Apella

Every module runs on the event stream the vision models produce. Four fixed cameras on the walls of each operating room feed a YOLO based detector that finds patients, staff and equipment, and a transformer model reads those detections over time to timestamp up to 14 events, from wheels in and draping to cleanup and room setup. Apella Pulse turns the events into live room status, delay predictions and text alerts, Apella Horizon into case duration forecasts and block utilization, and Apella Insight into retrospective analytics, while the same timestamps flow back to the hospital's record system.

A mobile app launched in September 2026 puts case tracking, live room views and lock screen notifications on surgeons' and anesthesiologists' phones, all driven by the detected events.

Detection runs on its own: the models timestamp events in real time and can write them to Epic without staff entering anything, and the BMJ paper describes the system working without oversight from operating room staff. Review sits with Apella. Lower confidence detections and unusual patterns go to an in house data quality team, which works out what actually happened and feeds the difficult cases into future training; at the study site, domain experts first annotated a random 10 percent of events and later nearly all of them. Staff can check any timestamp against live or retrospective video, and staffing and scheduling decisions stay with charge nurses, surgeons and perioperative leaders.

Apella does not say what confidence level sends an event to review, or whether a flagged event is held out of the record until someone checks it.

Two models do the detection work. A YOLO based detector finds patients, staff and equipment in images from four wall cameras, and a transformer based model reads the sequence of detections to infer events; the 2025 BMJ paper says the system was trained and tested on 137,517 cases from 315 operating rooms and 1,853 surgeons across multiple sites. Apella builds these models itself on real hospital video rather than synthetic data, and separate predictive models forecast case duration, turnover and delays from live events plus record history by surgeon, procedure and time of day.

Neither the model versions nor a release history showing when a model changes is available, a gap the Houston Methodist study also notes.

Apella builds its own computer vision models rather than adapting general purpose ones, trains them on video from real hospitals, and runs them at a small fraction of the cost of larger general models; a dedicated in house team handles data quality and annotation. Cameras and other room equipment are supplied and maintained by Apella, and hospitals connect through their own interface engines and networks. Google Analytics, on the website, is the only outside party the privacy policy names.

The cloud provider and any other subprocessors are not listed.

A 2026 study in the Journal of Imaging by Houston Methodist researchers built a synthetic control from 11 sister sites and 121,515 cases and found that the 15 room cardiothoracic suite running Apella added about 25 cases a month (95 percent CI 8.3 to 41.0), a 7 percent rise, with case mix and total operative minutes stable. Houston Methodist holds a minority equity stake in Apella and buys its camera hardware, and an Apella economist ran the original analysis.

A 2025 paper in BMJ Health & Care Informatics, written by Apella staff with a Houston Methodist coauthor, scored event detection against expert annotations at one health system (F1 of 0.98 to 0.99 across five events) and used the timestamps to compare 2,502 knee replacements by provider. A 2026 paper in the Journal of Vascular Surgery Vascular Insights mapped phase level variation in angiography across five hospitals. The platform wide figure of two extra cases per operating room per month comes from the company's own analysis across health systems.

Operating room video is encrypted, kept only for the retention period agreed with each hospital and never longer than 28 days, and "used only for model training and quality assurance", according to Apella's help center. The models identify events and objects rather than people, with no facial recognition; live views include patient privacy blurring, video is never tied to patient records, and footage is reviewed only inside an approved quality workflow. Each hospital owns its data and provisions access, and the training set already spans more than 137,000 cases from 315 operating rooms.

How a missed or false detection that reaches the record is handled as a safety event is not described.

As a business associate of its hospital customers, Apella says so in both its website privacy policy and its help center security page. The help center sets out how the service handles operating room data: encrypted video kept no longer than an agreed retention period of at most 28 days, footage reviewed only in approved quality workflows, access provisioned by the hospital, data owned by the hospital, and annual HIPAA and security training for every employee.

The business associate agreement itself, including whether it binds the cloud host and other subcontractors, is not posted.

