Abstract
Artificial intelligence (AI)-aided decision support has developed rapidly to meet the needs for effective analysis of substantial data sets from electronic medical records and medical images generated daily, and computer-assisted intelligent drug design. In clinical practice, paediatricians make medical decisions after obtaining a large amount of information about symptoms, physical examinations, laboratory test indicators, special examinations and treatments. This information is used in combination with paediatricians’ knowledge and experience to form the basis of clinical decisions. This diagnosis and therapeutic strategy development based on large amounts of information storage can be applied to both large clinical databases and data for individual patients. To date, AI applications have been of great value in intelligent diagnosis and treatment, intelligent image recognition, research and development of intelligent drugs and intelligent health management. This review aims to summarize recent advances in the research and clinical use of AI in paediatrics.
Keywords: Artificial intelligence, paediatrics, clinical diagnosis, decision making, medical imaging, electronic health record, machine learning
Introduction
Artificial intelligence (AI) has developed rapidly owing to its broad application to many fields, such as clinical medicine. Paediatric clinical diagnosis, an important part of clinical medicine, has benefited much from AI in various ways.
In this study, we searched influential journals to identify and review the latest articles on AI applications in the field of paediatric clinical diagnosis. After identifying several cutting-edge applications, we categorized them into the following segmentation scenarios: physical examination, medical imaging, comprehensive analysis and diagnosis, and the use of electronic health record (EHR) systems. This review summarizes recent advances in paediatric clinical diagnosis and provides researchers with information about the future prospects of AI in paediatrics.
Artificial intelligence-assisted auscultation
During the basic process of clinical diagnosis and treatment, doctors gradually collect patient information.1 This is combined with their own knowledge to make decisions about the next clinical stage, and is a process based on analysis and comparison.
Physical examination is one of the most effective methods by which doctors obtain patient information using simple and standard skills. However, the doctor’s judgment about positive signs can be affected by the particular situation, sensory abilities and practical experience, and sometimes leads to misjudgement. For example, heart murmurs are important in the diagnosis of congenital valvular heart disease and a variety of congenital heart diseases. However, simple physiological murmurs in children and young people are common, making it difficult to distinguish pathological murmurs. Although experienced paediatricians can use simple stethoscopes to identify pathological murmurs with sensitivity and specificity of more than 90%, it is more difficult for young paediatricians and primary care practitioners to rely solely on auscultation to identify cardiac murmurs, which makes it increasingly difficult to screen for heart diseases using physical examinations.2 Artificial intelligence-assisted auscultation (AIAA) is designed to improve screening validity. However, despite advances in heart sound classification algorithms and signal processing, this technology has not been widely adopted in clinical practice.
Clinical experiments involving noise detection algorithms are needed to compare performance with traditional auscultation and other algorithms, and to increase understanding of the potential of AIAA in clinical paediatric practice.3 The Cardiac Auscultation Record Card (CARD) database at the Johns Hopkins Children’s Medical Center contains a large amount of data from cases with pathological murmurs and with no murmurs, as well as noise-free cardiac auscultation data. Several test algorithms have been developed based on this auscultation database, and 3180 records of heart murmur cases have been analysed using an automated batch protocol. This AIAA system has been tested for measurement accuracy, patient age, heart rate, noise intensity, chest recording position and various pathological diagnostic criteria. The results show that the sensitivity and specificity of the AI system for identifying pathological cases is 93% and 81%, respectively.4
AI-based noise detection algorithms have been evaluated objectively and comprehensively in paediatric clinics, and have been found to operate well in clinical experiments.5 This approach could be used to compare the efficiency of other algorithms with the same data sets, which could increase understanding of the potential clinical applications of AIAA.
