Dataset Features, Measurement Scales, and Descriptive Statistics. Feature Type Scale Range / Categories Descriptive Statistics Age Numerical Continuous 18–80 years Mean = 47.3; SD = 17.1 BMI Numerical Continuous Kg/m² Mean = 26.4; SD = 4.8 Daily_Milk_Intake_ml Numerical Continuous mL/day Mean = 310; SD = 145 Cholesterol_mg_dl Numerical Continuous mg/dL Mean = 195; SD = 38 Calcium_Level_mg_dl Numerical Continuous mg/dL Mean = 9.4; SD = 0.9 Physical_Activity_Level Numerical Continuous Hours/week Mean = 3.2; SD = 2.1 Gender Categorical Nominal Male / Female Binary Country Categorical Nominal 10 categories Multi-class Milk_Type Categorical Nominal Whole / Low-fat / Skim / Plant-based Multi-class Lactose_Intolerance Categorical Binary Yes / No Binary Smoking Categorical Binary Yes / No Binary Diabetes Categorical Binary Yes / No Binary Age_Group (engineered) Categorical Ordinal Child / Young Adult / Middle-Aged / Senior Derived feature Doctor_Recommendation (target) Categorical Multi-class Maintain / Increase / Reduce Three-class outcome Note. SD = standard deviation. Descriptive statistics are reported for the full 10,000-record synthetic dataset. The Age_Group variable is a derived feature created during preprocessing. Doctor_Recommendation constitutes the three-class target variable (Maintain / Increase / Reduce). 2.2 Preprocessing Pipeline (i) Missing Value Handling: A systematic null-value audit confirmed the absence of missing values across all 10,000 records and 17 features; accordingly, no imputation procedures were required. (ii) Categorical Encoding: All nominal categorical variables — Gender, Country, Milk_Type, Lactose_Intolerance, Smoking, and Diabetes — were transformed using one-hot encoding (OHE). OHE was selected in preference to label encoding to preclude the introduction of spurious ordinal relationships that can systematically bias the learning dynamics of distance-based and linear classification algorithms [8]. (iii) Feature Engineering: A clinically meaningful Age_Group feature was derived by categorizing participants into four age strata: Child (0–17 years), Young Adult (18–34 years), Middle-Aged Adult (35–64 years), and Senior (65+ years). This binning strategy corrects the previously incomplete age segmentation reported in an earlier iteration of this analysis, which omitted the 35–64-year range — the largest clinical demographic subgroup. (iv) Feature Scaling: Standardization (zero mean, unit variance) was applied to all continuous numerical features — BMI, Daily_Milk_Intake_ml, Cholesterol_mg_dl, Calcium_Level_mg_dl, and Physical_Activity_Level — prior to model training. Scaling is critical for distance-sensitive algorithms (KNN, Logistic Regression) and was applied exclusively within each training fold of the cross-validation procedure to prevent data leakage. (v) Outlier Management: Outliers in continuous features were identified using the interquartile range (IQR) criterion and univariate histograms. Feature values exceeding three standard deviations from the respective feature mean were winsorised to their boundary values, mitigating the distorting influence of extreme observations without discarding valid records. 2.3 Evaluation Strategy Stratified 10-fold cross-validation (K = 10) was employed as the primary model evaluation approach. Stratification ensures that the class distribution of the full dataset is preserved within each fold, yielding a more robust and representative estimate of out-of-sample performance relative to a single train-test partition [12]. To prevent data leakage, all preprocessing steps — including feature scaling, OHE, and age group binning — were applied within the training portion of each fold prior to validation-set scoring. A complementary held-out 20% test set was reserved and preprocessed independently to report final point-estimate metrics. Performance was assessed using: accuracy, macro-averaged precision, macro-averaged recall, macro-averaged F1-score, one-vs-rest ROC-AUC, and full per-class confusion matrices for each model. A majority-class dummy classifier was included as a lower-bound reference, returning an accuracy of 0.382, against which all ML classifiers are benchmarked. 2.4 Model Architectures and Hyperparameter Configuration Five supervised classification algorithms were implemented and evaluated. Table 2 presents a concise summary of each model, its key hyperparameters, and its principal methodological advantage. Hyperparameters for Random Forest, Light GBM, and KNN were determined through five-fold cross-validated grid search on the training data as show in Figure 2; Logistic Regression employed default regularization. All models were implemented using the scikit-learn and Light GBM Python libraries. Table 2. Summary of Model Architectures, Hyperparameter Configurations, and Methodological Advantages. Model Description Key Hyperparameters Primary Advantage Logistic Regression Linear classifier using the softmax function for multi-class probability estimation C = 1.0; L2 regularisation; solver = lbfgs High interpretability; serves as a linear baseline Random Forest Bagging ensemble of decision trees; final prediction by majority vote / probability averaging n_estimators = 200; max_depth = 15; min_samples_split = 5; class_weight = balanced Robust to overfitting; handles non-linear feature interactions Light-GBM Gradient boosting framework using histogram-based leaf-wise tree growth for efficient training on tabular data learning_rate = 0.05; num_leaves = 50; n_estimators = 300; subsample = 0.8 Superior speed on large structured datasets; captures complex interactions K-Nearest Neighbors (KNN) Non-parametric classifier based on majority vote among k nearest neighbours in feature space k = 7; determined via grid search cross-validation; distance metric = Euclidean Simple; non-parametric; no training phase required Voting Classifier Soft-voting ensemble combining predicted class probabilities from LR, RF, Light-GBM, and KNN Soft voting; equal weights; constituent models as above Reduces individual model errors; leverages complementary strengths Note. LR = Logistic Regression; RF = Random Forest; LGBM = Light-GBM; KNN = K-Nearest Neighbors; OvR = one-vs-rest. Hyperparameters for RF, LGBM, and KNN were selected via five-fold cross-validated grid search. 2.5 SHAP Explainability Analysis SHAP (SHapley Additive exPlanations) analysis was applied to the two best-performing models — Random Forest and Light-GBM — to provide post-hoc, feature-level interpretability [13]. The Tree-Explainer algorithm was employed for efficient SHAP value computation, given the tree-based architecture of both models. SHAP summary plots and mean absolute SHAP value rankings were generated to identify the features exerting the greatest influence on model predictions across all three recommendation classes. This approach addresses a recognized limitation of the earlier iteration of this study, which had applied SHAP exclusively to Logistic Regression — a model whose coefficient-based interpretation is already transparent by design — thereby omitting interpretability analysis for the substantively more informative tree-based classifiers. 3. Results and Discussion 3.1 Cross-Validation Performance Table 3 reports mean performance metrics and their standard deviations across the 10 stratified folds. All five ML classifiers substantially and consistently outperformed the majority-class dummy baseline (accuracy = 0.382), confirming that each model captures genuine predictive signal rather than reflecting trivial majority-class assignment. Random Forest achieved the highest mean cross-validation accuracy (0.982 ± 0.006), with the Voting Classifier marginally below at 0.981 ± 0.007. Light-GBM performed comparably at 0.978 ± 0.008. Logistic Regression attained 0.964 ± 0.010, reflecting the strong linear separability of the synthetic feature space, while KNN returned the lowest cross-validation accuracy (0.948 ± 0.012), consistent with the elevated dimensionality induced by one-hot encoding. The narrow standard deviation ranges across all models indicate consistent, stable performance across folds, with no evidence of excessive fold-to-fold variance. Table 3. Stratified 10-Fold Cross-Validation Performance (Mean ± Standard Deviation). Model Accuracy Precision (Macro) Recall (Macro) F1-Score (Macro) ROC-AUC (OvR) Dummy Baseline 0.382 0.127 0.333 0.187 0.500 Random Forest 0.982 ± 0.006 0.981 ± 0.007 0.982 ± 0.006 0.981 ± 0.006 0.998 ± 0.002 Voting Classifier 0.981 ± 0.007 0.980 ± 0.008 0.981 ± 0.007 0.980 ± 0.007 0.997 ± 0.002 Light-GBM 0.978 ± 0.008 0.977 ± 0.009 0.978 ± 0.008 0.977 ± 0.008 0.997 ± 0.002 Logistic Regression 0.964 ± 0.010 0.963 ± 0.011 0.964 ± 0.010 0.963 ± 0.010 0.994 ± 0.003 KNN (k = 7) 0.948 ± 0.012 0.947 ± 0.013 0.948 ± 0.012 0.947 ± 0.013 0.981 ± 0.005 Note. OvR = one-vs-rest multiclass formulation. Precision, recall, and F1-score are reported as macro-averages. The Dummy Baseline reflects a majority-class classifier and does not compute standard deviations. Cross-validation results represent the primary generalizability estimate; held-out test results are presented separately in Table 4. 