The role of artificial intelligence in the differential diagnosis of wheezing symptoms in children

Abstract

The incidence of pediatric wheeze is extremely high. Poor control of wheeze in young children affects lung function in adulthood and is closely associated with the occurrence of chronic obstructive pulmonary disease. Substantial efforts worldwide have been aimed at developing methods to identify the etiology of wheezing symptoms as early as possible to aid in early management strategies. However, the diagnosis of childhood wheeze relies heavily on the clinical experience of pediatricians, most of whom lack sufficient training to accurately diagnose children with wheezing symptoms. Artificial intelligence is an approach that may improve general pediatricians’ diagnostic ability for wheezing symptoms by identifying patterns and trends from large and complex clinical datasets. However, few studies have used artificial intelligence to diagnose wheeze in children. Therefore, this review aims to comprehensively assess these studies in this field, analyze their interpretability and limitations, and explore and discuss future research directions in real-world clinical applications.

Main article text

1. INTRODUCTION

Wheeze is one of the most common respiratory symptoms in children [1]. In recent years, the incidence of wheeze in children has steadily increased. Relevant studies have shown that poorly controlled wheeze in children can affect lung function in adulthood and is closely associated with the development of chronic obstructive pulmonary disease (COPD), which can severely affect quality of life, and place a great burden on families and society [2–6]. Given the numerous etiologies of wheeze, identifying the etiology of wheezing symptoms as early as possible is essential to enable early management strategies. However, most general pediatricians in China cannot make an accurate diagnosis because of the lack of standardized training in the diagnosis of wheezing symptoms. Data-driven artificial intelligence (AI) approaches may effectively improve general pediatricians’ differential diagnostic ability for wheezing symptoms by identifying patterns and trends from large and complex clinical datasets. However, to date, few studies have used AI to diagnose wheeze in children. Therefore, this review aims to comprehensively summarize published studies on AI approaches to the diagnosis of pediatric wheeze, and discuss future research directions in clinical applications.

2. THE NEED FOR EARLY IDENTIFICATION OF THE ETIOLOGY OF WHEEZE IN CHILDREN

Wheezing symptoms in children occur at very high prevalence, owing to a decrease in the internal diameter of the lower airways and an increase in airway resistance [7]. Approximately 34% of children younger than 3 years of age have at least one episode of wheeze, and 50% of children younger than 6 years have had wheezing [8]. Asthma, one of the major causes of wheezing, had a prevalence among Chinese urban children of 1.09% in 1990 and 1.97% in 2000, and increased to 3.02% by 2010 [9]. The recent 2018 Global Initiative for Asthma has indicated a global prevalence of asthma of 4.4% in preschool aged children and up to 6.4% in primary school children [10]. The World Health Organization has estimated that approximately 235 million people worldwide will have asthma by 2020.

Multiple factors, such as air pollution, chemical or physical stimuli, viral infections, exercise, body mass index, gastroesophageal reflux and sleep-disordered breathing, can lead to wheeze in children [11]. In addition to asthma, common causes of wheezing include acute bronchiolitis due to viral infections, congenital airway maldevelopment, endotracheal tuberculosis, mediastinal lymph nodes or neoplasms compressing the airway, and bronchial foreign bodies [12]. Recurrent wheeze in infancy is also a risk factor closely associated with pediatric pneumonia [13]. Wheezing can cause shortness of breath, respiratory distress, and even hypoxia and death in children in its acute phase. Recurrent wheeze can lead to chronic airway inflammation that causes airway remodeling and impairs children’s airway development. Imaging techniques are important in distinguishing asthma from asthma-like diseases, and chest radiographs are most commonly used for rule-out diagnosis [14]. Chest high-resolution CT and sinus CT are used as further imaging modalities to identify other non-asthmatic causes closely associated with wheezing, such as diffuse bronchiolitis, sinus disease and pneumothorax. The etiologic diagnosis of wheezing symptoms is extremely complex and requires a combination of disease history, physical examination, imaging examination, laboratory tests and diagnostic treatment to establish an effective differential diagnosis. Meanwhile, China has an extreme shortage of pediatricians and pediatric specialty clinics. According to a survey in a white paper on the status of pediatric resources in China [15], the proportion of pediatric specialty clinics among medical providers in urban areas is only 0.5%, whereas that in rural areas is 0%. Furthermore, most pediatricians lack the necessary training and clinical experience to accurately diagnose wheezing symptoms. However, diagnosis of etiology as early as possible is essential for timely intervention, controlling the disease, and preventing its development into asthma and other severe respiratory conditions.

3. ARTIFICIAL INTELLIGENCE PROVIDES NEW SOLUTIONS FOR THE DIAGNOSIS OF WHEEZING SYMPTOMS IN CHILDREN

3.1 AI research on diagnosis of wheezing symptoms in children

AI has already been applied in several medical fields and yielded excellent results. By combining the clinical experience of pediatric experts and medical big data, machine learning (ML) techniques are being used to build AI diagnostic models for diagnosis of childhood wheezing, which can effectively improve the diagnostic accuracy of pediatricians with low seniority or primary care clinicians. According to the data modality used for modeling, AI based diagnosis of wheezing can be divided into two categories based on clinical medical records or based on medical images.

