The challenge
Medivia runs four hospitals and a dozen or so clinics. On-duty doctors on internal medicine wards watch dozens of patients at once. Signs of deterioration, such as the onset of sepsis, can be subtle and spread over time: a slightly higher temperature, a drop in blood pressure, a change in lab results.
The medical board wanted to support doctors with a system that spots such patterns earlier. The condition was clear: no automatic decisions, full transparency, and compliance with medical data regulations and with the EU AI Act, which classifies this type of system as high-risk.
The approach
We built a model that assesses the risk of deterioration for every hospitalised patient each hour, based on vital signs, test results and treatment history. The output is not a diagnosis but a signal: "this patient needs attention", together with the factors behind it.
From day one we designed the system around the AI Act: technical documentation, event logs, human oversight, post-deployment monitoring. The doctor always sees the reasoning and can dismiss the alert, and every dismissal feeds into the model's quality review.
The rollout
Data stays in Medivia's own infrastructure, on a closed network. The training data was pseudonymised, and only named people have access to it. Before going live, the model ran in silent mode for three months: it calculated risk, but doctors did not see the alerts. This let us check accuracy on real cases without affecting treatment.
We then launched on two pilot wards and, after six months, extended to all four hospitals. In training sessions with staff we focused mostly on limitations: when the model gets it wrong and why it should not be trusted blindly.
The results
The model reached an AUROC of 0.83 on validation data, a value typical of published early warning systems. In cases where the doctor confirmed the alert within three hours, the median time to the first antibiotic dose fell by 1.7 hours.
We make no claims about mortality: that would take longer observation and a prospective study, which Medivia is running with an academic partner. The medical team points out that some alerts are false, so the sensitivity threshold is reviewed regularly.
