AgileMD

In eCART the model is the cleared device, and the filing says so: the primary functions of eCART are imparted by a gradient boosted machine learning algorithm that takes input from the EHR in real time (K233253). Remove the model and there is no eCART. That holds for eCART only.

AgileMD also sells Clinical Pathways, a content and workflow product with more than 4,800 pathways live, whose value is curated clinical logic rather than a model; eCART's workflow directs clinicians into those pathways, but Clinical Pathways on its own is content, not AI.

What the system may and may not do is written into the clearance (K233253). It scores adult ward patients from validated EHR data only: vitals confirmed by a registered nurse count, and unvalidated data streaming from monitors is not used until a professional confirms it. Its predictions are for reference only, and no therapeutic decision is to be made on the score alone. The default thresholds are published (93 for moderate and 97 for high risk on a 0 to 100 scale), each score is shown with the observed odds of deterioration in the next 24 hours, and the output drives clinicians into workflows their own institution sets. The company's intervention study describes one such protocol: at the high tier a physician assesses the patient for ICU admission, at the intermediate tier nurses take vitals every two hours and call a physician. There is no low risk output that releases a patient from attention.

The open question is volume. The moderate threshold fires on about 7.5 percent of scored observations, and nothing is published about how many alerts are dismissed or acted on at a typical site.

What produces the output is named, trained data is characterized, and the version is identified. The 510(k) names a gradient boosted machine and describes the score: the model's predicted probability of death or ICU transfer, scaled 0 to 100 on specificity, shown with the observed odds of deterioration in the next 24 hours, with default thresholds of 93 and 97. The development paper (Critical Care Explorations, 2025) describes the training data (901,491 admissions at seven hospitals in three health systems, 2006 to 2022) and the predictor groups (demographics, vital signs, documentation and laboratory values), and compares the model with its predecessor eCARTv2. The product is sold and cleared as eCARTv5, so the version in use is named, and the group has published earlier generations and a paper showing that age, heart rate and respiratory rate alone recover much of the performance (Resuscitation, 2021).

Gaps: the full variable list is not published, there is no model card, and the company says nothing about how and when the model is updated; the clearance carries no change control plan.

Structurally short at the model layer: eCART is the company's own gradient boosted model, cleared by FDA as a specific version, so it cannot be swapped quietly for another provider's model. The rest of the chain is not named. The platform is hosted by the company off site, so patient data leaves the hospital by design, but no cloud provider, region or subprocessor list is published; the privacy notice says service providers for hosting, analytics and IT support may receive information and that it may be processed anywhere in the world, and the terms for organizations mention partners outside the United States and University of Chicago as a licensor.

The company's headline claim is an outcome (decreased mortality, length of stay and cost), and it has published an outcome study as well as prospective multi site validation. Winslow and colleagues (Critical Care Medicine, 2022) ran a pragmatic before and after study across a four hospital community and academic health system: with scores shown and an escalation protocol, in hospital mortality among elevated risk ward patients fell from 13.9 to 8.8 percent (adjusted odds ratio 0.60, 95 percent CI 0.52 to 0.71), with more and earlier ICU transfers. The current model was developed and validated in Critical Care Explorations (2025): 901,491 development admissions, then retrospective validation on 1,769,461 admissions and prospective validation on 205,946 admissions across 21 hospitals in three health systems, AUROC 0.834 retrospectively, better than NEWS, MEWS and its predecessor. A Yale New Haven comparison of six early warning scores on 362,926 encounters (JAMA Network Open, 2024) found eCART had the highest AUROC (0.895) and the highest positive predictive value at matched sensitivity.

The intervention study was not randomized, used an earlier version, and saw mortality fall in the average risk group the protocol did not touch, which points to other changes over the period. The company's founders co-author most of the literature and hold equity or royalty interests, declared in each paper, including the Yale study. The featured item on the company's publication page is a conference abstract, and Yale's reported 13 percent relative mortality reduction comes from a trade press interview, not a paper.

The assurances are general, with no specific commitments behind them. The home page says the cloud engine is SOC 2 and HIPAA compliant, hosted with no on site hardware, behind layers of protection, encryption and network security, with audit logging and monitoring. Nothing published states how long patient data is kept, whether it trains or updates the model, or how it is de-identified.

The company's privacy notice lists among its uses improving the services through artificial intelligence, developing new products and creating de-identified or aggregated information, and says information may be processed and stored anywhere in the world; it reads as covering account holders rather than patients, and the terms for organizations say the company does not knowingly collect individually identifiable health information, which does not describe how a cloud engine scoring EHR data works. eCART is on its fifth version, so models have been rebuilt, and nothing says from what.

Compliance is claimed and not supported by a published document. The home page says the cloud engine is SOC 2 and HIPAA compliant, but no business associate status, agreement, template or summary of terms is published. The documents that are published do not reach eCART's data: the privacy notice covers account holders, and the terms for organizations say the company does not knowingly collect individually identifiable health information and prohibits customers from providing it, which cannot be the arrangement for a cloud engine that scores EHR data.

A recognized attestation is claimed in the company's own material without the artifact. The home page says the cloud engine is SOC 2 compliant, without saying whether the report is Type 1 (controls at a point in time) or Type 2 (controls tested over a period), its date or its auditor, and no trust portal or request path is published.