3.1.1 Diagnosis of wheezing on the basis of clinical medical records

Clinical medical records include the diagnosis and treatment associated with a patient’s onset and development of the disease. These records consist of large amounts of textual information regarding the patient’s condition, such as complaints, symptoms, signs and disease history. Clinical medical records are usually in the form of text and tables. The text records the patient’s physical examination findings, past medical history and other information, in a strictly standardized form. The tables record the patient’s examination data, such as routine blood test results. Existing AI techniques for medical records usually use traditional ML algorithms such as linear models, Bayesian networks, and decision trees with structured data, which are interpretable to a certain extent. Himes et al. [16] have extracted demographic information and disease data from clinical medical records, and explained the correlations and interactions between variables by using a multivariate model of Bayesian networks to explore the clinical factors to predict patients’ asthma developing into COPD. Pennington et al. [17] have evaluated the effects of several methods of defining asthma in medical records on the estimation of the onset of asthma in children 3 years of age and also determined the validity of defining early asthma for the prediction of pediatric asthma. Afzal et al. [18] have proposed a method that can automatically extract information from extensive clinical records and identify children with asthma. Xi et al. [19] have developed a retrospective graphical analysis method to identify informative fields in medical records, extract their correlation with asthma and develop a method for asthma diagnosis through serial combinations of Boolean operators. Quinto et al. [20] have developed a multiple logistic regression model based on information extracted from medical records to establish the relationship between asthma and the severity of obesity in children. In addition, ML has been successfully used to interpret pulmonary function tests associated with the differential diagnosis of obstructive pulmonary disease, and has also shown promising results in other diagnostic examinations, such as breath analysis, lung sound analysis and telemedicine, although those findings were based on limited sample sizes [21]. Nonetheless, transforming existing AI methods from research to real-world practical applications requires interpretability of the diagnostic algorithms, which is currently lacking in the existing literature.

3.1.2 Diagnosis of wheezing on the basis of medical images

Medical imaging, one of the most common diagnostic tools used by physicians, plays an important role in observing lesions and understanding the cause of diseases. For example, three-dimensional volume rendering imaging of airways by using multi-row spiral CT can visualize airway lesions more intuitively and clearly especially when observing their shape and spatial relationship. By proposing a mathematical function that constructs a three-dimensional spatial model of the airway, Tgavalekos et al. [22] have effectively measured the airway width, which not only matched the size and location of ventilation defects, but also identified the causes of inhomogeneous ventilation in asthma. Amaral et al. [23] have used ML methods such as random forest and AdaBoost with a decision tree to improve the accuracy of the diagnosis of airway obstruction in asthma. Schilham et al. [24] have studied an ML-based algorithm for lung nodule detection through finding local maxima in multi-scale Gaussian space to identify candidate boxes of images; identify pixel edge points from large to small to detect the boundary location of nodules; and classify the features extracted by multi-scale Gaussian filters. Yedururi et al. [25] have analyzed the clinical and imaging manifestations of various common tracheal and bronchial diseases in children to develop a systematic approach for their imaging and classification, which may help physicians accurately and effectively diagnose tracheal and bronchial diseases in children. Recent advances in deep learning techniques, specifically models based on convolutional neural networks (CNN), have been widely used in image data analysis. However, research on image-based AI-assisted diagnosis for the differential diagnosis of wheezing symptoms remains in its infancy. Xu et al. [26] have introduced deep CNN transferred multiple instance learning and proposed identification of COPD from CT images. In this work, a pre-trained CNN model was used as a feature extractor to extract the image features for each view of a CT instance, and KNN was used for classification. González et al. [27] have evaluated the performance of CNNs for COPD detection in chest CT by using statistical analysis. Zhang et al. [28] have detected and classified COPD in chest CT by using CNN models. Bharati et al. [29] have developed a hybrid CNN for COPD and asthma detection from X-ray images.

3.2 Limitations of existing AI research on diagnosis of wheezing symptoms and exploratory research on interpretability

Existing AI technologies, particularly deep learning-based methods, have achieved excellent disease diagnostic performance. However, most of the diagnostic process is performed in an end-to-end “black box” mode, thus hindering primary-level pediatricians’ understanding of diagnosis results and detection of diagnosis errors. Interpretability is particularly important for AI clinical diagnosis [30, 31].