3.2 Held-Out Test Set Performance and Confusion Matrices Table 4 presents model performance on the held-out 20% test partition. Random Forest and the Voting Classifier each attained perfect scores (accuracy = 1.000) across all metrics. These values are attributable to the deterministic, rule-based structure of the synthetic target variable, in which class labels are a direct function of input features without stochastic noise. This behaviour is both expected and structurally inevitable in rule-generated synthetic data; it is explicitly not indicative of overfitting or data leakage, as all preprocessing transformations were applied within cross-validation folds and no feature selection was conducted on the full dataset. These scores should not be extrapolated to real-world clinical settings. The cross-validation results presented in Table 3 and Figure 3 constitute the more informative and conservative generalizability estimate. Table 4. Held-Out Test Set Performance (20% Partition). Model Accuracy Precision (Macro) Recall (Macro) F1-Score (Macro) ROC-AUC (OvR) Random Forest 1.0000* 1.0000* 1.0000* 1.0000* 1.0000* Voting Classifier 1.0000* 1.0000* 1.0000* 1.0000* 1.0000* Logistic Regression 0.9970 0.9950 0.9970 0.9960 0.9999 Light-GBM 0.9960 0.9964 0.9960 0.9962 1.0000* KNN (k = 7) 0.9765 0.9744 0.9765 0.9753 0.9931 Note. * Perfect scores (1.000) reflect the deterministic, rule-based labelling of the synthetic dataset and are not indicative of overfitting or data leakage. Cross-validation results (Table 3) provide the appropriate generalization estimate. OvR = one-vs-rest. 3.3 Qualitative Model Analysis Random Forest: The highest-performing classifier across both cross-validation and held-out evaluation. Ensemble aggregation of 200 decision trees reduces prediction variance and accommodates complex non-linear feature interactions, including higher-order dependencies that are beyond the representational capacity of linear models. The balanced class-weight parameter further ensures appropriate sensitivity to the minority recommendation class. Voting Classifier: Marginally below Random Forest in cross-validation, the soft-voting ensemble combines the complementary inductive biases of its four constituent classifiers — Logistic Regression, Random Forest, Light-GBM, and KNN. Probability averaging across constituent models moderates individual model errors and provides robust performance even when individual classifiers exhibit localized weaknesses. Light-GBM: Achieved cross-validation accuracy comparable to Random Forest with substantially reduced training time, attributable to its histogram-based gradient boosting algorithm and leaf-wise tree growth strategy. Light-GBM’s native handling of structured tabular data and efficient treatment of categorical features makes it particularly well-suited to the dataset characteristics employed in this study. Logistic Regression: Performed competitively (cross-validation accuracy = 0.964), a result consistent with the strong linear separability of the synthetic feature space. Its relatively lower performance compared with tree-based ensembles reflects the inherent inability of a linear boundary to capture non-linear interaction terms. Nevertheless, its full interpretability via model coefficients renders it a valuable clinical communication tool. K-Nearest Neighbors (KNN): The lowest-performing classifier, consistent with anticipated degradation under the elevated feature dimensionality introduced by one-hot encoding. The curse of dimensionality attenuates meaningful distance distinctions in high-dimensional spaces, reducing the discriminative power of majority-vote neighborhood assignment despite cross-validation optimization of k. 3.4 SHAP Explainability Analysis SHAP analysis applied to Random Forest and Light-GBM revealed a highly consistent feature importance ranking across both models, with Age, Calcium_Level_mg_dl, and Lactose_Intolerance emerging as the three most influential predictors. Table 5 presents the SHAP-derived feature ranking alongside clinical interpretations; Figure 3 depicts the SHAP summary plot for the Random Forest classifier. Higher Age values are strongly associated with a Reduce recommendation, consistent with the progressive decline in lactase activity and reduced gastrointestinal calcium absorption efficiency observed in older adults. Elevated serum calcium reduces the model’s predicted probability of an Increase recommendation, reflecting clinician caution regarding hypercalcaemia risk in the context of supplementary dairy intake. Lactose Intolerance constitutes the third most influential predictor, with its presence strongly shifting predictions towards a Reduce classification — an association well-grounded in established dietary management protocols Secondary contributors include BMI, Daily_Milk_Intake_ml, Cholesterol_mg_dl, and Diabetes status, each exerting directional SHAP effects consistent with accepted cardiometabolic dietary guidance as show in Figure 4. Table 5. SHAP Feature Importance Ranking for Random Forest: Mean Absolute SHAP Values and Clinical Interpretations. Rank Feature Mean |SHAP| (RF) Clinical Interpretation 1 Age 0.412 Older adults show higher likelihood of a Reduce recommendation, reflecting declining lactase activity and altered calcium metabolism 2 Calcium_Level_mg_dl 0.338 Elevated serum calcium reduces the probability of an Increase recommendation; supports clinical guidance on avoiding hypercalcaemia 3 Lactose_Intolerance 0.271 Strong predictor of a Reduce recommendation; consistent with clinical avoidance of lactose-containing dairy in intolerant individuals 4 BMI 0.184 Higher BMI values contribute to Reduce recommendations; reflects cardiometabolic risk considerations in dietary counselling 5 Daily_Milk_Intake_ml 0.156 Baseline intake level modulates recommendation direction; very low current intake increases the probability of an Increase recommendation 6 Cholesterol_mg_dl 0.127 Elevated cholesterol shifts predictions towards Reduce; aligned with guidance on limiting saturated fat from full-fat dairy 7 Diabetes 0.098 Presence of diabetes marginally increases Reduce probability; consistent with glycaemic management recommendations Note. Mean |SHAP| values represent the average absolute SHAP contribution across all test instances and all three recommendation classes. Rankings were consistent between Random Forest and Light-GBM; RF values are presented as the primary reference given its superior cross-validation accuracy. 3.5 Sensitivity Analysis: Class Imbalance Although the dataset approximates class balance across the three recommendation categories, SMOTE (Synthetic Minority Over-sampling Technique) was applied within the training folds of the cross-validation procedure to assess sensitivity to class distribution perturbations. Performance improvements relative to the un-resampled baseline were marginal (< 0.3% in macro-averaged F1-score), confirming that class imbalance does not constitute a primary source of bias in the current dataset. However, real-world clinical datasets frequently exhibit pronounced class imbalance in recommendation outcomes; in such contexts, SMOTE application within each training fold — rather than prior to the cross-validation split — is recommended as standard practice to prevent label leakage. 3.6 Limitations This study is subject to several fundamental limitations that must be considered in the interpretation of the reported findings. (i) Synthetic Dataset: The primary and most consequential limitation of this study is its exclusive reliance on a computationally generated synthetic dataset. Although the dataset was designed to simulate realistic clinical distributions, it cannot replicate the distributional heterogeneity, measurement noise, confounding factor structure, missing data patterns, and population-specific variation inherent in real clinical data. All reported performance metrics should be interpreted as illustrative upper bounds that are unlikely to generalize to authentic clinical settings without material degradation. (ii) Target Variable Construction: The study predicts physician-generated dietary recommendations rather than directly measured health outcomes. Critically, the recommendation labels in the dataset were assigned through deterministic rule-based logic rather than extracted from recorded clinical decisions. This construction limits the ecological validity of the predictive task and may not reflect the nuanced, contextual judgement of practice clinicians. (iii) External Validity: The trained models have not been evaluated on external real-world datasets. Their generalizability across diverse clinical populations, healthcare systems, and geographic contexts remains entirely unestablished. The perfect held-out scores observed for Random Forest and the Voting Classifier are direct artefacts of the synthetic data generation mechanism and should not be interpreted as evidence of deployable clinical utility. (iv) Feature Space Completeness: The dataset omits potentially important clinical determinants of dairy intake recommendations, including gut microbiome composition, genetic markers for lactase persistence (notably the LCT gene variants), longitudinal dietary recall data, renal function indices, and medication history. The absence of these variables may limit both the predictive completeness of the framework and the clinical face validity of its outputs. 3.7 Future Research Directions and Recommendations (i) Prospective Real-World Validation: Future work should prioritize the validation of these models using prospectively collected real-world clinical datasets sourced from multi-site healthcare facilities, ideally spanning multiple countries and demographic strata, to rigorously assess generalizability and practical clinical utility. (ii) Extended Feature Engineering: Incorporating additional biologically and clinically relevant variables — including genetic markers for lactase persistence, gut microbiome indices, longitudinal dietary records, and renal function parameters — would likely improve both the predictive accuracy and the mechanistic interpretability of future iterations of the framework. (iii) Systematic Hyperparameter Optimization: Bayesian hyperparameter optimization frameworks — such as Optuna or Hyperopt — applied to Light-GBM and KNN could yield incremental performance improvements and provide a more principled search across the hyperparameter space than grid-based approaches. (iv) Clinical Integration and Co-Design: Collaborative development with registered dietitians, clinical nutritionists, and practising physicians is strongly recommended to align model outputs with current evidence-based dietary guidelines, ensure clinical face validity, and facilitate responsible integration into nutritional decision-support workflows. 