Two things point to a working security function: a published vulnerability disclosure policy with a PGP key and a defined scope (the main site, the API and the demo environment), and a public system status page.

Cleared, identifiable and marketed inside its indication. K233253 (eCARTv5 Clinical Deterioration Suite, decided 21 June 2024; 21 CFR 870.2210, adjunctive predictive cardiovascular indicator, Class II, product code QNL; predicate CLEWICU, K200717; reference device PeraServer and PeraTrend, K172959) clears the gradient boosted model itself, for risk stratification and early warning of death or ICU transfer in hospitalized ward patients 18 and older, from nurse validated EHR data, as an aid that does not replace clinical judgment. The company's pages call eCART FDA cleared AI for clinical deterioration, which matches. Clearance rested on retrospective (1,769,461 encounters) and prospective (205,946 encounters) validation in three health systems.

No predetermined change control plan is described, so how the model may change after purchase is not governed by a published, regulator reviewed plan. The company also lists a pediatric machine learning early warning outcome study among its eCART publications, while the clearance covers adults only, and it does not say whether a pediatric version is in use or on what basis.

Subgroup performance is published by race and by clinical condition. Table D of the 510(k) summary (K233253) reports AUROC, sensitivity, specificity, positive and negative predictive value at both default thresholds for five race groups in the retrospective validation, with patient counts: American Indian or Alaska Native (3,282 patients), Asian or Mideast Indian (17,657), Black or African American (116,642), Native Hawaiian or Other Pacific Islander (1,427) and White (735,514). Tables B and C report the same measures for congestive heart failure, COVID-19, chronic pulmonary disease and sepsis. The 2025 development paper (Critical Care Explorations) reports AUROC of at least 0.80 across demographic and clinical subgroups.

The company says all five race groups met its performance thresholds. At the moderate threshold, sensitivity ranges from 46.6 percent for American Indian or Alaska Native patients and 48.1 percent for Black patients to 56.5 percent for Native Hawaiian or Other Pacific Islander patients, against 52.3 percent for White patients, and AUROC from 0.814 to 0.862; the two smallest groups have wide intervals. The outcome is death or ICU transfer, and transfer is a clinical decision, so any difference in how readily particular patients are transferred is learned as signal, an exposure the company does not address.

The company has put its error rates on the record at the operating points it sells. The 510(k) summary (K233253, decided 21 June 2024), the manufacturer's own filing, reports at the default moderate threshold of 93: sensitivity 51.8 percent retrospectively (902,951 of 1,744,044 deterioration events) and 48.8 percent prospectively (139,043 of 284,678), specificity about 93 percent, and positive predictive value 9.0 and 8.9 percent, with alerts on about 7.5 percent of observations; at the high threshold of 97, sensitivity 38.6 and 33.7 percent and specificity about 97 percent. The retrospective cohort is 1,769,461 encounters (2009 to 2023) and the prospective one 205,946 encounters (2023 to 2024) in three health systems; the outcome is death or ICU transfer within 24 hours. The observed event rate at each score is also shown to the clinician as odds of deterioration in the next 24 hours.

Nothing stands behind the output beyond that. The company's terms say output may be inaccurate, that it has no liability for loss arising from output, and that clinicians are solely responsible for verifying it; no warranty, indemnity or remediation commitment is published.

Integration is described by what it does, and no record system is named. The company says eCART and its pathways embed in native EHR workflows, and that pathways let clinicians order directly, align order sets, pull in live labs and vitals, generate documentation and use calculators. The 510(k) says the health system must supply an EHR connection and data interfaces.

No EHR vendor, interface standard, marketplace listing or integration document is named on the company's pages, although the Yale New Haven comparison study shows eCART running beside the Epic Deterioration Index at seven hospitals.

A single hosted option, with the location not committed. The 510(k) describes eCART as cloud based software that the health system connects to through its own EHR connection and data interfaces, and the company says its hosted platform needs no on site hardware and deploys in days to weeks.

No cloud provider, region or residency commitment is published; the privacy notice says information may be processed and stored anywhere in the world, and the terms for organizations mention partners outside the United States. The company reports more than 400 hospitals and 150,000 clinical users; named users on its news page include Yale New Haven Health's seven hospitals and John Muir Health.

No price is published: no rate card, unit of charge, band or implementation fee, and every route leads to a demo request. What is published is how the product is delivered (a hosted cloud engine with no on site hardware, deployed in days to weeks, with reporting at system, facility, unit and provider level) and that eCART and Clinical Pathways are separate products.

The company has also published a conference abstract measuring the return on investment of the product (Circulation, 2023) and markets reduced mortality, length of stay and direct variable costs, so the benefit side is quantified while the cost side is not.

The cleared scope is precise: hospitalized ward patients 18 and older, which excludes children and places the product on the general ward rather than in intensive care. Validation behind it is broad, including 21 hospitals in three health systems for the current model, a four hospital community and academic system for the outcome study, and Yale New Haven's seven hospitals, with subgroup results for sepsis, heart failure, chronic pulmonary disease and COVID-19; earlier versions were studied in surgical inpatients and long term acute care hospitals.

What is sold is less clearly bounded than what is cleared. The company lists a pediatric machine learning early warning outcome study (Pediatric Critical Care Medicine, 2025) among its eCART publications without saying whether a pediatric version is offered, and its home page says the model was trained and tested on 28 hospitals across six health systems, a wider base than the clearance describes.