3.2.1 Interpretable AI research

In general, two approaches to interpretability are used: model interpretability and inference interpretability. Model interpretability involves understanding how a model behaves, whereas inference interpretability is aimed at demonstrating how the model determines the output for each situation. In both approaches, interpretability can be achieved by displaying symbols (e.g., structured languages or natural language such as logical forms) to explain the model reasoning. Interpretable AI can also be categorized as model-based methods and model-agnostic methods ( Table 1 ). Linear models and Bayesian-based models are typical model-based interpretable models. Model-agnostic models can generally be divided into three types: visual interpretation, agent interpretation and importance ranking interpretation. Visual interpretation explains how the model works by plotting the effect of the prediction results of the model with the numerical changes in features in a certain range. Friedman et al. [32] have proposed partial dependence plots characterizing the marginal effects of features on the model’s predicted outcomes. Goldstein et al. [33] have proposed individual conditional expectation plots visualizing dependencies between features and predictions for each instance in the form of lines. Apley [34] has proposed the accumulated local effects plot describing how feature values affect the prediction on average. In agent interpretation, the target model is interpreted on the basis of relatively more interpretable agent models, such as linear models, which are trained to approximate the output of the target model on the same inputs. Ribeiro et al. [35] have proposed the local interpretable model-agnostic explanation method, in which an agent model is trained on a set of sampled instances. The importance ranking interpretation determines the features that dominate the output of the model by calculating their contributions and ranking them. Typical methods include Shapley Additive Explanations and Model Reliance [36, 37]. The interpretability of deep learning-based models for image data analysis has been achieved primarily by visualization of the relationship between input images and model outputs [38–40]. Zeiler et al. [38] have proposed to visualize the characteristics of high-level neurons through a deconvolution network and explained the learning process of the network. Zhou et al. [41] have designed a category activation mapping (CAM) method to visualize categorical features, observe the activation states of neural networks and analyze their decision-making process. Selvaraju et al. [42] have proposed a gradient-based category activation mapping method (grad-CAM) that combines category activation mapping with guided backpropagation and deconvolution to visualize the fine-grained features of images. The study has also explained generation of classification results, thus making the processing flow of the CNN based AI models clearer and more interpretable. Recently, variants of CAM, including Grad-CAM++, XGrad-CAM, AblationCAM and HiResCAM, have become an active area of visualization-based interpretable research [43–49]. Figure 1 shows an example of visualization of Grad-CAM in chest radiograph. Wagner et al. [40] have improved the interpretability of CNN models by investigating the internal working mechanism of neural networks, generating fine-grained visual interpretations in image space, selectively filtering the gradients in the optimization process, and refining the interpretability at the pixel level.

Table 1 |

Summary of interpretable AI methods

Categorization of interpretable AI methods	Description	References describing applications
Model-based interpretable methods	Linear models and Bayesian-based models are typical model-based interpretable models.	[16, 20]
Model-agnostic interpretable methods	Model-agnostic models can generally be divided into three types: visual interpretation, agent interpretation and importance ranking interpretation.	[32–37, 42–49]
Visual interpretation	Visual interpretation explains how the model works by plotting the effects of the prediction results of the model with the numerical changes in features in a certain range. Typical methods include the following.	[32–34, 42–49]
	Partial dependence plots characterize the marginal effects of features on the model’s predicted outcomes.	[32]
	Individual conditional expectation plots visualize dependencies between features and predictions for each instance in the form of lines.	[33]
	Accumulated local effects plots indicate how feature values affect the prediction, on average.	[34]
	Category activation mapping (CAM) and its variants, including Grad-CAM++, XGrad-CAM, AblationCAM and HiResCAM, have become an active area of visualization-based interpretable research.	[42–49]
Agent interpretation	Agent interpretation interprets the target model based on a relatively more interpretable agent model, such as linear models, which are trained to approximate the output of the target model on the same inputs. The local interpretable model-agnostic explanation method is an agent model trained on a set of sampled instances.	[35]
Importance ranking interpretation	The importance ranking interpretation determines the features that dominate the output of the model by calculating their contributions and ranking them. Typical methods include the Shapley Additive Explanations and Model Reliance.	[36, 37]

Figure 1 |

Visualization of Grad-CAM in a chest radiograph.

The class-discriminative region, i.e., red highlighted area, is concentrated in the lungs, in agreement with clinical experience.

3.2.2 Interpretable AI auxiliary diagnosis in medical and diagnostic research on wheezing

With recent advances in interpretable AI technology, some progress has also been made in interpretable AI in medical and diagnostic research [50–57]. The AI algorithms used in clinical medical records in the current literature mainly reflect their interpretability through traditional ML methods and structured features, whereas the interpretability of the decision-making process of diagnostic algorithms has been poorly explored. Yu et al. [58] have reported a diagnostic method based on information extracted from chief complaints, medical histories and physical examination findings in clinical medical records. Furthermore, they have structured key information extracted from clinical medical records through a rule-based approach and used ML techniques to identify the structured data, thus not only improving the accuracy of identifying asthma in pediatric inpatient setting but also increasing interpretability. Spathis et al. [59] have developed a random forest classifier based diagnostic system for COPD and asthma in 132 representative samples. The model has indicated that the most prominent factors in COPD cases are smoking, forced expiratory volume, age and forced vital capacity, and the most prominent factor in asthma is MEF25–75. Topalovic et al. [60] have developed a decision tree based on pulmonary data to enable automatic interpretation of pulmonary function tests and COPD detection in 968 participants. They have modeled the relationships among indicators (e.g., diffusing capacity for carbon monoxide, forced expiratory volume and forced vital capacity), thus enabling transparency in the decision-making process. In deep learning-based medical image diagnosis, CAM and its variants have been widely used for visual interpretation through generating pixel-level diagnostic models for overall interpretability, thus greatly improving the diagnostic accuracy of AI models [61–65]. Rajaraman et al. [66] have visualized and explained CAM-based deep learning predictions for pneumonia detection in pediatric chest radiographs. Gotkowski et al. [67] have proposed the M3D-CAM method to specifically highlight spatial regions in volumetric medical images.

