This paper is in the following e-collection/theme issue:

Prevention and Health Promotion (1167) Tools, Programs and Algorithms (335) Machine Learning (1825) Clinical Information and Decision Making (1722) Obesity and Nutrition as Public Health Problem (357) Infant Feeding Education, Lactation & Breastfeeding Support (39) Pregnancy Information, Education and Lifestyle Interventions (363)

Published on in Vol 11 (2025)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/69220, first published .
Development of Machine Learning–Based Risk Prediction Models to Predict Rapid Weight Gain in Infants: Analysis of Seven Cohorts

Development of Machine Learning–Based Risk Prediction Models to Predict Rapid Weight Gain in Infants: Analysis of Seven Cohorts

Development of Machine Learning–Based Risk Prediction Models to Predict Rapid Weight Gain in Infants: Analysis of Seven Cohorts

Authors of this article:

Miaobing Zheng1, 2 Author Orcid Image ;   Yuxin Zhang1 Author Orcid Image ;   Rachel A Laws1 Author Orcid Image ;   Peter Vuillermin3 Author Orcid Image ;   Jodie Dodd4 Author Orcid Image ;   Li Ming Wen5 Author Orcid Image ;   Louise A Baur5 Author Orcid Image ;   Rachael Taylor6 Author Orcid Image ;   Rebecca Byrne7 Author Orcid Image ;   Anne-Louise Ponsonby8 Author Orcid Image ;   Kylie D Hesketh1 Author Orcid Image
Article Authors Cited by Tweetations Metrics

    1Institute for Physical Activity and Nutrition, School of Exercise and Nutrition Sciences, Deakin University, Geelong, Australia

    2School of Health Sciences, Faculty of Health & Medicine, UNSW Sydney, Wallace Wurth Building, Kensington, Australia

    3Barwon Health, Geelong, Australia

    4Discipline of Obstetrics and Gynaecology, The Robinson Research Institute, The University of Adelaide, Adelaide, Australia

    5School of Public Health and Sydney Medical School, The University of Sydney, Sydney, Australia

    6Department of Medicine, University of Otago, Dunedin, New Zealand

    7School of Exercise and Nutrition Sciences, Faculty of Health, Queensland University of Technology, Kelvin Grove, Australia

    8The Florey Institute of Neuroscience and Mental Health, Murdoch Children's Research Institute, Royal Children's Hospital, The University of Melbourne, Parkville, Australia

    Corresponding Author:

    Miaobing Zheng, PhD


    Background: Rapid weight gain (RWG) during infancy, defined as an upward crossing of one centile line on a weight growth chart, is highly predictive of subsequent obesity risk. Identification of infant RWG could facilitate obesity risk assessment from infancy.

    Objective: Leveraging machine learning (ML) algorithms, this study aimed to develop and validate risk prediction models to identify infant RWG by the age of 1 year.

    Methods: Data from 7 Australian and New Zealand cohorts were pooled for risk model development and validation (n=5233). A total of 8 ML algorithms predicted infant RWG using routinely available prenatal and early postnatal factors, including maternal prepregnancy weight status, maternal smoking during pregnancy, gestational age, parity, infant sex, birth weight, any breastfeeding and timing of solids introduction at the age of 6 months. Pooled data were randomly split into a training dataset (70%) and a test dataset (30%) for model training and validation, respectively. Model consistency was evaluated using 5-fold cross-validation. Model predictive performance was evaluated by area under the receiver operating characteristic (ROC) curve (AUC), accuracy, precision, sensitivity, specificity, and Cohen κ.

    Results: The average prevalence of infant RWG was 27%. In the training dataset, all ML algorithms showed acceptable to excellent discrimination with AUCs ranging from 0.75 to 0.86. Accuracy, which indicates the overall correctness of the model, ranged from 0.69 to 0.78. Precision, which measures the model’s ability to avoid false positives, ranged from 0.68 to 0.77. The spread of sensitivity, specificity, and Cohen κ of all models was 0.68‐0.80, 0.65‐0.78, and 0.38‐0.56, respectively. Of the 8 algorithms, the Gradient Boosting model showed the most favorable predictive accuracy. Validation of the Gradient Boosting model in the testing dataset exhibited excellent discrimination (AUC 0.3‐0.6) and good ability to make accurate predictions, particularly true positive cases (with accuracy and sensitivity>0.75), but modest performance for precision (0.57‐0.60) and Cohen κ (0.47‐0.52).

    Conclusions: This study developed the first set of ML-based risk prediction models to identify infants’ risk of experiencing RWG by the age of 1 year with acceptable accuracy. The models could be feasibly integrated into routine child growth monitoring and may facilitate population-wide early obesity risk assessment in primary health care.

    JMIR Public Health Surveill 2025;11:e69220

    doi:10.2196/69220

    Keywords

    infants; rapid weight gain; risk prediction; machine learning; childhood obesity; pooled analysis 


    Overweight and obesity are an intractable global challenge with substantial health consequences [1]. Compelling evidence suggests that obesity risk originates early in life and is hard to reverse once developed [2]. World Health Organization (WHO) global estimates indicate that a high prevalence of overweight and obesity is already evident in children younger than 5 years, with 37 million cases being reported in 2022 [3]. The WHO has thus recognized “Lowering the prevalence of overweight in children under the age of five” as one of the key outcome targets in the acceleration plan to stop obesity [4]. Identification of obesity risk from early life along with timely delivery of tailored interventions, are therefore strongly recommended to achieve effective and sustained impact.

    Extensive research has attempted to develop obesity risk prediction models to identify high-risk populations [5], mostly in older children and adults when obesity and modifiable risk factors are already established [6,7]. Despite some models having been developed to predict early childhood obesity risk, these models are largely derived from a single cohort using statistical approaches and are often constrained to linear relationship assumptions, resulting in suboptimal model prediction accuracy and low generalisability [8]. A key consideration for developing clinically useful risk prediction models is clinical impact and ease of clinical integration [9]. It is not possible to evaluate the clinical impact of these models due to the costs and ethical reasons associated with long-term follow-up and allowing children to develop obesity without intervention. Moreover, widespread obesity stigma and the reluctance to label children as having obesity further impede the clinical integration of existing obesity risk models [10,11]. Thus, novel approaches to facilitate early obesity risk identification, communication, and model integration in clinical practice are needed.

