Skip to content Skip to sidebar Skip to footer

AI predicts ICU mortality with near-perfect accuracy

New research indicates AI could help clinicians identify high-risk intensive care patients sooner and better understand the factors influencing mortality risk
A ceiling mounted hospital directional sign highlighting the way towards the intensive care unit - 3D render

Artificial intelligence (AI) could soon help clinicians identify the patients most at risk of dying in intensive care (ICU), with Australian researchers finding some machine learning models are able to predict mortality with more than 98 per cent accuracy.

The study, led by academics from the Australian Catholic University (ACU) and Charles Darwin University (CDU), tested eight machine learning algorithms using 600 intensive care patient records from the Medical Information Mart for Intensive Care III (MIMIC-III) dataset to assess their ability to predict mortality and identify the factors most strongly associated with patient outcomes.

Two models proved particularly effective, correctly predicting patient outcomes in 98.33 and 98.23 per cent of cases.

Lead author Professor Niusha Shafiabady, head of discipline for IT and director of the Women in AI for Social Good Lab at ACU and CDU Adjunct Professor, said explainability was critical to building clinician confidence.

“These systems can assist clinicians in identifying high-risk patients who require urgent attention or targeted interventions,” she said.

“Such systems enable continuous monitoring of at-risk patients, supporting proactive care and early intervention to prevent deterioration or adverse events.”

The researchers said machine learning has the potential to overcome some of the limitations associated with traditional mortality prediction tools such as the Acute Physiology and Chronic Health Evaluation (APACHE) and Simplified Acute Physiology Score (SAPS), which provide static assessments and can require regular recalibration.

Related: ED nurses face systemic ‘moral distress’: Study | Medical AI advancing faster than safety checks | OpenAI launches ChatGPT Health | First National Health Literacy Framework launched 

A major barrier to the adoption of AI in healthcare identified by the study is the so-called ‘black box’ problem, where AI algorithms generate predictions without clearly explaining how a decision was reached. To address this, the researchers applied explainable AI techniques, to identify the variables influencing predictions.

For example, rather than simply saying “this patient has a high risk of mortality”, an explainable AI model may follow this prediction up by outlining its reasoning, such as: “this prediction was primarily influenced by the patient’s hypertension, tumour diagnosis and cardiovascular disease.”

The analysis found hypertension, tumours, endocrine disease, digestive disease and cardiovascular disease were among the strongest predictors of mortality.

Professor Shafiabady said the findings demonstrated how explainable AI could be integrated into healthcare without replacing clinical judgement.

“By embedding interpretable findings into clinical decision-support systems, this study supports the advancement of ML tools that are both accurate and clinically meaningful, ensuring they complement rather than complicate frontline healthcare delivery,” she said.

The study concludes that machine learning models have significant potential to support ICU decision-making, patient triage and resource allocation, but noted that further clinical validation is required.

Future research will examine how the models perform using larger datasets and in different healthcare environments.

Leave a comment