By identifying infection risks earlier, hospitals can target screening, prioritise investigations and focus prevention where it matters most.
Ashleigh Myall
Infection Prevention and Control teams already know how to respond to infection risk
They can screen patients, introduce contact precautions, investigate possible transmission and strengthen infection-control measures. The problem is that they are often forced to act after an infection has already been detected.
By then, other patients may have been exposed and the opportunity for prevention may have passed.
Predictive surveillance is about creating an earlier window for action.
Hospitals Still See Too Much Infection Risk Too Late
Healthcare-associated infections remain one of the most persistent threats to patient safety.
The latest European point-prevalence survey estimates that 4.3 million patients acquire at least one healthcare-associated infection in EU and EEA hospitals each year. One in three organisms identified in these infections was resistant to an important antibiotic.
The same report estimates that at least 20% of infections could be prevented through sustained, multifaceted Infection Prevention and Control programmes. The World Health Organization reports that well-implemented programmes can reduce healthcare-associated infections by as much as 70%.
But prevention depends on timing. Teams need to know where risks are emerging while there is still an opportunity to intervene.
The Limits of Reactive Surveillance
Hospitals use several approaches to identify and manage infection risk:
Point-prevalence surveys provide valuable estimates of infection burden, but only offer a snapshot in time.
Universal screening can detect otherwise hidden carriage, but is expensive and difficult to sustain across an entire hospital.
Screening based on clinical suspicion depends on an infection becoming visible enough to trigger concern.
Contact tracing helps reconstruct possible exposure, but often begins after a positive result has been reported.
Risk-factor screening is more targeted, but commonly relies on static criteria that may not reflect how a patient’s risk changes during their admission.
Each approach has an important role. The problem is the gap between them.
Patients move between beds, wards, diagnostic areas and operating theatres. Their clinical condition changes. They undergo procedures, receive antibiotics and share environments with different people.
Infection risk is dynamic, while many surveillance processes remain retrospective or fixed.
What Is Predictive Surveillance?
Predictive surveillance uses routinely collected hospital data to estimate where infection risk may be emerging next.
Depending on the use case, this can include:
Microbiology and virology results.
Admissions, discharges and transfers.
Patient bed movements and shared locations.
Previous healthcare exposure.
Antibiotic use and invasive procedures.
Clinical observations and underlying conditions.
Changes in local organism prevalence.
Machine-learning models can analyse how these factors interact over time and produce updated estimates of risk at patient, ward or hospital level.
This is not the same as diagnosing an infection or confirming transmission. A risk score cannot tell an Infection Prevention and Control team that a patient definitely has an infection, or that one patient infected another.
It can show where closer review may be needed.
The purpose is to help clinical teams direct their attention earlier—not replace their judgement.
What Earlier Intelligence Makes Possible
A useful prediction must connect to a practical action. Depending on the organism, patient and setting, predictive surveillance can support:
More Targeted Screening. Patients can be prioritised according to their current risk rather than broad or static admission criteria. This may help identify hidden carriage while reducing low-value testing.
Earlier Clinical Review. IPC teams can review high-risk patients before a routine screening result or obvious cluster brings them to attention.
Proportionate Precautions. Predictions can support earlier consideration of screening, isolation or other contact precautions. These decisions should remain subject to clinical review and local policy.
Faster Investigation. Patient movements, shared locations and potential exposures can be brought together before teams begin manually reconstructing them.
Better Use of Resources. IPC time, isolation capacity, environmental cleaning and laboratory testing can be focused on the patients and locations where earlier action may have the greatest value.
The goal is not to intervene everywhere. It is to identify where intervention is most likely to matter.
Evidence That Infection Risk Can Be Predicted
The underlying capability is increasingly well established.
Our study, published in The Lancet Digital Health, evaluated more than 100,000 hospital admissions across the UK and Switzerland. The models achieved an AUROC of up to 0.89 when predicting hospital-onset COVID-19 infection.
Importantly, patient-contact networks were among the strongest predictors. Infection risk was shaped not only by a patient’s individual characteristics, but by their changing connections to other patients and locations across the hospital.