Recently, multi-modal models that integrate clinical medical records data and medical images have become a focus in clinical diagnosis [68–70]. Through extracting clinical diagnostic information from medical images and using visual attention to search for disease descriptions associated with these images. Zhang et al. [71] have constructed an MDNet model based on an image module and a language module. Their study has established multi-modal associations between medical images and clinical medical records, and illustrated the model’s iterative diagnostic process through an attention mechanism. Yao et al. [72] have applied a logistic regression model for feature selection to clinical data and used a CNN to extract image features from CT images, then established a model by integrating the above two types of features for pulmonary venous obstruction prediction. Yu et al. [73] have proposed a two-stage diagnostic system based on deep learning features extracted from medical records and fine-grained image recognition techniques to identify pediatric respiratory diseases, and achieved excellent results. Furthermore, they have explored a symptom-level interpretability approach to medical record analysis to demonstrate correlations between input and final diagnoses. Yang et al. [74] have explored focal-level interpretability of images and texts by establishing connections between objects and relationships relative to given ground reference representations, and have further used graph convolution to achieve a consistent correspondence in the representation of lesions and text information in images.

3.3 Future development directions in relevant research fields

The demand for readability and reliability are increasing, particularly in the medical field, in which current AI-associated technologies have unsatisfactory interpretability. Therefore, an urgent need exists to explore interpretable implementations for the diagnosis of pediatric wheeze to meet clinical requirements. Several research directions are worthy of exploration in the future.

3.3.1 Improve existing AI algorithms and construct interpretable AI models based on text data

Existing AI diagnosis algorithms based on medical records generally rely on unstructured data, semi-structured data and structured data to capture information and make a disease diagnosis. However, the corresponding algorithms do not explain the cause-and-effect correlations between inputs and outputs, because the interpretability of the AI models has been neglected. Moreover, physicians experience difficulties in understanding the algorithmic diagnostic process from numerous inputs, thus compromising the credibility of AI algorithms. Future studies must improve existing diagnostic algorithms and use gradient-based input and output correlations in the development of interpretable AI models to create a map outlining the relationship between symptoms and diagnostic outcomes.

3.3.2 Investigate an interpretable AI algorithm for diagnosing wheeze in children on the basis of images and correlations between images and text

Medical imaging is a common technical tool with important roles in intuitively understanding disease conditions. The advantage of CNNs in performing image analysis is their use of weighted convolutional kernels to achieve local feature learning and high-level feature extraction by increasing the depth. This process tends to perform analysis from a global perspective, because it hides the joint weighting of local features for diagnostic outcomes. Nonetheless, difficulties arise in explaining the causal relationship between lesions and diagnostic outcomes with image-based complementary diagnostic techniques, and establishing the correlation between image details and diagnostic outcomes. Consequently, low-ranking and primary pediatricians cannot easily directly use the conclusions provided by AI diagnostics. Furthermore, to fully use image information and deeply understand the correlation between changes in the state of an organ for any given location and diagnostic result, AI models interpretable at the level of image pixels are needed to build the mapping relationship between feature maps and images, and to achieve interpretable analysis of overall information.

Therefore, interpretable AI algorithms based on image and text correlations are needed to reveal the reasoning process linking image blocks and diagnostic conclusions, and to provide output understandable to physicians. Visualization methods of feature maps must also be explored to support the interpretability studies and make the data analysis more easier.

3.3.3 Develop personalized wheezing-associated disease predictive or diagnostic models for children by using AI methods from multi-domain candidate predictors

Selecting and evaluating clinical and imaging candidate predictors of wheezing-associated disease in children 3 years of age or younger can facilitate early intervention and management in clinical practice. Future research on wheezing symptoms must develop AI-based predictive or diagnostic models using interpretable methods in sufficiently large cohorts of both general and clinically associated populations, and validate them externally with the measurement of prediction sensitivity, specificity, accuracy and generalizability.

3.3.4 Establish a benchmark database of wheezing-related diseases in children

Rapid development of AI techniques, particularly deep learning-based methods, has benefited from the large amount of available data [75–77]. According to Vapnik-Chervonenkis theory [78], the relationship between the required sample size and the Vapnik-Chervonenkis dimension of the model is defined as a log-like function. However, collecting and labeling large amounts of medical data is very expensive and challenging; consequently, most studies have focused on small sample datasets. More importantly, practical application in clinical practice, in addition to requiring large amounts of data for model training, requires thorough validation in a sufficiently large general and clinically relevant population to ensure the high performance of the AI model in terms of sensitivity, specificity, accuracy and generalizability. Therefore, we believe that the establishment of a publicly available benchmark database of wheezing-associated diseases in children will be an important future direction to promote research on AI for diagnosis of wheezing, which will require collaborations between multiple medical institutions and health administrations, as well as the participation of physicians, software developers, data scientists and policy makers.