    Convincing evidence suggests that accelerated physical growth in early life is highly predictive of later obesity and cardiometabolic outcomes [12,13]. Several systematic reviews concluded that rapid weight gain (RWG) from birth to the age of 1 year, defined as the upward centile crossing in a weight growth chart, was associated with a fourfold higher risk of later obesity than those without infant RWG [14,15]. As such, infant RWG is considered a sensitive proxy marker that denotes future obesity risk. A risk prediction model to predict infant RWG could facilitate early obesity risk identification from infancy and the timely delivery of interventions before adverse health behaviors are established. Furthermore, such a model could be easily evaluated and integrated in primary health care where infant growth is already routinely monitored. To improve risk prediction accuracy, we propose the use of machine learning (ML), a type of artificial intelligence technique that is powerful at making accurate predictions. Reviews have indicated that ML models outperform statistical models for obesity risk prediction [16,17]. Using data from 7 Australian and New Zealand cohorts, we used ML to develop and validate risk prediction models to identify infants at risk of experiencing RWG by the age of 1 year.


    Data Source and Variables

    Data from 7 Australian and New Zealand cohorts: Barwon Infant Study (BIS) [18], Healthy Beginnings [19,20], Infant Feeding Activity and Nutrition Trial (InFANT) [21,22], InFANT Extend [23], LIMIT [24,25], NOURISH [26,27], and Prevention of Overweight in Infancy (POI) were used [28,29]. Baseline recruitment of all studies was undertaken either antenatally (15‐34 wk gestation) or postnatally (between ages 3‐4 m) from 2007 to 2012. The 7 studies include 6423 parent-infant dyads from diverse geographical areas at the study baseline. BIS is a prospective birth cohort. The remaining studies are intervention studies targeting nutrition and physical activity in infants or pregnant women. Detailed study characteristics of the cohorts were published elsewhere [30,31]. None of the intervention studies showed differences in infant growth outcomes between intervention and control groups by the age of 1 year [32], and thus groups were combined to study prognosis [33].

    Rapid Weight Gain During Infancy

    All cohorts collected child anthropometrics at birth and around 1 year of age. Child length and weight at birth were transcribed from child health records. Child length and weight around the age of 1 year were measured by trained staff. WHO growth standards were used to calculate age- and sex-specific weight-for-age z scores at birth and around the age of 1 year, and the difference between the two time points was calculated. RWG during infancy was defined as a change in weight-for-age z score>0.67, which is clinically equivalent to upward crossing one centile line in a weight growth chart [14,15].

    Selection of Predictor Variables

    A range of prenatal and early postnatal factors were collected across all the cohorts. Prenatal factors include maternal country of birth (native-born vs overseas-born), an education level (tertiary vs nontertiary), smoking during pregnancy (smoker vs nonsmoker), and pre or early-pregnancy BMI (kg/m2). Child early postnatal factors included sex (boys vs girls), birth weight (kg), gestational age (weeks), parity (≥1 sibling vs no sibling), any breastfeeding (yes vs no) at age 6 months, and whether introduced solids (yes vs no) at age 6 months.

    The selection of predictor variables for risk prediction model development was informed by previous findings on determinants of infant RWG, the availability of data or ease of data collection at primary health care visits, and the level of data missingness. Our previous statistical findings in the 7 cohorts assessed maternal prepregnancy BMI, smoking during pregnancy, country of birth, gestational age, parity, infant sex, birth weight, breastfeeding, and solids introduction at age 6 months as potential predictors of RWG by the age of 1 year [30], and these were also considered as potential predictors for the current risk prediction model development. Continuous variables capture more nuanced information compared with categorical variables, hence, continuous forms of maternal prepregnancy BMI, gestational age, and birth weight were used for risk prediction whenever possible. Maternal country of birth was not included as a predictor as it had missing data >40%, and missing data imputation would not be reliable. A total of 3 models were purposely chosen. Model 1 is limited to variables that are collected as part of the antenatal and child health records in Australia or can be easily collected at the primary health care visits: maternal prepregnancy BMI, smoking during pregnancy, gestational age, parity, infant sex, and birth weight. Infant feeding at age 6 months is not routinely collected in Australia. However, given the vital contribution of infant feeding, especially breastfeeding, in determining the risk of infant RWG, we conducted 2 additional models to further incorporate the 2 infant feeding variables. Model 2 included model 1 variables plus any breastfeeding at the age of 6 months. Model 3 additionally included solids introduction at age 6 months, upon model 2. The 3 models also enable the comparison of risk prediction accuracy at birth versus at 6 months postnatal.

    Model Development and Validation

    Detailed model development and validation processes are summarised in Supplementary Figure 1 in Multimedia Appendix 1. All ML models were trained and validated using Python 3.11.8 and Scikit Learning Tool Kit library 1.6.1. Data from 7 cohorts were pooled and randomly split into a training dataset (70%) and a testing dataset (30%) [34]. The training dataset was used for model development and internal validation, from which the best-performing model was selected. The testing dataset was reserved for external validation to assess how the best-performing model performed with the new data. Missing data imputation was conducted to impute predictors with missing data: maternal prepregnancy BMI, smoking, gestational age, breastfeeding, and solids introduction. Maternal smoking during pregnancy (yes vs no) was imputed by mode imputation [35], whereby missing values were replaced with the most frequent category observed in the dataset. Median imputation was used to impute missing data on maternal prepregnancy BMI and gestational age. For breastfeeding and solids introduction, we adopted ML algorithms to impute missing values using infant sex, birth weight, and parity as predictors and those with complete data as the training dataset for missing data prediction. Detailed missing data imputation methods are provided in Supplementary Table 1 in Multimedia Appendix 1. All predictor variables were standardized before risk prediction, ensuring all variables were in the same scale to facilitate convergence speed and interpretability.

