Tuberculosis: a portable lung ultrasound model reaches an AUC of 0.95 in Uganda, but on 73 participants recruited from separate groups

A team from Rutgers, Emory and Makerere University trained a neural network to recognise pulmonary tuberculosis from twelve chest ultrasound images taken with a portable probe plugged into a tablet, alone or combined with twelve simple clinical variables, in 316 participants recruited in Kampala, Uganda. On 73 participants held out for testing, the combined model reaches an AUC of 0.95, and a shortened six-zone version keeps most of the performance, above the minimum thresholds the WHO sets for a screening test. The signal is promising for places without X-ray, but the test set is small, the confidence intervals are very wide, and people with and without TB were recruited from separate groups, which probably flatters the numbers.

The context

Tuberculosis remains one of the deadliest infectious diseases in the world, and a large share of cases are never diagnosed. The bottleneck is triage: deciding which people with a cough, or which relatives of a patient, should get a microbiological sputum test (Xpert MTB/RIF Ultra, a rapid molecular test, or culture, slower but the reference). The World Health Organization recommends chest radiography for this, increasingly read by computer-aided detection (CAD) software. But an X-ray requires a machine, electricity, radiation shielding and a room, all of which many peripheral health centres lack.

Point-of-care lung ultrasound (POCUS) ticks the opposite boxes: a phone-sized probe, no radiation, an examination of a few minutes. Signs suggestive of tuberculosis have been described, such as small lesions just under the pleura or consolidations, but there are no standardised interpretation criteria, and reading depends heavily on the operator. Hence the idea of handing the reading to an algorithm. Two recent studies, cited by the authors, paved the way: in Benin, Suttels and colleagues reported an AUC of 0.93 (sensitivity 93%, specificity 81%) in 504 symptomatic adults; in Zambia, a network processing video clips from 497 participants reached 0.89 to 0.92. This preprint adds a practical question: do basic clinical data already do most of the job, and can the number of zones to scan be reduced?

The method

The study is prospective: participants were enrolled between 2023 and 2025 in Kampala, at existing tuberculosis study sites and city health facilities. Three groups were recruited: index cases (people with confirmed tuberculosis who had received at most 15 days of treatment), their household contacts, and clinic patients evaluated for tuberculosis but ultimately found to have another respiratory condition. Of 318 enrolled, 316 were analysed: 118 confirmed TB cases (only 4 of them household contacts), 96 people without respiratory disease and 102 patients with another respiratory condition. Median age is 32, 56% are men, 20% live with HIV.

The reference standard is microbiological: Xpert Ultra plus liquid culture (MGIT) for index cases and contacts, Xpert Ultra alone for clinic patients. Two trained and certified technicians performed the scans with a Philips Lumify probe connected to an Android tablet, blinded to laboratory results and radiography. Each participant was scanned in twelve chest zones (front, side and back, on both sides), with a six-second clip per zone.

For the model, the authors keep a single image per zone, the first frame of the clip, a rule fixed in advance. Each image goes through a ResNet18, a convolutional neural network (a classic architecture that extracts increasingly abstract visual patterns), which summarises it as a vector of numbers. The twelve vectors are then combined by attention-based multiple-instance learning: the patient is treated as a "bag" of twelve images, and the model learns by itself how much weight to give each zone, without being told where the lesion is. Three models are compared: a clinical-only model (a random forest, an ensemble of decision trees, fed with twelve variables such as age, HIV status, BMI, cough duration and weight-loss duration), an ultrasound-only model, and a combined model that adds the clinical variables to the image summary.

Participants were split at random: 223 for training, 20 for validation (used to decide when to stop training) and 73 for testing (28 with TB, 22 without respiratory disease, 23 with another condition). Finally, a shortened six-zone protocol was derived from the mean attention weights among TB-positive training participants: R5, L5, L3, R1, R3 and L1.

The results

The AUC (area under the ROC curve) measures a model's ability to rank a sick person above a healthy one: 0.5 is chance, 1 is perfect. On the 73 test participants, the clinical-only model already reaches 0.88 (95% confidence interval 0.80 to 0.95). Ultrasound alone rises to 0.93 (0.86 to 0.97), the combined model to 0.95 (0.89 to 0.99). With six zones, AUCs are 0.91 for ultrasound alone and 0.93 for the combined model.

For a triage tool, the real question is: if the model is set to miss few patients, how many healthy people are sent for testing for nothing? At 90% sensitivity (the share of patients detected), specificity (the share of non-patients correctly ruled out) is 66.2% for the clinical model, 73.3% for ultrasound alone, 82.2% for the twelve-zone combined model (CI 34.9 to 97.8%) and 91.1% for the six-zone combined model (CI 25.0 to 98.0%). These operating points are interpolated from the ROC curve; the authors state they are not deployment thresholds. On point estimates, all twelve-zone models and the six-zone combined model exceed the WHO 2025 minimum for a screening test (at least 90% sensitivity and 60% specificity).