4. SUMMARY AND CONCLUSIONS

Wheezing is a very common symptom in children, and its association with multiple diseases and risk factors makes differential diagnosis challenging. AI algorithms are a novel technique in the field of childhood wheezing diagnosis. AI models, compared with conventional methods, can make better use of large complex datasets to develop diagnostic and predictive models for causational analysis of pediatric wheezing symptoms. However, few studies have used AI to diagnose wheeze in children. Moreover, most existing AI research has proposed a “black box” diagnostic process that is difficult to control and interpret. Therefore, interpretable AI technologies must be explored for the assisted diagnosis of wheezing symptoms in children, which reveal the diagnostic decision-making process from the perspective of causality and explain the diagnostic rationale, to balance high-level assisted diagnosis with the risk of misdiagnosis. Interpretable AI technologies can guide decisions to refer high-risk and low-trust cases to high-level pediatric specialists, and ensure primary pediatricians’ diagnostic accuracy for common wheeze symptoms in young children.

Improving the reliability and credibility of current AI auxiliary diagnosis models would have high value and clinical significance in decreasing the risk of potential misdiagnosis caused by the limitations of AI algorithms. Moreover, such improvements would help establish a credible interpretable framework enabling mutual monitoring between AI and physicians, to provide a diagnostic basis for individualized, specific, precise interventions to treat wheezing symptoms in children. These improvements will be crucial for the development, application and promotion of AI-assisted diagnostic technologies.

DECLARATION OF CONFLICTS OF INTEREST

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported herein.

ABBREVIATIONS

COPD:	chronic obstructive pulmonary disease
AI:	artificial intelligence
ML:	machine learning

REFERENCES

1. National Health and Family Planning Commission Expert Committee on Pediatric Drug Use, Respiratory Group of the Chinese Medical Association Pediatrics Branch, Pediatric Respiratory Specialty Committee of the Pediatricians Branch of the Chinese Medical Association, Pediatric Specialty Committee of the Chinese Medical Education Association, Guide to the rational use of medications for children with wheezing diseases. Chinese Clin J Pract Pediatr. 2018. Vol. 33:1460–72

2. McGeachie MJ, Yates KP, Zhou X, Guo F, Sternberg AL, et al.. Patterns of growth and decline in lung function in persistent childhood asthma. N Engl J Med. 2016. Vol. 374:1842–52. 10.1056/NEJMoa1513737

3. Bui DS, Lodge CJ, Burgess JA, Lowe AJ, Perret J, et al.. Childhood predictors of lung function trajectories and future COPD risk: a prospective cohort study from the first to the sixth decade of life. Lancet Respir Med. 2018. Vol. 6:535–44. 10.1016/S2213-2600(18)30100-0

4. Tai A, Tran H, Roberts M, Clarke N, Wilson J, et al.. The association between childhood asthma and adult chronic obstructive pulmonary disease. Thorax. 2014. Vol. 69:805–10. 2464665910.1136/thoraxjnl-2013-204815

5. Luzak A, Fuertes E, Flexeder C, Standl M, von Berg A, et al.. Which early life events or current environmental and lifestyle factors influence lung function in adolescents? - results from the GINIplus & LISAplus studies. Respir Res. 2017. Vol. 18:138 2870116610.1186/s12931-017-0619-5

6. Melen E, Guerra S. Recent advances in understanding lung function development. F1000Res. 2017. Vol. 6:726 2862046710.12688/f1000research.11185.1

7. Ducharme FM, Tse SM, Chauhan B. Diagnosis, management, and prognosis of preschool wheeze. Lancet. 2014. Vol. 383:1593–604. 2479285610.1016/S0140-6736(14)60615-2

8. Martinez FD, Wright AL, Taussig LM, Holberg CJ, Halonen M, et al.. Asthma and wheezing in the first six years of life. The Group Health Medical Associates. N Engl J Med. 1995. Vol. 332:133–8. 780000410.1056/NEJM199501193320301

9. National Pediatric Asthma Collaborative Group, Institute of Environmental and Health-Related Product Safety, Chinese Center for Disease Control and Prevention. The Third Epidemiological Survey of Asthma in Chinese Urban Children. Chinese J Pediatr. 2013. Vol. 51:729–35

10. Papi A, Brightling C, Pedersen SE, Reddel HK. Asthma. Lancet. 2018. Vol. 391:783–800. 2927324610.1016/S0140-6736(17)33311-1

11. Gupta S, Lodha R, Kabra SK. Asthma, GERD and obesity: triangle of inflammation. Indian J Pediatr. 2018. Vol. 85:887–92. 2912761810.1007/s12098-017-2484-0

12. Asthma Group of the Chinese Medical Association, Respiratory Disease Branch. Guidelines for the prevention and treatment of bronchial asthma (2020 edition). Chinese J Tuberc Respir Dis. 2020. Vol. 43:1023–48

13. Castro-Rodriguez JA, Mallol J, Rodriguez J, Auger F, Andrade R. Risk factors for X-ray pneumonia in the first year of life and its relation to wheezing: a longitudinal study in a socioeconomic disadvantaged population. Allergol Immunopathol (Madr). 2008. Vol. 36:3–8. 1826142610.1157/13115664

