From validated AI-decision support to clinical evidence

Estimated Reading Time: 3 min

CE- and FDA-certification is a starting point, not a finish line. At Healthplus.ai, we believe that deploying a clinical AI system responsibly means building a continuous and rigorous chain of evidence, from initial validation through to real-world clinical impact. In this article, we explain how we approach this for PERISCOPE®, our AI clinical decision support system for predicting postoperative bacterial infection risk.

Step 1: Retrospective validation as part of CE certification

Before PERISCOPE® could be deployed in any hospital, we validated the model extensively on retrospective patient populations across multiple centres and surgical departments. This work, conducted in line with CE-MDR Class IIa requirements, demonstrated strong discriminative performance and formed the evidentiary foundation of our certification. The results were published in Lancet Regional Health – Europe (van der Meijden et al., 2025), providing independent scientific visibility into the model's performance characteristics.

Step 2: Under-water testing before go-live

Retrospective performance is necessary but not sufficient. Before activating PERISCOPE® in a new hospital environment, we run a "silent mode" validation phase: the model generates predictions on live patient data without surfacing them to clinicians. This allows us to verify that the model performs as expected on local infrastructure, patient case mix, and data pipelines, and to detect any distributional shifts before clinicians rely on the output.

Step 3: Real-world evidence on adoption, usability, and behaviour change

Once live, we collect structured real-world evidence to understand how PERISCOPE® affects clinical practice. This goes beyond performance metrics. We examine whether clinicians actually adopt the system, how it influences decision-making, and whether it changes targeted screening or peri-operative management. Outcomes are assessed using a pre-post design comparing matched periods, with correction for known confounders.

Step 4: Stepped-wedge randomised controlled trial

For the highest level of evidence, we are moving beyond observational designs. Rather than a classic parallel-group RCT, which is logistically and ethically challenging for digital health solutions, we will conduct stepped-wedge randomised controlled trials (SW-RCT). In this design, departments sequentially cross over from control to intervention in a randomised order. Every department eventually receives the intervention, which improves ethical acceptability and statistical power while controlling for time trends.

Primary outcomes under investigation include the time-to-diagnosis of infection, hospital readmissions, length of stay, and ICU admission rates. Secondary outcomes include clinician adoption, diagnostic and therapeutic behaviour, and cost-effectiveness, reflecting both the clinical and operational value PERISCOPE® aims to deliver.

Why this matters

The path from a validated, certified model to trusted clinical evidence is not a single leap; it is a structured path. Each step builds on the last, and each provides a different lens on what it means for AI to work in practice. This is how we at Healthplus.ai ensure that PERISCOPE® doesn't just perform well on paper, but genuinely improves patient outcomes.