Artisight

Models drive the clinical and operational modules: computer vision and voice recognition detect room events, falls, bed exits and surgical milestones, ambient documentation turns conversations and procedural events into EHR entries, and predictive models forecast case length, turnover, staffing and discharge readiness. In the operating room, automated timestamps, milestone alerts and drafted operative notes all depend on that detection running continuously on the edge device.

A large part of the platform also works as communication infrastructure. Virtual nursing is remote nurses handling admissions, discharges, reassessments and documentation over two way audio and video, and teleconsult, location tracking of staff and equipment, and recorded case review are sold alongside the models. At Bryan Health the AI sits behind a human first remote observation model.

People act on what the system detects. A fall risk or bed exit alert goes to a virtual observer or the bedside team, virtual nurses do the documentation and patient contact, and in the operating room the system records timestamps and sends milestone notifications while staff run the case. Surgeons review drafted operative notes, and the five level rollout reserves full workflow automation for the last stage, after visibility, coordination, assistance and prediction are in place.

Alert sensitivity and the expected false alarm rate for fall detection are not stated.

The technical stack is described at the component level. Each device pairs dual 4K cameras, a noise canceling microphone array and radar with an NVIDIA Jetson edge processor, and the perception stack fuses computer vision, voice recognition, indoor positioning accurate to about a foot and contactless vital signs. A large language model is trained in the United States and refined inside each hospital, and the surgical models are said to be trained on more than 500,000 procedures across more than 50 specialties.

Model architecture, training data composition and detection accuracy for any task are not published.

Most of the chain sits inside the hospital. Models run on NVIDIA Jetson processors in the device on site, cloud components run on Microsoft Azure or AWS only at a client's request, and no external streaming relay is used. Retraining data is de identified and synthetic, made under an expert determination, and protected health information is never stored permanently. Third party service providers sign business associate agreements and get no access to patient personal information.

No model or model provider is named for the detection tasks or the large language model, and no subprocessor list covers the cloud path.

Outcome figures come from named health systems. Bryan Health redeployed ten full time staff from one to one observation, cut falls with harm by 40 percent and was handling 375 virtual nursing calls a day within weeks of launch. Operating room results include a 16 percent sustained productivity gain after deployment and more than $7 million saved by automating administrative work and expanding remote nursing. Case studies also cover The Guthrie Clinic, WellSpan Health, Sanford Health and ECU Health.

The figures carry no measurement period, baseline or control, and no peer reviewed study of the platform has appeared; some of the reporting hospitals are also investors.

Patient data handling is specific. Protected health information is held in temporary databases for less than 24 hours and never stored permanently, so Artisight cannot locate a patient record after discharge. Algorithms are built without protected health information: each site retrains on de identified and synthetic data, and a third party audit of privacy procedures and an expert determination cover how that data is made. In surgery the platform analyzes sensor data in real time and stores no video or audio.

Processing stays on hardware inside the hospital firewall, cloud components run only at the client's request, and the device carries a physical shutter and audio cutoff that anyone in the room can use. Artisight states it does not sell, share or rent personal information. How the system handles a missed or false detection is not described.

Artisight treats the patient data its software collects as protected health information and enters into HIPAA business associate agreements with its hospital customers. Any third party service provider must sign a business associate agreement with Artisight and is barred from access to patient personal information. Behind the agreement sit the operating commitments: data held for less than 24 hours, no protected health information in model building, de identification by expert determination, and processing on hospital premises.

Which modules and which optional cloud components the agreement covers is not set out.

Systems are SOC 2 Type 2 compliant in every deployment configuration, cloud, on premises and hybrid, the company states, so the claim covers installations inside the hospital as well as cloud services. The platform specification lists AES-256 encryption at rest, TLS 1.3 in transit, audit logging and role based access, while the privacy policy still names AES-128 and TLS 1.2. Hardware sits behind the customer's firewall and physical security.

No trust center or downloadable audit report is offered.

The platform is sold as workflow, communication and observation software, and none of its modules carries FDA clearance. Most modules sit outside device rules: operating room timestamps, case length prediction, virtual nursing and teleconsults inform scheduling and staffing rather than diagnosis.

Two modules sit closer to patient monitoring. Touchless vital signs are pitched as supporting continuous monitoring, and fall and line removal alerts can take the place of an in room sitter, as at Bryan Health, which moved ten staff off one to one observation. No intended use statement is published for either.

Governance centers on the data that feeds the models. Algorithms are built without protected health information, each installation retrains on de identified and synthetic data reflecting its own environment, and a third party audit and an expert determination cover how that data is created. New algorithms and features deploy automatically across a customer's network as part of the subscription.

Detection performance across skin tones, body sizes, lighting and patient positions, the conditions that matter most for fall and bed exit alerts on high risk patients, is not reported.

Every alert passes through a person before anything happens at the bedside. A virtual nurse or observer looks at the room, the bedside team responds, surgeons review drafted operative notes, and anyone in the room can close the physical shutter or cut the audio. Synchronized event timelines support quality review and teaching after the fact.

No detection accuracy, false alarm rate, miss rate or warranty is published for any task.

Integration runs both ways with Epic and Cerner in real time, with no interface fees. The device connects directly to Epic, real time location systems, BD infusion pumps and in room televisions, and the software pulls patient context such as names, preferred language and planned procedures from the record over FHIR and other standard interfaces. In the operating room, timestamps and milestones flow into perioperative documentation and drafted operative notes go into Epic; in patient rooms, ambient documentation fills EMR fields.

Integration timelines and interface documentation are not published.

The customer chooses where the platform runs: on premises, in the cloud or hybrid. Processing hardware sits inside the hospital behind its own firewall, so inference on video of its patients stays in the building, and cloud components run on Microsoft Azure or AWS only when the client asks. A proprietary WebRTC stack carries audio and video without an outside relay, and existing cameras can be brought onto the platform through the Artisight Connect Module.

Protected health information is held for less than 24 hours, so data residency follows the customer's own infrastructure. Cloud regions for the optional components are not named.

Artisight sells one annual subscription that covers devices, software updates and support, with hardware refreshed every two years and new algorithms deployed automatically, so there is no capital refresh cycle. Hospitals with cameras already installed can use the Artisight Connect Module instead of new devices. Rollout moves through five levels, visibility, coordination, assistance, prediction and automation, with expansion room by room or department by department.

The subscription's price and its unit, whether per room, bed or module, are not published.

One platform covers the inpatient floor and the operating room, and it is live in hundreds of hospitals. Patient room work spans virtual nursing, observation, fall prevention, documentation and teleconsults; operating room work spans case coordination, timeouts and debriefs, op note drafting, case length prediction and asset tracking, with surgical models trained across more than 50 specialties. Named systems range from ECU Health in rural eastern North Carolina to academic and regional systems such as The Guthrie Clinic, WellSpan Health, Sanford Health and Bryan Health.

The buyer is a health system or hospital rather than a department. Ambulatory practices without rooms to instrument are outside its reach.