Stephen I. Feller , 2025-04-15 20:50:00
April 15, 2025
2 min read
Key takeaways:
- The system outperformed human analyses by 9% and exceeded WHO requirements for non-sputum triage tests for TB.
- It combines an ultrasound linked to a smartphone using a suite of deep learning AI models.
VIENNA — A point-of-care lung ultrasound system connected to a smartphone combined with a suite of deep learning models outperformed human analyses as well as WHO requirements for non-sputum tuberculosis tests.
In a study presented at ESCMID Global 2025, data showed that end-users could potentially, with minimal training, diagnose and treat TB in places where chest X-ray facilities and trained radiologists are not available, researchers told Healio.
Data derived from Suttels V, et al. Abstract 00252. Presented at: ESCMID Global; April 11-15, 2025; Vienna.
“Current WHO recommendations propose chest X-ray or molecular tests for TB triage,” said lead study author Veronique Suttels, MD, PhD, of the University of Lausanne. “However, these tests are virtually unavailable at the primary care level in low- and middle-income countries, where most TB patients first present.”
According to WHO, TB killed 1.25 million people globally in 2023 and returned to its perch as the world’s top cause of death from an infectious disease — after being briefly overtaken by COVID-19 during the acute phase of the pandemic.
“It is estimated that one-third of TB cases remain undiagnosed,” Suttels said. “Therefore, it is essential to investigate true point-of-care tests with minimal technical requirements. To address this, we evaluated lung ultrasound (LUS) connected to simple smartphone, with automatic interpretation powered by AI.”
The researchers conducted a prospective cohort study of 504 adults in Benin, West Africa, with respiratory symptoms, using a standardized 14-point LUS protocol applied to images taken with a point-of-care ultrasound connected to a smartphone, using ULTR-AI, a suite of deep learning models designed to automate TB risk assessment.
Among the study population, 61% were men, the median age was 40 years (interquartile range [IQR], 30-52), 13% had a TB history and 15% were HIV-positive (median CD4 count, 92 cells/mm³; IQR, 43-358).
Overall, 192 (38%) had confirmed pulmonary TB, of whom 66% had underweight (BMI < 18.5 kg/m²), 31% had a high quick organ failure score and 21% were hospitalized by day 28. Also, 4% of patients with confirmed TB died.
According to the study, ULTR-AI showed 93% sensitivity and 81% specificity (area under the receiving operating characteristic curve, 0.93; 95% CI, 0.92-0.95), surpassing WHO targets of 90% sensitivity and 70% specificity for non-sputum-based TB tests. The AI-guided LUS also outperformed human experts by 9%, the researchers reported.
Suttels said the lung ultrasound system already is a promising, cheap and accessible triage tool for pulmonary TB where chest X-ray is not available — adding that because the device connects to a smartphone it has potential for wide adoption.
“This study is the first to prospectively assess AI-assisted LUS interpretation against robust microbiological reference standards, showing potential to improve diagnostic accuracy and scalability,” Suttels said. “AI-guided LUS could help close critical diagnostic gaps in resource-limited settings, though further external validation is needed.”
References:
For more information:
Veronique Suttels, MD, can be reached by email at veronique.suttels@outlook.com.
Perspective
Rapid screening and diagnosis are critical components of WHO’s “End TB Strategy.” This is particularly challenging in low- and middle-income countries.
This study, presented at ESCMID 2025, assessed an AI-guided point-of-care LUS examination with a smart phone app interpretation compared with expert interpretation in a cohort of 504 patients in Benin. The AI-guided lung ultrasound was 93% sensitive and 81% specific — exceeding the standards set by WHO.
This is a potentially huge step forward for the rapid diagnosis and triage of TB in high-risk populations, especially as initial sputum analysis is not required [and] applicability and cost appear reasonable for widespread adoption. If performance of this technology is equally efficacious across populations with varied TB pre-test probabilities, its global implementation should not be delayed.
Gonzalo Bearman, MD, MPH, FACP, FSHEA, FIDSA
Professor of medicine
Virginia Commonwealth University
Disclosures: Bearman reports no relevant financial disclosures.
Sources/Disclosures
Collapse
Suttels V, et al. Abstract 00252. Presented at: ESCMID Global; April 11-15, 2025; Vienna.
Disclosures:
Suttels reports no relevant financial disclosures.