    For model development in the training dataset, 8 ML algorithms widely used in study prognosis in health research were explored: Logistic Regression, Decision Tree, Random Forest, AdaBoost, Gradient Boosting, K-Neighbours, Support Vector, and Multi-Layer Perceptron [16]. The hyperparameter (configuration setting) of each model were optimized using the grid search cross-validation technique, which systematically finds the best combination of hyperparameter for a ML model [36]. The grid search cross-validation technique reduces overfitting and provides more reliable hyperparameter selection than manual tuning. Detailed model hyperparameters are presented in Supplementary Table 2 in Multimedia Appendix 1. For model training and internal validation, 5-fold cross-validation was conducted, where the training dataset was randomly split into 5 blocks of equal size (fold), and 5 independent iterations of training and internal validation were conducted, so every block of data was used for both training and internal validation [37]. Imbalanced outcome data (ie, higher proportion of infants without RWG and vs with RWG) can result in biased risk prediction models with better performance in predicting infants with no RWG. We used the Synthetic Minority Over-sampling Technique (SMOTE) [38], a commonly used method in ML, to address imbalanced outcome data. SMOTE generates synthetic samples of infants with RWG that are combinations of existing RWG infants, thereby balancing the outcome distribution and improving the prediction of infants with RWG. SMOTE was exclusively performed on the training folds, which prevents data leakage, eliminates the risk of overestimating the model’s performance, and maintains a proper evaluation of the model’s generalisability (Supplementary Figure 2 in Multimedia Appendix 1). The performance metrics of five iterations were averaged to obtain the mean performance metrics of each ML model. The average model performance of 8 ML models was compared, from which the best-performing models were selected for subsequent external validation in the testing dataset.

    The predictive performance of the models against the observed prevalence of infant RWG in both internal and external validation was assessed by various metrics, including receiver operating characteristic (ROC) plot, precision-recall curve (PRC) plot, accuracy, precision, sensitivity, specificity, F1-score, and Cohen κ. A ROC curve plots false positive rate (1-specificity) as the x-axis against true positive rate (sensitivity) as the y-axis at various thresholds (defined as the cutoff value at which the model determines a positive or negative prediction). The area under the ROC curve (AUC) summarizes the model’s overall ability to discriminate between infants with or without RWG across various thresholds. An AUC of 0.70‐0.79 and ≥0.80 is considered acceptable and excellent discrimination, respectively [39]. Accuracy evaluates the overall correctness of the model (both true positives and negatives). Precision examines the model’s ability to predict true positives while minimizing false positives. Sensitivity (recall) assesses the model’s ability to correctly classify true positives. The PRC plots recall (y-axis) against precision (x-axis) over different thresholds, which is particularly useful for a classification model with imbalanced outcome data. The closer the curve toward the top right corner (high precision and high recall) or the larger area under the PRC (AUPRC), the better model performance with >0.7 is generally considered as decent model performance [40]. Likewise, F1-score evaluates the balanced performance of the model that accounts for both precision and sensitivity, which is particularly useful for risk prediction with the outcome that has an uneven distribution. Specificity measures the model’s ability to correctly identify true negatives. Higher values of these metrics indicate better model performance. Cohen κ assesses agreement between predicted and observed outcomes, accounting for agreement by chance, and is also useful for imbalanced outcome data (0.61‐1.0: perfect agreement; 0.41‐0.60 moderate agreement; 0.21‐0.40 fair agreement; ≤0.20 slight or poor agreement). Detailed interpretation and calculation of these metrics are provided in Supplementary Table 3 in Multimedia Appendix 1. To enhance model transparency and interpretability, additional Shapley Additive Explanations (SHAP) analyses were performed to quantify the contribution of each predictor in the risk prediction of infant RWG.

    Ethical Considerations

    Each study received ethical approval from relevant institutional ethics committees and informed written consent was obtained from participants. The original consent covers secondary analysis without additional consent. The current analysis uses deidentified data and received an ethics exemption from the Deakin University Human Research Ethics Committee (2023-033).


    Of 6423 parent-infant dyads from the 7 cohorts at baseline, 81% (n=5233) with all required data were included in analyses. The mean prepregnancy BMI was 27.7 kg/m2. The majority of women did not smoke during pregnancy (89.1%). The mean gestational age and infant birth weight were 39 weeks and 3.5 kg, respectively. The pooled sample included an even proportion of males and females with 64.1% being first-born child. Most infants were breastfed beyond age 6 months (59.5%) and started solids at or after age 6 months (79.9%). Across the 7 cohorts, the mean age (SD) at RWG assessment ranged from 7.0 (1.9) to 13.1 (1) months [30]. The average prevalence of RWG across 7 cohorts spanned from 21.6% to 41.3% [30]. The mean age and prevalence of RWG of 7 cohorts were 10.3 (SD 3) months and 27.4%, respectively (Table 1).

    Table 1. Characteristics of the pooled sample from 7 cohorts (N=5233).
    Maternal and infant characteristicsValues
    Maternal prepregnancy BMI (kg/m2), n; mean (SD)5065; 27.7 (6.5)
    Maternal smoking during pregnancy
    Smoker, n (%)466 (10.9)
    Nonsmoker, n (%)3818 (89.1)
    Gestational age (weeks), n; mean (SD)4185; 39.3 (1.8)
    Infant sex
    Males, n (%)2649 (50.6)
    Females, n (%)2584 (49.4)
    Infant birth weight (kg), n; mean (SD)5233; 3.5 (0.5)
    Parity, n (%)
    No sibling3352 (64.1)
    ≥1 sibling1877 (35.9)
    Breastfeeding duration, n (%)
    <6 months1975 (40.5)
    ≥6 months2904 (59.5)
    Timing of solids introduction, n (%)
    Before 6 months695 (20.1)
    At or after 6 months2615 (79.9)
    Rapid weight gain around the age of 1 year, n (%)
    Yes1433 (27.4)
    No3800 (72.6)

    Internal Validation

    Visual inspection showed that ROC curves for all models (Figure 1) deviated from the diagonal line and were arched toward the top-left corner of the plot, indicating good discriminative ability. All 3 models showed similar AUCs with values ranging from 0.76‐0.85 (model 1), 0.75‐0.86 (model 2), and 0.76‐0.86 (model 3). Among the 8 ML algorithms, the Gradient Boosting classifier showed the best performance in terms of AUC (0.84‐0.86) in all models, suggesting an “excellent” ability to distinguish between those with and without RWG. Other predictive performance metrics of the models are shown in Table 2. The accuracy of models 1, 2, and 3 ranged from 0.69‐0.76, 0.70‐0.77, and 0.70‐0.78, respectively, which indicates that the models correctly predicted the outcomes for 69%‐78% of infants. Likewise, the 3 sets of models demonstrated similar precision of 0.68‐0.75 (model 1), 0.68‐0.76 (model 2), and 0.67‐0.77 (model 3), indicating that the models correctly predicted 67%‐77% of positive cases while showing good ability to minimize false positives. In addition, models showed good ability to identify infants with RWG as indicated by sensitivity of 0.68‐0.79 (model 1), 0.69‐0.80 (model 2), 0.69‐0.80 (model 3). The AUC range of PRC for models 1 to 3 was 0.51‐0.68, 0.53‐0.71, and 0.54‐0.71, respectively (Figure 2). For F1-score, the respective range for all models was 0.68‐0.76 (model 1), 0.69‐0.77 (model 2), and 0.69‐0.78 (model 3), indicating that the models exhibited a good balance between precision and sensitivity and suggesting good overall model performance with low prediction errors. Models 1, 2, and 3 had a specificity of 0.65‐0.75, 0.67‐0.76, and 0.65‐0.78, respectively, indicating a good ability to identify infants with no RWG. The range of Cohen κ for all models was 0.38‐0.53 (model 1), 0.39‐0.55 (model 2), and 0.40‐0.56 (model 3). The addition of infant feeding variables at age 6 months in models 2 and 3 showed 0.01‐0.03 higher increase across all predictive performance metrics when compared with model 1. Consistent with AUC results, the Gradient Boosting classifier showed the best performance in terms of accuracy, precision, sensitivity, AUPRC, F1-score, specificity, and Cohen κ.

    ‎
    Figure 1. Receiver operating characteristics (ROC) curves of eight machine learning algorithms in predicting infants at risk of rapid weight gain by the age of 1 year in the training dataset (internal validation). (A) Model 1 included maternal prepregnancy BMI, smoking during pregnancy, gestational age, parity, infant sex, and birth weight as predictors. (B) Model 2 included model 1 predictors plus any breastfeeding (yes or no) at age 6 months. (C) Model 3 included model 2 predictors plus solids introduction (yes or no) at age 6 months. AUC: area under the receiver operating characteristic curve; MLP: multilayer perceptron; SVC: support vector classifier.
    ‎
    Figure 2. Precision recall curves (PRC) of 8 machine learning algorithms in predicting infants at risk of rapid weight gain by the age of 1 year in the training dataset (internal validation). (A) Model 1 included maternal prepregnancy BMI, smoking during pregnancy, gestational age, parity, infant sex, and birth weight as predictors. (B) Model 2 included model 1 predictors plus any breastfeeding (yes or no) at age 6 months. (C) Model 3 included model 2 predictors plus solids introduction (yes or no) at age 6 months. AUC: area under the receiver operating characteristic curve; MLP: multilayer perceptron; SVC: support vector classifier.
    Table 2. Predictive performance metrics of 8 machine learning algorithms in predicting infants at risk of rapid weight gain by the age of 1 year in the training dataset (internal validation).
    Machine learning algorithmsAccuracyPrecisionSensitivityF1-scoreSpecificityCohen κ
    Model 1a, mean (SD)b
    Logistic regression0.69 (0.01)0.69 (0.03)0.68 (0.04)0.68 (0.01)0.70 (0.03)0.38 (0.02)
    Decision tree classifier0.70 (0.01)0.68 (0.02)0.75 (0.03)0.71 (0.01)0.65 (0.02)0.40 (0.01)
    Random forest classifier0.76 (0.01)0.74 (0.02)0.77 (0.03)0.76 (0.02)0.74 (0.01)0.51 (0.02)
    AdaBoost classifier0.69 (0.01)0.67 (0.03)0.73 (0.03)0.70 (0.01)0.65 (0.04)0.38 (0.02)
    Gradient boosting classifier0.76 (0.01)0.75 (0.02)0.78 (0.03)0.77 (0.02)0.75 (0.02)0.53 (0.03)
    Support vector classifier0.70 (0.01)0.68 (0.02)0.74 (0.02)0.71 (0.01)0.67 (0.02)0.41 (0.02)
    K-neighbors classifier0.73 (0.01)0.70 (0.02)0.79 (0.03)0.74 (0.02)0.68 (0.02)0.47 (0.02)
    Multilayer perception classifier0.71 (0.01)0.69 (0.02)0.74 (0.03)0.71 (0.01)0.68 (0.03)0.42 (0.02)
    Model 2c, mean (SD)
    Logistic regression0.70 (0.01)0.70 (0.02)0.69 (0.02)0.69 (0.01)0.71 (0.02)0.40 (0.01)
    Decision tree classifier0.70 (0.01)0.68 (0.02)0.72 (0.02)0.70 (0.01)0.67 (0.02)0.39 (0.02)
    Random forest classifier0.76 (0.01)0.75 (0.02)0.77 (0.03)0.76 (0.01)0.76 (0.02)0.52 (0.02)
    AdaBoost classifier0.71 (0.01)0.69 (0.03)0.73 (0.03)0.71 (0.01)0.68 (0.04)0.41 (0.02)
    Gradient boosting classifier0.77 (0.01)0.77 (0.02)0.78 (0.02)0.77 (0.01)0.77 (0.02)0.55 (0.01)
    Support vector classifier0.72 (0.01)0.70 (0.03)0.75 (0.02)0.72 (0.02)0.69 (0.02)0.44 (0.03)
    K-neighbors classifier0.75 (0.01)0.72 (0.02)0.80 (0.03)0.76 (0.01)0.70 (0.02)0.50 (0.02)
    Multilayer perception classifier0.73 (0.01)0.71 (0.02)0.74 (0.02)0.73 (0.02)0.71 (0.01)0.46 (0.02)
    Model 3d, mean (SD)
    Logistic regression0.70 (0.01)0.70 (0.02)0.69 (0.02)0.69 (0.01)0.71 (0.02)0.40 (0.01)
    Decision tree classifier0.70 (0.01)0.67 (0.02)0.74 (0.02)0.71 (0.01)0.65 (0.01)0.40 (0.01)
    Random forest classifier0.77 (0.00)0.76 (0.02)0.77 (0.02)0.76 (0.01)0.76 (0.02)0.53 (0.01)
    AdaBoost classifier0.71 (0.01)0.69 (0.03)0.74 (0.02)0.71 (0.01)0.68 (0.04)0.41 (0.02)
    Gradient boosting classifier0.78 (0.00)0.77 (0.02)0.79 (0.01)0.78 (0.01)0.77 (0.01)0.56 (0.01)
    Support vector classifier0.71 (0.01)0.69 (0.02)0.75 (0.02)0.72 (0.02)0.68 (0.02)0.43 (0.03)
    K-neighbors classifier0.74 (0.01)0.71 (0.02)0.80 (0.03)0.75 (0.01)0.69 (0.02)0.49 (0.02)
    Multilayer perception classifier0.73 (0.01)0.72 (0.02)0.73 (0.02)0.72 (0.01)0.72 (0.02)0.45 (0.02)

    aModel 1 included maternal prepregnancy BMI, smoking during pregnancy, gestational age, parity, infant sex, and birth weight as predictors.

    bValues are presented as mean (SD) and are calculated as the average performance metrics of five iterations from the 5-fold cross validation

    cModel 2 included model 1 predictors plus breastfeeding (yes or no) at age 6 months.

    dModel 3 included model 2 predictors plus solids introduction (yes or no) at age 6 months.