Two secondary results matter. When the first image of each clip is replaced by the middle one, the AUC drops to 0.85 for ultrasound alone and 0.90 for the combined model. And subgroup analyses (HIV, sex, previous TB) show no difference, but they rest on 20 people living with HIV and 3 with previous TB in the test set.

Clinical translation, by our calculation and with an assumed prevalence: among 1,000 symptomatic people of whom 10% have tuberculosis, a model with 90% sensitivity and 82% specificity would detect 90 patients, miss 10, and send about 160 people without TB for an Xpert test. If the true specificity sat at the lower bound of the interval (35%), that would be about 585 unnecessary tests, a triage step that barely triages.

What is good

First, the reference standard is strong for tuberculosis. Diagnosis rests on Xpert Ultra and, for most participants, culture; ultrasound operators were blinded to results; enrolment was prospective, with an acquisition protocol defined in advance and the frame-selection rule fixed a priori.

Second, the comparator is fair. Rather than pitting the algorithm against a rushed human, the authors compare it with what free clinical data already predict (AUC 0.88). This shows what ultrasound actually adds, a few points of AUC and above all specificity. Confidence intervals are reported throughout, and the unfavourable sensitivity analysis (middle frame) is in the paper rather than left out.

Third, the design targets real use. A low-cost portable probe, a population in which 20% of participants live with HIV, a shortened six-zone protocol to save time, public funding (NIH, NIAID, grant R21AI171755), no declared conflicts of interest, and a commitment to deposit the data in a public repository upon publication.

What is less good

First, population bias in its most classic diagnostic form, spectrum bias. The people with TB are almost all index cases already diagnosed at TB centres, often with a high bacterial load (76% smear-positive), while nearly half of the people without TB are household contacts with no respiratory symptoms at all. Separating these two groups is easier than triaging a queue of patients who are all coughing. Only four household contacts had TB, none asymptomatic: subclinical tuberculosis, the target of active case-finding, could not be evaluated. On top of this comes differential verification: clinic patients had Xpert only, without culture. The authors themselves acknowledge this recruitment may inflate performance.

Second, a misleading metric due to sample size. The test covers 73 people, 28 with TB, from a single random split. A specificity "of 91.1%" whose interval runs from 25 to 98% does not allow anyone to conclude it meets the WHO threshold. As a sign of this instability, the six-zone model shows better specificity than the twelve-zone model despite a lower AUC. No calibration, decision curve or repeated cross-validation backs up the numbers.

Third, untested generalisation and a risk of shortcut learning. One site, one device, two operators trained for the study, and no external validation. The AUC drop from 0.93 to 0.85 simply by changing frame within the same clip shows the model depends on fragile acquisition details; we do not know exactly what it learns, nor how it would behave with other probes or less trained health workers. There is no comparison with chest X-ray read by CAD, the current standard, nor with human ultrasound readers. Code release is not mentioned.

What it changes

For the research community, this preprint confirms a signal already seen in Benin and Zambia: lung ultrasound read by a neural network carries useful information about tuberculosis. It brings two ideas worth reusing: always compare against a clinical-only model, and shorten the acquisition protocol. The next step is known, and the authors name it: external validation in the intended-use population (consecutive people presenting for screening), across several sites and devices, head-to-head with chest X-ray and CAD.

For clinicians, nothing changes today. This is a research prototype with no regulatory status, and the result does not justify replacing an available X-ray. It is also a reminder that a well-built clinical score (weight loss, cough duration, BMI) already reaches an AUC of 0.88 in this population.

For patients and the public, the medium-term promise is concrete: tuberculosis triage that a technician can carry out with a probe and a tablet, where no X-ray exists. But a triage tool with too little specificity burns through test cartridges, and one with too little sensitivity sends contagious patients home. Only a real-world study will tell which way this one leans.

Further reading

The preprint: Deep learning of point-of-care lung ultrasound and clinical data for pulmonary tuberculosis triage in Uganda, Sharma, Nakiyingi, Jovanovska Janeva et al., Emory University, Rutgers New Jersey Medical School, Makerere University, Mulago National Referral Hospital and Georgia Institute of Technology, medRxiv, posted 17 November 2026, DOI 10.64898/2026.11.14.26368125. Not peer reviewed.

On Tatakoto: lung ultrasound and clinical data for tuberculosis in Benin and an audit of vision-language models on tuberculosis chest X-rays.

Editorial transparency note: this article was produced by Tatakoto's automated editorial pipeline from version 1 of the preprint (medRxiv PDF, main text). The "out of 1,000 people" conversion is our own calculation, with a 10% prevalence assumed for illustration, not a result of the study. The English, Spanish and Chinese versions are AI-generated. Spotted an error? Contact us.