PeriGen

Split product, and the split is real. The learned components are PeriCALM Patterns, which uses LSTM neural networks to detect and label fetal heart rate baseline, accelerations and decelerations and to detect uterine contractions (cleared as Patterns 3.0, K241009), and the multifactor labor curve, a statistical model fitted to population data that places a patient's dilation against a reference population adjusted for contraction frequency, epidural use and parity.

PeriWatch Vigilance, the early warning product the company leads with, is built on both: it sorts and color codes patients and notifies clinicians when fetal heart rate patterns, labor progress or maternal vital signs cross limits the hospital sets.

The company also sells PeriWatch Surveillance, a central fetal monitoring, documentation and archiving system that stands on its own, along with quality reporting and virtual care configurations, so removing the models would still leave a working surveillance system. B because the early warning product does not exist without the models.

Adjunctive by design and by clearance. The Patterns indication is additional secondary information for qualified clinicians, with an FDA warning that patient management decisions should not rest on its annotations alone, and the company says AI is meant to complement clinical expertise, never replace it. Nothing is acted on automatically. Vigilance notifies clinicians when maternal vital signs, fetal heart rate patterns or labor progress cross limits the hospital sets, and LaborWatch escalates persisting concerns up a chain of command the hospital configures. A clinician can delete an annotation the algorithm made.

What is needed to set those limits is missing: no default thresholds, no miss rate at them, and no guidance on what happens when an alert is not acted on. Automatic escalation to leaders also raises a question the company does not address, about what the escalation record means for the clinician who saw the first alert.

What produces the output is named, with its version and update discipline. The fetal heart rate component, PeriCALM Patterns 3.0, uses long short term memory (LSTM) neural networks to find baseline, accelerations, decelerations and uninterpretable segments, and detects contractions the same way as its predecessor, per the 510(k) summary (K241009); the company's publication list traces the method to its papers on neural network pattern detection (2005) and LSTM fetal state classification (2017). Training is characterized: over 1,800 hours of curated tracings covering a wide range of clinical conditions, stratified by gestational age, maternal BMI, Apgar score, cord pH, birthweight and acquisition device, and the FDA test set was 70 tracings (30 preterm, 40 term) from GE, Philips and Neoventa monitors. The model is locked. The company says it is static once trained and does not learn in the clinical environment, changes arrive as cleared versions, and the product page says which gestational range the current release covers and that the 32 week clearance comes in a future version, so a change to the system is visible.

Two gaps remain: the source institutions and population of the training tracings are not named, and the version and method of the labor curve in the current product are not stated. The company's 2024 research describes new multifactor labor curves built with Gaussian processes, and whether those are in the product is not said.

The chain is structurally short. The models are the company's own, built in house and cleared by FDA as locked versions, so they cannot be swapped quietly, and no outside model provider sits between the tracing and the output. The servers run in the hospital's environment on VMware or Citrix.

The one outside service named is identity: mobile and remote sign in uses the company's Microsoft Entra ID B2C service, federated to the hospital's identity provider, with the external URL published through the hospital's own Azure App Proxy or an equivalent. No subprocessor list is published, and the website privacy policy's reference to third party hosting and analytics providers covers the site rather than the service.

The evidence is real and falls short of the outcome claim the company leads with, early recognition to prevent avoidable harm in childbirth. On performance, three experts from the National Institute of Child Health and Human Development agreed with the software's analysis of 100 tracings in over 97 percent of assessments (2013, supplement abstract), and Parer and Hamilton (American Journal of Obstetrics and Gynecology, 2010), with the company's founder as second author, found the software's agreement with the majority of five experts on 769 tracing segments (proportion 0.87, weighted kappa 0.52) no different from the experts' own (0.83 and 0.58). Company-led papers in the same journal report the multifactor labor curve's discrimination for cesarean delivery (AUC 0.78 to 0.93 on 4,703 births, 2016) and its prediction error for dilation (mean absolute error 0.83 cm on 8,022 births, externally validated, 2024).

On outcomes, the evidence is observational. Chiweza and colleagues (BJOG, 2022, Baylor College of Medicine, no company authors) reported fewer intrapartum stillbirths and early neonatal deaths in the six months after Vigilance was introduced at a hospital in Lilongwe, Malawi, with a small rise in cesarean delivery, and called for a prospective trial; the company's impact page adds a further 82 percent reduction there without a source. A Lee Health study (MCN, 2026) found no change in how nurses managed uterine tachysystole after the alerts alone, and an improvement once an education module was added, which about a third of the nurses completed; the company's summary reports only the combined result. Customer reports include Hoag's cesarean rate falling from about 29 percent to 24 to 25 percent (Los Angeles Times, 2025) and a health system, cited by NYC Health + Hospitals, with 12 percent fewer unexpected complications in term babies over two years. The company also lists an Ochsner study of centralized remote fetal monitoring across five hospitals whose paper does not name the product. No controlled or prospective multi site outcome study is published.