Automated image-based bilirubin analysis
Automated image-based bilirubin (AIB) has been used to monitor neonatal and infant jaundice. Xuzhou Affiliated Hospital of Southeast University, China, conducted a prospective study of 194 cases of neonatal jaundice. Data on total serum bilirubin (TSB) level, transcutaneous bilirubin (TCB) concentration, AIB and clinical signs were collected when newborns were admitted to hospital, and before and after phototherapy.6 The results showed a correlation between AIB and TSB (r = 0.824), with a high level of consistency (96.5% of the sample fell within the 95% limits of agreement). Furthermore, the strongest correlation and greatest consistency were found for a subgroup of newborns with TSB > 10 mg/dl and ≤20 mg/dl. In 97.5% neonates, TSB indicated jaundice levels lower than AIB plus 3.80 mg/dl. The main advantages of AIB are that it does not rely on the exclusive detection of bilirubin level, is non-invasive and is inexpensive. However, external environmental factors, such as insufficient light, light reflection from colorimetric cards and trembling in child patients, have a much greater impact on AIB than on TCB. These methods have a range of potential applications for family use and dynamic monitoring of moderate neonatal and infant jaundice, but comprehensive understanding of their diagnostic significance for neonatal jaundice requires further multicentre studies involving detailed evaluation.6
AI-based analysis of medical imaging
In addition to physical examinations, medical imaging (an important auxiliary examination method) can provide disease information.7 Medical images, such as X-ray images, computed tomography images, pathological images and ophthalmoscope images, can play an important role in disease diagnosis. The use and analysis of medical images is another important application of AI.8 Deep neural networks, which aim to identify the rules, modes and hidden relationships in data, have been used for image analysis.9 Bone age radiography is commonly used in paediatric departments and enables paediatricians to estimate children’s skeletal development using X-ray photographs. An AI model that includes 12,611 paediatric skeletal radiographs has been established based on machine learning and deep neural networks developed by the US National Institutes of Health. The clinical testing interpretations of four senior imaging specialists were used as a control. The root mean square of the interpretation differences for bone age in the AI model was 0.93 to 1.17 years. In terms of the stability of duplicated diagnoses, the specialists’ diagnosis discrepancy rate was 14% to 18.5%, and the error rate of the AI model was 15%. The authors concluded that their deep neural networks could estimate the rate of skeletal maturity with an accuracy similar to that obtained by radiologists and existing automated methods.10 AI-assisted image technology, such as computed tomography and magnetic resonance imaging, is one of the 30 most rapidly developing domains in cancer research.11 In addition, one review of AI applications for depression found that papers on electroencephalography-based diagnosis accounted for the largest percentage of publications reviewed.12
AI-based comprehensive analysis of multiple information
The diseases mentioned above can be clearly diagnosed from specific signs and typical imaging information. However, some diseases must be diagnosed through comprehensive analysis of multiple information sources, such as patient signs, symptoms, test results, radiological images and therapeutic responses. Each patient is unique and shows different symptoms. It is important for doctors to consider whether results are consistent with the patient’s self-reports. However, it is inevitable that doctors will occasionally make diagnostic mistakes in the course of their careers. Many factors can lead to such mistakes, including doctors’ knowledge, experience and psychological and emotional states. One machine-learning approach, the multilayer perceptron method, can handle the complicated properties of different types of data.13 The application of AI to diabetes diagnosis is attracting much interest.14 This requires the comprehensive analysis of several parameters that include heart rate variability and arterial blood glucose alterations. A machine-learning algorithm has been designed to dichotomize subjects as diabetic or non-diabetic.15
Artificial intelligence and electronic health records
The adoption and use of EHR systems is considered an integral part of solving the problem of medical errors. The rapid development of EHR systems has presented the challenge of AI management of large amounts of data and text information.16 The use of EHR systems ensures that reliable information about different patients is integrated and utilized in patient care, and promotes automated error detection and clinical diagnosis.17
Machine learning can provide objective, comprehensive and accurate results from database analysis.18 The application of machine learning in EHR is of great value. First, predictive modelling with EHR data can save labour, enhance efficiency, increase accuracy19,20 and promote personalized medicine.21 Second, machine-learning techniques can transform a large amount of textual information and unstructured data from patient records into structured data.22,23
Summary
Owing to the large age range and substantial individual differences among paediatric patients, paediatric EHR data often lacks wide coverage and has a high distribution imbalance.24 In addition, expert diagnosis may be affected by subjective factors such as fatigue level and diagnosis time limits.25 AI-assisted diagnosis is essential to provide complex diagnostic mechanisms and to increase the efficiency and accuracy of paediatric clinical diagnostic services.
