AI for Healthcare-Associated Infection prediction in adult ICUs: prognostic models systematic review

Background

Identifying patients at high risk of Healthcare-Associated Infections (HAIs) is still a major challenge in Intensive Care Units (ICUs). Prognostic prediction models based on Machine Learning and Artificial Intelligence (ML/AI) methods can enhance individual-patient HAI risk estimation and enable preventive interventions. This systematic review (protocol:osf.io/bzqfp) aims to identify ML/AI-based prognostic models for estimating the HAI risk in adult ICU-patients and investigate their methodological characteristics.

Methods

PubMed, Scopus, and Web of Science were searched to identify studies developing and/or evaluating ML/AI prognostic models of HAI prediction in adult ICUs. Data on study characteristics, data sources, HAI types, setting, and performance assessment were extracted and synthesized narratively.

Results

Out of 3930 unique screened records, 33 studies (2013-2025) were included, primarily from China (60.6%), USA (15.2%), and Europe (15.2%). Data sources were approximately half public repositories (e.g., MIMIC-III/IV, eICU) (45.5%), and half primary medical records, surveillance or cohort study data (54.5%), mostly from general/mixed ICUs (81.8%). Bloodstream infections (BSIs) (54.4%) and ventilator-associated/hospital-acquired pneumonias (21.2%) were the most common HAI outcomes. Most studies aimed to only develop ML/AI models (72.7%), and only 27.3% performed external evaluation. On average 5 models per study were developed (range: 1-12), with tree-based methods being the most frequently used (66.7% Random Forests and 63.6% XGBoost). All studies reported model discrimination assessment (Area Under the Curve), but only 12.1% assessed calibration.

Conclusions

Different methodological gaps emerged among ML/AI-based models for HAI prediction in the ICU, that indicate potential overoptimism in model performance, such as the widespread use of a few public datasets, limited external evaluation, and a lack of model calibration assessment.

Key messages

• Despite growing interest in ML/AI-based methods for HAI prediction in the ICU, methodological gaps could limit the applicability of these models in clinical settings.

• Future research should shift focus from model proliferation to rigorous external evaluation and appropriate assessment of model reliability across diverse clinical environments.