4. Conclusion This study designed, implemented, and comparatively evaluated a suite of machine learning classifiers for predicting physician-recommended dairy milk intake adjustments from clinical and lifestyle features within a controlled synthetic data environment. Random Forest and the Voting Classifier demonstrated the strongest generalization performance under stratified 10-fold cross-validation, attaining mean accuracies of 0.982 and 0.981, respectively. Light-GBM performed comparably at 0.978, while Logistic Regression and KNN provided informative competitive baselines at 0.964 and 0.948. The perfect held-out test scores recorded for Random Forest and the Voting Classifier are explicitly attributable to the deterministic, rule-based structure of the synthetic target variable and are not indicative of overfitting, data leakage, or deployable clinical accuracy. Cross-validation results constitute the appropriate and more conservative generalizability estimate. SHAP analysis applied to the two best-performing models identified Age, Calcium Level, and Lactose Intolerance as the most clinically influential predictors — a ranking consistent with established mechanisms governing dairy intake guidance in clinical nutrition practice. The corrected one-hot encoding pipeline, complete age group binning, stratified k-fold evaluation strategy, extended per-class confusion matrix diagnostics, and SHAP analysis applied to tree-based models collectively address the methodological concerns identified through peer review. Notwithstanding these contributions, the synthetic origin of the dataset represents a fundamental constraint on the conclusions that may be drawn. Future work must prioritize prospective validation on real-world clinical datasets to establish the practical utility and clinical transferability of this framework before any consideration of integration into dietary decision-support systems. Declarations Conflict of Interest: The authors declare no conflict of interest. Author Contributions: Conceptualization: M.M.H. and R.S.; Literature Review: M.M.H. and R.S.; Methodology: M.M.H.; Figures and Table Design: M.M.H.; Experiments: M.M.H.; Writing — Original Draft: M.M.H.; Writing — Review and Editing: R.S. Data Availability: Data are publicly available at: https://www.kaggle.com/datasets/hammadansari7/milks-effect-on-human-health Funding: No funding was received for this research. Ethical Considerations: The dataset is publicly available and synthetic in nature, requiring no institutional ethical approval for research use. References Kaggle Dataset: Milk’s Effect on Human Health. Available at: https://www.kaggle.com/datasets/hammadansari7/milks-effect-on-human-health/data Zhang, X., Chen, X., Xu, Y., Yang, J., Du, L., Li, K., & Zhou, Y. (2021). Milk consumption and multiple health outcomes: Umbrella review of systematic reviews and meta-analyses in humans. Nutrition & Metabolism, 18(7). https://doi.org/10.1186/s12986-020-00527-y Sharifan, P., Roustaee, R., Shafiee, M., Longworth, Z.L., Keshavarz, P., Davies, I.G., Webb, R.J., Mazidi, M., & Vatanparast, H. (2025). Dairy milk consumption and risk of cardiovascular and bone health outcomes in adults: An umbrella review and updated meta-analyses. Nutrients, 17, 2723. https://doi.org/10.3390/nu17172723 Lamarche, B., et al. (2025). The effects of dairy milk consumption on cardiovascular and stroke risk. The American Journal of Clinical Nutrition, 121(5), 956–964. https://doi.org/10.1093/ajcn/nqz276 Poppitt, S.D. (2020). Cow’s milk and dairy milk consumption: Is there now consensus for cardiometabolic health? Frontiers in Nutrition, 7, 574725. https://doi.org/10.3389/fnut.2020.574725 Sanjulian, L., et al. (2025). The role of dairy milk in human nutrition: Myths and realities. Nutrients, 17, 646. https://doi.org/10.3390/nu17040646 Amin, T., et al. (2025). Clinical evidence and mechanistic pathways of human milk oligosaccharide supplementation for health benefits: An updated review. Frontiers in Nutrition, 12, 1599678. https://doi.org/10.3389/fnut.2025.1599678 Shen, S., Song, X., Ding, H., Cui, X., Xie, Z., & Huang, H. (2026). Data-driven soft sensing for raw milk ethanol stability prediction. Sensors, 26(3), 903. https://doi.org/10.3390/s26030903 Loforte, Y., Zanzan, M., de Almeida, A., Cadavez, V., & Gonzales-Barron, U. (2026). Growth of Listeria monocytogenes in goat’s pasteurised milk cheese during maturation. Applied Microbiology, 6, 16. https://doi.org/10.3390/applmicrobiol6010016 Ma, S., et al. (2026). Fast and accurate prediction of milk yield in dairy goats using deep learning. Journal of Dairy Science. https://doi.org/10.3168/jds.2025-27905 Villoz, F., et al. (2024). Dairy intake and risk of cognitive decline and dementia: A systematic review and dose-response meta-analysis. Advances in Nutrition, 15, 100160. https://doi.org/10.1016/j.advnut.2023.100160 Molle, A., et al. (2026). Targeted iPLS for the prediction of cheese-making traits from individual milk spectra. Food Chemistry, 504, 148030. https://doi.org/10.1016/j.foodchem.2026.148030 Lundberg, S.M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30, 4765–4774. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 01 May, 2026 Reviewers agreed at journal 21 Apr, 2026 Reviewers invited by journal 31 Mar, 2026 Editor assigned by journal 21 Mar, 2026 Submission checks completed at journal 21 Mar, 2026 First submitted to journal 20 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8927461","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":627252219,"identity":"6387085a-7f24-4bc0-b574-8cb4583be384","order_by":0,"name":"Mian Muhammad Hamza","email":"data:image/png;base64,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","orcid":"","institution":"Government College University, Faisalabad","correspondingAuthor":true,"prefix":"","firstName":"Mian","middleName":"Muhammad","lastName":"Hamza","suffix":""},{"id":627252222,"identity":"99e3ed4b-501e-4960-a5c9-de7f1a8e26a8","order_by":1,"name":"Ruhma Shahbaz","email":"","orcid":"","institution":"Faisalabad Medical University","correspondingAuthor":false,"prefix":"","firstName":"Ruhma","middleName":"","lastName":"Shahbaz","suffix":""}],"badges":[],"createdAt":"2026-02-20 15:53:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8927461/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8927461/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107972328,"identity":"964c61ca-2d66-48ff-85b6-b79da0bc69be","added_by":"auto","created_at":"2026-04-28 07:01:18","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":73562,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eMethodology flowchart illustrating the end-to-end pipeline: dataset loading, preprocessing, stratified cross-validation, model training, evaluation, and SHAP explainability analysis.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8927461/v1/47e707c176aa39930e294c33.jpg"},{"id":108007769,"identity":"063442fc-57c2-4742-ae96-d7d2a5c9e38d","added_by":"auto","created_at":"2026-04-28 13:01:49","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":195175,"visible":true,"origin":"","legend":"\u003cp\u003eModel Architecture Diagram\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8927461/v1/fd2f9506f3c1a3a214fa2916.jpg"},{"id":108007463,"identity":"e5e4f348-fdc6-4067-82a4-9f9811579a72","added_by":"auto","created_at":"2026-04-28 13:00:10","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":116466,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eConfusion matrices for all five classifiers evaluated on the held-out 20% test set (rows = true class; columns = predicted class). Classes: M = Maintain; I = Increase; R = Reduce.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8927461/v1/3a3ec033259e7f1629560409.jpg"},{"id":108007138,"identity":"678bbb84-7f05-466b-9ebe-0cb58d824b8e","added_by":"auto","created_at":"2026-04-28 12:58:41","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":40601,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eSHAP summary plot for the Random Forest classifier illustrating feature importance and directional contribution across all three recommendation classes (Maintain, Increase, Reduce). Color intensity indicates feature value magnitude; horizontal position reflects the direction and magnitude of the SHAP contribution to the predicted outcome.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8927461/v1/e4725a229e566145383d82f2.jpg"},{"id":108009011,"identity":"3ccc71b5-c387-4fc3-bc4f-a731cdc6b996","added_by":"auto","created_at":"2026-04-28 13:09:02","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":773366,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8927461/v1/6cdf0623-75e0-439f-8039-805879c98e10.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Predicting Dairy Milk Intake Recommendations Using Machine Learning Models: A Comparative Analysis","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eDairy milk and dairy products have constituted a cornerstone of human dietary practice across millennia, furnishing a concentrated array of essential nutrients \u0026mdash; proteins, calcium, phosphorus, riboflavin, and vitamins D and B12 \u0026mdash; that collectively sustain bone mineralization, neuromuscular function, and systemic metabolic processes [1,2]. Despite this established nutritional profile, the epidemiological evidence linking dairy consumption to long-term health outcomes presents a considerably more nuanced picture than historical dietary guidelines have acknowledged.\u003c/p\u003e\n\u003cp\u003eProspective studies and meta-analyses report that moderate dairy milk intake is inversely associated with the risk of stroke and certain manifestations of cardiovascular disease (CVD); however, the magnitude and direction of this association vary markedly according to dairy subtype, fat content, and the specific cardiovascular endpoint under investigation [3,4]. Total dairy consumption \u0026mdash; encompassing both liquid milk and fermented products such as yogurt \u0026mdash; has been correlated with a modestly reduced risk of overall CVD, while high-fat dairy products have been implicated in adverse cardiometabolic outcomes in selected cohorts [3,5]. Similarly, low-fat dairy variants have demonstrated favorable associations with blood pressure regulation, whereas full-fat products may confer neutral or marginally deleterious effects on cardiovascular mortality in certain populations [4]. In parallel, the precise roles of dairy milk in modulating the risk of type 2 diabetes mellitus and hormonally sensitive cancers remain subjects of active scientific inquiry, with published findings indicating inconsistent and population-specific associations [6].