Tuberculosis: does adding routine clinical data to lung ultrasound improve screening? A Beninese preprint on 502 patients and a 101-patient test set

A South African, Beninese and Swiss team trained a neural network to read point-of-care lung ultrasound, then added a second model that uses routine clinical data (age, fever, night sweats, HIV status, etc.) to screen for tuberculosis. On a test set of 101 patients, averaging the two scores gives an AUROC of 0.95 versus 0.91 for imaging alone, with sensitivity 0.93 and specificity 0.76. The gain is plausible but not demonstrated: the authors themselves say it is not statistically conclusive (p = 0.10), and everything comes from a single hospital.

The context

Tuberculosis remains one of the deadliest infectious diseases worldwide, and screening is often the bottleneck: in many health centres in sub-Saharan Africa there is no radiologist and sometimes no X-ray machine. Point-of-care ultrasound (POCUS) of the lungs is an attractive option: portable probes are inexpensive, radiation-free, and an algorithm can assist a minimally trained operator. Earlier work has shown that deep learning can pick up signs of tuberculosis on these images.

This study asks a more modest but very practical question: since a clinician has basic patient information anyway, how much does the algorithm gain by using it on top of the image? The authors work on the cohort their group already published, this time with a set of clinical variables and a broader definition of tuberculosis cases.

The method

The data. 502 patients with respiratory symptoms were enrolled at the CNHU-PPC university hospital in Cotonou (Benin), a single site. Each underwent a standardised 14-point chest ultrasound protocol (on average 27 acquisitions per patient, from 10 to 33). The reference standard is the Xpert MTB/RIF Ultra molecular test on sputum (which detects the bacillus's DNA); patients diagnosed clinically (symptoms, imaging, outcome) are also counted as positive. In total 224 patients are labelled tuberculosis. Data are split into five cross-validation folds (about 80 patients each; a technique that rotates which subset is used for training and validation) and a fixed test set of 101 patients, 43 of whom have tuberculosis.

The image model. It is a ResNet-18, a standard convolutional network, applied to 224 × 224 greyscale images. A multi-instance learning scheme treats a patient's images as a set: each of the 14 anatomical locations gets its own representation, then an additive attention mechanism weights the images before a single decision layer per patient.

The clinical variables. They fall into three families: 28 self-reported variables (cough, weight loss, fever, night sweats, previous tuberculosis, etc.), 13 measured variables (oxygen saturation, heart rate, blood pressure, temperature, lymph nodes, etc.) and 3 from rapid point-of-care tests (HIV, malaria rapid test and its done/not-done status). A greedy sequential selection keeps 18 variables for the best model.

The fusion. Three strategies are compared: concatenating the representations before the decision, averaging the two probabilities, or training a logistic regression on the two models' outputs. The main metric is the AUROC (area under the ROC curve: the probability that the model gives a higher score to a patient with tuberculosis than to a randomly chosen patient without).

The results

ML metrics. On the test set, the image-only model reaches an AUROC of 0.91 [95% CI 0.86–0.96]. Clinical-only models get 0.88 to 0.89 with variable selection. The best fusion (variables from all three families, selection, score averaging) reaches 0.95 [0.91–0.99]; logistic regression gives 0.94 and concatenation 0.93. At a threshold set for 90% sensitivity during development, the fused model shows on test a sensitivity of 0.93 [0.81–0.98] and a specificity of 0.76 [0.63–0.85]. The gain over imaging alone is about 4 AUROC points, with p = 0.10: chance cannot be ruled out. Variables from point-of-care rapid tests add little, and on their own have almost no discriminative power (development AUROC 0.57).

Clinical translation. Apply these figures to 1,000 patients with the same tuberculosis proportion as the test sample (43%, i.e. 426 cases): about 396 would be detected and 30 missed (false negatives), while among the 574 without tuberculosis about 138 would be wrongly referred for confirmatory testing (false positives). But the sensitivity confidence interval (0.81–0.98) allows anywhere between 9 and 81 missed cases. In a mass-screening setting with lower prevalence, say 10% (an illustration, not from the paper), the same rates would give 7 missed cases but about 216 false positives per 1,000 people: false positives would become the real operational question. The authors state that these values meet WHO targets for a triage test, but those targets are normally assessed with much narrower intervals than a 101-patient test can provide.

What is good

A pragmatic question. Rather than hunting for a new model, the study measures what information already collected in the consultation is worth, and which is useless. The selection result (18 variables suffice, saving the entry of 27 more values) is directly usable to design a lightweight tool.

Simple, compared fusions. Three methods are tested, and the simplest (score averaging) is the best or as good as the others. This is an honest result that avoids needless complexity. The authors also note that other variants (maximum, minimum, calibration) did not help.

Transparency on data and ethics. The study is registered (NCT05423847), approved by an ethics committee, with written consent from each participant. Data are available for research under a data use agreement via Hugging Face. The equipment maker (Butterfly) states it took no part in funding, collection, analysis or conclusions; funding comes from the European Union (EDCTP3), Switzerland and the Swiss Lung Association.

What is less good

A tiny test set and an unestablished gain (misleading metric). 101 patients, 43 with tuberculosis: confidence intervals are wide and the 4-point AUROC gain does not cross the significance threshold (p = 0.10). The decision threshold is also set on development data, and the operating performance (sensitivity 0.93, specificity 0.76) rests on very few cases.

One hospital, no external validation (population bias). The 502 patients come from a tuberculosis referral centre in Cotonou: high prevalence (45% in the cohort), patients referred for respiratory symptoms, a single probe type. Nothing guarantees the model works in a primary-care health centre, with another operator or another prevalence. The authors report no external validation, and no code is mentioned.

A shifting case definition and variables possibly entangled with the reference standard (information leakage and noisy labels). Patients "diagnosed clinically" are counted as positive: that diagnosis may have been made from the very symptoms and imaging the model uses, which can make clinical variables artificially predictive. The paper does not analyse this. In addition, this definition differs from the group's earlier work on the same cohort (which counted only bacteriologically confirmed cases, AUROC 0.91), complicating any comparison. Finally, the paper has no dedicated limitations section and does not explicitly state that the split is made at patient level.

What it changes

For the research community. The paper offers a useful comparison framework (ultrasound alone, clinical alone, fusion) and an accessible dataset. It will need a larger sample, several sites, and ideally a prospective evaluation where the model is frozen before testing.

For clinicians. Nothing changes today: this is a non-peer-reviewed preprint without external validation. The practical lesson is mainly that a few consultation items (weight, fever, sweats, history, HIV status) seem to complement the image, a reminder that the algorithm does not replace the clinical history.

For patients and the public. AI-assisted lung ultrasound could one day ease triage where there is no radiologist, but it is not yet a validated tool. A positive triage result is not a diagnosis: it points to a confirmatory test.

Going further

Preprint: arXiv:2610.06034 (submitted 5 October 2026, version 1, not peer-reviewed). Authors: Joshua M. Jansen van Vüren, Christiaan M. Geldenhuys, Devendra S. Parihar, Véronique Suttels, Trevor Brokowski, Ablo P. Wachinou, Mary-Anne Hartley, Rensu P. Theart, Grant Theron and Thomas R. Niesler. Dataset: AI-POCUS-COP/ultrai-lung-ultrasound on Hugging Face (data use agreement required). Clinical study: NCT05423847. Related work from the same group: arXiv:2610.03290 (echocardiography pretraining for lung ultrasound).

Editorial transparency: this article was produced with the help of artificial intelligence tools from the preprint text; every figure comes from the paper, except the per-1,000-patient conversions, calculated by the editors. The English, Spanish and Chinese versions are machine translations.