In China, the application of AI in paediatric clinical diagnosis has mainly focused on accurate diagnosis and treatment, medical image recognition, intelligent health management and patient triage. These applications feature an increasing number of approaches. Future potential applications include classification of disease diagnosis, individualized treatment guidance, prognostic evaluation and clinical studies enrolment.
Although AI has shown great promise in paediatrics, substantial technical barriers remain to be overcome. The scarcity of high-quality data indicates the need for joint development of a package of schemes.26 Combined analysis using feature enrichment integrated with health informatics could strengthen training efficacy and diagnostic accuracy.
We hope that the advances summarized in this paper, and the challenges and future prospects identified, will inspire research in AI approaches for paediatrics.
Declaration of conflicting interest
The authors declare that there is no conflict of interest.
Funding
This research was funded by the National Natural Science Foundation of China (grant number 61902037), the Fundamental Research Funds for the Central Universities (grant number 500419804) and the China Postdoctoral Science Foundation (grant number 2018M641397).
ORCID iD
References
- 1.Burton-Jones A, Volkoff O. How can we develop contextualized theories of effective use? A demonstration in the context of community-care electronic health records. Inf Syst Res 2017; 28: 468–489. doi: 10.1287/isre.2017.0702. [Google Scholar]
- 2.Kumar K, Thompson WR. Evaluation of cardiac auscultation skills in pediatric residents. Clin Pediatr (Phila) 2013; 52: 66–73. doi: 10.1177/0009922812466584. [DOI] [PubMed] [Google Scholar]
- 3.Lai LS, Redington AN, Reinisch AJ, et al. Computerized automatic diagnosis of innocent and pathologic murmurs in pediatrics: a pilot study. Congenit Heart Dis 2016; 11: 386–395. doi: 10.1111/chd.12328. [DOI] [PubMed] [Google Scholar]
- 4.Thompson WR, Reinisch AJ, Unterberger MJ, et al. Artificial intelligence-assisted auscultation of heart murmurs: validation by virtual clinical trial. Pediatr Cardiol 2019; 40: 623–629. doi: 10.1007/s00246-018-2036-z. [DOI] [PubMed] [Google Scholar]
- 5.Zia MK, Griffel B, Fridman V, et al. Noise detection in heart sound recordings. Conf Proc IEEE Eng Med Biol Soc 2011; 2011: 5880–5883. doi: 10.1109/IEMBS.2011.6091454. [DOI] [PubMed] [Google Scholar]
- 6.Yang B, Huang D, Gao XY, et al. Neonatal and early infantile jaundice: assessment by the use of the smartphone. Chinese Journal of Neonatology 2018; 033: 277–282. (in Chinese). doi: 10.3760/cma.j.issn.2096-2932.2018.04.009. [Google Scholar]
- 7.Erickson BJ, Korfiatis P, Akkus Z, et al. Machine learning for medical imaging. Radiographics 2017; 37: 505–515. doi: 10.1148/rg.2017160130. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Summers RM. Deep learning lends a hand to pediatric radiology. Radiology 2018; 287: 323–325. doi: 10.1148/radiol.2018172898. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Li YW, Jang WF, Yang L, et al. On neural networks and learning systems for business computing. Neurocomputing 2018; 275: 1150–1159. doi:10.1016/j.neucom.2017.09.054 [Google Scholar]
- 10.Tajmir SH, Lee H, Shailam R, et al. Artificial intelligence-assisted interpretation of bone age radiographs improves accuracy and decreases variability. Skeletal Radiol 2019; 48: 275–283. doi: 10.1007/s00256-018-3033-2. [DOI] [PubMed] [Google Scholar]
