Introduction

Asthma is a chronic lung disease portrayed by deadly asthma attacks that result in wheezing, breathlessness, and chest tightness [1]. Compared to those without asthma, children and adolescents with asthma had additional $3,362.56 in annual medical expenditures [2]. Currently, the only protective measures to combat asthma and asthma attacks are medications and avoidance of triggers [3,4]. The American Lung Association identified several factors that may cause asthma, such as family history, occupational exposure, and smoking [5]. Furthermore, studies have shown that 80% of asthma cases arise during the first six years of a person’s life [6]. Due to the lack of a cure, it is crucial that asthma development in youth can be predicted and detected early so that prevention and early intervention may take place.

Recent studies on asthma epidemiology have identified some causes of asthma that align with those identified by the American Lung Association and observed the differences in asthma prevalence in different age groups [6,7]. One study on the epidemiology of asthma in children and adults found that asthma is more prevalent in boys among children while among adults asthma is more prevalent in women, which is triggered by similar causes, namely, family history, occupational exposures, and environmental triggers [7]. Another study revealed that childhood asthma has more common remission compared to adults, and asthma mortality rates are higher in adolescents and young adults compared to young children and older adults [6]. The higher asthma mortality rate among adolescents underscores the need to address asthma with preventative measures.

The recent development of artificial intelligence techniques (especially machine learning and deep learning models), as well as the data availability of detailed patient information, empowers the prediction and risk assessment of various chronic diseases. For example, machine learning has been employed in predicting cardiovascular diseases [8–10]. The use of machine learning and artificial intelligence in asthma diagnosis has been touched upon and strongly suggested to improve clinical practices, especially in low-income areas [11]. Numerous studies utilized clinical data to diagnose asthma patients using simple machine learning models such as support vector machines, random forests, decision trees, and logistic regression [12–14]. These studies only had access to datasets with hundreds of patients while few had access to datasets with more than 1,000 patients [12,15]. While many machine learning models have been successfully built for predicting asthma exacerbations among asthmatic patients [16–24], fewer studies have been conducted on predicting asthma risk using various datasets especially publicly available data, such as the 2019 Michigan BRFSS data and a small dataset of 202 children from Ibn Sina Hospital Center in Morocco [8,9]. In these studies, the dataset was either too small, limited to a specific region, or relied heavily on clinical factors, which may limit model performance and implementation [9]. This is evident by the relatively poor model performance in the 2019 Michigan BRFFS data with the highest AUC score of 0.63 for the logistic regression. These predictive models’ performance could be improved by using a larger dataset and employing more advanced data processing and machine learning techniques.

This study aims to build predictive models for asthma development in youth using the combined National Health Interview Survey (NHIS) 2021 and 2022 survey data, which ensures a large sample size. Furthermore, we employed different sampling techniques, including oversampling, undersampling, and both, to address the class imbalance issue (most survey subjects had no asthma). Two complementary methods, Least Absolute Shrinkage and Selection Operator (LASSO) and random forest models, were used to identify factors associated with asthma development in youth. Finally, different machine learning models, such as logistic regression and neural networks, were used for predicting asthma development in youth, and their model performances were compared.

Materials and methods

Results

Predictive models for asthma development in youth

Table 1 summarizes the performances of the five predictive models on the 2021 and 2022 tested data, including random forest, XGBoost, neural networks, logistic regression, and support vector machine (SVM) with the linear kernel. Trained with the original unbalanced data without any sampling techniques, XGBoost showed the best model performance with AUC and F1 scores of 0.6796 and 0.2022. In contrast, while it has the highest accuracy, SVM (linear) had no precision and sensitivity, especially the F1 score, indicative of poor model performance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Model performance measures of machine learning models for asthma with different sampling techniques (Only the best performances are shown).

https://doi.org/10.1371/journal.pone.0336591.t001

We applied the undersampling technique to address the class imbalance issue (only 10.3% of subjects with asthma) in the dataset, which randomly removed samples from the major class (subjects without asthma) to match the number of subjects as the minor class (subjects with asthma). As shown in Table 1, all machine learning models trained on the balanced dataset with undersampling performed much better than those trained on the original data, as evidenced by the model performance measures, specifically the F1 score (the most appropriate model performance measure for class imbalance). For example, while the AUC and F1 scores for the neural networks model trained on original data are 0.5449 and 0.1735, the model performance was significantly increased when the model was trained on the under-sampled dataset, with the AUC and F1 score of 0.6640 and 0.3430. In addition, trained on the balanced dataset, the logistic regression (AUC = 0.7654 and F1 = 0.3452) and support vector machine (AUC = 0.7600 and F1 = 0.3464) performed well based on their AUC and F1 scores.

To deal with the class imbalance, another sampling technique, oversampling, was applied to oversample the minority group (subjects with asthma). Based on the F1 score and AUC value, as shown in Table 1, the oversampling technique improved the model performance compared to the models trained on the original dataset. For example, the AUC and F1 scores for the neural networks model are 0.6146 and 0.3088. However, the performance of the models trained on the oversampled dataset was poorer than that of those trained on the under-sampled dataset. In addition, the combination of both oversampling and undersampling using SMOTE-Tomek technique resulted into similar model performance with that of those trained on the oversampling technique, but lower than that of those trained on the under-sampled dataset (Table 1).

Out of all the sampling techniques and models, the models trained on the undersampling of the major class yielded the best model performance. In addition, logistic regression and support vector machine were the two models with the best model performance based on the AUC and F1 scores.

To ensure the generalizability, reliability, and robustness of our machine-learning models trained on the 2021 and 2022 data, we utilized the NHIS 2023 dataset for external validation of these models. As shown in Table 2, the performance of each model on the 2023 dataset was similar but lower than the performance of each model tested on the 2021 and 2022 under-sampled dataset. For example, the AUC and F1 scores for logistic regression using 2023 NHIS data were 0.7464 and 0.3223, which are comparable to the model performance on the 2021 and 2022 under-sampled data, 0.7654 and 0.3452.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. External validation of machine-learning models for asthma using 2023 NHIS data.

https://doi.org/10.1371/journal.pone.0336591.t002

To assess differences in predictive model performance across age groups, we applied the models to three cohorts: 0–4, 5–12, and 13–17 years, using the undersampled dataset. S2 Table demonstrated that, for the 2021 and 2022 datasets, all five models achieved higher performance metrics, such as AUC and F1 scores, in the 5–12 age group compared to the 0–4 and 13–17 groups. S3 Table confirmed these findings with the validation dataset from 2023 BRFSS data. These results indicate that predictive accuracy for asthma is the greatest in the 5–12 age group.

Potential risk factors for asthma development in youth

While it is important to predict the risk of asthma development in youth using machine learning models, it is equally important to identify factors that might contribute to asthma development, which can be valuable for early prevention. Using Random Forest and LASSO, we identified two lists of variables that might be the potential risk factors for asthma development. To test how these candidate variables are associated with asthma development, we selected the intersection of variables between two lists to avoid possible false positives and tested for their effects on asthma development. As shown in Table 3, as indicated by the significant P-value (p < .001), the occurrence of asthma is highly correlated with sex, age, the child’s health status, number of times the child visited urgent care in the past twelve months, number of times the child visited the emergency room in the past twelve months, if the child took prescription medication in the past twelve months, received a flu vaccine in the past twelve months, the asthma status of the child’s parents, the poverty level of the child’s family, and the symptoms of COVID-19. Based on the data from all subjects shown in Table 3, males are more likely to develop asthma, children around age 11 are more likely to develop asthma than children around age 9, children with fair to poor health are more prone to asthma, children who visited the emergency room or urgent care more than 3 times are likely to have asthma, children whose parents had asthma are more likely to have asthma.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Summary statistics on the variables identified by both LASSO and random forest.

https://doi.org/10.1371/journal.pone.0336591.t003

Fig 1 is the SHAP plot of the most influential features on the prediction of the logistic regression, which depicts how different variations within each feature impacted the model prediction. In this study, no asthma was represented by 0 and 2 and asthma was represented by 1. Therefore, features with a large positive SHAP value pushed the prediction towards asthma while features with a large negative SHAP value pushed the prediction towards no asthma. As shown in Fig 1, took prescription medication in past 12 months, age, and general health status are the three most influential features in asthma risk prediction. In addition, we utilized the mean SHAP values for all selected variables to examine the contribution of each variable to the asthma prediction. As shown in S1 Table, the variables “took prescription medication in past 12 months”, “age” and “general health status” had the highest magnitude mean SHAP value of 0.094132, 0.076407 and 0.042564. Therefore, these variables are the most important variables in our models contributing to the prediction of asthma.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. SHAP summary plot of top 20 variables in the asthma risk prediction model.