Other researchers have since demonstrated predictive models across urinary tract infections, bloodstream infections, surgical-site infections and multidrug-resistant organisms. For example, a 2025 study of ICU patients developed an interpretable model for multidrug-resistant organism infection with an AUROC of 0.83.
The research is promising, but model accuracy alone is not enough. Many studies remain retrospective or limited to a single hospital. The important next step is showing that predictions remain accurate, useful and safe when placed into real clinical workflows.
Listen to The Lancet Digital Health podcast about the original research →
20 JULY 2022
19 MIN
THE LANCET DIGITAL HEALTH
Ashleigh Myall on predicting hospital-onset COVID-19 infections
Ashleigh Myall joins Diana Samuel to discuss a new machine-learning framework that integrates dynamic patient-contact networks with patient clinical variables and contextual hospital variables to predict hospital-onset COVID-19 infections.
Listen to Podcast
From Static Screening to Precision Surveillance
Most screening strategies make a trade-off.
Universal screening can identify more cases but requires considerable laboratory capacity. Selective screening uses fewer resources but may miss patients whose risk is not captured by fixed criteria.
Predictive surveillance offers a more dynamic approach. Rather than assessing risk once at admission, it can update as new information becomes available.
A patient’s risk may rise because they:
Move into a ward where infection prevalence is increasing.
Share a location with a newly identified case.
Undergo an invasive procedure.
Begin a course of broad-spectrum antibiotics.
Remain in hospital longer than expected.
Develop new clinical signs or laboratory abnormalities.
This allows surveillance strategies to respond to what is actually happening inside the hospital.
Evidence from NEX deployments indicates that active targeted surveillance can reduce the volume of screening required while identifying high-risk acquisitions before routine hospital screening. We are continuing to evaluate how this translates into clinical, operational and economic impact across different settings.
Prediction Must Earn Its Place in the Workflow
Predictive surveillance should not become another dashboard that teams have to check or another source of poorly targeted alerts.
To be useful, it must be:
Timely. Predictions should arrive early enough for an appropriate intervention to change the outcome.
Locally validated. A model developed in one hospital may not perform in the same way somewhere else. Performance and calibration must be assessed in the population where it will be used.
Explainable. Teams need to understand the main factors contributing to a risk estimate and be able to review the supporting data.
Actionable. Every alert should connect to a defined workflow, such as reviewing a patient, requesting a test or investigating a possible exposure.
Continuously monitored. Infection patterns, clinical practice and hospital populations change. Performance, false alerts and model drift need ongoing review.
Clinically governed. Predictive outputs should support professional judgement, not operate as unreviewed diagnoses or automatic instructions.
Standards including TRIPOD+AI and DECIDE-AI provide useful frameworks for transparent reporting and real-world clinical evaluation.
Building a Connected Infection Intelligence System
Predictive surveillance is one part of a wider shift towards Infection Intelligence.
As hospital data becomes more connected, different technologies can provide complementary views of infection risk:
Genomic sequencing can help determine whether apparently related cases are genetically connected.
Natural-language processing can identify relevant information held in clinical notes.
Rapid diagnostics can shorten the time between suspicion and confirmation.
Patient-movement and spatial models can reveal possible pathways of exposure.
Antimicrobial stewardship systems can support more timely review of prescribing decisions.
Research has already shown the value of combining routinely collected hospital information with contact networks. More recent work has demonstrated how real-time genomic surveillance can identify previously unrecognised hospital transmission, while digital tools are increasingly being used to support antimicrobial prescribing and stewardship.
The opportunity is not to place more disconnected systems into hospitals. It is to bring these signals together and translate them into clear, timely actions for clinical teams.
Earlier Intelligence, Not More Alerts
Predictive surveillance will not prevent infections by itself.
Its value comes from giving Infection Prevention and Control teams more time: time to review a patient, request a test, investigate a possible cluster or introduce proportionate precautions before more people are exposed.
The future of infection prevention is not simply more data or more AI.
It is intelligence that helps hospitals see infection risks earlier, act earlier and prevent avoidable harm.
