What's New in Pulmonology AI — The State of the Field in Mid-2026

Several recent studies show how quickly AI is developing in pulmonology — from drug discovery and imaging to wearables and mobile health. Most of the data presented here are early feasibility or diagnostic validation findings, not evidence of improved patient outcomes. That distinction matters.

⏱️ Reading time: ~3 minutes

💊 AI-Generated Drug for IPF: Early Clinical Signals

Idiopathic pulmonary fibrosis (IPF) has limited treatment options. In June 2025, Nature Medicine published the first Phase 2a multicenter, double-blind, randomised trial of a drug identified by generative AI — a TNIK (TRAF2 and NCK-interacting kinase) inhibitor designed computationally and advanced to human testing.[1] Phase 2a data are early, and this is not a cure. It is, however, an important proof of feasibility: an AI-generated molecule has now completed the full arc from computational design to randomised human trial. Whether the compound will demonstrate meaningful clinical efficacy requires further trials.

🫁 Digital Lung Twins: Promising, but Ex Vivo

A team at the University of Toronto built digital twins of individual human lungs from ex vivo lung perfusion (EVLP) data — integrating more than 75 parameters across physiology, biochemistry, radiography, and multi-omics — and accurately predicted individual therapeutic responses.[2] Important caveat: these are lungs outside a living body, used in transplant medicine. Digital twins built from routine clinical data in living patients do not yet exist. As a proof of concept for personalised lung pharmacology, it is compelling; clinical application remains some distance away.

📊 Open CXR Foundation Model: Useful Approach, Open Implementation Questions

Ark+, published in Nature in July 2025, is a chest X-ray (CXR) foundation model built to work across clinical settings — not just where it was trained.[3] The authors report that code and pretrained models have been released publicly, which allows institutions to test and adapt it without a licensing arrangement. For institutions considering AI-assisted CXR interpretation, this reduces vendor dependency and allows local validation. Whether "open" extends to commercial use and what infrastructure is required for deployment should be verified directly before any implementation decision. Full performance figures are in the paper.

🔧 AI-Guided Lung Ultrasound for Non-Expert Staff

A prospective multicenter study in JAMA Cardiology tested AI-guided lung ultrasound in 176 patients presenting with dyspnoea (shortness of breath): 98.3% of examinations by non-physician staff met diagnostic quality standards — on par with expert-acquired images (96.6%).[4] The study population had a high prevalence of cardiogenic causes; performance in primarily pulmonary presentations may differ. A direct comparison between AI-assisted and unassisted acquisition by the same staff was not performed — a limitation the authors acknowledge. These findings suggest a potential role for AI-guided ultrasound in expanding access, but external validation in broader dyspnoea populations is needed before drawing wider conclusions.

📡 A Wearable Patch That Listens to Your Lungs

A miniature MEMS (micro-electromechanical systems) seismometer patch — taped to the chest wall — simultaneously quantifies work of breathing (WoB) and detects crackles, wheeze, and normal breath sounds via deep learning.[5] In 96 patients, the model achieved 93% accuracy and 97% specificity, with an AUC (area under the precision-recall curve) of 0.97. With n=124 total across a single US centre, this is early-stage research. The concept — continuous respiratory monitoring outside a hospital — is clinically relevant, but larger multicentre studies are needed.

📱 Screening for COPD Exacerbations with a Smartphone

A prospective study across 13 hospitals in China tested an AI system that uses a standard smartphone microphone to detect acute exacerbations of COPD (AECOPD).[6] Patients place the phone on the chest and perform deep-breathing and coughing manoeuvres; the AI analyses the audio without requiring patient-reported symptoms. In 256 analysed patients, it achieved an AUC of 0.955 (95% CI: 0.929–0.976), with 89.9% sensitivity and 93.8% specificity — consistent across age, disease severity, and noise environments. The study is China-based; cost data and healthcare context differ substantially from Switzerland. Whether performance holds in European primary care settings has not been confirmed.

⚠️ What AI in Pulmonology Still Cannot Prove

The studies above were primarily validated on diagnostic accuracy — how well a system detects or classifies a condition. Whether more accurate or earlier detection translates into fewer hospitalisations, slower disease progression, or reduced mortality was not the primary endpoint of any of these studies. Secondary or surrogate endpoints may have been reported; the full papers should be reviewed. Diagnostic accuracy is a necessary first step, not evidence of clinical benefit.

🇨🇭 A Note for Swiss Practice

All studies were conducted outside Switzerland. For market access in Switzerland, CE marking under the EU MDR (EU Medical Device Regulation) is generally the basis — Switzerland recognises CE-marked devices under its own medical device ordinance. Specific registration obligations depend on the product category, the software classification, and applicable transitional rules; they are not uniform across all AI tools. Swissmedic’s device database (Swissdamed) is the authoritative reference for current status. For what is in use at your institution, your clinical informatics or respiratory medicine team is the right starting point.

📌 In Brief

• An AI-generated drug for IPF reached a Phase 2a RCT — an important feasibility milestone; whether it demonstrates clinical efficacy requires further trials
• Digital lung twins predicted therapeutic responses in ex vivo lungs — compelling proof of concept, not yet applicable in living patients
• Ark+, a chest X-ray foundation model with publicly released code and weights, addresses generalisability — commercial and deployment terms should be verified before use
• Non-expert staff achieved diagnostic-quality ultrasound in 98.3% of cases with AI guidance — findings need external validation across broader dyspnoea populations
• A wearable MEMS patch detected crackles and wheeze with 97% specificity in 96 patients — early-stage, single-centre; larger studies are needed
• A smartphone AI detected COPD exacerbations with AUC 0.955 in a 13-hospital Chinese study — performance in European settings has not been confirmed
• None of these studies had patient outcomes as a primary endpoint — diagnostic accuracy is a necessary first step, not evidence of clinical benefit

📚 Sources

1. Xu Z et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nat Med. 2025;31:2602–2610. Link
2. Zhou X, Wang B et al. Digital twins of ex vivo human lungs enable accurate and personalized evaluation of therapeutic efficacy. Nat Biotechnol. 2026. Link
3. Ma D, Pang J, Gotway MB, Liang J. A fully open AI foundation model applied to chest radiography. Nature. 2025;643(8071):488–498. Link
4. Baloescu C et al. Artificial Intelligence–Guided Lung Ultrasound by Non-experts. JAMA Cardiol. 2025. Link
5. Sang B et al. A MEMS seismometer respiratory monitor for work of breathing assessment and adventitious lung sounds detection via deep learning. Sci Rep. 2025;15:9015. https://doi.org/10.1038/s41598-025-93011-7
6. Gong Y et al. AI-driven smartphone screening for acute COPD exacerbations: enhancing health equity in developing regions. npj Digit Med. 2025;8:715. https://doi.org/10.1038/s41746-025-02086-z

Disclosure: No conflict of interest. Sources independently identified by SwissMedAI from peer-reviewed literature.