CLEW Medical

The models are the regulated device. FDA cleared the CLEWICU System under product code QNL, common name Future Health Condition Prediction Software, and the substance of that device is two machine learning models, CLEWHI and CLEWLR. The company's identity, its evidence base and its entire differentiation rest on those models rather than on the surrounding software. Caveat recorded for buyers: commercial positioning has drifted toward a virtual ICU platform, and the platform layer includes a notification tool called NotifyMe that is user configured rules rather than machine learning. Establish which configuration is being quoted and whether the cleared models are included in it.

The cleared indications for use carry an explicit adjunctive constraint, stating that predictions are for reference only and that no therapeutic decision should be made solely on them. Nothing is auto closed, auto escalated or auto actioned. THE UNDER EXAMINED HALF IS THE LOW RISK MODEL, and it is where this category's real oversight exposure sits. CLEWLR labels patients as unlikely to deteriorate and is marketed as supporting step down and discharge decisions. Its cleared specificity is about 90 percent, so roughly one patient in ten carrying that label does not meet the low risk definition, while its sensitivity of 35.5 to 47 percent means it identifies well under half of genuinely low risk patients, which is the conservative direction. A wrong high risk alert is seen and dismissed. A wrong low risk label quietly removes a patient from attention and nobody is positioned to notice. Ask what the deployment guidance says about acting on CLEWLR and who is permitted to act on it. Held below A because the realised lead time distribution and post deployment alert response behaviour are not published.

Best in this category by a wide margin, and the direct contrast with the Epic Sepsis Model, which is graded D on this axis. Published: what each model predicts and over what horizon, the input sources, the feature count and its reduction from 80 inputs to 50, the pre specified performance targets, and validation results with 95 percent confidence intervals. The peer reviewed CHEST paper adds a SHAP analysis identifying the ten most impactful features across more than 15 million model label comparisons, and describes model features as expert identified collections of data points carrying information about specific organ dysfunction, plus temporal features tracking change over time. Held at B rather than A because the feature list itself, the model class and the training population demographics are not disclosed, and no model card exists.

THE FIRST DIRECT MEASUREMENT OF THIS CATEGORY'S DEFINING FAILURE MODE. Lilly et al., CHEST, doi 10.1016/j.chest.2023.10.036, compared model notifications against telemedicine system alerts and bedside monitor alarms for predicting intubation and vasopressor initiation at UMass Memorial and WakeMed. Reported accuracy was 0.87 to 0.94 for the models, 0.28 to 0.53 for telemedicine alerts and 0.019 to 0.028 for bedside monitor alarms, with roughly 50 fold lower alarm burden than telemedicine alerts and overall test performance more than five times higher. Alarm burden was treated as a first class outcome rather than a footnote, which no other vendor assessed in this category has done. Separately the 510(k) validation ran on two independently evaluated datasets, the UMass eICU dataset of 6,534 patient stays and the public MIMIC III benchmark of 5,069 patient stays. Held at B because there is no randomised trial, no evaluation free of vendor authorship, and no published patient outcome or clinician response data from live deployment. Conflicts are fully disclosed in the paper: four authors are CLEW employees and two are consultants, while the corresponding academic author and three others declared none. A would require an independent study with no vendor authorship, or prospective evidence that the alarm reduction changes what clinicians actually do.

The 510(k) states that the server and interface components are software only and run on infrastructure the user provides, which allows a hospital to keep processing inside its own environment. That is the same class of answer that earned credit for OmniMD and 314e Dexit. The basis is the regulatory filing rather than customer facing material: no privacy policy, data processing addendum or retention statement was located. Open and unaddressed: whether customer data is used for model retraining, which the Predetermined Change Control Plan turns into a live question rather than a theoretical one, and cross border handling given Israeli headquarters serving US health systems.

No public statement on business associate agreements, execution terms or cost was located. Not Rated means no public evidence was found, not evidence of weakness. The company names US health system deployments, so agreements plainly exist, but nothing about them is published. Same handling as the Epic In Basket record.

No SOC 2 of any type, no HITRUST, no ISO 27001, no trust centre and no report request path was located. Worth stating plainly because it is a common buyer confusion in this category: A 510(k) CLEARANCE IS A SAFETY AND EFFECTIVENESS DETERMINATION, NOT AN INFORMATION SECURITY ATTESTATION. The same distinction this index recorded against ISO 13485 applies here. A company holding two FDA clearances and selling into critical care publishes nothing about its information security posture.