\u003c/p\u003e\n\u003cp\u003eThese heterogeneous outcomes underscore the inadequacy of uniform dietary recommendations and highlight an urgent clinical need for individualized guidance grounded in patient-specific clinical and lifestyle profiles. The convergence of high-dimensional electronic health records, clinical biomarker data, and advanced machine learning (ML) methodologies offers a promising avenue for meeting this need [4,5]. Supervised classification algorithms, particularly ensemble-based methods such as Random Forest and gradient-boosted frameworks, have demonstrated robust performance in predicting cardiometabolic risk and treatment outcomes from structured clinical data, frequently surpassing traditional statistical models in both accuracy and generalizability [5,13].\u003c/p\u003e\n\u003cp\u003eIn the domain of ML interpretability, SHAP (SHapley Additive exPlanations) has emerged as the de facto standard for attributing feature-level contributions in complex, non-linear models [13]. Applications of SHAP to clinical prediction tools have been shown to enhance clinician trust, facilitate regulatory scrutiny, and support the responsible deployment of AI-driven decision-support systems. The utility of SHAP is particularly pronounced in the context of tree-based ensembles, where non-linear feature interactions and higher-order dependencies cannot be adequately characterised through coefficient-based approaches alone.\u003c/p\u003e\n\u003cp\u003eStratified k-fold cross-validation represents the methodological standard for evaluating ML classifiers on datasets of moderate size, providing substantially more stable and conservative performance estimates than single train-test partitions, particularly in the presence of class imbalance [12]. One-hot encoding is the established preprocessing strategy for nominal categorical variables \u0026mdash; such as country of origin and milk type \u0026mdash; as it preserves their nominal character without introducing artificial ordinal constraints that may systematically bias distance-based or linear classification algorithms [8].\u003c/p\u003e\n\u003cp\u003eThe present study addresses the aforementioned gap by constructing an ML-based classification framework for the prediction of physician-recommended dairy milk intake adjustments. In contrast to prior work focused on agricultural yield forecasting or pathological disease detection, the clinical focus of this study \u0026mdash; determining whether a clinician would advise a patient to maintain, increase, or reduce dairy milk consumption \u0026mdash; positions it squarely within the domain of personalized nutritional decision support. This focus is informed by recent umbrella reviews that emphasize the heterogeneity of dairy-specific health effects across subtypes, age groups, and geographic populations [3,11].\u003c/p\u003e\n\u003ch2\u003e1.1 Study Objectives\u003c/h2\u003e\n\u003cp\u003e(i) \u0026nbsp; \u0026nbsp;To develop and comparatively evaluate ML classifiers \u0026mdash; encompassing Random Forest, Logistic Regression, Light-GBM, KNN, and a Voting Classifier ensemble \u0026mdash; for predicting physician-recommended dairy milk intake adjustments.\u003c/p\u003e\n\u003cp\u003e(ii) \u0026nbsp; To apply stratified k-fold cross-validation as the primary evaluation framework, thereby ensuring reliable and unbiased performance estimates.\u003c/p\u003e\n\u003cp\u003e(iii) To employ SHAP analysis on the two best-performing models to identify and clinically interpret the most influential predictive features.\u003c/p\u003e\n\u003cp\u003e(iv) \u0026nbsp; To critically assess the limitations inherent to synthetic data and to outline a prospective validation roadmap using real-world clinical datasets.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1.2 Research Questions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u0026ldquo;How can be the multiple ML models be developed for prediction of possible milk effect on humans, belongs to multiple regions.?\u0026rdquo;\u003c/p\u003e\n\u003cp\u003e\u0026ldquo;How can SHAP interpretability be integrated with predictive methodology in other to ensure practical clinical interpretabilities.?\u0026rdquo;\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003ch2\u003e2.1 Dataset Characteristics and Synthetic Nature\u003c/h2\u003e\n\u003cp\u003eThe dataset employed in this study is publicly accessible via the Kaggle repository (https://www.kaggle.com/datasets/hammadansari7/milks-effect-on-human-health) and comprises 10,000 synthetic records [1]. The dataset was computationally generated to simulate a diverse clinical cohort by sampling feature values from realistic marginal distributions. The target variable (Doctor_Recommendation) was assigned through a rule-based decision logic intended to approximate clinical guidelines governing dairy intake recommendations.\u003c/p\u003e\n\u003cp\u003eThe dataset encompasses 17 features spanning demographic, clinical, and lifestyle domains. Table 1 provides a comprehensive summary of all features, their measurement scales, value ranges, and descriptive statistics. Numerical features include: Age (18\u0026ndash;80 years; mean = 47.3, SD = 17.1), BMI (kg/m\u0026sup2;; mean = 26.4, SD = 4.8), Daily_Milk_Intake_ml (mean = 310 mL, SD = 145), Cholesterol_mg_dl (mean = 195 mg/dL, SD = 38), Calcium_Level_mg_dl (mean = 9.4 mg/dL, SD = 0.9), and Physical_Activity_Level (hours/week; mean = 3.2, SD = 2.1). Categorical features encompass Gender, Country (10 categories), Milk_Type (Whole, Low-fat, Skim, Plant-based), Lactose_Intolerance, Smoking, and Diabetes.\u003c/p\u003e\n\u003cp\u003eIt is explicitly acknowledged that synthetic datasets generated via statistical sampling and rule-based label assignment cannot fully replicate the distributional complexity, measurement noise, confounding structure, and inter-individual variability inherent in real-world clinical data. Accordingly, all reported performance metrics should be interpreted as upper-bound estimates that are unlikely to be sustained under external real-world validation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Summary of Dataset Features, Measurement Scales, and Descriptive Statistics.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFeature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eType\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eScale\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRange / Categories\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDescriptive Statistics\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003e18\u0026ndash;80 years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eMean = 47.3; SD = 17.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBMI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eKg/m\u0026sup2;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eMean = 26.4; SD = 4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDaily_Milk_Intake_ml\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003emL/day\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eMean = 310; SD = 145\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCholesterol_mg_dl\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003emg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eMean = 195; SD = 38\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCalcium_Level_mg_dl\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003emg/dL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eMean = 9.4; SD = 0.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePhysical_Activity_Level\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eHours/week\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eMean = 3.2; SD = 2.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGender\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eMale / Female\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eBinary\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCountry\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003e10 categories\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eMulti-class\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMilk_Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eWhole / Low-fat / Skim / Plant-based\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eMulti-class\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLactose_Intolerance\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eBinary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eYes / No\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eBinary\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSmoking\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eBinary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eYes / No\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eBinary\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDiabetes\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eBinary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eYes / No\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eBinary\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge_Group (engineered)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eOrdinal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eChild / Young Adult / Middle-Aged / Senior\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eDerived feature\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDoctor_Recommendation (target)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\n \u003cp\u003eMulti-class\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 147px;\"\u003e\n \u003cp\u003eMaintain / Increase / Reduce\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 144px;\"\u003e\n \u003cp\u003eThree-class outcome\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eSD = standard deviation. Descriptive statistics are reported for the full 10,000-record synthetic dataset. The Age_Group variable is a derived feature created during preprocessing. Doctor_Recommendation constitutes the three-class target variable (Maintain / Increase / Reduce).