    External Validation

    The Gradient Boosting classifier was used to predict infant RWG in the testing dataset. The ability of the Gradient Boosting classifier to distinguish between infants with or without RWG was excellent in the testing dataset with ROC AUC of 0.83‐0.86 for 3 models, respectively (Figure 3). The AUPRC of 3 models spanned from 0.70 to 0.73, indicating a decent model performance (Figure 4). All 3 models (model 1, 2, 3) performed well in terms of accuracy (0.75, 0.78, and 0.78), sensitivity (0.75, 0.78, and 0.80) and specificity (0.76, 0.77, and 0.77), but performed modestly for precision (0.57, 0.60, and 0.60), F1-score (0.65, 0.68, and 0.69), and Cohen κ (0.47, 0.51, and 0.52). Consistent with the results of the internal validation, model 2 with the addition of breastfeeding improved the predictive performance of model 1 by 0.01-0.04. However, the performance of models 2 and 3 was similar.

    ‎
    Figure 3. Receiver operating characteristics (ROC) curves of Gradient boosting classifier (identified in the internal validation) in predicting infants at risk of rapid weight gain by the age of 1 year in the testing dataset (external validation). (A) Model 1 included maternal prepregnancy BMI, smoking during pregnancy, gestational age, parity, infant sex, and birth weight as predictors. (B) Model 2 included model 1 predictors plus any breastfeeding (yes or no) at age 6 months. (C) Model 3 included model 2 predictors plus solids introduction (yes or no) at age 6 months. AUC: area under the curve.
    ‎
    Figure 4. Precision recall curves (PRC) of Gradient Boosting classifier (identified in the internal validation) in predicting infants at risk of rapid weight gain by the age of 1 year in the testing dataset (external validation). (A) Model 1 included maternal prepregnancy BMI, smoking during pregnancy, gestational age, parity, infant sex, and birth weight as predictors. (B) Model 2 included model 1 predictors plus any breastfeeding (yes or no) at age 6 months. (C) Model 3 included model 2 predictors plus solids introduction (yes or no) at age 6 months.

    Additional Analyses

    The SHAP analyses revealed that birth weight was the most influential factor underlying the prediction of infant RWG, followed by breastfeeding, gestational age, and maternal prepregnancy BMI. In contrast, parity, infant sex, timing of solids introduction, and maternal smoking had relatively less impact on the prediction (Supplementary Figure 3 in Multimedia Appendix 1).


    Principal Findings

    Harnessing high-quality prenatal and early postnatal data from 7 Australian and New Zealand cohorts, we developed the first ML models to predict the risk of infant RWG by the age of 1 year (a proxy marker for later obesity risk) using only measures routinely or feasibly collected in primary health care. Our models enable risk prediction in infancy from birth to age 6 months, a critical window to intervene before the development of modifiable obesity-contributing behaviors such as unhealthy dietary intake and low physical activity. The resulting models demonstrated acceptable performance in predicting the risk of infant RWG.

    Given the strong association between infant RWG and elevated obesity risk [12-15], early identification of infants at risk of RWG is a crucial step toward early obesity prevention. While research has assessed factors associated with infant RWG [30,41], here we present the first risk prediction models to predict RWG by the age of 1 year. The Gradient Boosting classifier exhibited excellent discrimination (>80%) to distinguish between infants with or without RWG, and performed well in making accurate predictions, particularly true positive cases (with accuracy and sensitivity all exceeding 75% confidence). However, the performance of the model in minimizing false positives was modest as indicated by accuracy, F1-score, AUPRC, and Cohen κ. Notably, the addition of breastfeeding at age 6 months in risk prediction (model 2) improved model predictive accuracy by 3%‐4% across all performance metrics. However, the inclusion of solids introduction showed minimal impact on model performance. This suggests that acceptable risk prediction can be obtained at birth with very few routinely collected prenatal and birth variables, but the addition of postnatal breastfeeding would further enhance the risk prediction.

    Comparison With Previous Work

    Several statistically based risk prediction models have been developed to identify early obesity risk at ages 2‐5 years using predictors collected from birth to age 2 years among children from the United Kingdom [42-45], the United States [46], Israel [47], and New Zealand [48]. Furthermore, various ML-based early childhood obesity risk prediction models have also been developed [49-51]. However, the clinical utility of these models is questionable. For example, the statistical-based models were often derived from a small or a single dataset with external validation being conducted in only 1 or 2 datasets [42-44,48], both of which limit the model predictive accuracy and generalizability. For previous ML-based childhood obesity risk prediction models [49-51], although large eHealth data were used, these models are not clinically feasible as they require over a hundred predictors for risk prediction, undermining utility in primary health care settings.

    Studies to evaluate the real-world impact of risk prediction models are the essential next step in the translation pathways [33]. There are, however, important economic, ethical, and clinical challenges for impact evaluation of existing obesity risk prediction models [7]. We proposed a novel and pragmatic solution by using infant RWG as a more acceptable and implementable alternative marker for obesity risk. Apart from systematic reviews [14,15], childhood obesity risk prediction models also consistently found that infant RWG is the most influential predictor of obesity outcomes [43,44,46], providing further evidence to support infant RWG as a valid infancy marker of future obesity risk. Our previous analysis in the current 7 cohorts showed that infant RWG was associated with 4.5 times increased risk of child overweight or obesity by the age of 5 years [31]. This translates to a population attributable fraction of 48.6%, suggesting that about half of child overweight or obesity were attributed to infant RWG with the prevalence of 27%.