- 11.Tran BX, Latkin CA, Sharafeldin N, et al. Characterizing artificial intelligence applications in cancer research: a Latent Dirichlet Allocation analysis. JMIR Med Inform 2019; 7: e14401. doi:10.2196/14401. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Tran BX, McIntyre RS, Latkin CA, et al. The current research landscape on the artificial intelligence application in the management of depressive disorders: a bibliometric analysis. Int J Environ Res Public Health 2019; 16: 2150. doi:10.3390/ijerph16122150 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Li YW, Tang GC, Du JM, et al. Multilayer perceptron method to estimate real-world fuel consumption rate of light duty vehicles. IEEE Access 2019; 7: 63395–63402. doi:10.1109/access.2019.2914378 [Google Scholar]
- 14.Vu GT, Tran BX, McIntyre RS, et al. Modeling the research landscapes of artificial intelligence applications in diabetes (GAPRESEARCH). Int J Environ Res Public Health 2020; 17: 1982. doi:10.3390/ijerph17061982 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Jong GJ, Huang CS, Yu GJ, et al. Artificial neural network expert system for integrated heart rate variability. Wireless Personal Communications 2014; 75: 483–509. doi:10.1007/s11277-013-1373-8 [Google Scholar]
- 16.Liang H, Tsui BY, Ni H, et al. Evaluation and accurate diagnoses of pediatric diseases using artificial intelligence. Nat Med 2019; 25: 433–438. doi: 10.1038/s41591-018-0335-9. [DOI] [PubMed] [Google Scholar]
- 17.Hornberger J. Electronic health records: a guide for clinicians and administrators. JAMA 2009; 301: 110. doi: 10.1001/jama.2008.910. [Google Scholar]
- 18.Li YW, Yang L, Yang BH, et al. Application of interpretable machine learning models for the intelligent decision. Neurocomputing 2019; 333: 273–283. doi:10.1016/j.neucom.2018.12.012 [Google Scholar]
- 19.Kumar V, Liao K, Cheng SC, et al. Natural language processing improves phenotypic accuracy in an electronic medical record cohort of type 2 diabetes and cardiovascular disease. J Am Coll Cardiol 2014; 63: A1359. doi: 10.1016/S0735-1097(14)61359-0. [Google Scholar]
- 20.Tsopra R, Peckham D, Beirne P, et al. The impact of three discharge coding methods on the accuracy of diagnostic coding and hospital reimbursement for inpatient medical care. Int J Med Inform 2018; 115: 35–42. doi: 10.1016/j.ijmedinf.2018.03.015. [DOI] [PubMed] [Google Scholar]
- 21.Rajkomar A, Oren E, Chen K, et al. Scalable and accurate deep learning with electronic health records. NPJ Digit Med 2018; 1: 18–27. doi: 10.1038/s41746-018-0029-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Assale M, Dui LG, Cina A, et al. The revival of the notes field: leveraging the unstructured content in electronic health records. Front Med (Lausanne) 2019; 6: 1–23. doi:10.3389/fmed.2019.00066. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Vest JR, Grannis SJ, Haut DP, et al. Using structured and unstructured data to identify patients' need for services that address the social determinants of health. Int J Med Inform 2017; 107: 101–106. doi: 10.1016/j.ijmedinf.2017.09.008. [DOI] [PubMed] [Google Scholar]
- 24.Li XZ, Wang HZ, Xiong YJ, et al. Application of convolutional neural network in pediatric disease prediction. China Digital Medicine 2018; 13: 11–13. doi:10.3969/j.issn.1673-7571.2018.10.004. [Google Scholar]
- 25.Liu MQ, Lan J, Chen X, et al. Bone age assessment model based on multi-dimensional feature fusion using deep learning. Academic Journal of Second Military Medical University 2018; 39: 909–916. doi:10.16781/j.0258-879x.2018.08.0909 [Google Scholar]
- 26.Liu B, Chi WH, Li X, et al. Evolving the pulmonary nodules diagnosis from classical approaches to deep learning-aided decision support: three decades’ development course and future prospect. J Cancer Res Clin Oncol 2020; 146: 153–185. doi: 10.1007/s00432-019-03098-5. [DOI] [PMC free article] [PubMed] [Google Scholar]