\u003c/p\u003e\n\u003ch2\u003e2.2 Preprocessing Pipeline\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(i) Missing Value Handling:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eA systematic null-value audit confirmed the absence of missing values across all 10,000 records and 17 features; accordingly, no imputation procedures were required.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(ii) Categorical Encoding:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eAll nominal categorical variables \u0026mdash; Gender, Country, Milk_Type, Lactose_Intolerance, Smoking, and Diabetes \u0026mdash; were transformed using one-hot encoding (OHE). OHE was selected in preference to label encoding to preclude the introduction of spurious ordinal relationships that can systematically bias the learning dynamics of distance-based and linear classification algorithms [8].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(iii) Feature Engineering:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eA clinically meaningful Age_Group feature was derived by categorizing participants into four age strata: Child (0\u0026ndash;17 years), Young Adult (18\u0026ndash;34 years), Middle-Aged Adult (35\u0026ndash;64 years), and Senior (65+ years). This binning strategy corrects the previously incomplete age segmentation reported in an earlier iteration of this analysis, which omitted the 35\u0026ndash;64-year range \u0026mdash; the largest clinical demographic subgroup.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(iv) Feature Scaling: Standardization\u003c/em\u003e\u003c/strong\u003e (zero mean, unit variance) was applied to all continuous numerical features \u0026mdash; BMI, Daily_Milk_Intake_ml, Cholesterol_mg_dl, Calcium_Level_mg_dl, and Physical_Activity_Level \u0026mdash; prior to model training. Scaling is critical for distance-sensitive algorithms (KNN, Logistic Regression) and was applied exclusively within each training fold of the cross-validation procedure to prevent data leakage.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(v) Outlier Management:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eOutliers in continuous features were identified using the interquartile range (IQR) criterion and univariate histograms. Feature values exceeding three standard deviations from the respective feature mean were winsorised to their boundary values, mitigating the distorting influence of extreme observations without discarding valid records.\u003c/p\u003e\n\u003ch2\u003e2.3 Evaluation Strategy\u003c/h2\u003e\n\u003cp\u003eStratified 10-fold cross-validation (K = 10) was employed as the primary model evaluation approach. Stratification ensures that the class distribution of the full dataset is preserved within each fold, yielding a more robust and representative estimate of out-of-sample performance relative to a single train-test partition [12]. To prevent data leakage, all preprocessing steps \u0026mdash; including feature scaling, OHE, and age group binning \u0026mdash; were applied within the training portion of each fold prior to validation-set scoring.\u003c/p\u003e\n\u003cp\u003eA complementary held-out 20% test set was reserved and preprocessed independently to report final point-estimate metrics. Performance was assessed using: accuracy, macro-averaged precision, macro-averaged recall, macro-averaged F1-score, one-vs-rest ROC-AUC, and full per-class confusion matrices for each model. A majority-class dummy classifier was included as a lower-bound reference, returning an accuracy of 0.382, against which all ML classifiers are benchmarked.\u003c/p\u003e\n\u003ch2\u003e2.4 Model Architectures and Hyperparameter Configuration\u003c/h2\u003e\n\u003cp\u003eFive supervised classification algorithms were implemented and evaluated. Table 2 presents a concise summary of each model, its key hyperparameters, and its principal methodological advantage. Hyperparameters for Random Forest, Light GBM, and KNN were determined through five-fold cross-validated grid search on the training data as show in Figure 2; Logistic Regression employed default regularization. All models were implemented using the scikit-learn and Light GBM Python libraries.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Summary of Model Architectures, Hyperparameter Configurations, and Methodological Advantages.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDescription\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eKey Hyperparameters\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrimary Advantage\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLogistic Regression\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003eLinear classifier using the softmax function for multi-class probability estimation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eC = 1.0; L2 regularisation; solver = lbfgs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eHigh interpretability; serves as a linear baseline\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRandom Forest\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003eBagging ensemble of decision trees; final prediction by majority vote / probability averaging\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003en_estimators = 200; max_depth = 15; min_samples_split = 5; class_weight = balanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eRobust to overfitting; handles non-linear feature interactions\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLight-GBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003eGradient boosting framework using histogram-based leaf-wise tree growth for efficient training on tabular data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003elearning_rate = 0.05; num_leaves = 50; n_estimators = 300; subsample = 0.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eSuperior speed on large structured datasets; captures complex interactions\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eK-Nearest Neighbors (KNN)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003eNon-parametric classifier based on majority vote among k nearest neighbours in feature space\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003ek = 7; determined via grid search cross-validation; distance metric = Euclidean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eSimple; non-parametric; no training phase required\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 120px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVoting Classifier\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 184px;\"\u003e\n \u003cp\u003eSoft-voting ensemble combining predicted class probabilities from LR, RF, Light-GBM, and KNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eSoft voting; equal weights; constituent models as above\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 160px;\"\u003e\n \u003cp\u003eReduces individual model errors; leverages complementary strengths\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eLR = Logistic Regression; RF = Random Forest; LGBM = Light-GBM; KNN = K-Nearest Neighbors; OvR = one-vs-rest. Hyperparameters for RF, LGBM, and KNN were selected via five-fold cross-validated grid search.\u003c/p\u003e\n\u003ch2\u003e2.5 SHAP Explainability Analysis\u003c/h2\u003e\n\u003cp\u003eSHAP (SHapley Additive exPlanations) analysis was applied to the two best-performing models \u0026mdash; Random Forest and Light-GBM \u0026mdash; to provide post-hoc, feature-level interpretability [13]. The Tree-Explainer algorithm was employed for efficient SHAP value computation, given the tree-based architecture of both models. SHAP summary plots and mean absolute SHAP value rankings were generated to identify the features exerting the greatest influence on model predictions across all three recommendation classes. This approach addresses a recognized limitation of the earlier iteration of this study, which had applied SHAP exclusively to Logistic Regression \u0026mdash; a model whose coefficient-based interpretation is already transparent by design \u0026mdash; thereby omitting interpretability analysis for the substantively more informative tree-based classifiers.\u003c/p\u003e"},{"header":"3. Results and Discussion","content":"\u003ch2\u003e3.1 Cross-Validation Performance\u003c/h2\u003e\n\u003cp\u003eTable 3 reports mean performance metrics and their standard deviations across the 10 stratified folds. All five ML classifiers substantially and consistently outperformed the majority-class dummy baseline (accuracy = 0.382), confirming that each model captures genuine predictive signal rather than reflecting trivial majority-class assignment.\u003c/p\u003e\n\u003cp\u003eRandom Forest achieved the highest mean cross-validation accuracy (0.982 \u0026plusmn; 0.006), with the Voting Classifier marginally below at 0.981 \u0026plusmn; 0.007. Light-GBM performed comparably at 0.978 \u0026plusmn; 0.008. Logistic Regression attained 0.964 \u0026plusmn; 0.010, reflecting the strong linear separability of the synthetic feature space, while KNN returned the lowest cross-validation accuracy (0.948 \u0026plusmn; 0.012), consistent with the elevated dimensionality induced by one-hot encoding. The narrow standard deviation ranges across all models indicate consistent, stable performance across folds, with no evidence of excessive fold-to-fold variance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Stratified 10-Fold Cross-Validation Performance (Mean \u0026plusmn; Standard Deviation).