Healthplus.ai

The models are the product. PERISCOPE is certified in its own right as software as a medical device, the company sells nothing else, and removing the prediction leaves nothing to buy. The learned component is plain to see: gradient boosted tree models (XGBoost) trained on pre and intraoperative record data give each adult surgical patient a probability of postoperative infection within 7 and within 30 days, and they are retrained and recalibrated on each hospital's own historical data before use. No platform layer, data business or services wrapper sits around the prediction.

Adjunctive by design and by certification. After surgery the system gives each patient a probability of infection within 7 and 30 days, shown inside the record as a risk category with the top contributing factors, and the company says plainly that clinicians decide what action to take. The published study states that the tool does not recommend specific actions and leaves decisions to clinical judgment, with each department's clinical leadership setting its own protocol around the categories. The company also says each hospital's model is used only once it meets quality benchmarks at that site, and that performance is continuously monitored.

The missing part is the low risk end. A clinician quoted on the company's site says low risk means more confidence when discharging, and the company markets safer discharge directly. A false low risk label produces no visible error: it shortens a stay, and the infection surfaces later, possibly at another hospital. The boundaries of the risk categories are not published, nor is the miss rate at the boundary used for discharge, nor what the company does when a site's monitoring shows performance has dropped.

What produces the output is named in detail. The models are XGBoost gradient boosted trees, one for infection within 7 days and one within 30 days, built on roughly 50 pre and intraoperative parameters from the record (demographics, comorbidities, vital signs, biomarkers). The development and updating method, imputation, hyperparameter tuning and exclusion criteria are published in The Lancet Regional Health Europe paper (December 2024) and its supplement. Training data is characterized: 253,010 procedures from 2014 to 2023 at three hospitals, with 2022 and 2023 held out for temporal validation, and the company says its models have been trained on more than one million patient records. Each prediction shows its top contributing factors, and the study reports which features drove each hospital's model. The update method is published too: at each new hospital the same pipeline and feature set are retrained on local historical data and recalibrated where needed, and the study says performance and data drift will be monitored after go live.

Nothing published tells a hospital when the model running there has changed. No version identifier, release notes or change notice is published, and the company does not say how often a site's model is retrained after go live or whether the site is told. The training pipeline is proprietary; the study's analysis code is public.

The chain is structurally short at the model layer. The models are built in house, trained on each hospital's own data with the company's own proprietary pipeline, and no outside model provider sits between the record and the prediction. For hosting, the partners page names Microsoft Azure and Amazon Web Services as cloud infrastructure partners, and a hospital can choose an on premises installation instead, in which case no transfer to a cloud host occurs.

The company's privacy statement, which covers its own contacts rather than the service, lists categories of other recipients (external IT service providers, notified bodies, its administration tool supplier, email and contact form providers) without naming them, and no subprocessor list for the service is published.

The headline claim is prediction: the company says PERISCOPE predicts which surgical patients are at risk of postoperative infection, and the published evidence answers that claim directly. van der Meijden and colleagues, The Lancet Regional Health Europe (published online 5 December 2024, doi 10.1016/j.lanepe.2024.101163), developed the models at Leiden University Medical Center, externally validated them at Radboud University Medical Center and at Ziekenhuis Oost-Limburg in Genk, Belgium, a non academic hospital, and then updated them on each site's own data. The design is retrospective: 253,010 procedures from 2014 to 2023, with 2022 and 2023 held out for temporal validation. Before updating, the 30 day models reached an AUROC of 0.77 at Radboud and 0.85 at Genk; after local updating they reached 0.82 and 0.91, and 0.82 at Leiden. The study reports calibration (slopes of 0.85 to 0.95 after updating), precision and recall, sensitivity, specificity and predictive values at a stated cut off, subgroup results, and decision curve analysis, which found net benefit above treating everyone or no one over a wide range of the thresholds studied (0 to 30 percent) at all three sites.

What it does not show is that using the tool changes outcomes. The paper says clinical trials of patient outcomes and costs are planned, while the company's pages say hospitals using the product report earlier identification of high risk patients and safer discharge, and its funding announcement cites a negative predictive value above 98 percent, which the study reaches at one of three hospitals (0.94 to 0.95 for 30 days at the two Dutch academic centers). The company describes the work as independent validation. The first author is its head of clinical affairs and the chief executive, a coauthor, is a major shareholder, and the study was funded by a REACT EU grant from the European Regional Development Fund, all declared.

The privacy statement (updated March 2026) says the company acts as a data processor when hospitals use the product, has appointed a data protection officer, holds ISO 27001 certification, and uses encryption, two factor authentication and access control. That is a general assurance. It does not say what patient data is retained, for how long, or whether data from one hospital is used beyond that hospital's own model, and the rest of the statement covers website visitors, customers' contract details, job applicants and newsletters rather than the service. Safety engineering, meaning how a model error or safety event is handled, is not published either.

What is known about training use comes from the research rather than from a commitment. The published method retrains and recalibrates the models on each hospital's own historical record data before use, and the product draws only on data the record already holds, with no new data entry. Where that retraining runs is not stated, and the company offers both cloud and on premises installation. An investor quoted in the company's own funding announcement says every new hospital makes the product more accurate, which points to data from one site improving the product for others; the company has not said whether that is so.

HIPAA does not apply to the European deployments, so the European equivalent is what counts here. The status is stated: the company's privacy statement says it acts as a data processor when hospitals use the product, names a data protection officer, and points patients to their hospital as the party responsible for their data. The document behind that status is missing. The privacy statement covers the company's own contacts, customers' contract details, job applicants and newsletter subscribers, not the processing done in the service, and no data processing agreement, template or summary of its terms is published.

The company is preparing to enter the United States. Once it serves a US hospital, a business associate agreement is required and European compliance will not substitute for it.

The company's own material names a recognized certification. The privacy statement says Healthplus.ai is ISO 27001 certified and names encryption, two factor authentication and access control among its controls, and the product FAQ says the company meets ISO 13485 and ISO 27001. No certificate, certifying body, scope, version of the standard or expiry date is published, and there is no trust portal.

Separately, Class IIa certification under the EU Medical Device Regulation means a notified body examined the device's IT security as part of its safety and performance requirements; that covers the device rather than the company's operations.

PERISCOPE is CE marked as a Class IIa medical device under the EU Medical Device Regulation; the company announced approval for the European market in March 2024 and states that it meets ISO 13485 for quality management. It has no FDA clearance. The company's September 2026 funding announcement describes a path toward FDA approval in the United States beginning with a multicenter clinical trial, and it names Wake Forest Baptist as a clinical validation partner for the US market, so the US pathway is stated and at an early stage. No notified body or certificate number is named, so the European certificate cannot be matched to a version and intended use from the company's pages.

The European route is not a lesser one on evidence. Class IIa certification under the MDR requires a clinical evaluation, and the development study describes its protocol as part of that evaluation.

Subgroup performance is published and the weak spots are named. The development and validation study (van der Meijden and colleagues, The Lancet Regional Health Europe, published online December 2024, doi 10.1016/j.lanepe.2024.101163) evaluated the locally updated 30 day models at each of three hospitals by sex, age, surgery priority, surgical specialty, procedure location, open or laparoscopic surgery, admission type, and whether an infection was already present at the time of surgery, measuring discrimination, calibration and net benefit in each group. At Leiden University Medical Center most subgroups reached an AUROC above 0.70 with net benefit above the default strategies, but two did not: emergency surgery (AUROC 0.66, 95 percent CI 0.61 to 0.72) and patients with an ongoing infection at the time of surgery (0.68, 0.65 to 0.70), where acting on the model was no better than treating everyone or no one. The paper attributes this partly to the high infection rate in both groups, 40 to 43 percent. At Radboud University Medical Center and Ziekenhuis Oost-Limburg every subgroup evaluated reached an AUROC above 0.70 with positive net benefit.

A separate fairness paper by the same first author (medRxiv preprint, 2025, doi 10.1101/2025.03.24.25324500, since published in Lancet Digital Health; figures read in the preprint) reports PERISCOPE by sex on a test set of about 9,000 procedures: AUROC 0.80 (0.78 to 0.81) for women and 0.82 (0.80 to 0.83) for men, calibration slopes of 0.81 and 0.92, a slight overestimation of risk in women, and positive net benefit in both groups. It concludes that no further bias mitigation was needed for sex.

Both papers are the company's own work. The first and corresponding author is its head of clinical affairs, another author is an employee, and the chief executive, a coauthor, is a major shareholder, all declared. The company presents the research on its own pages and says it checks each hospital's model across 16 subgroups, such as sex and age, and uses it only when the results meet quality benchmarks. Race, ethnicity and socioeconomic status were not available in the study data, so no result exists for them. One structural caveat is open: the outcome label is a registered infectious complication, more than 72 hours of non prophylactic antibiotics, or surgery to treat an infection, so the model learns from what clinicians recorded and did, which can carry differences in how infection is recognized between groups.

