Nearly a third of the over 10 million people who fell ill with tuberculosis (TB) in 2022 went undetected. Artificial intelligence (AI) has the potential to help close this detection gap. Prolonged cough (>2 weeks) is a characteristic symptom of pulmonary TB; however, the sensitivity of cough screening is limited as descriptions are subjective, rely on recall, and symptoms are nonspecific.
AI solutions have recently been developed to better harness this acoustic biomarker to aid in cough reporting, initiating health seeking action, and interpretation. To date, two use cases have been explored for TB care:
-
AI-powered monitoring of cough counts as a prediction of TB disease and/or treatment progression
-
AI-powered classification of cough sounds for TB screening
A Stop TB literature mapping conducted in April 2024 identified an increasing interest in the application of AI for cough sound classification in recent years, with published between 2018-2024. There were four named software available for cough classification in the literature, alongside five un-named algorithms:
- Health Acoustic Representations (HeAR)
- Swaasa AI
- TBscreen
- Timbre
These algorithms were evaluated using data from high TB burden countries, including India, South Africa, Kenya, and Zambia, and a combined dataset from Rapid Research in Diagnostics Development for TB (R2D2) network countries (India, Philippines, South Africa, Uganda, Vietnam, Tanzania, and Madagascar). The majority of studies were internal validations, that evaluated AI performance on data from the same source as the training dataset.
Current literature generally tested AI classification on elicited coughs collected from adults in health facilities using microphones or smartphones. Often, recordings took place in quiet, controlled environments and underwent subsequent processing using noise filtering technology to reduce background noise.
AI analyses cough sounds to predict whether TB was present or absent. To evaluate AI, studies then compared AI's prediction to microbiological and clinical reference standards that indicated if individuals truly had TB or not.
Our mapping identified high performance of AI for TB cough classification, with the area under the receiver operating characteristic curve, an indication of overall accuracy, ranging from 0.61 [± 0.14] (TBscreen in Kenya) to 0.95 (un-named algorithm in South Africa). In studies where it was reported, sensitivity was generally high, but ranged from 0.16 ± 0.11 to 0.95.
|
Title (Year) |
AI algorithm |
Validation type |
Performance |
|
HeAR |
Internal |
AUC: 0.652 (0.520, 0.784). |
|
|
Detection of tuberculosis by automatic cough sound analysis (2018) |
Un-named AI |
Internal |
AUC: 0.95 |
|
Various, un-named AI |
Internal |
AUC ranged between 0.689 (0.647,0.732)- 0.743 (0.703.0.780) across algorithms |
|
|
Predicting tuberculosis from real-world cough audio recordings and meta data (2024) |
Various, un-named AI |
Internal |
Average AUC across algorithms: 0.70± 0.05 |
|
Automatic cough classification for tuberculosis screening in a real-world environment (2021) |
Un-named AI |
Internal |
AUC: 0.94 |
|
Acoustic epidemiology of pulmonary tuberculosis and COVID-19 leveraging explainable AI/ML (2022) |
Timbre |
External |
Sensitivity: 80-83% against microbiological reference standard |
|
TBscreen: A passive cough classifier for tuberculosis screening with a controlled dataset (2024) |
TBscreen |
Internal |
AUC ranged from: 0.61 (± 0.14) to 0.86 (±0.03), depending on the dataset. Sensitivity ranged from: 0.16 (± 0.11) to 0.80 (± 0.03), depending on thedataset |
|
Un-named AI |
Internal |
AUC: 88% |
|
|
Swaasa |
External |
AUC: 0.90 Sensitivity: 75% |
Although current literature was limited by the high prevalence of internal validations, sub-optimal reference standards, limited sample sizes, and potential for bias due to AI developer involvement, it demonstrated the potential of AI cough tools to supplement TB diagnostic pathways.
Key Findings:
