Neurology · Acute stroke triage on non-contrast CT
JLK-NCCT
JLK, Inc.
A March 2026 clearance that puts JLK's haemorrhage and occlusion models on the first scan a stroke patient gets. The FDA summary reports 78.5 per cent sensitivity and 90.3 per cent specificity for occlusion on plain CT, and 97 per cent or better for haemorrhage. The published evidence is retrospective, company-authored and, for the occlusion model, split across cohorts that behave differently.
Performance Metrics
Clinical Evidence
The device bundles two tasks with two different evidence records. For haemorrhage it reuses the JLK-ICH model cleared as K243363 in January 2025. That clearance's summary reports 97.3 per cent [[sensitivity]] (95 per cent CI 94.8 to 99.5) and 97.9 per cent [[specificity]] (95.5 to 99.5), AUC 0.978, on 376 US scans (188 with haemorrhage) read by two US board-certified neuroradiologists with a third as tie-breaker. Sensitivity held at 0.92 or above in every haemorrhage subtype except three extradural cases (0.67), and fell to 0.80 for bleeds under 1 mL (n=20), 0.90 in Black or African American patients (n=60) and 0.83 in the eight Asian and Pacific Islander patients. The peer-reviewed counterpart is Kang and colleagues (Neuroradiology 2025): 1,370 scans from one Korean tertiary centre, 98.7 per cent sensitivity (97.8 to 99.3), 88.5 per cent specificity (83.6 to 92.3), AUROC 0.936, with six non-expert readers improving from AUROC 0.953 to 0.967 (P=0.009) with the model's help. The specificity gap between the two datasets, 97.9 against 88.5 per cent, is the number to hold in mind. Occlusion on plain CT is the harder problem and the newer claim. The K252421 standalone test, the only validation the 510(k) summary reports for this device, used 288 cases (144 with LVO): 78.5 per cent sensitivity (71.9 to 84.7), 90.3 per cent specificity (85.1 to 94.7), AUC 0.880 (0.837 to 0.920). In the accompanying reader comparison the software's sensitivity of 0.792 exceeded the unaided average of 0.568 across neuroradiologists and general radiologists (difference 0.224, 95 per cent CI 0.144 to 0.306), and its specificity of 0.933 exceeded their 0.840 (difference 0.093, 0.038 to 0.150; general radiologists 0.796, neuroradiologists 0.869). The study was set up to show non-inferiority to neuroradiologists and superiority to general radiologists, and the summary says both were met. That is a comparison of the software against readers working alone, not a study of readers using the software. The published record for JLK's non-contrast occlusion algorithm, reported under the name JLK-CTL, is company-authored throughout and the 510(k) summary does not say whether the cleared model is version-identical to it. Kim and colleagues (Stroke 2024) describe the method: handcrafted hemispheric density and volume asymmetries plus a [[deep-learning]] clot sign, trained on 2,463 patients from five hospitals, with 80.1 per cent sensitivity and 88.6 per cent specificity (AUC 0.888) in a 95-patient external set. Chung and colleagues (Stroke: Vascular and Interventional Neurology 2025) tested it in 534 consecutive patients at six Korean centres: AUC 0.859, sensitivity 0.787, specificity 0.832, and 0.898 specificity in 274 non-stroke subjects. Sunwoo, Ryu and colleagues (JNIS 2026) added a multinational set: AUC 0.963 in 723 consecutive Korean patients but 0.899 in a 240-patient US case-control set, and an eight-reader crossover study in which AI assistance raised pooled sensitivity from 46.6 to 63.7 per cent and specificity from 91.9 to 94.9 per cent. No prospective or randomised study of the device exists, so the evidence grade is retrospective. One FDA clearance and a published literature earn a single AurisMark; two marks would need a prospective study or documented broad uptake, and the product has neither yet.
| Study | Design | n | Sensitivity | Specificity | AUC | Published |
|---|---|---|---|---|---|---|
| JLK, Inc. (K252421 standalone LVO validation; manufacturer-sponsored) | RetrospectiveRetrospective | 288 | 78.5% (95% CI 71.9–84.7) | 90.3% (95% CI 85.1–94.7) | 0.880 (95% CI 0.837–0.920) | FDA 510(k) summary K252421 (March 2026), https://www.accessdata.fda.gov/cdrh_docs/pdf25/K252421.pdf; 144 LVO-positive and 144 negative NCCT cases, 2+1 US board-certified neuroradiologist ground truth; reader comparison: software sens 0.792 / spec 0.933 vs unaided readers 0.568 / 0.840. Not peer-reviewed |
| JLK, Inc. (K243363 standalone ICH validation; the ICH model reused unchanged in JLK-NCCT) | RetrospectiveRetrospective | 376 | 97.3% (95% CI 94.8–99.5) | 97.9% (95% CI 95.5–99.5) | 0.978 (95% CI 0.959–0.992) | FDA 510(k) summary K243363 (January 2025), https://www.accessdata.fda.gov/cdrh_docs/pdf24/K243363.pdf; 188 ICH-positive and 188 negative US scans; subgroup sensitivity 0.80 for bleeds <1 mL (n=20), 0.67 for extradural (n=3), 0.83 in Asian and Pacific Islander patients (n=8). Not peer-reviewed |
| Kim PE, Yang H, Kim D, et al. (JLK-CTL development and external validation; company co-authors) | RetrospectiveRetrospective | 95 | 80.1% (95% CI 72.0–88.1) | 88.6% (95% CI 84.7–92.5) | 0.888 (95% CI 0.850–0.925) | Stroke 2024;55(6):1609–1618 (PMID 38787932). Seven Korean hospitals; 2,463 training, 275 internal and 95 external validation patients (LVO prevalence 26.3% in the external set); ExtraTrees model on hemispheric NCCT features plus a deep-learning clot sign |
| Chung JW, Lee M, Ha SY, et al. (JLK-CTL multicentre validation; company co-authors) | RetrospectiveRetrospective | 534 | 78.7% (95% CI 71.6–84.7) | 83.2% (95% CI 79.0–86.9) | 0.859 (95% CI 0.827–0.887) | Stroke Vasc Interv Neurol 2025;5(5):e001788 (PMID 41573318). Consecutive patients with NCCT and CTA within 24 h at six Korean stroke centres, 30.7% with LVO; specificity 0.898 in 274 non-stroke subjects; LVO score correlated with follow-up infarct volume (r=0.54) |
| Sunwoo L, Ryu WS, Buch K, et al. (multinational validation and reader study; company authors, first author is JLK's vice CMO) | RetrospectiveRetrospective | 963 | — | — | 0.963 (Korea, n=723) / 0.899 (US, n=240) | J Neurointerv Surg 2026, online 2 June (PMID 42173676). Consecutive Korean cohort (127 LVO) and US case-control cohort (120 LVO); eight-reader crossover: AI assistance raised pooled AUC 0.718 to 0.852, sensitivity 46.6% to 63.7%, specificity 91.9% to 94.9% (all P<0.001); NRI 5.5%, NNS 18.2 |
| Kang DW, Kim M, Park GH, et al. (JLK-ICH pivotal validation and reader study; company co-authors) | RetrospectiveRetrospective | 1,370 | 98.7% (95% CI 97.8–99.3) | 88.5% (95% CI 83.6–92.3) | 0.936 (95% CI 0.915–0.957) | Neuroradiology 2025;67(6):1511–1519 (PMID 40116947). Single Korean tertiary centre, 2+1 expert ground truth, additional external validation on a multi-ethnic US dataset; six non-expert readers on 800 scans improved from AUROC 0.953 to 0.967 with assistance (P=0.009) |
Inside the algorithm
Editorial featureHow JLK-NCCT reaches a notification.
Five stages — from raw input to verdict — drawn from manufacturer documentation and the public regulatory record.
- INGEST
- NORMALISE
- DETECT
- LOCALISE
- VERDICT
Stage 01 · INGEST
The first scan, forwarded as it is acquired.
The input is the non-contrast head CT, the first study almost every suspected stroke patient gets. Studies are forwarded as objects from the scanner to a JLK server that runs on premises or in the cloud. integration is optional. Training covered scanners from GE, Siemens, Philips and Toshiba and 3,067 cases from US and South Korean hospitals; the labelling limits use to adults and excludes scans degraded by motion.
Input
Non-contrast head CT
Inference
Per study, on receipt
Inside the Auris+ Listing
Five more sections complete this device’s Auris+ Listing.
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
Safety Record
The device has been on the US market since March 2026 and the record is correspondingly short. openFDA returned no recall, enforcement or MAUDE adverse-event report for JLK as manufacturer or brand as of 7 September 2026. The labelling bounds the risk: the device is notification-only, does not mark the image, and states that all cases should still undergo CT angiography. The practical hazard is the [[false-negative]]: at 78.5 per cent standalone sensitivity roughly one occlusion in five will not trigger a notification, and the labelling says an absent notification must not be read as exclusion. Excessive motion, haemorrhagic transformation and very thin or absent ventricles are listed contraindications.
Intended Use & Indications
Per the FDA 510(k) record (K252421), JLK-NCCT is a radiological computer-aided triage and notification software designed for analysing non-contrast head CT (NCCT) images. It assists hospital networks and trained clinicians by flagging findings suggestive of (1) intracranial haemorrhage (ICH) and (2) large vessel occlusion (LVO) involving the internal carotid artery (ICA), middle cerebral artery M1 (MCA-M1) and middle cerebral artery M2 (MCA-M2). Analysis runs on an on-premises or cloud-based JLK server in parallel with the ongoing standard-of-care interpretation. Users receive notifications for suspected ICH and LVO on mobile devices; the compressed preview images are for information only. The device does not modify the original image and is not a primary diagnostic device. It is for adults only. The labelling states that cases without a "Suspected ICH" or "Suspected LVO" notification must not be assumed to exclude either condition, and that all cases should undergo CT angiography as part of the standard stroke work-up.