\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eROC-AUC (OvR)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDummy Baseline\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.382\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.127\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.333\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.187\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.500\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRandom Forest\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.982 \u0026plusmn; 0.006\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.981 \u0026plusmn; 0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.982 \u0026plusmn; 0.006\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.981 \u0026plusmn; 0.006\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.998 \u0026plusmn; 0.002\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVoting Classifier\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.981 \u0026plusmn; 0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.980 \u0026plusmn; 0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.981 \u0026plusmn; 0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.980 \u0026plusmn; 0.007\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.997 \u0026plusmn; 0.002\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLight-GBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.978 \u0026plusmn; 0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.977 \u0026plusmn; 0.009\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.978 \u0026plusmn; 0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.977 \u0026plusmn; 0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.997 \u0026plusmn; 0.002\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLogistic Regression\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.964 \u0026plusmn; 0.010\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.963 \u0026plusmn; 0.011\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.964 \u0026plusmn; 0.010\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.963 \u0026plusmn; 0.010\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.994 \u0026plusmn; 0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eKNN (k = 7)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.948 \u0026plusmn; 0.012\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.947 \u0026plusmn; 0.013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.948 \u0026plusmn; 0.012\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.947 \u0026plusmn; 0.013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.981 \u0026plusmn; 0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eOvR = one-vs-rest multiclass formulation. Precision, recall, and F1-score are reported as macro-averages. The Dummy Baseline reflects a majority-class classifier and does not compute standard deviations. Cross-validation results represent the primary generalizability estimate; held-out test results are presented separately in Table 4.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e3.2 Held-Out Test Set Performance and Confusion Matrices\u003c/h2\u003e\n\u003cp\u003eTable 4 presents model performance on the held-out 20% test partition. Random Forest and the Voting Classifier each attained perfect scores (accuracy = 1.000) across all metrics. These values are attributable to the deterministic, rule-based structure of the synthetic target variable, in which class labels are a direct function of input features without stochastic noise. This behaviour is both expected and structurally inevitable in rule-generated synthetic data; it is explicitly not indicative of overfitting or data leakage, as all preprocessing transformations were applied within cross-validation folds and no feature selection was conducted on the full dataset. These scores should not be extrapolated to real-world clinical settings. The cross-validation results presented in Table 3 and Figure 3 constitute the more informative and conservative generalizability estimate.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4. Held-Out Test Set Performance (20% Partition).\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1-Score (Macro)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eROC-AUC (OvR)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRandom Forest\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVoting Classifier\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLogistic Regression\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9970\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9950\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9970\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9960\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9999\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLight-GBM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9960\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9964\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9960\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9962\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e1.0000*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 113px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eKNN (k = 7)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9765\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9744\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9765\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9753\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 102px;\"\u003e\n \u003cp\u003e0.9931\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003e* Perfect scores (1.000) reflect the deterministic, rule-based labelling of the synthetic dataset and are not indicative of overfitting or data leakage. Cross-validation results (Table 3) provide the appropriate generalization estimate. OvR = one-vs-rest.\u003c/p\u003e\n\u003ch2\u003e3.3 Qualitative Model Analysis\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003eRandom Forest:\u003c/strong\u003e The highest-performing classifier across both cross-validation and held-out evaluation. Ensemble aggregation of 200 decision trees reduces prediction variance and accommodates complex non-linear feature interactions, including higher-order dependencies that are beyond the representational capacity of linear models. The balanced class-weight parameter further ensures appropriate sensitivity to the minority recommendation class.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVoting Classifier:\u003c/strong\u003e Marginally below Random Forest in cross-validation, the soft-voting ensemble combines the complementary inductive biases of its four constituent classifiers \u0026mdash; Logistic Regression, Random Forest, Light-GBM, and KNN. Probability averaging across constituent models moderates individual model errors and provides robust performance even when individual classifiers exhibit localized weaknesses.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLight-GBM:\u003c/strong\u003e Achieved cross-validation accuracy comparable to Random Forest with substantially reduced training time, attributable to its histogram-based gradient boosting algorithm and leaf-wise tree growth strategy. Light-GBM\u0026rsquo;s native handling of structured tabular data and efficient treatment of categorical features makes it particularly well-suited to the dataset characteristics employed in this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLogistic Regression:\u003c/strong\u003e Performed competitively (cross-validation accuracy = 0.964), a result consistent with the strong linear separability of the synthetic feature space. Its relatively lower performance compared with tree-based ensembles reflects the inherent inability of a linear boundary to capture non-linear interaction terms. Nevertheless, its full interpretability via model coefficients renders it a valuable clinical communication tool.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eK-Nearest Neighbors (KNN):\u003c/strong\u003e The lowest-performing classifier, consistent with anticipated degradation under the elevated feature dimensionality introduced by one-hot encoding. The curse of dimensionality attenuates meaningful distance distinctions in high-dimensional spaces, reducing the discriminative power of majority-vote neighborhood assignment despite cross-validation optimization of k.\u003c/p\u003e\n\u003ch2\u003e3.4 SHAP Explainability Analysis\u003c/h2\u003e\n\u003cp\u003eSHAP analysis applied to Random Forest and Light-GBM revealed a highly consistent feature importance ranking across both models, with Age, Calcium_Level_mg_dl, and Lactose_Intolerance emerging as the three most influential predictors. Table 5 presents the SHAP-derived feature ranking alongside clinical interpretations; Figure 3 depicts the SHAP summary plot for the Random Forest classifier.\u003c/p\u003e\n\u003cp\u003eHigher Age values are strongly associated with a Reduce recommendation, consistent with the progressive decline in lactase activity and reduced gastrointestinal calcium absorption efficiency observed in older adults. Elevated serum calcium reduces the model\u0026rsquo;s predicted probability of an Increase recommendation, reflecting clinician caution regarding hypercalcaemia risk in the context of supplementary dairy intake. Lactose Intolerance constitutes the third most influential predictor, with its presence strongly shifting predictions towards a Reduce classification \u0026mdash; an association well-grounded in established dietary management protocols Secondary contributors include BMI, Daily_Milk_Intake_ml, Cholesterol_mg_dl, and Diabetes status, each exerting directional SHAP effects consistent with accepted cardiometabolic dietary guidance as show in Figure 4.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 5. SHAP Feature Importance Ranking for Random Forest: Mean Absolute SHAP Values and Clinical Interpretations.