No warranty, indemnity or remediation obligation attaches to the output. What the company does stand behind is a published error rate. The development and validation study it presents on its own evidence page (van der Meijden and colleagues, The Lancet Regional Health Europe, December 2024, led by the company's head of clinical affairs) reports sensitivity, specificity and predictive values on temporal validation data from 2022 and 2023 at three hospitals, at a cut off set equal to each site's infection rate, against an outcome taken from the record: a registered infectious complication, more than 72 hours of non prophylactic antibiotics, or surgery to treat an infection. For the 30 day models: Leiden University Medical Center sensitivity 0.75, specificity 0.74, positive predictive value 0.32, negative predictive value 0.95; Radboud University Medical Center 0.68, 0.80, 0.36 and 0.94; Ziekenhuis Oost-Limburg 0.82, 0.86, 0.19 and 0.99. The 7 day models reach negative predictive values of 0.97, 0.97 and 0.99. Calibration slopes after local updating ranged from 0.85 to 0.95, a slight overestimation of risk. The design is retrospective, on 253,010 procedures in all.

The figures are narrower than they look in two ways. The cut off in the paper is a single threshold, while the product shows risk categories whose boundaries are not published, so these are not the error rates of the display as sold. And the company's September 2026 funding announcement cites a negative predictive value above 98 percent, which the study reaches at one of its three hospitals, the one with a 4 percent infection rate; at the two Dutch academic centers, where about 14 percent of patients developed an infection within 30 days, it was 0.94 to 0.95.

Four record systems are named on the partners page: ChipSoft HiX, Epic, Oracle Health (Cerner) and Nexuzhealth. The product page says PERISCOPE works directly inside Epic and is built to integrate with HiX, showing the risk category and contributing factors alongside patient data in the clinician's existing screen. The product reads pre and intraoperative data from the record and returns a prediction into it within an hour of the end of surgery, with no new data entry.

None of these integrations can be checked in a marketplace listing or integration documentation linked from the company's pages, no FHIR or other standard interface is described, and the company does not say which system runs at which named hospital.

Deployment options are stated: cloud or on premises. The partners page names Microsoft Azure and Amazon Web Services as cloud infrastructure partners, with AWS described as serving hospitals in multiple regions, but no region is committed for hospital data, so residency is settled only for an on premises installation, where the data stays in the hospital. Implementation typically takes 8 to 12 weeks depending on the record system, and the models are calibrated to each hospital before use.

Named sites: the three study hospitals (Leiden University Medical Center, Radboud University Medical Center, and Ziekenhuis Oost-Limburg in Belgium), clinicians quoted from Amsterdam UMC and Deventer Ziekenhuis, Rigshospitalet in Denmark as an implementation partner, and Wake Forest Baptist as a clinical validation partner for the United States. Tenant isolation, and where local retraining runs, are not stated.

No price is published: no rate card, unit of charge, band or implementation fee. What is published is how the product is delivered: cloud or on premises installation, integration with HiX, Epic, Oracle Health (Cerner) and Nexuzhealth, and an implementation of typically 8 to 12 weeks in which the company handles the technical integration.

The site also offers a return on investment calculator whose method is published (postoperative infection rates, extra bed days and treatment costs from Dutch public sources), though results are released only after a visitor enters a name, organization and email address. It is not stated whether retraining and recalibration on the hospital's data at go live, and any later recalibration, are included in the license or charged separately. The company's own research argues local updating is needed for the models to perform at a new hospital, so the question bears on total cost.

Coverage is named: adult inpatient surgery across specialties, with a prediction issued within an hour of the end of surgery for infection within 7 and 30 days. The outcome is postoperative bacterial infection broadly, including surgical site infection, pneumonia and urinary tract infection, rather than surgical site infection alone. The development study (The Lancet Regional Health Europe, December 2024) excluded patients under 18, admissions whose main procedure treated an infection, cardiological and radiological interventions, radiation therapy, diagnostic endoscopy, biopsies, eye surgery and procedures on pregnant patients, and reported results by specialty at each of three hospitals in the Netherlands and Belgium.

Validation covers only part of what is sold. The company markets the product for surgical patients generally, while its own subgroup results at Leiden University Medical Center show no net benefit for emergency surgery or for patients already being treated for an infection at the time of surgery, and the company's pages do not carve these groups out. The 30 day horizon runs past discharge, so part of what the model predicts happens after the patient has gone home, and who follows up then is a question for the deploying hospital. Sites named beyond the study hospitals, in Denmark and the United States, have no published validation yet.