The only privacy policy published (updated November 2025) covers the website and customer portal, not the service, so nothing is published on what patient data the product retains, for how long, or whether customer data is used in development. What is stated about the model is narrow: the fetal heart rate algorithm was trained on over 1,800 hours of curated tracings and is static once trained, so it does not keep learning in the clinical environment. Where those tracings came from, and whether data from customer hospitals feeds later versions, is not stated. Safety engineering, meaning how a missed pattern or a false alert is handled as a safety event, is not published either.

The retention question is specific to this product. PeriWatch Surveillance captures and archives fetal tracings and documentation, with a collection architecture built to ride through network or system downtime, and the enterprise servers run on the hospital's own VMware or Citrix infrastructure. Fetal monitoring tracings are primary evidence in birth injury claims, and the limitation period for a claim on behalf of a child often does not start until the child reaches adulthood, so a hospital may need a legible tracing two decades after the birth.

A status is claimed without the document behind it. The PeriWatch Surveillance page describes a HIPAA compliant patient census view, which is a compliance claim, but no business associate status is stated, no agreement or summary of its terms is published, and the only privacy policy covers the website and customer portal.

Three features make the scope of any agreement matter: the product takes data from the record and continuously from the fetal monitoring network, it archives tracings for the long periods birth injury claims require, and mobile access runs through the company's identity service.

The company has been part of Halma plc, a UK listed group, since 2021, and its software has been used in Malawi through Baylor College of Medicine and Texas Children's Hospital, where HIPAA does not apply, and nothing says which PeriGen entity contracts and signs.

The company displays the AICPA SOC for Service Organizations mark under Certifications in its site footer, which indicates a SOC examination by an independent auditor, but it names neither the report type (SOC 1, SOC 2 or SOC 3), the period covered nor the auditor, and no trust portal or request path is published.

Separately, the Patterns 3.0 clearance records a cybersecurity evaluation under FDA's 2023 premarket guidance, which covers the device software rather than the company's operations. The mobile app's security design is published in outline: sign in delegated to the hospital's identity provider through OpenID Connect, role based, facility level access passed through OAuth 2.0 token claims, and an external URL the hospital controls.

The AI is cleared, and the product sold goes beyond the clearance. PeriCALM Patterns 3.0 (K241009, decided 10 January 2025, product code HGM, submitted by PeriGen Solutions in Israel) is cleared as additional secondary information for qualified clinicians during antepartum or intrapartum monitoring at 32 weeks or later, to annotate and summarize fetal heart rate baseline, accelerations and decelerations and uterine contractions, with a warning that management decisions should not rest on its annotations alone. It succeeds CALM Patterns (K040788, 2005, 36 weeks and later). The labor curve traces to CALM Curve (K052715, 2005, singleton cephalic pregnancies beyond 35 weeks), cleared to LMS Medical Systems, the earlier owner of the technology. The company states PeriWatch Surveillance is 510(k) cleared without naming the number.

The company is careful with the label: it calls the algorithms cleared and says Vigilance incorporates them, and its product page says Vigilance is cleared from 36 weeks today, with 32 weeks coming in a future version. What Vigilance is sold to do (identify patients at risk of adverse outcomes, sort and color code them, grade tracings on three or five tier scales, and notify and escalate) is broader than annotation and summary, and no clearance is cited for it. The company also describes its algorithms as the only FDA cleared ones for fetal heart rate interpretation available in the United States; the clearance covers annotation of baseline, accelerations and decelerations, not a full interpretation including variability and category.

No subgroup results are published. The company describes a process: its fetal heart rate algorithm uses no demographic inputs, its training data was stratified by gestational age, maternal BMI, Apgar score, umbilical cord pH, birthweight and acquisition device to confirm these did not affect the output, and it monitors for bias through post market surveillance. None of those analyses is published with figures, and monitoring under the device regime is a quality system commitment rather than a published evaluation. The 510(k) summary for Patterns 3.0 says the software was tested separately in term (40 tracings) and preterm (30 tracings, 32 to 35 weeks) groups against obstetrician readers and passed every non inferiority criterion, but gives no sensitivity, specificity or agreement figures for either group.