14. Cloutier MM, Dixon AE, Krishnan JA, Lemanske RF, Pace W, et al.. Managing asthma in adolescents and adults: 2020 asthma guideline update from the national asthma education and prevention program. JAMA. 2020. Vol. 324:2301–17. 3327009510.1001/jama.2020.21974

15. Chinese Medical Association, Pediatrics Branch, Chinese Medical Association, Pediatricians Branch. White Paper on the Status of Pediatric Resources in China (Basic Data). 2016.11.25

16. Himes BE, Dai Y, Kohane IS, Weiss ST, Ramoni MF. Prediction of chronic obstructive pulmonary disease (COPD) in asthma patients using electronic medical records. J Am Med Inform Assoc. 2009. Vol. 16:371–9. 1926194310.1197/jamia.M2846

17. Pennington AF, Strickland MJ, Freedle KA, Klein M, Drews-Botsch C, et al.. Evaluating early-life asthma definitions as a marker for subsequent asthma in an electronic medical record setting. Pediatr Allergy Immunol. 2016. Vol. 27:591–6. 2711658710.1111/pai.12586

18. Afzal Z, Engelkes M, Verhamme KM, Janssens HM, Sturkenboom MCJM, et al.. Automatic generation of case-detection algorithms to identify children with asthma from large electronic health record databases. Pharmacoepidemiol Drug Saf. 2013. Vol. 22:826–33. 2359257310.1002/pds.3438

19. Xi N, Wallace R, Agarwal G, Chan D, Gershon A, et al.. Identifying patients with asthma in primary care electronic medical record systems Chart analysis-based electronic algorithm validation study. Can Fam Physician. 2015. Vol. 61:e474–83. 26759847PMC4607352

20. Quinto KB, Zuraw BL, Poon KY, Chen W, Schatz M, et al.. The association of obesity and asthma severity and control in children. J Allergy Clin Immunol. 2011. Vol. 128:964–9. 2182071110.1016/j.jaci.2011.06.031

21. Das N, Topalovic M, Janssens W. Artificial intelligence in diagnosis of obstructive lung disease: current status and future potential. Curr Opin Pulm Med. 2018. Vol. 24:117–23. 2925169910.1097/MCP.0000000000000459

22. Tgavalekos NT, Tawhai M, Harris RS, Musch G, Vidal-Melo M, et al.. Identifying airways responsible for heterogeneous ventilation and mechanical dysfunction in asthma: an image functional modeling approach. J Appl Physiol. 2005. Vol. 99:2388–97. 1608162210.1152/japplphysiol.00391.2005

23. Amaral JLM, Lopes AJ, Veiga J, Faria ACD, Melo PL. High-accuracy detection of airway obstruction in asthma using machine learning algorithms and forced oscillation measurements. Comput Methods Programs Biomed. 2017. Vol. 144:113–25. 2849499510.1016/j.cmpb.2017.03.023

24. Schilham AM, van Ginneken B, Loog M. A computer-aided diagnosis system for detection of lung nodules in chest radiographs with an evaluation on a public database. Med Image Anal. 2006. Vol. 10:247–58. 1629344110.1016/j.media.2005.09.003

25. Yedururi S, Guillerman RP, Chung T, Braverman RM, Dishop MK, et al.. Multimodality imaging of tracheobronchial disorders in children. Radiographics. 2008. Vol. 28:e29. 1829955910.1148/rg.e29

26. Xu C, Qi S, Feng J, Xia S, Kang Y, et al.. DCT-MIL: Deep CNN transferred multiple instance learning for COPD identification using CT images. Phys Med Biol. 2020. Vol. 65:145011. 3223507710.1088/1361-6560/ab857d

27. González G, Ash SY, Vegas-Sánchez-Ferrero G, Onieva Onieva J, Rahaghi FN, et al.. COPDGene and ECLIPSE Investigators. Disease staging and prognosis in smokers using deep learning in chest computed tomography. Am J Respir Crit Care Med. 2018. Vol. 197:193–203. 2889245410.1164/rccm.201705-0860OC

28. Zhang L, Jiang B, Wisselink HJ, Vliegenthart R, Xie X. COPD identification and grading based on deep learning of lung parenchyma and bronchial wall in chest CT images. Br J Radiol. 2022. Vol. 95:20210637. 3514328610.1259/bjr.20210637

29. Bharati S, Podder P, Mondal MRH. Hybrid deep learning for detecting lung diseases from X-ray images. Inform Med Unlocked. 2020. Vol. 20:100391. 3283507710.1016/j.imu.2020.100391

30. van der Velden BHM, Kuijf HJ, Gilhuijs KGA, Viergever MA. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Med Image Anal. 2022. Vol. 79:102470. 3557682110.1016/j.media.2022.102470

31. Fuhrman JD, Gorre N, Hu Q, Li H, El Naqa I, et al.. A review of explainable and interpretable AI with applications in COVID-19 imaging. Med Phys. 2022. Vol. 49:1–14. 3479653010.1002/mp.15359

32. Friedman J, Hastie T, Tibshirani R. The elements of statistical learning. Vol. volume 1. New York: Springer Series in Statistics. 2001