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRank\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFeature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMean |SHAP| (RF)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eClinical Interpretation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e0.412\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eOlder adults show higher likelihood of a Reduce recommendation, reflecting declining lactase activity and altered calcium metabolism\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCalcium_Level_mg_dl\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e0.338\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eElevated serum calcium reduces the probability of an Increase recommendation; supports clinical guidance on avoiding hypercalcaemia\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLactose_Intolerance\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e0.271\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eStrong predictor of a Reduce recommendation; consistent with clinical avoidance of lactose-containing dairy in intolerant individuals\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e4\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBMI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e0.184\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eHigher BMI values contribute to Reduce recommendations; reflects cardiometabolic risk considerations in dietary counselling\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e5\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDaily_Milk_Intake_ml\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e0.156\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eBaseline intake level modulates recommendation direction; very low current intake increases the probability of an Increase recommendation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e6\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCholesterol_mg_dl\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e0.127\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003eElevated cholesterol shifts predictions towards Reduce; aligned with guidance on limiting saturated fat from full-fat dairy\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 93px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 160px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDiabetes\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e0.098\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 192px;\"\u003e\n \u003cp\u003ePresence of diabetes marginally increases Reduce probability; consistent with glycaemic management recommendations\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eMean |SHAP| values represent the average absolute SHAP contribution across all test instances and all three recommendation classes. Rankings were consistent between Random Forest and Light-GBM; RF values are presented as the primary reference given its superior cross-validation accuracy.\u003c/p\u003e\n\u003ch2\u003e3.5 Sensitivity Analysis: Class Imbalance\u003c/h2\u003e\n\u003cp\u003eAlthough the dataset approximates class balance across the three recommendation categories, SMOTE (Synthetic Minority Over-sampling Technique) was applied within the training folds of the cross-validation procedure to assess sensitivity to class distribution perturbations. Performance improvements relative to the un-resampled baseline were marginal (\u0026lt; 0.3% in macro-averaged F1-score), confirming that class imbalance does not constitute a primary source of bias in the current dataset. However, real-world clinical datasets frequently exhibit pronounced class imbalance in recommendation outcomes; in such contexts, SMOTE application within each training fold \u0026mdash; rather than prior to the cross-validation split \u0026mdash; is recommended as standard practice to prevent label leakage.\u003c/p\u003e\n\u003ch2\u003e3.6 Limitations\u003c/h2\u003e\n\u003cp\u003eThis study is subject to several fundamental limitations that must be considered in the interpretation of the reported findings.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(i) Synthetic Dataset:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eThe primary and most consequential limitation of this study is its exclusive reliance on a computationally generated synthetic dataset. Although the dataset was designed to simulate realistic clinical distributions, it cannot replicate the distributional heterogeneity, measurement noise, confounding factor structure, missing data patterns, and population-specific variation inherent in real clinical data. All reported performance metrics should be interpreted as illustrative upper bounds that are unlikely to generalize to authentic clinical settings without material degradation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(ii) Target Variable Construction:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eThe study predicts physician-generated dietary recommendations rather than directly measured health outcomes. Critically, the recommendation labels in the dataset were assigned through deterministic rule-based logic rather than extracted from recorded clinical decisions. This construction limits the ecological validity of the predictive task and may not reflect the nuanced, contextual judgement of practice clinicians.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(iii) External Validity:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eThe trained models have not been evaluated on external real-world datasets. Their generalizability across diverse clinical populations, healthcare systems, and geographic contexts remains entirely unestablished. The perfect held-out scores observed for Random Forest and the Voting Classifier are direct artefacts of the synthetic data generation mechanism and should not be interpreted as evidence of deployable clinical utility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(iv) Feature Space Completeness:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eThe dataset omits potentially important clinical determinants of dairy intake recommendations, including gut microbiome composition, genetic markers for lactase persistence (notably the LCT gene variants), longitudinal dietary recall data, renal function indices, and medication history. The absence of these variables may limit both the predictive completeness of the framework and the clinical face validity of its outputs.\u003c/p\u003e\n\u003ch2\u003e3.7 Future Research Directions and Recommendations\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(i) Prospective Real-World Validation:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eFuture work should prioritize the validation of these models using prospectively collected real-world clinical datasets sourced from multi-site healthcare facilities, ideally spanning multiple countries and demographic strata, to rigorously assess generalizability and practical clinical utility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(ii) Extended Feature Engineering:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eIncorporating additional biologically and clinically relevant variables \u0026mdash; including genetic markers for lactase persistence, gut microbiome indices, longitudinal dietary records, and renal function parameters \u0026mdash; would likely improve both the predictive accuracy and the mechanistic interpretability of future iterations of the framework.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(iii) Systematic Hyperparameter Optimization:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eBayesian hyperparameter optimization frameworks \u0026mdash; such as Optuna or Hyperopt \u0026mdash; applied to Light-GBM and KNN could yield incremental performance improvements and provide a more principled search across the hyperparameter space than grid-based approaches.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e(iv) Clinical Integration and Co-Design:\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eCollaborative development with registered dietitians, clinical nutritionists, and practising physicians is strongly recommended to align model outputs with current evidence-based dietary guidelines, ensure clinical face validity, and facilitate responsible integration into nutritional decision-support workflows.\u003c/p\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eThis study designed, implemented, and comparatively evaluated a suite of machine learning classifiers for predicting physician-recommended dairy milk intake adjustments from clinical and lifestyle features within a controlled synthetic data environment. Random Forest and the Voting Classifier demonstrated the strongest generalization performance under stratified 10-fold cross-validation, attaining mean accuracies of 0.982 and 0.981, respectively. Light-GBM performed comparably at 0.978, while Logistic Regression and KNN provided informative competitive baselines at 0.964 and 0.948.\u003c/p\u003e\n\u003cp\u003eThe perfect held-out test scores recorded for Random Forest and the Voting Classifier are explicitly attributable to the deterministic, rule-based structure of the synthetic target variable and are not indicative of overfitting, data leakage, or deployable clinical accuracy. Cross-validation results constitute the appropriate and more conservative generalizability estimate.\u003c/p\u003e\n\u003cp\u003eSHAP analysis applied to the two best-performing models identified Age, Calcium Level, and Lactose Intolerance as the most clinically influential predictors \u0026mdash; a ranking consistent with established mechanisms governing dairy intake guidance in clinical nutrition practice. The corrected one-hot encoding pipeline, complete age group binning, stratified k-fold evaluation strategy, extended per-class confusion matrix diagnostics, and SHAP analysis applied to tree-based models collectively address the methodological concerns identified through peer review.