Comparative Performance of Machine Learning Models Using Food Intake Frequency Versus Vegetable Intake Data to Predict Problematic Mealtime Behaviour in Japanese Preschool Children

Variables	Control (n = 121)	Selective eating (n = 162)	P value
Demographic variables Age, years Male, %	4.6 ± 1.0 59.5%	4.6 ± 0.9 57.4%	0.562 0.808
FFQ
Rice	12.7 ± 3.1	12.4 ± 3.1	0.403
Bread	5.1 ± 2.9	5.2 ± 2.6	0.887
Noodle	1.8 ±1.1	1.7 ± 1.5	0.963
Fish	2.6 ± 1.6	2.5 ± 2.0	0.677
Meat	5.5 ± 3.2	5.0 ± 2.7	0.158
Eggs	3.6 ± 2.3	3.2 ± 2.6	0.109
Beans/tofu	3.7 ± 2.6	3.1 ± 2.7	0.063
Vegetable	11.2 ± 3.9	9.0 ± 5.0	<0.001^*
Fruit	5.5 ± 4.0	4.0 ± 3.2	0.001^*
Dairy products	8.1 ± 4.3	8.1 ± 4.3	0.931
Tea	12.9 ± 3.4	12.7 ± 3.0	0.682
Sweetened beverage	2.1 ± 2.3	3.5 ± 3.3	<0.001^*
Snack	4.9 ± 2.9	5.3 ± 3.6	0.323
Instant foods	0.4 ± 0.5	0.5 ± 0.6	0.846
Fast foods	0.5 ± 0.4	0.7 ± 0.7	0.015^*
Designated vegetables
Cabbage	2.6 ± 1.9	1.7 ± 1.5	<0.001^*
Cucumber	3.0 ± 2.4	2.2 ± 1.8	0.001^*
Taro	0.7 ± 0.7	0.3 ± 0.5	<0.001^*
Daikon radish	1.8 ± 1.5	1.3 ± 1.4	0.002^*
Tomato	3.9 ± 3.3	2.2 ± 2.5	<0.001^*
Eggplant	1.4 ± 1.4	0.8 ± 1.1	<0.001^*
Carrot	4.4 ± 2.7	3.8 ± 2.6	0.033^*
Green onion	1.9 ± 1.7	1.1 ± 1.4	<0.001^*
Chinese cabbage	1.6 ± 1.6	1.0 ± 1.2	<0.0001^*
Green pepper	1.5 ± 1.3	0.9 ± 1.2	<0.001^*
Lettuce	1.9 ± 2.2	1.0 ± 1.4	<0.001^*
Onion	4.7 ± 2.7	3.6 ± 2.5	0.001^*
Potato	3.1 ± 2.1	2.2 ± 1.7	<0.001^*
Spinach	2.3 ± 1.6	1.6 ± 1.6	<0.001^*
Broccoli.	3.0 ± 2.0	2.1 ± 2.0	<0.001^*

Mean±standard deviation (SD) or percentage (%).

* P < 0.05.

Table 2.

Comparison of machine learning algorithms for predicting selective eating in preschool children using food frequency data (5-fold cross-validation).

Model	ROC-AUC	Accuracy	Precision	Recall	F1 Score
Random Forest	0.522	0.527	0.581	0.611	0.592
SVM	0.543	0.559	0.599	0.722	0.645
Naïve Bayes	0.619	0.578	0.673	0.553	0.578
Neural Network	0.553	0.535	0.584	0.614	0.558
k-NN	0.519	0.527	0.583	0.597	0.585
XGBoost	0.511	0.520	0.575	0.604	0.585
XGBoost (limited depth)	0.521	0.523	0.577	0.612	0.590
SGDClassifier	0.544	0.537	0.595	0.612	0.585
L1 LogisticRegression	0.618	0.585	0.622	0.731	0.663
Logistic Regression	0.615	0.582	0.621	0.718	0.658
LightGBM	0.514	0.527	0.575	0.650	0.605

[i] k-NN, k-Nearest Neighbours; LightGBM, Light Gradient Boosting Machine; ROC-AUC, area under the receiver operating characteristic curve; SGDClassifier, Stochastic Gradient Descent Classifier; SVM, Support Vector Machine; XGBoost, Extreme Gradient Boosting.

