Researchers at Dutch Leiden University Medical Center (LUMC) and the Leiden Institute of Advanced Computer Science (LIACS) have developed an AI model that can support doctors in deciding when a patient can safely leave the intensive care unit (ICU). The model not only predicts the risk of ICU readmission but also shows which patient characteristics underpin that prediction.
Determining the right moment for ICU discharge is a complex balancing act. Discharging a patient too early increases the risk of readmission, which is associated with complications and higher mortality. Keeping patients in intensive care unnecessarily long also carries risks, including hospital-acquired infections, while ICU capacity and specialised staff are scarce.
Existing predictive models do not always provide the support clinicians need. Some can accurately identify high-risk patients without explaining why, while other models provide explanations that clinicians do not always consider medically plausible.
Five risk profiles
The Dutch researchers therefore developed a model designed to combine accurate predictions with a small and understandable set of criteria. The algorithm was trained using ten years of ICU patient data and divides patients into five groups, each with a distinct risk profile.
One group, for example, consists of patients whose white blood cell count rises substantially during the final 24 hours before planned discharge. This can indicate an emerging infection. Patients with this profile were found to be more than three times as likely to be readmitted to intensive care as the average patient.
For doctors, such a warning could be a reason to order additional tests or consider continuing ICU treatment. “This could allow an emerging infection to be identified and treated earlier, rather than the patient deteriorating after discharge from the ICU,” explains Siri van der Meijden, technical physician at LUMC.
The model is not yet used in clinical practice. According to the researchers, the study primarily demonstrates that it is possible to develop AI risk predictions that allow doctors to quickly understand the factors behind them. The final decision remains with the clinician.
Combining AI and clinical expertise
That combination is essential, stresses researcher Lincen Yang. An algorithm can recognise patterns across ten years of patient data – far more information than a doctor could remember and assess simultaneously. But an algorithm does not actually see the individual patient. Experienced ICU doctors can also base their judgement on observations at the bedside.
The aim is therefore not to let AI make discharge decisions, but to combine the strengths of data analysis with clinical expertise. “It is important that users understand what is happening and that every step can be traced,” Yang says. Transparency of algorithms, he argues, is important not only for healthcare professionals but also for governments and society.
Collaboration across disciplines
The research also highlights the importance of close collaboration between medical and technical disciplines. Yang initially expected applying an existing AI model to medical data to be relatively straightforward. In practice, mutual understanding and the ability to work with each other's systems proved equally important.
“At first I thought: this is easy. My model, their data, done. But interdisciplinary research also requires mutual understanding and the ability to work with each other's systems. And that takes time,” Yang says.
By combining computer science expertise with the clinical knowledge of ICU professionals, the researchers ultimately developed a model whose predictions are not only technically grounded but also understandable and potentially useful to clinicians.
AI already supporting ICU decisions
The use of AI to support ICU discharge decisions is not entirely new in the Netherlands. Amsterdam UMC clinicians previously worked with Dutch health-tech company Pacmed on Pacmed Critical, decision-support software that uses machine learning to estimate the risk of a patient being readmitted to the ICU or dying unexpectedly within seven days of discharge. The prediction provides additional information to intensivists, while responsibility for the decision remains with the physician.
Since 2023, Pacmed Critical has also been introduced within Santeon, a collaboration of seven Dutch teaching hospitals. A subsequent phase launched in 2024 focused on scaling up the technology across the hospital network and learning how AI can be implemented responsibly and structurally in intensive care.
AI is also being explored at other Dutch hospitals to predict admissions to intensive care. Such models use hospital data to identify patients at increased risk of deterioration and anticipate demand for ICU care. This can support capacity and workforce planning.
Together, these initiatives show how AI is increasingly being explored across ICU admission, treatment and discharge. The new Leiden research adds a particular emphasis on explainability: predictions should not only be accurate, but also make clear to clinicians which patient characteristics are driving the predicted risk.