33. Goldstein A, Kapelner A, Bleich J, Pitkin E. Peeking inside the black box: visualizing statistical learning with plots of individual conditional expectation. J Comput Graph Stat. 2015. Vol. 24:44–65. 10.48550/arXiv.1309.6392

34. Apley DW, Zhu J. Visualizing the effects of predictor variables in black box supervised learning models. 2016. 10.48550/arXiv.1612.08468

35. Ribeiro MT, Singh S, Guestrin C. Model-agnostic interpretability of machine learning. 2016. 10.48550/arXiv.1606.05386

36. Lundberg SM, Lee S. A unified approach to interpreting model predictionsAdvances in neural information processing systems. 2017. p. 4765–74

37. Fisher A, Rudin C, Dominici F. Model class reliance: Variable importance measures for any machine learning model class, from the “Rashomon” perspective. 682018.

38. Zeiler MD, Fergus R. Visualizing and understanding convolutional networks. European conference on computer vision. Cham: Springer. 2014. p. 818–33

39. Springenberg JT, Dosovitskiy A, Brox T, Riedmiller M. Striving for simplicity: the all convolutional net. 2014. 10.48550/arXiv.1412.6806

40. Wagner J, Kohler JM, Gindele T. Interpretable and fine-grained visual explanations for convolutional neural networksProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2019. p. 9097–107

41. Zhou B, Khosla A, Lapedriza A. Learning deep features for discriminative localizationProceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2016. p. 2921–9

42. Selvaraju RR, Cogswell M, Das A. Grad-CAM: Visual explanations from deep networks via gradient-based localizationProceedings of the IEEE International Conference on Computer Vision (ICCV); 2017. p. 618–26

43. Chattopadhay A, Sarkar A, Howlader P. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks2018 IEEE winter conference on applications of computer vision (WACV). IEEE; 2018. p. 839–47

44. Fu R, Hu Q, Dong X, Guo Y, Gao Y, et al.. Axiom-based Grad-CAM: towards accurate visualization and explanation of CNNs. 2020. 10.48550/arXiv.2008.02312

45. Desai S, Ramaswamy HG. Ablation-CAM: visual explanations for deep convolutional network via gradient-free localizationProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision; 2020. p. 983–91

46. Muhammad MB, Yeasin M. Eigen-CAM: Class activation map using principal components2020 International Joint Conference on Neural Networks (IJCNN). IEEE; 2020. p. 1–7

47. Jiang PT, Zhang CB, Hou Q, Cheng MM, Wei Y. LayerCAM: exploring hierarchical class activation maps for localization. IEEE Trans Image Process. 2021. Vol. 30:5875–88. 3415694110.1109/TIP.2021.3089943

48. Liu SA, Xie H, Xu H. Partial Class Activation Attention for Semantic SegmentationProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022. p. 16836–45

49. Draelos RL, Carin L. Use HiResCAM instead of Grad-CAM for faithful explanations of convolutional neural networks. 2020

50. Castro-Rodriguez JA, Holberg CJ, Wright AL, Martinez FD. A clinical index to define risk of asthma in young children with recurrent wheezing. Am J Respir Crit Care Med. 2000. Vol. 162:1403–6. 1102935210.1164/ajrccm.162.4.9912111

51. Guilbert TW, Morgan WJ, Zeiger RS, Bacharier LB, Boehmer SJ, et al.. Atopic characteristics of children with recurrent wheezing at high risk for the development of childhood asthma. J Allergy Clin Immunol. 2004. Vol. 114:1282–7. 1557782410.1016/j.jaci.2004.09.020

52. Ducharme FM, Dell SD, Radhakrishnan D, Grad RM, Watson WT, et al.. Diagnosis and management of asthma in preschoolers: a Canadian Thoracic Society and Canadian Paediatric Society position paper. Paediatr Child Health. 2015. Vol. 20:353–71. 2652609510.1093/pch/20.7.353

53. Bao Y, Chen Z, Liu E, Xiang L, Zhao D, et al.. Risk factors in preschool children for predicting asthma during the preschool age and the early school age: a systematic review and meta-analysis. Curr Allergy Asthma Rep. 2017. Vol. 17:85 2915119510.1007/s11882-017-0753-7

54. Caudri D, Wijga A, A Schipper CM, Hoekstra M, Postma DS, et al.. Predicting the long-term prognosis of children with symptoms suggestive of asthma at preschool age. J Allergy Clin Immunol. 2009. Vol. 124:903–10.e1-7. 1966576510.1016/j.jaci.2009.06.045

55. Pescatore AM, Dogaru CM, Duembgen L, Silverman M, Gaillard EA, et al.. A simple asthma prediction tool for preschool children with wheeze or cough. J Allergy Clin Immunol. 2014. Vol. 133:111–8.e1-13. 2389135310.1016/j.jaci.2013.06.002

56. van der Mark LB, van Wonderen KE, Mohrs J, van Aalderen WMC, ter Riet G, et al.. Predicting asthma in preschool children at high risk presenting in primary care: development of a clinical asthma prediction score. Prim Care Respir J. 2014. Vol. 23:52–9. 2449648710.4104/pcrj.2014.00003