Distribution of ROC–AUC values by machine-learning model.
*Note*: Each box represents the interquartile range (IQR) of ROC–AUC values across all configurations for each model, with horizontal lines indicating the median and whiskers extending to 1.5× IQR. The horizontal axis shows ROC–AUC (0–1), and models are arranged from lowest to highest median AUC. Naive Bayes and Neural Network achieved the highest central performance, whereas Random Forest and k-NN showed relatively lower discriminative ability.

Feature occurrence across best-performing models using (A) Food Frequency Questionnaire (FFQ) categories and (B) individual vegetable intake variables.
*Note:* Each cell indicates whether a given feature was selected as important (black = included; white = excluded) in the best-performing model for each algorithm. Vegetable and sweetened beverage intake frequently appeared among FFQ-based models, while taro, spinach, and green onion were consistently selected in vegetable-based models, indicating robust predictive value across algorithms.

Permutation importance in (A) FFQ-based model (Naïve Bayes) and (B) vegetable intake model (L1 logistic regression).
*Note:* Bars indicate the mean decrease in model performance when each variable was permuted, representing its relative contribution to prediction accuracy. In the FFQ-based model (A), vegetable and sweetened beverage intake were the most important, whereas in the vegetable intake-based model (B), tomato, green onion, and taro were key predictors.

Table 3.

Comparison of machine learning algorithms for predicting selective eating in preschool children using vegetable frequency data (5-fold cross-validation).

Model	ROC-AUC	Accuracy	Precision	Recall	F1 Score
Random Forest	0.668	0.624	0.673	0.674	0.681
SVM	0.676	0.623	0.691	0.653	0.747
Naïve Bayes	0.702	0.638	0.722	0.648	0.823
Neural Network	0.656	0.602	0.633	0.652	0.678
k-NN	0.633	0.603	0.657	0.653	0.673
XGBoost	0.649	0.612	0.659	0.664	0.663
XGBoost (limited depth)	0.660	0.614	0.662	0.666	0.667
SGDClassifier	0.630	0.595	0.638	0.647	0.661
L1 LogisticRegression	0.717	0.647	0.722	0.659	0.808
Logistic Regression	0.715	0.646	0.719	0.661	0.798
LightGBM	0.695	0.639	0.685	0.690	0.691

[i] Abbreviations: k-NN, k-Nearest Neighbours; LightGBM, Light Gradient Boosting Machine; ROC-AUC, area under the receiver operating characteristic curve; SGDClassifier, Stochastic Gradient Descent Classifier; SVM, Support Vector Machine; XGBoost, Extreme Gradient Boosting.

References

Authors

Metrics

Articles in this issue

DOI: https://doi.org/10.34763/jmotherandchild.20263001.d-25-00036 | Journal eISSN: 2719-535X | Journal ISSN: 2719-6488

Journal RSS Feed

Language: English

Page range: 106 - 115

Submitted on: Sep 10, 2025

Accepted on: Dec 5, 2025

Published on: Jun 8, 2026

Published by: Institute of Mother and Child

In partnership with: Paradigm Publishing Services

Publication frequency: 1 issue per year

Keywords:

food,

Related subjects:

Pediatrics and juvenile medicine,

Paediatric haematology and oncology,

Public health

© 2026 Naoki Sakane, Yaeko Kawaguchi, Junichiro Somei, Akiko Suganuma, Masayuki Domichi, published by Institute of Mother and Child
This work is licensed under the Creative Commons Attribution 4.0 License.

Volume 30 (2026): Issue 1 (January 2026)

Comparative Performance of Machine Learning Models Using Food Intake Frequency Versus Vegetable Intake Data to Predict Problematic Mealtime Behaviour in Japanese Preschool Children

Figures & Tables

Table 1.

Table 2.

Figure 1.

Figure 2.

Figure 3.

Table 3.

Paradigm

My account