Artificial intelligence could help clinicians identify intensive care patients at increased risk of death more accurately than conventional scoring methods. Researchers from Australian Catholic University (ACU) and Charles Darwin University (CDU) have tested machine learning models that combine high predictive accuracy with techniques designed to explain which factors influence their conclusions.
The study addresses an important limitation of AI in clinical practice. Although machine learning algorithms have previously performed well in predicting ICU outcomes, their lack of transparency can make it difficult for healthcare professionals to understand and trust their predictions.
Beyond traditional ICU risk scores
Clinicians currently use tools such as the Acute Physiology and Chronic Health Evaluation (APACHE) and Simplified Acute Physiology Score (SAPS) to estimate mortality risk among ICU patients. According to the researchers, these models have limitations when dealing with changing patient conditions. They can also require considerable time for validation and need regular recalibration.
Machine learning offers an alternative because algorithms can analyze complex relationships between multiple clinical variables. Previous research has shown that such models can outperform traditional scoring systems. However, strong predictive performance alone is not necessarily sufficient for clinical adoption. Clinicians also need insight into why an algorithm classifies a patient as being at high or low risk.
The researchers therefore combined machine learning with methods for explaining the resulting predictions. Two algorithms performed particularly well. An extra trees (ET) model achieved an accuracy of 98.33 percent in predicting ICU mortality, while a gradient boosting (GB) model reached 98.23 percent.
Making AI predictions understandable
Because the extra trees model achieved the highest accuracy, the researchers subjected its predictions to additional explanatory analyses. These were used to determine which patient characteristics contributed most strongly to mortality predictions and whether the results corresponded with existing medical knowledge. The analysis identified hypertension, tumors, endocrine diseases, digestive diseases and cardiovascular diseases as important factors in the model's predictions.
Lead author Niusha Shafiabady, adjunct professor at CDU and head of discipline for IT at ACU, believes explainability could make machine learning more useful to healthcare professionals. By understanding which factors influence a prediction, clinicians may be better able to assess, trust and act on the information generated by an AI system. Such systems could potentially support the identification of high-risk ICU patients who require urgent attention or targeted interventions. Continuous analysis could also help healthcare teams recognize deterioration earlier and take preventive action.
Clinical relevance
The researchers emphasize that AI should support rather than complicate clinical decision making. Integrating interpretable predictions into clinical decision support systems could help bridge the gap between high-performing algorithms and their practical use at the bedside.
Further research is needed before these models can be widely deployed. The researchers plan to evaluate the algorithms using larger datasets and across different healthcare environments. This will be important for determining whether the reported performance can be reproduced in more diverse patient populations and clinical settings. The study also involved researchers from Amirkabir University of Technology, the University of New England, University of Technology Sydney and Western Sydney University.
Dutch study
Last month, researchers at Leiden University Medical Center (LUMC) and the Leiden Institute of Advanced Computer Science (LIACS) developed an AI model to support decisions about when patients can safely leave intensive care. The model predicts the risk of ICU readmission while also explaining which patient characteristics influence its assessment.
Trained on ten years of ICU data, the algorithm divides patients into five risk profiles using a limited number of understandable criteria. One profile includes patients whose white blood cell count rises sharply during the final 24 hours before planned discharge, potentially indicating infection. These patients were more than three times as likely to require ICU readmission as the average patient.
The Dutch model is not yet used clinically. Rather than replacing medical judgement, the researchers aim to combine AI’s ability to identify patterns in large datasets with clinicians’ bedside observations and expertise. Transparency and traceability are considered essential for making such AI tools clinically useful and trustworthy.
References
BMJ Health & Care Informatics (research)
Add ICT&health on Google
Show more content from ICT&health in Google Search.