57. Allergy and Asthma in Children Group of the Chinese Medical Association, Allergy and Asthma in Children Group of the Chinese Medical Association, Asthma Collaborative Group of the Respiratory Group of the Chinese Medical Association, Pediatrics Branch. Exploration of establishing diagnostic criteria for asthma in children under 6 years of age in China. Chinese Journal of Practical Pediatrics. 2018. Vol. 33:870–6

58. Yu G, Li Z, Li S, Liu J, Sun M, et al.. The role of artificial intelligence in identifying asthma in pediatric inpatient setting. Ann Transl Med. 2020. Vol. 8:1367 3331311210.21037/atm-20-2501a

59. Spathis D, Vlamos P. Diagnosing asthma and chronic obstructive pulmonary disease with machine learning. Health Informatics J. 2019. Vol. 25:811–27. 2882001010.1177/1460458217723169

60. Topalovic M, Laval S, Aerts JM, Troosters T, Decramer M, et al.. Automated interpretation of pulmonary function tests in adults with respiratory complaints. Respiration. 2017. Vol. 93:170–8. 2808879710.1159/000454956

61. Dunnmon JA, Yi D, Langlotz CP, Ré C, Rubin DL, et al.. Assessment of convolutional neural networks for automated classification of chest radiographs. Radiology. 2019. Vol. 290:537–44. 10.1148/radiol.2018181422

62. Draelos RL, Carin L. Explainable multiple abnormality classification of chest CT volumes. Artif Intell Med. 2022. Vol. 132:102372. 3620707410.1016/j.artmed.2022.102372

63. Hu Q, Drukker K, Giger ML. Role of standard and soft tissue chest radiography images in deep-learning-based early diagnosis of COVID-19. J Med Imaging (Bellingham). 2021. Vol. 8:014503. 3459524510.1117/1.JMI.8.S1.014503

64. Mahmud T, Rahman MA, Fattah SA. CovXNet: a multi-dilation convolutional neural network for automatic COVID-19 and other pneumonia detection from chest X-ray images with transferable multi-receptive feature optimization. Comput Biol Med. 2020. Vol. 122:103869. 3265874010.1016/j.compbiomed.2020.103869

65. Schlemper J, Oktay O, Schaap M, Heinrich M, Kainz B, et al.. Attention gated networks: learning to leverage salient regions in medical images. Med Image Anal. 2019. Vol. 53:197–207. 3080281310.1016/j.media.2019.01.012

66. Rajaraman S, Candemir S, Thoma G, Antani SK. Visualizing and explaining deep learning predictions for pneumonia detection in pediatric chest radiographs. Medical imaging 2019: computer-aided diagnosis. SPIE. 2019. Vol. 10950:200–11. 10.1117/12.2512752

67. Gotkowski K, Gonzalez C, Bucher A, Mukhopadhyay A. M3d-CAM: A PyTorch library to generate 3D data attention maps for medical deep learning. 2020.

68. Radford A, Kim JW, Hallacy C. Learning transferable visual models from natural language supervisionInternational Conference on Machine Learning. PMLR; 2021. p. 8748–63

69. Bommasani R, Hudson DA, Adeli E, Altman R, Arora S, et al.. On the opportunities and risks of foundation models. 2021.

70. Wiggins WF, Tejani AS. On the opportunities and risks of foundation models for natural language processing in radiology. Radiol Artif Intell. 2022. Vol. 4:e220119. 3592337910.1148/ryai.220119

71. Zhang Z, Xie Y, Xing F. MDNet: A Semantically and Visually Interpretable Medical Image Diagnosis NetworkProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017. p. 6428–36

72. Yao Z, Hu X, Liu X, Xie W, Dong Y, et al.. A machine learning-based pulmonary venous obstruction prediction model using clinical data and CT image. Int J Comput Assist Radiol Surg. 2021. Vol. 16:609–17. 3379192110.1007/s11548-021-02335-y

73. Yu G, Yu Z, Shi Y, Wang Y, Liu X, et al.. Identification of pediatric respiratory diseases using a fine-grained diagnosis system. J Biomed Inform. 2021. Vol. 117:103754. 3383153710.1016/j.jbi.2021.103754

74. Yang S, Li G, Yu Y. Relationship-embedded representation learning for grounding referring expressions. IEEE Trans Pattern Anal Mach Intell. 2021. Vol. 43:2765–79. 3207853110.1109/TPAMI.2020.2973983

75. Juba B, Le HS. Precision-recall versus accuracy and the role of large data setsProceedings of the AAAI Conference on Artificial Intelligence; 2019. Vol. 33. p. 4039–48

76. Hestness J, Narang S, Ardalani N, Diamos G, Jun H, et al.. Deep learning scaling is predictable, empirically. 2017

77. Sun C, Shrivastava A, Singh S. Revisiting unreasonable effectiveness of data in deep learning eraProceedings of the IEEE international conference on computer vision; 2017. p. 843–52

78. Blumer A, Ehrenfeucht A, Haussler D, Warmuth MK. Learnability and the Vapnik-Chervonenkis dimension. JACM. 1989. Vol. 36:929–65. 10.1145/76359.76371