More children die of pneumonia than any other infectious disease worldwide, with more than 700,000 children under 5 succumbing to the condition every year. The highest concentration of cases occurs in South Asia and sub-Saharan Africa, where finding tools to help diagnose respiratory conditions in low-resource settings is of paramount importance.

That’s where the Johns Hopkins Children’s Center Global Program in Pediatric Respiratory Sciences — which works to improve the respiratory health of children in low- and middle-income countries — comes in. Director and pediatric pulmonologist Eric McCollum and pediatric pulmonologist Sunaina Kapoor are testing whether smart stethoscopes with built-in AI algorithms can accurately diagnose respiratory illnesses in pediatric patients.

“The AI-enabled stethoscopes can identify the presence of crackles, wheezes or other abnormal lung sounds that can be suggestive of conditions like pneumonia or asthma,” Kapoor says. “AI can really fit into this space and has the potential to address this gap really well.” In the low-resource areas where the program works, there may not be physicians or other healthcare practitioners trained in auscultation — which is difficult to teach, particularly in children — or access to chest X-rays.

The AI stethoscope used in the studies is a commercially available product originally designed for home use by caregivers. McCollum, Kapoor and their team calibrated the stethoscope’s algorithms to fit the low-resource use case. Pneumonia is the biggest target of this work, though the team is also studying its effectiveness in diagnosing bronchiolitis.

Current Research Studies

In the BLAAAST (Bangladesh Lung Auscultation AI for Antibiotic Stewardship) project, a first-of-its-kind randomized controlled trial in its third of five years, the AI stethoscope will be used to evaluate 2,500 patients under age 5 presenting to a rural health clinic with pneumonia who would be treated with the World Health Organization’s integrated management of childhood illness (IMCI) strategy. IMCI, developed to treat patients in low-resource countries, relies heavily on antibiotics for pneumonia. Kapoor notes that with better vaccination rates in these countries in recent years, children presenting with respiratory symptoms more often have viral infections, rather than bacterial.

To address overuse of antibiotics, BLAAAST will treat children in the control group with amoxicillin syrup, as IMCI recommends, and the intervention group will receive either amoxicillin syrup or a placebo based on the result of a digital stethoscope test that uses an AI algorithm. The study will assess whether the AI stethoscope is an effective diagnostic tool and can safely reduce unnecessary antibiotic use. Funded by the National Institutes of Health, it is the first clinical trial evaluating the health outcomes of patients examined with an AI-enabled digital stethoscope in a low-resource setting. 

The LaLeLa Lung Study, also NIH-funded, is looking at diagnostic accuracy of the AI stethoscopes in children presenting to primary care clinics in South Africa with a cough or difficulty breathing. The triple-blind study will have patients undergo imaging, the digital stethoscope test and a clinical evaluation by physicians. A panel of pediatricians will evaluate the diagnostic accuracy of the stethoscope when integrated with WHO’s IMCI strategy.

Another diagnostic accuracy study, the IMCI-Plus study, supported by the Thrasher Research Fund, is evaluating the accuracy of AI stethoscopes in diagnosing children in outpatient settings, using lung ultrasound as the reference standard. This study is taking place at two sites in South Africa, as well as in Tanzania and Senegal.

“There’s still a large gap in terms of the AI models and the data quality as it pertains to children,” McCollum says. “BLAAAST is a highly controlled study trying to address if we can safely reduce antibiotic use in these kids. The LaLeLa Study relaxes the standards slightly, and we’re evaluating the accuracy of the model. And then, in the IMCI-Plus study, it’s more of a real-world environment. They’re all separate studies, but are interrelated and addressing different questions within the AI stethoscope application.”

McCollum and Kapoor also see similar domestic uses for the devices in the future, particularly in rural areas with fewer medical resources.

Continuing Development

Kapoor has an early career award from the Thrasher Research Fund to improve the diagnostic accuracy of the AI stethoscopes by integrating a child’s respiratory rate into the algorithm. This work builds on the group’s decade-plus of prior research in AI lung sound classification, and includes the Xilonga project, which is aimed at developing more accurate pediatric pneumonia classification algorithms. In this project, physicians captured lung sound recordings and performed conventional stethoscope auscultation simultaneously, subsequently classifying the recorded sounds against the stethoscope reference standard and then using those classifications to train an algorithm. As part of that study, the group developed Auscultation Lab, a novel annotation platform that uses spectrograms and half-breath annotations to improve the algorithms and serves as a tool to teach auscultation to healthcare practitioners.

McCollum’s work with smart stethoscopes dates back to 2013, during his fellowship. That led to collaborations with Johns Hopkins University Whiting School of Engineering professors Mounya Elhilali and James West, with whom McCollum and others developed the methodology for interpreting lung sounds and designed a smart stethoscope that was later commercialized. That work started with duct-taping a microphone onto a stethoscope and plugging it into a sound recorder; the product that was later commercialized, Feelix by Sonavi Labs, filtered out distracting noise and gave a suggested diagnosis on an LED screen.

While McCollum and Kapoor are currently using a different device in the field, they are still working with Whiting engineers on refining the algorithms they use in their work, with the eventual goal of creating device-agnostic software that could work with any smart stethoscope.

“Our inherent mission is really to help the diagnosis of respiratory illness in children,” Kapoor says. “I think our development work is key because being device-agnostic is going to be really important to eventually unlock devices that are more financially sustainable and feasible to implement on a larger global scale.”