This matters most in obstetrics. Maternal outcomes in the United States differ sharply by race, and NYC Health + Hospitals, a customer, describes its adoption as part of closing equity gaps for Black and Brown women. An early warning system that fires on objective thresholds could narrow differences in how warning signs are recognized, while a pattern model trained to agree with expert readers inherits whatever those readers do. Which effect dominates has not been answered in public.

No warranty, indemnity or remediation commitment attaches to the output. What the company does offer is an agreement rate it has published. Its January 2013 announcement, still on its site, reports an independent study in which three experts from the Eunice Kennedy Shriver National Institute of Child Health and Human Development reviewed the analysis produced by PeriCALM Patterns for the final hour of 100 tracings and agreed with it in over 97 percent of assessments (Wiener and colleagues, American Journal of Obstetrics and Gynecology supplement abstract, 2013). That is an agreement rate against a stated reference, with the method and the number of tracings given.

Three limits for a buyer. The experts reviewed the software's analysis rather than reading the tracings independently first, which tends to raise agreement. The version tested was the one in use in 2013, not Patterns 3.0; the 3.0 510(k) summary says the software met non inferiority margins against obstetrician readers on 70 tracings for sensitivity, specificity and positive predictive value in detecting accelerations and decelerations, but publishes no figures. And agreement across all assessments says little about the rare patterns that matter most.

Integration is claimed broadly and nothing is named. The company says its products work with all industry standard EMRs, with bidirectional interfaces that sync documentation between PeriWatch Surveillance and the record in near real time, admission and discharge feeds, and single sign on through Active Directory or the EMR, and that Vigilance reads from the record and from whichever central fetal monitoring system the hospital runs.

No record system, interface standard or marketplace listing is named on the company's pages, so none of it can be checked before contact, and supported fetal monitor models are available only on request. The 2016 acquisition of the WatchChild system from Hill-Rom brought interfacing libraries and record integration work, and named deployments at large health systems show the product runs in production somewhere, though which record systems it connects to is still not stated.

Deployment options are published: bedside, facility virtual care, enterprise virtual care across a health system, or a hybrid, on an enterprise server architecture that supports VMware and Citrix, which places the servers in the hospital's own environment.

Mobile and remote access runs through the hospital's identity provider: the company's own Microsoft Entra ID B2C service handles OpenID Connect sign in and delegates authentication to the hospital, and the hospital publishes an external URL through Microsoft Azure App Proxy or an equivalent inside its own tenant. Vigilance does not require the company's central monitoring system and works alongside the one a hospital already has.

Named sites include Mount Sinai Medical Center in Miami Beach, NYC Health + Hospitals (live at North Central Bronx, with all 11 hospitals planned), Community Health Systems hospitals, Hoag, Lee Health, the University of Maryland Medical System and OhioHealth, and the company reports more than 500 hospitals in the United States and Canada.

Where data rests is implied by the architecture rather than stated, and nothing is published on remote support access, where the Vigilance data warehouse sits, or whether any access comes from the company's operations in Israel.

No price is published: no rate card, unit of charge, band or implementation fee. What a buyer can find are customer figures. NYC Health + Hospitals announced a $2.75 million investment to bring Vigilance to its 11 hospitals, which deliver about 15,000 babies a year, and Hoag raised nearly $300,000 from donors for Vigilance at its Newport Beach and Irvine campuses. These are disclosed contract and funding values rather than the vendor's price, and neither states the term or what it covers.

The product line is published in enough detail to scope a quote: Surveillance, Vigilance, LaborWatch, the mobile app, virtual care configurations, and quality reports in standard, enhanced and custom packages.

Single specialty and deliberately so: obstetrics, covering antepartum and intrapartum monitoring, with maternal vital signs, fetal heart rate patterns and labor progress in one system. The limits are published. The fetal heart rate analysis is cleared from 32 weeks in Patterns 3.0, while the product page says the version in use today covers 36 weeks and later; the labor curve is indicated for singleton, cephalic pregnancies beyond 35 weeks. Customers range from standalone hospitals to large systems, including the public system NYC Health + Hospitals.

Validation covers only part of what is sold. NYC Health + Hospitals describes using Vigilance in obstetric triage, postpartum units, obstetric recovery and the emergency department, and the company's virtual care page names postpartum units, while the published validation concerns fetal heart rate analysis and labor progress before birth, not maternal vital sign alerting afterward. The fetal heart rate analysis is not indicated before 32 weeks, and the labor curve is not indicated for twins or non cephalic presentations.