https://doi.org/10.1371/journal.pone.0336591.g001

Discussion

This study utilizes the data from the 2021 and 2022 youth and parent National Health Interview Survey to build machine-learning models to predict asthma development in youth. The NHIS includes questions ranging from pediatric to demographic and socioeconomic information. We linked the youth survey data with their parent survey data to account for the contribution of parents’ information to asthma development in youth. We applied several sampling techniques to deal with the class imbalance issue (most subjects without asthma), including oversampling, undersampling, and both. We built five different machine learning models to predict asthma development in youth. By comparison, predictive models trained on the dataset with the undersampling of the majority class yielded the best model performance overall. Undersampling outperformed no sampling and oversampling by balancing the training dataset without causing overfitting or creating synthetic data points that may not represent real-world scenarios. Among the models tested, the logistic regression and support vector machine with a linear kernel were the best-performing models for predicting asthma development in youth, based on the AUC score and especially the F1 score. This may be due to the relatively small size of the NHIS dataset after pre-processing and the goal of this study being binary classification, in which these two models can perform well. In this study, the relatively poor model performance of neural network and XGBoost were due to underfitting. In contrast, the poor model performance of random forest largely resulted from overfitting. Therefore, in this study, with the NHIS dataset, the simple models, logistic regression and SVM (linear), might be more suitable for predicting the risk of asthma in youth than other more complex machine learning models. In addition, we found that several factors are significantly associated with asthma development in youth, such as age, gender, and family poverty ratio.

In this study, by employing different sampling techniques to address the class imbalance issue, we have successfully built multiple machine learning models to predict asthma development in youth aged 0–17 with good model performance. One previous study used machine learning to predict asthma amongst children aged 7 months to 12 years at a Morocco hospital, which achieved high performances with F1 scores around 0.80 [9]. However, prenatal, perinatal, postnatal, and environmental factors cannot directly apply to the US population and children aged 13–17, which might limit its generalization. A recent study used Canadian Healthy Infant Longitudinal Development birth cohort data to predict pediatric asthma [29]. The dataset was focused on children up to 4 years of age and included family medical history, clinical data, and environmental factors for young children. They used machine-learning models with 1,484 children and 132 variables and achieved great model performance with an AUC of 0.99 when predicting asthma in children at age 4 [29]. However, their model cannot apply to children aged 5–17. Further, their data heavily relied on pediatric and prenatal clinical data extracted from patient health records and some environmental factors, which are difficult to obtain in real life, especially for those with low socioeconomic status. The dataset used in our study is national survey data containing basic demographics on youth aged 0–17, which is easily accessible, making it more practical. In this study, we examined model performance across age groups. The predictive models for asthma performed better for ages 5–12 compared to ages 0–4 and 13–17, which might be due to the relatively larger sample size in the 5–12 age group or the data from the 5–12 age group predicting asthma risk more effectively. The exact reason for this finding is still unknown, which requires further investigation.

In another asthma study conducted on the 2019 Michigan Behavioral Risk Factor Surveillance System (BRFSS) data, the group employed similar pre-processing techniques and used SMOTE and ROSE to deal with the class imbalance in the data [8]. That study also employed similar machine-learning models on their dataset, yielding decent performance and identifying many risk factors. However, their dataset focuses on adults in Michigan rather than youth in the United States. Furthermore, the performance of the models in the Michigan study had the highest AUC score of 0.629 and F1 score of 0.287. In our study, the best model has an AUC score of 0.7632 and an F1 score of 0.3416. Therefore, the predictive models in our study performed much better in predicting asthma development in youth, which might be due to the improved pre-processing techniques and a more encompassing data source in this study. Some of the significant variables identified in the Michigan study were also identified in our study, such as flu vaccine and income. While the Michigan study showed that females have a higher risk for asthma development in adults, our study showed that males have a higher risk for asthma development in youth. This disparity is likely due to the different datasets utilized between this study and the Michigan study. The Michigan study used the Michigan BRFSS data, which only includes participants from Michigan who were 18 years or older. In contrast, our study utilizes a national survey mainly focused on participants younger than 18 years of age.