\u003c/p\u003e\n\u003cp\u003eNotwithstanding these contributions, the synthetic origin of the dataset represents a fundamental constraint on the conclusions that may be drawn. Future work must prioritize prospective validation on real-world clinical datasets to establish the practical utility and clinical transferability of this framework before any consideration of integration into dietary decision-support systems.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflict of Interest:\u0026nbsp;\u003c/strong\u003eThe authors declare no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions:\u0026nbsp;\u003c/strong\u003eConceptualization: M.M.H. and R.S.; Literature Review: M.M.H. and R.S.; Methodology: M.M.H.; Figures and Table Design: M.M.H.; Experiments: M.M.H.; Writing \u0026mdash; Original Draft: M.M.H.; Writing \u0026mdash; Review and Editing: R.S.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability:\u0026nbsp;\u003c/strong\u003eData are publicly available at: https://www.kaggle.com/datasets/hammadansari7/milks-effect-on-human-health\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u0026nbsp;\u003c/strong\u003eNo funding was received for this research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Considerations:\u0026nbsp;\u003c/strong\u003eThe dataset is publicly available and synthetic in nature, requiring no institutional ethical approval for research use.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eKaggle Dataset: Milk\u0026rsquo;s Effect on Human Health. Available at: https://www.kaggle.com/datasets/hammadansari7/milks-effect-on-human-health/data\u003c/li\u003e\n\u003cli\u003eZhang, X., Chen, X., Xu, Y., Yang, J., Du, L., Li, K., \u0026amp; Zhou, Y. (2021). Milk consumption and multiple health outcomes: Umbrella review of systematic reviews and meta-analyses in humans. Nutrition \u0026amp; Metabolism, 18(7). https://doi.org/10.1186/s12986-020-00527-y\u003c/li\u003e\n\u003cli\u003eSharifan, P., Roustaee, R., Shafiee, M., Longworth, Z.L., Keshavarz, P., Davies, I.G., Webb, R.J., Mazidi, M., \u0026amp; Vatanparast, H. (2025). Dairy milk consumption and risk of cardiovascular and bone health outcomes in adults: An umbrella review and updated meta-analyses. Nutrients, 17, 2723. https://doi.org/10.3390/nu17172723\u003c/li\u003e\n\u003cli\u003eLamarche, B., et al. (2025). The effects of dairy milk consumption on cardiovascular and stroke risk. The American Journal of Clinical Nutrition, 121(5), 956\u0026ndash;964. https://doi.org/10.1093/ajcn/nqz276\u003c/li\u003e\n\u003cli\u003ePoppitt, S.D. (2020). Cow\u0026rsquo;s milk and dairy milk consumption: Is there now consensus for cardiometabolic health? Frontiers in Nutrition, 7, 574725. https://doi.org/10.3389/fnut.2020.574725\u003c/li\u003e\n\u003cli\u003eSanjulian, L., et al. (2025). The role of dairy milk in human nutrition: Myths and realities. Nutrients, 17, 646. https://doi.org/10.3390/nu17040646\u003c/li\u003e\n\u003cli\u003eAmin, T., et al. (2025). Clinical evidence and mechanistic pathways of human milk oligosaccharide supplementation for health benefits: An updated review. Frontiers in Nutrition, 12, 1599678. https://doi.org/10.3389/fnut.2025.1599678\u003c/li\u003e\n\u003cli\u003eShen, S., Song, X., Ding, H., Cui, X., Xie, Z., \u0026amp; Huang, H. (2026). Data-driven soft sensing for raw milk ethanol stability prediction. Sensors, 26(3), 903. https://doi.org/10.3390/s26030903\u003c/li\u003e\n\u003cli\u003eLoforte, Y., Zanzan, M., de Almeida, A., Cadavez, V., \u0026amp; Gonzales-Barron, U. (2026). Growth of Listeria monocytogenes in goat\u0026rsquo;s pasteurised milk cheese during maturation. Applied Microbiology, 6, 16. https://doi.org/10.3390/applmicrobiol6010016\u003c/li\u003e\n\u003cli\u003eMa, S., et al. (2026). Fast and accurate prediction of milk yield in dairy goats using deep learning. Journal of Dairy Science. https://doi.org/10.3168/jds.2025-27905\u003c/li\u003e\n\u003cli\u003eVilloz, F., et al. (2024). Dairy intake and risk of cognitive decline and dementia: A systematic review and dose-response meta-analysis. Advances in Nutrition, 15, 100160. https://doi.org/10.1016/j.advnut.2023.100160\u003c/li\u003e\n\u003cli\u003eMolle, A., et al. (2026). Targeted iPLS for the prediction of cheese-making traits from individual milk spectra. Food Chemistry, 504, 148030. https://doi.org/10.1016/j.foodchem.2026.148030\u003c/li\u003e\n\u003cli\u003eLundberg, S.M., \u0026amp; Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30, 4765\u0026ndash;4774.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-umm-alqura-university-for-medical-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Journal of Umm Al-Qura University for Medical Science](https://link.springer.com/journal/44361)","snPcode":"44361","submissionUrl":"https://submission.springernature.com/new-submission/44361/3","title":"Journal of Umm Al-Qura University for Medical Science","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Open","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Machine Learning, Dairy Milk Intake, Physician Recommendation Prediction, SHAP (SHapley Additive exPlanations), Random Forest, Light-GBM, Cross-Validation, Clinical Decision Support","lastPublishedDoi":"10.21203/rs.3.rs-8927461/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8927461/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground: \u003c/strong\u003eDairy milk occupies a central position in human nutrition, supplying indispensable micronutrients — including calcium, vitamins D and B12, and high-quality proteins — that sustain bone integrity, cardiovascular function, and metabolic homeostasis. Despite its nutritional value, the optimal level of dairy milk consumption remains subject to ongoing debate, owing to inter-individual variability in lactase persistence, cardiometabolic risk profiles, and the heterogeneous health effects of different dairy subtypes. Clinical guidelines have yet to converge on universally applicable intake thresholds, underscoring the need for personalized, data-driven approaches to dietary counselling.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eA publicly available synthetic dataset comprising 10,000 records and 17 clinico-demographic features was employed. Nominal categorical variables were encoded via one-hot encoding to preclude the imposition of spurious ordinal relationships/ Five supervised classification algorithms were trained and rigorously evaluated: Logistic Regression, Random Forest, Light-GBM, K-Nearest Neighbors (KNN), and a soft-voting Voting Classifier ensemble. Model assessment relied upon stratified 10-fold cross-validation as the primary evaluation paradigm, supplemented by a held-out 20% test set. Performance was quantified across accuracy, macro-averaged precision, recall, F1-score, one-vs-rest ROC-AUC, and per-class confusion matrices. SHAP (SHapley Additive exPlanations) analysis was subsequently applied to the two best-performing models to identify the most clinically influential predictive features.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eUnder stratified 10-fold cross-validation, Random Forest and the Voting Classifier attained mean accuracies of 0.982 ± 0.006 and 0.981 ± 0.007, respectively, while Light-GBM reached 0.978 ± 0.008. Logistic Regression and KNN achieved 0.964 ± 0.010 and 0.948 ± 0.012, respectively; all models substantially outperformed the majority-class dummy baseline (accuracy = 0.382). Perfect held-out test scores observed for Random Forest and the Voting Classifier are attributable to the deterministic, rule-based structure of the synthetic target variable and do not reflect real-world generalizability. SHAP analysis identified Age, Calcium Level, and Lactose Intolerance as the three most influential predictors, consistent with established clinical determinants of dairy intake guidance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eMachine learning classifiers demonstrated strong predictive performance for physician-recommended dairy milk intake adjustments within a controlled synthetic setting. SHAP-based interpretability enhanced transparency by elucidating the clinical basis of model predictions. Nonetheless, the synthetic origin of the dataset constitutes a fundamental limitation; prospective validation on real-world clinical cohorts is a prerequisite before these models could be considered for integration into dietary decision-support systems.\u003c/p\u003e","manuscriptTitle":"Predicting Dairy Milk Intake Recommendations Using Machine Learning Models: A Comparative Analysis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-28 07:01:09","doi":"10.21203/rs.3.rs-8927461/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-05-01T19:58:35+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"172887095394575162714264492895011283782","date":"2026-04-21T18:32:58+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-31T05:42:06+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-21T22:28:21+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-21T07:49:41+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Umm Al-Qura University for Medical Science","date":"2026-03-20T11:08:53+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-umm-alqura-university-for-medical-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Journal of Umm Al-Qura University for Medical Science](https://link.springer.com/journal/44361)","snPcode":"44361","submissionUrl":"https://submission.springernature.com/new-submission/44361/3","title":"Journal of Umm Al-Qura University for Medical Science","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Open","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4f5f99ae-b6d6-49f5-9a67-7bfaec677cbd","owner":[],"postedDate":"April 28th, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-01T19:58:35+00:00","index":41,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-28T07:01:09+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-28 07:01:09","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8927461","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8927461","identity":"rs-8927461","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}