Developing artificial intelligence technology for pediatric pulmonology: Lessons from COVID-19.

To the Editor

The health crisis ignited by the coronavirus disease 2019 (COVID‐19) pandemic challenged clinicians with an unprecedented amount of workload and lack of resources to care for patients. As intensive care units and emergency departments filled up, accurate risk assessment at the point of care became essential. In 2020, novel artificial intelligence (AI) approaches emerged as rapid decision tools for COVID‐19 patients based on the integration of multidimensional data, including lung imaging and electronic health records. However, as older patients with severe acute respiratory syndrome coronavirus 2 (SARS‐CoV‐2) were at higher risk, AI technology has primarily been developed for adults. In this comment, we address three outstanding questions on the future impact of AI on the health of children with respiratory disorders.

Have we left children behind? Many studies have shown that AI can be implemented to improve clinical outcomes in adults with COVID‐19. For instance, a large AI study by Bennett et al. including 174,568 adults with SARS‐CoV‐2 infection from 34 different sites showed that first‐day machine learning models accurately predicted clinical severity with an area under the receiver operating curve (AUC) of 0.87. The investigators used 64 health metrics to construct predictive models of severe clinical course that were stable over time. Unfortunately, similar studies have yet to be conducted in the pediatric population. The gap in AI technology development for children and adults has remained large even as epidemiological observations in 2021–22 demonstrated that the initial perception that COVID‐19 only affects older patients is not true. In 2022, pediatric hospitals are still overwhelmed and admitting more children with COVID‐19 than ever. Pediatric COVID‐19 cases currently represent 18.9% of the total cases and 21.9% of the new infections in the country. Furthermore, the pandemic shifted the epidemiology and impact of other respiratory viruses for the pediatric population. For instance, respiratory syncytial virus, an aggressive pathogen in infants, reached an unprecedented peak in the summer–fall of 2021. It is important to remember that even before the pandemic, respiratory illnesses were the top cause of hospitalization and death in children. Going forward, we should consider new strategies to fight pediatric COVID‐19 and future pandemics, including the development of pediatric‐centered AI technology for early identification, risk stratification, and outcome prediction of COVID‐19, and other respiratory illnesses in children.

What is the feasibility of AI for pediatric pulmonology? The clinical and lung imaging heterogeneity of viral infections in newborns, infants, children, and adolescents , has been a barrier for the adaptation of AI tools developed for adults. Notably, a recent study by Dayan et al. using federated learning (also known as collaborative learning) showed promise for the pediatric population. Federated learning differs from traditional, centralized AI methods because it trains an algorithm on multiple decentralized data sets, greatly accelerating development and deployment. A major advantage of federated learning is that the privacy of the data is preserved at each site, while different sites can securely train and contribute to a global AI model by sharing only partial model weights. Thus, federated learning also accelerates collaborations making large studies possible in almost real‐time. The study by Dayan et al., the largest federated learning initiative to date, integrated data from 16,148 patients from 20 international hospitals, including lung imaging and clinical information. Only 102 cases were from a pediatric hospital, a reflection of the scarcity of data in the design of AI technology for children. Despite the small number, the federated model predicted the need of oxygen support for children after 24 h from hospital admission with 0.97 AUC. The model achieved 0.72 AUC when trained only on the limited pediatric data in isolation from the large adult cohort.

The study by Dayan et al. demonstrated that the transfer of knowledge from adult models to children can substantially improve the accuracy of deep learning technology for pediatric AI healthcare. Deep learning methods are notoriously dependent on large amounts of data and this knowledge may only be transferable when the adult and pediatric populations share clinical manifestations. This is generally true for COVID‐19, although differences between how the disease affects children and adults are well known. , In fact, COVID‐19 may not be the ideal model as the disease is generally mild in children and a large database of severely affected children will be required to develop an algorithm to predict a severe outcome. Even among children, lung imaging features of COVID‐19 vary by age, for example, ground glass opacities are more common in older individuals, whereas perihilar markings are more common in younger subjects. , These age‐related changes reflect the different pathogenesis of viral respiratory infections in children compared with older people. Although we remain excited about the potential of learning from adult data how to better treat children, conducting large pediatric‐centered studies is essential to provide even better insight into the distinctive needs of treating children. Novel methods such as federated learning approaches can facilitate collaborations between isolated pediatric centers that will benefit from sharing data and technological resources to create robust, generalizable and more accurate models that could not be obtained only based on their local data.

Do we need AI in pediatric pulmonary medicine? The use of AI in pediatric pulmonology is significantly underdeveloped compared to adult respiratory medicine. Clinical and research communities must work together to prevent the widening gap between adult and pediatric healthcare. Most efforts in developing AI tools are still centered on adults and federated models such as the above6 may only work when the adult and pediatric populations share clinical manifestations like in COVID‐19. Nevertheless, millions of children with respiratory conditions like asthma, bronchopulmonary dysplasia, pediatric sleep apnea, sepsis or pneumonia should also benefit from AI technologies in the future. Aside from severity quantification and outcome prediction in respiratory illnesses, AI can provide a new dimension to pediatric disease phenotyping and precision medicine approaches. Notably, a challenge for the development of AI tools is the need to generate good quality and large learning databases through multi‐institutional cooperative studies. Novel methods like federated learning approaches can facilitate collaborations between isolated pediatric centers that will benefit from sharing data and technological resources to create robust, generalizable, and more accurate models that could not be obtained only based on their local data. Thus, we urge the scientific community to embrace data sharing and federated learning to generate pediatric‐centered efforts that improve clinical outcomes in this vulnerable age group. With new technological advances, health equity for children may be in near sight.

AUTHOR CONTRIBUTIONS

Gustavo Nino: Conceptualization (equal); writing—review and editing (equal). Marius G. Linguraru: Conceptualization (equal); writing—review and editing (equal).

CONFLICTS OF INTEREST

The authors declare no conflicts of interest.