When conducting external validation of the models using the 2023 NHIS data, we noticed a slight decrease in the model performance compared to the 2021 and 2022 data. One potential reason is that six variables included in the trained models were missing in the 2023 NHIS data, which likely contributed to the slightly lower performance of the models using the 2023 NHIS data. These variables such as “Symptoms of COVID-19” and “Ever dazed or memory gap”, which were excluded in the 2023 dataset, had a relatively high SHAP value, indicating its influence in the model prediction. Nevertheless, similar performance of models (trained on the 2021 and 2022 under-sampled data) on the 2023 NHIS data further validated the robustness and generalization of our machine learning models in predicting the risk of asthma in youth.

Besides building predictive models for asthma development in youth, another aim of this study is to identify potential risk factors for asthma development in youth, which might help with early asthma detection and prevention. Our study has identified several well-known risk factors for asthma development in youth, such as the presence of asthma in the parent, gender, socioeconomic status, and maternal smoking [39–42]. It is well-known that family history is one of the major causes of asthma [5,39]. This could be due to the shared environment between the parents and child that may influence the development of asthma [43]. Our study also found that boys are at a higher risk for asthma development than girls, which is also consistent with previous studies [40,41]. The increased risk for asthma development in boys might be due to the increased allergic inflammation and serum IgE levels in boys, as well as the smaller airway diameter relative to lung volumes in boys than girls [40,41]. Children living in families with lower socioeconomic status have a higher risk of asthma [42]. A previous review has shown that low socioeconomic status could be a risk factor for pediatric asthma as children living in families with low socioeconomic status have greater exposure to indoor and outdoor hazards (such as molds, mice, second-hand smoke, chemicals, and air pollutants) associated with increased risk of asthma [44]. Additionally, maternal smoking was another factor identified that could influence asthma development in both the mother and, to a greater degree, the child [43]. A review study indicated the influence of maternal smoking on childhood asthma that leads to increased asthma risk in children through direct damage on fetal lung development, fetal growth, and neuronal differentiation and later on modulated airway hyperreactivity and reduced lung function [45].

Based on the SHAP plot and values, the most influential variables in predicting asthma risk among youth include those who took prescription medication in the past 12 months, age, and general health status. It is reasonable to observe the significant association between high asthma risk with taking a prescription medication and poor health status. Asthma patients might be taking medication to help with their asthma and have bad general health status, or the individuals who took medication due to bad health conditions might have a higher risk of asthma. We speculate that those who are older among youth might be more likely to be exposed to environmental risks, which leads to a higher risk of asthma. These variables and other identified variables in this study need to be further investigated and confirmed in future studies.

An interesting finding was that more people with asthma took a flu shot in the past 12 months than those who did not. As the Michigan BRFSS study mentioned, this could be due to children with asthma being more prone to receiving the flu shot, as their parents may be aware of the dangers that arise if a child with asthma gets the flu [8]. Another interesting finding is the association of severe COVID-19 symptoms with asthma in youth. Currently, there are mixed results on the association of COVID-19 with asthma. A recent nationwide population-based cohort study with approximately 50 million people in South Korea showed that COVID-19 was associated with an increased risk of asthma onset [46]. A retrospective cohort study on 27,423 children in Children’s Hospital of Philadelphia Care Network ages 1–16 from March 1, 2020, to February 28, 2021, showed that the COVID-19 diagnosis has no significant impact on asthma onset [47]. However, that study did not examine whether severe symptoms of COVID-19 could impact asthma onset in children. Using national survey data, our study is the first to provide the association between severe COVID-19 symptoms and asthma risks in U.S. youth. Our observation that severe COVID-19 symptoms were associated with a higher risk of asthma (25.0% vs. 10.2%) underscores the potential long-term respiratory impacts of viral illness in children. These findings align with emerging evidence of post-viral airway remodeling [48], which deserves further investigation and highlights the importance of ongoing surveillance, preventive strategies, and public health interventions to mitigate pediatric asthma burden. In addition, we observed a modest decline in model performance when the COVID-19 variable was absent from the external validation dataset (NHIS 2023), highlighting its important contribution to asthma prediction in the pediatric population. Evidence from a recent multicenter retrospective cohort study of 274,803 adults in the U.S. Collaborative Network of the TriNetX database, a large real-world clinical dataset, demonstrated a strong association between COVID-19 infection and subsequent asthma onset [49]. Similar validation efforts could be extended to pediatric populations by leveraging real-world clinical health record data to further evaluate the relationship between COVID-19 and asthma onset in children.