    Predicting the risk of RWG in infancy offers a pragmatic and low-cost solution for conducting ethically acceptable impact studies to assess clinical impact, as child growth indicators are routinely collected. Our models can be feasibly integrated into routine growth monitoring. Clinicians providing infant care are familiar with the routine use of risk prediction instruments. For example, the neonatal sepsis calculator has been successfully implemented around the world to improve antibiotic stewardship over the past decade [52]. Assessing the risk of infant RWG will raise the awareness of both primary health care professionals and parents about the vital role of early growth in shaping later child health. Research has shown that primary health professionals require support to monitor growth effectively despite growth monitoring being part of the routine infant care [53]. Moreover, both health professionals and parents have poor awareness on the concept and adverse health impact of infant RWG [10,54]. Embedding risk assessment of RWG in routine growth monitoring may motivate families with infants to make behavioral changes and establish positive health behaviors from the beginning of life.

    Past studies primarily focused on exploring factors associated with childhood obesity but not infant RWG. Our study is novel in leveraging SHAP analyses to understand the relative contribution of prenatal and postnatal factors in infant RWG prediction. Our finding that birth weight was the most influential factor contributing to infant RWG prediction is supported by postnatal catch-up growth in body weight commonly observed in low-birth-weight infants [55]. It is well known that birth weight is influenced by a range of maternal factors from preconception to pregnancy such as maternal prepregnancy BMI, gestational weight gain, maternal health status, maternal nutrition, and other lifestyle behaviors during pregnancy [56]. Furthermore, our analyses showed that maternal prepregnancy BMI, gestational age, and breastfeeding also played a crucial role in infant RWG prediction, which further highlights the importance of targeting women during preconception and pregnancy to prevent infant RWG.

    Leveraging statistical findings, ML, and high-quality cohort data for risk prediction model development is a major strength of the current study, ensuring predictive accuracy, generalisability, and clinical interpretability of the resulting risk prediction models. Statistics focuses on inference, whereas ML focuses on making accurate predictions, it has been increasingly advocated that statistics and ML could be combined to build reliable and clinically useful risk prediction models [57]. The use of statistical inferences to guide our predictor selection improves the clinical interpretability and utility of ML based risk prediction models. On the other hand, ML enables accurate predictions by discovering nuanced relationships between variables and data dimensions beyond “linear” relationship that is often assumed in statistical prediction. Another strength of our study was the use of diverse high-quality data from 7 cohorts, which enhances the model generalisability. We also considered clinical integration during model development by purposely selecting predictor variables readily collectable in primary health care, promoting future clinical integration. Moreover, ML-based risk prediction models can be easily transformed into digital tools that allow automatic risk prediction and convenient integration with electronic management systems in primary health care. Self-assessment by families with newborns or infants is also possible.

    Limitations

    Notwithstanding, the limitations of our study should be noted. Some predictors had high levels of missing data. Although missing data imputation was conducted, further model validation with high quality data will be desirable to further enhance the predictive accuracy, generalisability, and clinical utility of our models. Moreover, further research incorporating eHealth records or clinical data for model calibration is needed before clinical integration. However, notably, the use of high-quality cohort data also has its benefit in building accurate models. Our models will serve as a foundation for future validation and calibration with more robust cohort and clinical data. The prevalence of infant RWG across 7 cohorts is comparable to those reported in previous research [14,15]. However, the utility of the resulting model in countries other than Australia and New Zealand requires further investigation. Finally, our model did not capture other genetic or environmental predictors of infant RWG, but it is important to note that we aim to build simple and practical models for clinical integration instead of complex models with superior predictive accuracy but limited clinical utility.

    Conclusions

    We developed the first set of ML-based risk prediction models to identify infant RWG, a potent proxy marker denoting obesity risk later in life. The resulting models using routinely collected prenatal and postnatal factors showed good predictive accuracy as evidenced by various predictive performance metrics with gradient boosting classifiers exhibiting the best performance. Our next step is to further calibrate the resulting models using routine clinical data and convert our ML models into a digital tool for use in primary health care or enable self-assessment for families with newborn or infants. Future clinical trials will be conducted to evaluate the real-world utility of our models in identifying RWG and assess the impact of our models along with early interventions in prevention of infant RWG and, in turn, long-term overweight, and obesity.

    Data Availability

    The datasets analyzed during this study are not publicly available due to ethical restrictions but are available from the corresponding authors along with relevant study chief investigators on reasonable request.

    Authors' Contributions

    MZ conceived the study and wrote the original draft. YZ developed the study methodology and conducted the formal analysis. KDH, RL, PV, JD, LMW, LAB, RT, RB, and ALP collected and curated the data. All authors critically reviewed and approved the final version of the paper. MZ obtained financial support for the project and MZ is supported by Australian Research Council Discovery Early Career Researcher Award (DE240100635).

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    Additional tables and figures.