Encryption covers data in transit end to end and at rest, log data included, and access defaults to none until granted. The help center also lists logging of every network request with alerts on suspicious activity, intrusion detection on production systems, regular vulnerability scanning, automated penetration tests every week, daily backups kept for 30 days, risk assessments before production changes and annual security training. Hospitals connect the cameras through their own firewall rules, VLANs and address allowlists, and Apella monitors and manages its hardware centrally.

Apella's trust center lists a SOC 2 Type II report and web and network penetration test reports, released on request, alongside the controls behind them.

The models detect when events happen in the room, such as wheels in, draping or cleanup, and the output feeds scheduling, delay alerts, analytics and cleaning audits rather than any assessment of the patient or the surgery. The company's summary pages say Apella is not a surgical robot and does not perform or assist procedures, and its help center adds that the system uses no facial recognition, tracks workflows rather than individuals and is not part of the medical record.

Detected timestamps can still be written into the perioperative documentation in Epic, where they become the times that schedules and reports rely on.

Model quality is managed through monitoring after go live. A data quality team reviews low confidence and unusual detections, checks outputs against verified ground truth and looks for recurring errors, and the company says it evaluates performance on the specific workflows of each new hospital. The 2025 BMJ paper reports detection at a single health system and says performance has stayed consistently high across diverse operating environments, naming room layout, lighting and occlusion by equipment or staff as the risks, with four cameras per room to limit blind spots.

No results are broken out by hospital, room type or procedure, including behind the home page line that Apella pinpoints bottlenecks free of bias.

Detection accuracy is measured and cited by the company. In the 2025 BMJ Health & Care Informatics paper led by Apella's data scientists, the system was scored against expert annotations at a Texas academic health system: wheels in produced 184,984 predictions against 184,172 annotated events with an F1 of 0.99, patient draped scored 0.98, and bed transfer, cleanup end and wheels out each scored 0.99, with median timing errors of 6.4 to 16.9 seconds. The company's blog and AI summary page both cite the paper, and its data quality team reviews flagged detections after go live.

Timestamps that land in the record can be checked against video, and no warranty, service credit or remedy for a wrong entry is offered.

Detected events are written back to Epic in real time, with Epic OpTime named, and the platform also reads schedules and staffing from the record system, working alongside both Epic and Oracle Health Cerner. Integration combines historical data files with live feeds through the hospital's interface engine and usually takes three to five months. Dayton Children's Hospital is cited with a 99 percent reduction in EHR data errors, and the Houston Methodist study describes automated, real time documentation of perioperative events independent of staff entries.

Apella has no listing in Epic's Showroom directory.

Four fixed cameras sit on the walls of each operating room, supplied, monitored and maintained by Apella, and sensitive data is stored in a cloud environment wherever possible. Staff reach live views, dashboards and analytics in a browser with single sign on, on the mobile app and through text alerts. Each hospital's network team sets up the secure tunnels, firewall rules and VLANs the system uses, and EHR feeds run through its interface engine.

Which cloud, which region and how customers are kept apart are not stated.

Apella provides and maintains the operating room cameras itself, and a surgical capacity calculator on its website estimates the added cases a hospital could run from its own inputs, with the outputs labeled illustrative. The company's headline return is an average of two additional cases per operating room per month, with up to a 15x return reported in some deployments.

A rollout pairs a three to five month EHR integration with camera installation, often two or three rooms a night, then a 30 to 45 day period in which the models learn the hospital's workflows. No price, unit of charge or implementation fee is posted.

Built for hospital operating rooms, the same models extend to other anesthesia areas such as interventional radiology, cath labs and endoscopy, trained on the procedure types a partner hospital chooses. Houston Methodist runs it across more than 200 operating rooms in nine hospitals, and Tampa General, MUSC, Hackensack Meridian Health, NorthBay Health and Dayton Children's Hospital are named customers.

Published studies cover specific specialties: total knee replacement at a Texas academic health system, a 15 room cardiothoracic suite at Houston Methodist, and lower extremity angiography across five Houston Methodist hospitals. The procedure areas outside the operating room have no named deployment yet.