Radiology · Comprehensive chest X-ray decision-support
Annalise CXR
Annalise.ai
Annalise CXR is the broadest chest X-ray model on the market — 124 findings under its CE Class IIb mark, against the single-digit label sets of most competitors — and its pivotal Lancet Digital Health reader study remains the largest of its kind, improving radiologist accuracy on 102 of 127 findings. A three-continent regulatory footprint (FDA, CE-MDR, TGA) and deployment at scale across 40-plus NHS trusts give it unusual real-world depth for a product this young. But the evidence base is multireader and observational rather than randomised, and the US clearance covers only a five-finding triage subset of the comprehensive product — which places the platform at two marks rather than three.
Performance Metrics
Clinical Evidence
Annalise CXR's defining feature is the breadth of its label set, and its pivotal evidence matches that shape. The multireader multicase study (Seah et al., *Lancet Digital Health* 2021;3:e496–e506) — funded by Annalise.ai and the largest MRMC study yet conducted by number of findings concurrently evaluated — had 20 radiologists read 2,568 enriched cases with and without model assistance across 127 findings. Assisted radiologists' macro-averaged AUC rose from 0.713 (95% CI 0.645–0.785) unassisted to 0.808 (0.763–0.839); accuracy improved statistically significantly for 102 of 127 findings (80%), was non-inferior for 19 (15%), and decreased for none. The underlying model was trained on 821,681 images from 284,649 patients. Real-world evidence followed at an Australian private teleradiology practice (Jones et al., *BMJ Open* 2021;11:e052902): across 2,972 consecutive cases reviewed with the model live, 92 cases (3.1%) had significant report changes attributable to the model, 43 (1.4%) had changed patient management, and radiologists fully agreed with the model in 86.5% of cases; 16 findings across 13 cases (0.5%) were judged missed by the model. The US triage clearances rest on independent standalone validation at four Mass General Brigham hospitals. For pneumothorax (Hillis et al., *JAMA Network Open* 2022; n=1,000 retrospective radiographs), the model detected simple pneumothorax with 94% sensitivity and 92% specificity, and tension pneumothorax with 94.5% sensitivity and 95.3% specificity — above the FDA's CADt performance benchmarks. For vertebral compression fracture (*JACR* 2024; n=596 consecutive radiographs with frontal and lateral projections from the same network), sensitivity was 89.3% and specificity 89.2%; notably, only 36.4% of the true-positive patients had a charted vertebral compression fracture diagnosis and only 33.1% were on disease-modifying osteoporosis medication, supporting the opportunistic-screening rationale behind the Bone Health and Fracture Liaison Service indication. The caveats are structural. No randomised controlled trial of Annalise CXR has been published as of the review date; the MRMC design used an enriched, retrospective dataset and the real-world study was observational and vendor-supported. The 124-finding comprehensive product evaluated in the Lancet study is not what is cleared in the US — the FDA footprint covers a five-finding triage subset — so published comprehensive-model performance should not be read across to the US product without that distinction.
| Study | Design | n | Sensitivity | Specificity | AUC | Published |
|---|---|---|---|---|---|---|
| Seah JCY, Tang CHM, Buchlak QD, et al. (MRMC, 20 radiologists) | RetrospectiveRetrospective | 2,568 | — | — | 0.808 assisted vs 0.713 unassisted (macro-averaged) | Lancet Digital Health, Jul 2021 |
| Jones CM, et al. (real-world observational, Australian teleradiology) | ProspectiveProspective | 2,972 | — | — | — | BMJ Open, Dec 2021; 3.1% significant report changes |
| Hillis JM, et al. (standalone, 4 Mass General Brigham hospitals) | RetrospectiveRetrospective | 1,000 | 94% | 92% | — | JAMA Network Open, Dec 2022; tension PTX 94.5% / 95.3% |
| VCF identification cohort — 4 US hospitals (Ghatak, Hillis, et al.) | RetrospectiveRetrospective | 596 | 89.3% | 89.2% | — | JACR, 2024; only 36.4% of true positives had a charted VCF diagnosis |
Clinical Pulse
Verified clinician sentiment
Verified clinician reviews launching · Apply to contribute →
Reviews are signed by clinicians verified against their regulator's public register.
Inside the Auris+ Listing
Six more sections complete this device’s Auris+ Listing.
Inside the algorithm
ProPro unlocks the five-stage editorial walkthrough of how this device's AI reaches a verdict — model architecture, training data, inference latency, and where the clinician stays in the loop.
Decision Ledger
ProPro unlocks a private, cross-vendor log of every case you read on this device — what the AI called, what you concluded, and a one-line reason if you overrode it. Ready for the EU AI Act's deployer logging obligations when they land in 2028.
Clinical Evidence Deep Dive
ProPro unlocks the structured clinical-evidence summary — study count, target patient population, and a tabular accuracy-metrics view drawn from peer-reviewed sources.
Peer-Reviewed Publications
ProPro unlocks the curated peer-reviewed publication list with PubMed cross-links — the citation backbone of every editorial verdict.
Post-Market & Regulatory Conditions
ProPro unlocks the post-market surveillance summary, recall record, and the conditions of approval that bound real-world use.
AI Algorithm Version History
ProPro unlocks the chronological record of algorithm version changes — what changed when, drawn from manufacturer changelogs and regulatory filings.
Regulatory Approvals
K250831
Class II
Class IIb
Safety Record
No Annalise.ai device recalls or FDA safety communications were identified in publicly available sources as of June 2026; the MAUDE adverse-event database could not be queried directly from this environment, so this should be read as "none found in public reporting" rather than an exhaustive audit. The products are decision-support and notification overlays — they do not alter images, remove studies from the reading queue, or drive treatment — which limits the harm surface. The principal documented risk is over-reliance on a negative output: published sensitivities (94% pneumothorax, 89.3% vertebral compression fracture in independent testing) leave a real false-negative floor, and the broader chest X-ray AI literature documents hidden-stratification effects (such as pneumothorax models keying on chest-drain presence), so the absence of a flag does not substitute for the radiologist's independent read.
Intended Use & Indications
Annalise CXR exists in two regulatory forms that procurement reviewers should keep distinct. The comprehensive decision-support product (Annalise Enterprise CXR), CE-marked Class IIb under the EU MDR and approved for clinical use in Australia and New Zealand, analyses adult chest X-rays and presents up to 124 findings with heatmap localisation alongside the radiologist's standard read. The US product (Annalise Triage CXR / Annalise Enterprise) is a narrower computer-aided triage and notification (CADt) device: it analyses adult chest X-rays in the background and raises case-level worklist flags for five cleared findings — pneumothorax, tension pneumothorax, pleural effusion, pneumoperitoneum, and vertebral compression fracture — without marking the image or providing a diagnosis. In both forms the system is a decision-support tool; final diagnostic and treatment decisions remain with the interpreting clinician.