    DOCX File, 495 KB
    1. Report of the commission on ending childhood obesity. World Health Organisation; 2016. URL: https://www.who.int/publications/i/item/9789241510066 [Accessed 2025-06-10]
    2. Simmonds M, Llewellyn A, Owen CG, et al. Predicting adult obesity from childhood obesity: a systematic review and meta-analysis. Obes Rev. Feb 2016;17(2):95-107. [CrossRef] [Medline]
    3. Obesity and overweight. World Health Organization. 2024. URL: https://www.who.int/news-room/fact-sheets/detail/obesity-and-overweight [Accessed 2025-05-19]
    4. WHO acceleration plan to stop obesity. World Health Organization. 2023. URL: https://www.who.int/publications/i/item/9789240075634 [Accessed 2025-05-31]
    5. Gou H, Song H, Tian Z, et al. Prediction models for children/adolescents with obesity/overweight: a systematic review and meta-analysis. Prev Med. Feb 2024;179:107823. [CrossRef] [Medline]
    6. Seyednasrollah F, Mäkelä J, Pitkänen N, et al. Prediction of adulthood obesity using genetic and childhood clinical risk factors in the cardiovascular risk in Young Finns study. Circ Cardiovasc Genet. Jun 2017;10(3):e001554. [CrossRef] [Medline]
    7. Butler ÉM, Derraik JGB, Taylor RW, et al. Prediction models for early childhood obesity: applicability and existing Issues. Horm Res Paediatr. 2018;90(6):358-367. [CrossRef] [Medline]
    8. Butler ÉM, Derraik JGB, Taylor RW, et al. Childhood obesity: how long should we wait to predict weight? J Pediatr Endocrinol Metab. Apr 25, 2018;31(5):497-501. [CrossRef] [Medline]
    9. Redsell SA, Rose J, Weng S, et al. Digital technology to facilitate proactive assessment of obesity risk during Infancy (ProAsk): a feasibility study. BMJ Open. Sep 6, 2017;7(9):e017694. [CrossRef] [Medline]
    10. Redsell SA, Atkinson P, Nathan D, et al. Parents’ beliefs about appropriate infant size, growth and feeding behaviour: implications for the prevention of childhood obesity. BMC Public Health. Nov 18, 2010;10:711. [CrossRef] [Medline]
    11. Redsell SA, Atkinson PJ, Nathan D, et al. Preventing childhood obesity during infancy in UK primary care: a mixed-methods study of HCPs’ knowledge, beliefs and practice. BMC Fam Pract. Jun 23, 2011;12:54. [CrossRef] [Medline]
    12. Singhal A. Long-term adverse effects of early growth acceleration or catch-up growth. Ann Nutr Metab. 2017;70(3):236-240. [CrossRef] [Medline]
    13. Arisaka O, Ichikawa G, Koyama S, et al. Childhood obesity: rapid weight gain in early childhood and subsequent cardiometabolic risk. Clin Pediatr Endocrinol. 2020;29(4):135-142. [CrossRef] [Medline]
    14. Zheng M, Lamb KE, Grimes C, et al. Rapid weight gain during infancy and subsequent adiposity: a systematic review and meta-analysis of evidence. Obes Rev. Mar 2018;19(3):321-332. [CrossRef] [Medline]
    15. Ong KK, Loos RJF. Rapid infancy weight gain and subsequent obesity: systematic reviews and hopeful suggestions. Acta Paediatr. Aug 2006;95(8):904-908. [CrossRef] [Medline]
    16. An R, Shen J, Xiao Y. Applications of artificial intelligence to obesity research: scoping review of methodologies. J Med Internet Res. Dec 7, 2022;24(12):e40589. [CrossRef] [Medline]
    17. Colmenarejo G. Machine learning models to predict childhood and adolescent obesity: a review. Nutrients. Aug 16, 2020;12(8):2466. [CrossRef] [Medline]
    18. Vuillermin P, Saffery R, Allen KJ, et al. Cohort profile: the Barwon infant study. Int J Epidemiol. Aug 2015;44(4):1148-1160. [CrossRef] [Medline]
    19. Wen LM, Baur LA, Simpson JM, et al. Effectiveness of an early intervention on infant feeding practices and “tummy time”: a randomized controlled trial. Arch Pediatr Adolesc Med. Aug 2011;165(8):701-707. [CrossRef] [Medline]
    20. Wen LM, Baur LA, Simpson JM, et al. Effectiveness of home based early intervention on children’s BMI at age 2: randomised controlled trial. BMJ. Jun 26, 2012;344:e3732. [CrossRef] [Medline]
    21. Campbell K, Hesketh K, Crawford D, et al. The INFANT Feeding Activity and Nutrition Trial (INFANT) an early intervention to prevent childhood obesity: cluster-randomised controlled trial. BMC Public Health. Mar 31, 2008;8:103. [CrossRef] [Medline]
    22. Campbell KJ, Lioret S, McNaughton SA, et al. A parent-focused intervention to reduce infant obesity risk behaviors: a randomized trial. Pediatrics. Apr 2013;131(4):652-660. [CrossRef] [Medline]
    23. Campbell KJ, Hesketh KD, McNaughton SA, et al. The extended infant feeding, activity and nutrition trial (InFANT Extend) program: a cluster-randomized controlled trial of an early intervention to prevent childhood obesity. BMC Public Health. Feb 18, 2016;16:166. [CrossRef] [Medline]
    24. Dodd JM. Dietary and lifestyle advice for pregnant women who are overweight or obese: the LIMIT randomized trial. Ann Nutr Metab. 2014;64(3-4):197-202. [CrossRef] [Medline]
    25. Dodd JM, Turnbull D, McPhee AJ, et al. Antenatal lifestyle advice for women who are overweight or obese: LIMIT randomised trial. BMJ. Feb 10, 2014;348:g1285. [CrossRef] [Medline]
    26. Daniels LA, Magarey A, Battistutta D, et al. The NOURISH randomised control trial: positive feeding practices and food preferences in early childhood - a primary prevention program for childhood obesity. BMC Public Health. Oct 14, 2009;9:387. [CrossRef] [Medline]
    27. Daniels LA, Mallan KM, Battistutta D, Nicholson JM, Perry R, Magarey A. Evaluation of an intervention to promote protective infant feeding practices to prevent childhood obesity: outcomes of the NOURISH RCT at 14 months of age and 6 months post the first of two intervention modules. Int J Obes (Lond). Oct 2012;36(10):1292-1298. [CrossRef] [Medline]
    28. Taylor BJ, Heath ALM, Galland BC, et al. Prevention of Overweight in Infancy (POI.nz) study: a randomised controlled trial of sleep, food and activity interventions for preventing overweight from birth. BMC Public Health. Dec 19, 2011;11:942. [CrossRef] [Medline]
    29. Taylor RW, Heath ALM, Galland BC, et al. Three-year follow-up of a randomised controlled trial to reduce excessive weight gain in the first two years of life: protocol for the POI follow-up study. BMC Public Health. Aug 11, 2016;16(1):771. [CrossRef] [Medline]
    30. Zheng M, Hesketh KD, Vuillermin P, et al. Determinants of rapid infant weight gain: a pooled analysis of seven cohorts. Pediatr Obes. Oct 2022;17(10):e12928. [CrossRef] [Medline]
    31. Zheng M, Hesketh KD, Vuillermin P, et al. Understanding the pathways between prenatal and postnatal factors and overweight outcomes in early childhood: a pooled analysis of seven cohorts. Int J Obes (Lond). Jul 2023;47(7):574-582. [CrossRef] [Medline]
    32. Askie LM, Espinoza D, Martin A, et al. Interventions commenced by early infancy to prevent childhood obesity-the EPOCH collaboration: an individual participant data prospective meta-analysis of four randomized controlled trials. Pediatr Obes. Jun 2020;15(6):e12618. [CrossRef] [Medline]
    33. Moons KGM, Royston P, Vergouwe Y, et al. Prognosis and prognostic research: what, why, and how? BMJ. Feb 23, 2009;338:b375. [CrossRef] [Medline]
    34. Kohavi R. A study of cross-validation and bootstrap for accuracy estimation and model selection. Presented at: International Joint Conference on Artificial Intelligence; Jul 2009:1137-1145; São Paulo, Brazil.
    35. Zhang Z. Missing data imputation: focusing on single imputation. Ann Transl Med. 2016;4(1):9. [CrossRef]
    36. Belete DM, Huchaiah MD. Grid search in hyperparameter optimization of machine learning models for prediction of HIV/AIDS test results. Int J Comput Appl. Sep 2, 2022;44(9):875-886. [CrossRef]
    37. Sampath R, Indumathi J. Earlier detection of Alzheimer disease using N-fold cross validation approach. J Med Syst. Oct 2, 2018;42(11):217. [CrossRef] [Medline]
    38. Chawla NV, Bowyer KW, Hall LO, et al. SMOTE: Synthetic minority over-sampling technique. J Artif Intell. 2002;16:321-357. [CrossRef]
    39. Alba AC, Agoritsas T, Walsh M, et al. Discrimination and calibration of clinical prediction models: Users’ guides to the medical literature. JAMA. Oct 10, 2017;318(14):1377-1384. [CrossRef] [Medline]
    40. Saito T, Rehmsmeier M. The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS One. 2015;10(3):e0118432. [CrossRef] [Medline]
    41. Mihrshahi S, Battistutta D, Magarey A, Daniels LA. Determinants of rapid weight gain during infancy: baseline results from the NOURISH randomised controlled trial. BMC Pediatr. Nov 7, 2011;11:99. [CrossRef] [Medline]
    42. Levine RS, Dahly DL, Rudolf MCJ. Identifying infants at risk of becoming obese: can we and should we? Public Health. Feb 2012;126(2):123-128. [CrossRef] [Medline]
    43. Santorelli G, Petherick ES, Wright J, et al. Developing prediction equations and a mobile phone application to identify infants at risk of obesity. PLoS One. 2013;8(8):e71183. [CrossRef] [Medline]
    44. Weng SF, Redsell SA, Nathan D, et al. Estimating overweight risk in childhood from predictors during infancy. Pediatrics. Aug 2013;132(2):e414-e421. [CrossRef] [Medline]
    45. Ziauddeen N, Wilding S, Roderick PJ, et al. Predicting the risk of childhood overweight and obesity at 4-5 years using population-level pregnancy and early-life healthcare data. BMC Med. May 11, 2020;18(1):105. [CrossRef] [Medline]
    46. Robson JO, Verstraete SG, Shiboski S, et al. A risk score for childhood obesity in an urban Latino cohort. J Pediatr. May 2016;172:29-34. [CrossRef] [Medline]
    47. Rossman H, Shilo S, Barbash-Hazan S, et al. Prediction of childhood obesity from nationwide health records. J Pediatr. Jun 2021;233:132-140. [CrossRef] [Medline]
    48. Butler ÉM, Pillai A, Morton SMB, et al. A prediction model for childhood obesity in New Zealand. Sci Rep. Mar 18, 2021;11(1):6380. [CrossRef] [Medline]
    49. Dugan TM, Mukhopadhyay S, Carroll A, et al. Machine learning techniques for prediction of early childhood obesity. Appl Clin Inform. 2015;6(3):506-520. [CrossRef] [Medline]
    50. Lingren T, Thaker V, Brady C, et al. Developing an algorithm to detect early childhood obesity in two tertiary Pediatric Medical Centers. Appl Clin Inform. Jul 20, 2016;7(3):693-706. [CrossRef] [Medline]
    51. Hammond R, Athanasiadou R, Curado S, et al. Predicting childhood obesity using electronic health records and publicly available data. PLOS ONE. 2019;14(4):e0215571. [CrossRef] [Medline]
    52. Deshmukh M, Mehta S, Patole S. Sepsis calculator for neonatal early onset sepsis - a systematic review and meta-analysis. J Matern Fetal Neonatal Med. Jun 2021;34(11):1832-1840. [CrossRef] [Medline]
    53. Rossiter C, Cheng H, Denney-Wilson E. Primary healthcare professionals’ role in monitoring infant growth: a scoping review. J Child Health Care. Dec 2024;28(4):880-897. [CrossRef] [Medline]
    54. Pesch MH, Levitt KJ, Danziger P, et al. Pediatrician’s beliefs and practices around rapid infant weight gain: a qualitative study. Glob Pediatr Health. 2021;8:2333794X21992164. [CrossRef] [Medline]
    55. Martin A, Connelly A, Bland RM, et al. Health impact of catch-up growth in low-birth weight infants: systematic review, evidence appraisal, and meta-analysis. Matern Child Nutr. Jan 2017;13(1). [CrossRef] [Medline]
    56. Yu Z, Han S, Zhu J, et al. Pre-pregnancy body mass index in relation to infant birth weight and offspring overweight/obesity: a systematic review and meta-analysis. PLoS ONE. 2013;8(4):e61627. [CrossRef]
    57. Finlayson SG, Beam AL, van Smeden M. Machine learning and statistics in clinical research articles-moving past the false dichotomy. JAMA Pediatr. May 1, 2023;177(5):448-450. [CrossRef] [Medline]