Based on easily accessible demographic and family history data, our predictive models have the potential to be implemented in an easily accessible online asthma risk calculator, which would allow any users to obtain a pre-screening before seeking professional medical diagnosis if at high risk of asthma. Furthermore, individuals at high risk of asthma as predicted by our models could be directed to a wearable respiratory monitoring device, which could more accurately diagnose or evaluate the risk of asthma for that individual [50–52]. Incorporating easily accessible, parent-reported variables—such as overall health status, urgent care visits, flu vaccination history, parental asthma history, and prescription medication use—into routine screening could strengthen early identification of children at risk for asthma in real-world settings. Leveraging these low-burden measures can enable large-scale, cost-effective screening approaches and guide resource allocation in high-volume primary care environments.

This study has several limitations. First, the NHIS does not contain every factor potentially associated with asthma, as its purpose is to get a general health overview of the public. This prevents us from identifying more significant variables associated with asthma development in youth. Including those variables might lead to better model performance. Another limitation is the lack of detail in the prescription medication question. The NHIS survey only asked whether the selected child had taken any prescription medication in the past 12 months, without specifying the type. Consequently, we cannot determine whether asthma medications were included among the reported prescriptions and are unable to exclude children taking asthma medications from our machine learning models. Second, undersampling could have removed many records, which might negatively influence the model performance. Third, although the predictive models have relatively better model performance than previous studies, the models need to be further improved, for example, by including more variables. Fourth, due to the cross-sectional nature of the dataset in this study, we are unable to determine the causal effect, the potential risk factors for asthma development in youth identified in this study must be further validated by longitudinal studies. Fifth, our use of LASSO in feature selection may not have been the best option as LASSO is a regression tool. In contrast, the logistic regression model with an L1 penalty is a classification tool, which may be more beneficial for identifying factors associated with asthma. Sixth, a small percentage of overlaps might exist between the 2021 and 2022 samples. However, the NIHS is a cross-sectional national survey study, and the study redesign after 2019 allows data users to pool multiple years of sample together to obtain reliable estimates [53]. Thus, we expect the potential overlaps won’t affect our analysis results. Lastly, employing sampling techniques in our study may lead to less calibrated models, which would undermine the interpretation of model predictions. When extra data becomes available in the future, we will perform the model calibration. The findings from this study need to be validated using an external dataset with similar variables if such a dataset becomes available in the future. More importantly, our results might be subject to self-reported bias due to the recall error in the survey study. For example, parent-reported asthma diagnoses were not validated by spirometry, and prior evidence suggests that approximately 30% of asthma diagnoses may be inaccurate, often reflecting conditions such as viral wheeze or gastroesophageal reflux disease (GERD), which could lead to an overestimation of pediatric asthma prevalence [54]. Future studies incorporating spirometry validation would strengthen diagnostic accuracy and reduce misclassification bias. Additionally, the absence of COVID-19–related variables in our external validation dataset may have introduced further bias into our findings.

Conclusions

Using the combined national survey data and applying different sampling techniques, this study successfully built several machine-learning models for predicting asthma development in youth, which will be very valuable for early screening. Out of these models, the logistic regression model and Support Vector Machine with a linear kernel combined with undersampling showed the best model performance. Furthermore, external validation of our models using the NHIS 2023 dataset shows that our models are generalizable, reliable, and robust. The identification of potential risk factors associated with asthma development in youth (such as general health status and having received flu vaccine in the past 12 months) will provide valuable insights for early asthma prevention.

Supporting information