The deepest regulatory position in this category and among the strongest in the index. Emergency Use Authorization for the respiratory deterioration model in June 2020. First clearance K200717 in 2021. Second clearance K233216 on 13 January 2024 under 21 CFR 870.2210, Adjunctive Predictive Cardiovascular Indicator, product code QNL, Class II, software validated to IEC 62304. Indications for use, the adjunctive limitation, the pre specified endpoints and the validation results are all public. WHAT THE SECOND CLEARANCE ACTUALLY WAS deserves reading precisely: press coverage described second generation models and a breakthrough, while the filing states its purpose as demonstrating that the same statistical criteria were met after retraining on a reduced input set of 50 features rather than the original 80, alongside broadening the use environment from the ICU alone to all hospital critical care areas at customer request. That is a real and useful advance, because it lowers the data feed bar for hospitals with thinner monitoring integration, but it is not a performance improvement and the announcement did not say so.

B is unusual on this axis and rests on one substantive artifact. The FDA cleared Predetermined Change Control Plan permits the models to be retrained for individual hospitals without a new submission, and it BINDS that retraining in public: training and testing datasets must be completely independent, validation must use patient data from at least three geographically diverse sites, no single site may contribute more than 50 percent of the validation dataset, and any retrained model must still meet minimum sensitivity of 0.6 and PPV of 0.1 for CLEWHI and minimum sensitivity of 0.25 and specificity of 0.9 for CLEWLR. THIS IS A DIRECT ANSWER TO THE PROBLEM THE EPIC SEPSIS MODEL RECORD IDENTIFIES, that a proprietary model performs materially differently at different hospitals and the burden of discovering that falls on the hospital. It is a regulator reviewed commitment about how the model may change after purchase, and nothing else in this index answers that question with a checkable document. Held below A because no subgroup performance breakdown by age, sex, race or admission diagnosis is published anywhere, and because MIMIC III is a single centre dataset whose demographic composition and documentation patterns are carried into the validation unexamined.

Ingests EHR data and patient monitoring device data over an HL7 connection, which is the correct interface choice for a product that needs bedside device streams that FHIR does not cover well, and a heavier integration than a FHIR application. Multiple EMRs are claimed and a One Click feature deep links back into the medical record without a separate log in. Held at B because no EHR partner is named, there is no marketplace listing or partner review credential of the kind that earned Elaborate an A, and the change control plan itself discloses that hospitals present differing input data types and frequencies, meaning integration depth varies by site in a way the vendor does not quantify.

Software only, running on hospital provided server or cloud infrastructure per the 510(k). Deployments named across both bedside and tele ICU settings: UMass Memorial across five facilities, WakeMed Health and Hospitals, EQUUM Medical which licensed it as its in house virtual ICU platform, Sheba Medical Center and Tel Aviv Sourasky Medical Center. THE FILING DISCLOSES A DEPLOYMENT REALITY THE COMMERCIAL MATERIAL DOES NOT: the change control plan anticipates that models may need retraining for hospitals presenting reduced input data types, reduced data frequency or additional input types, so the model a given hospital runs is not necessarily the model that was cleared. Ask directly whether your site receives the cleared model or a retrained variant, who validates it, and against what.

No pricing published at any level. No rate card, no stated unit of pricing, no indicative band, and no published implementation or integration fee. Enterprise quote on contact only. Standard for this category and materially behind the published ladders that earned Affineon a Commercial A in the clinical inbox category.

Scope is explicit and published, which is worth crediting. The cleared indication covers hospital critical care areas including the ICU, emergency department critical care and resuscitation, post anesthesia care, step down, post surgical recovery and specialised units such as cardiac, neurocritical and high dependency care, and it is limited to patients aged 18 and over. A published age limit is a negative scope statement of the kind this index asks detection vendors for, and it rules out paediatric and neonatal use unambiguously. Held at B because the evidence base is two academic eICU datasets plus MIMIC III, all academic or research settings, so community hospital and safety net performance is unestablished. That is precisely the gap the Epic Sepsis Model county emergency department study exposed, where sensitivity fell to 14.7 percent and median lead time was zero.