    ‎
    AUC: area under the receiver operating characteristic curve
    AUPRC: area under the precision-recall curve
    BIS: Barwon Infant Study
    InFANT: Infant Feeding Activity and Nutrition Trial
    ML: machine learning
    POI: Prevention of Overweight in Infancy
    PRC: precision-recall curve
    ROC: receiver operating characteristic
    RWG: rapid weight gain
    SHAP: Shapley Additive Explanations
    SMOTE: Synthetic Minority Over-sampling Technique
    WHO: World Health Organization

    Edited by Onicio Leal Neto; submitted 25.11.24; peer-reviewed by Chyntia Mayadewi, Takeshi Hase; final revised version received 18.03.25; accepted 25.03.25; published 18.06.25.

    Copyright

    © Miaobing Zheng, Yuxin Zhang, Rachel A Laws, Peter Vuillermin, Jodie Dodd, Li Ming Wen, Louise A Baur, Rachael Taylor, Rebecca Byrne, Anne-Louise Ponsonby, Kylie D Hesketh. Originally published in JMIR Public Health and Surveillance (https://publichealth.jmir.org), 18.6.2025.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Public Health and Surveillance, is properly cited. The complete bibliographic information, a link to the original publication on https://publichealth.jmir.org, as well as this copyright and license information must be included.