At Phramongkutklao Hospital in Bangkok, clinical teams have began using daily AI-generated risk predictions to identify inpatients who may benefit from additional screening for carbapenem-resistant Enterobacterales, or CRE. The work is designed around a patient safety premis: can identifying CRE risk earlier help the clinical team act sooner and protect other vulnerable patients? To our knowledge, this is among the first reported real-world uses of AI to guide active surveillance for CRE in a hospital. The deployment is at an early stage, with clinical oversight and ongoing monitoring at its centre.
NEX Health Intelligence
The Patient Safety Window Before Detection
CRE can colonise a patient without causing symptoms. A patient may therefore carry the organism before the hospital receives a positive result.
During this period, appropriate precautions may not yet be in place. This creates an opportunity for CRE to spread through shared staff, equipment or the hospital environment.
Earlier screening gives the clinical team more time to identify colonisation and consider appropriate action. The CDC recommends targeted measures including isolation and contact precautions for patients with CRE.
A pragmatic ICU trial also reported a lower acquisition rate during active surveillance testing and pre-emptive isolation (Jung et al., 2023).
How the Model Supports Earlier Screening
The model uses routinely collected hospital data to estimate which inpatients may be at increased risk of becoming CRE positive during the following 7 days.
It considers:
Where the patient has stayed and how they have moved through the hospital.
Connections to patients with known CRE.
Antimicrobial prescribing.
Medical devices and procedures.
Other clinical events associated with acquisition risk.
Risk can develop through several connected events over time. The model learns how these factors interact and presents the result for clinical review.
Recent research has developed retrospective machine-learning models to support CPE screening and called for prospective validation in clinical settings (Kim et al., 2026). We are now evaluating how this type of prediction can support real screening decisions.
From Prediction To Clinical Action
Each day, the model highlights patients whose risk may justify closer review.
The clinician examines the patient’s context and decides whether additional screening is appropriate. The model does not diagnose CRE or automatically determine isolation, treatment or other clinical action.
Clinical judgement remains central. The risk prediction helps the doctor focus attention, while the decision considers the patient, current guidance and the hospital’s available capacity.
Measuring What Matters for Patients
The model’s performance is checked daily. It is also configured and recalibrated regularly to account for changes in CRE prevalence and screening behaviour.
Technical performance is one part of the evaluation. The patient safety questions are whether the model helps identify additional patients earlier, supports timely precautions and reduces the period during which high-risk patients remain unrecognised.
217%
increase in screening yield
84.6%
identified before clinical culture
0.825
AUROC out-of-policy cohort
12 days
median CPE risk lead time
12 of 60
CRE-positive found outside screening policy
What Comes Next
The immediate priority is to understand more of how the system affects screening decisions and improves patient safety in routine care.
Next the team is assessing how it can support decisions about who should be screened, where screening may be lower priority, and how it can be used consistently by ward and Infection Prevention and Control teams.
This work will help determine whether AI-guided active surveillance can become a practical part of protecting patients from CRE.
