Risk Prediction of ICU Admission and Artificial Intelligence: A Narrative Review

Postoperative ICU admission is a complex clinical decision influenced by multiple factors, including patient characteristics, surgical type, intraoperative variables and organisational constraints. Traditional risk stratification tools have limitations in accuracy and consistency. Artificial intelligence models demonstrate promising predictive capability for ICU admission risk across various surgical contexts, with several retrospective studies reporting high accuracy.

Every year worldwide, over 300 million patients undergo noncardiac surgical procedures, with further increases expected in the coming years (Weiser et al. 2015). Postoperative morbidity and mortality can vary depending on the type of surgery and individual patient characteristics. Scientific evidence suggested that frailty is associated with higher rates of postoperative mortality across all noncardiac surgical specialities (George et al. 2020). Postoperative monitoring plays a crucial role in detecting and preventing potential complications. In a randomised trial including 400 patients, continuous vital sign monitoring with real-time alerting to staff reduced the episodes of desaturation and overall adverse events within 30 days compared with the standard-of-care group (Mølgaard et al. 2025). Admission to the Intensive Care Unit (ICU) after major noncardiac surgery can ensure continuous high-quality postoperative care. However, this clinical decision must consider multiple complex factors, seeking to avoid unnecessary ICU admissions and, on the other hand, possible subsequent late and unplanned ICU admissions, which are associated with increased mortality and morbidity (Khanna et al. 2023). Scientific interest in the role of Artificial Intelligence (AI) in supporting medical practice is constantly growing, as demonstrated by the publication of over 300,000 articles on the topic on PubMed, nearly 60,000 of which in the last year. The primary objective of this narrative review is to describe the main publications analysing the role of AI in supporting physicians in predicting the risk of ICU admission after major noncardiac surgery.

Traditional Criteria to Determine ICU Admission After Noncardiac Surgery

Several different factors can contribute to determining a planned ICU admission after major noncardiac surgery (Lavezzo et al. 2025). First, patient-specific characteristics, including Body Mass Index (BMI), male gender and pre-existing comorbidities (e.g., diabetes, cardiac and respiratory problems, sarcopenia). The association between non-modifiable preoperative comorbidities and perioperative risk can be better objectified using specific assessment tools. Scientific evidence suggests a higher risk of postoperative ICU admission among patients with American Society of Anesthesiologists (ASA) physical status ≥3 (Huang et al. 2021) or a Charlson Comorbidity Index ≥2 (Jerath et al. 2018) or a Revised Cardiac Index (Lee Criteria) ≥2 (Fayed et al. 2023). The type of surgical procedure plays a key role in determining perioperative risk. The ACS NSQIP (American College of Surgeons National Surgical Quality Improvement Program) risk calculator requires the collection of 40 preoperative variables (28 related to patient characteristics and comorbidities and 12 to laboratory values). Although this score does not include ICU admission as a possible outcome, data from ACS NSQIP database were evaluated to predict the need for ICU (Kongkaewpaisan et al. 2019).

Meguid (Meguid et al. 2016) developed the Surgical Risk Preoperative Assessment System (SURPAS) using eight variables (four patient-related and four intervention-related) from the ACS NSQIP database (over six million surgical cases in nine surgical specialities). The SURPAS score demonstrated good predictive power for postoperative morbidity and mortality. In a subsequent large retrospective analysis, Rozeboom et al. (2022) demonstrated that the SURPAS model accurately predicts ICU admission in diverse surgical populations.

The progressive increase in the average age of the surgical population and the negative consequences of neoadjuvant treatments among oncological patients underscore the role of preoperative frailty in worsening postoperative outcomes. The level of frailty, as assessed using the Clinical Frailty Scale (CFS), was associated with a significant increase in ICU mortality in older patients (Bruno et al. 2023).

A predefined preoperative risk, determined by analysing patient- and surgery-related factors, may be modified by the intraoperative course (e.g., change in preplanned surgery, long duration of intervention, late surgery end, high intraoperative blood loss or serious general intraoperative problems). The Surgical Apgar Score (SAS) includes a collection of variables related to three intraoperative parameters (blood loss, the lowest mean arterial pressure, and the lowest heart rate), with lower scores associated with worse outcomes (Gawande et al. 2007). In a retrospective cohort study of 8,501 adult patients undergoing major abdominal surgery at a single medical centre, the SAS was strongly associated with immediate ICU admission, highlighting the importance of intraoperative haemodynamic in the ICU triage process (Sobol et al. 2013).

The CARES (Combined Assessment of Risk Encountered in Surgery) model was developed by retrospectively analysing 90,785 electronic medical records of noncardiac/non-neurosurgical patients managed at the Singapore General Hospital (Chen et al. 2018). The CARES risk calculator provided accurate predictions for both mortality and ICU admission (the Area Under the Receiver Operating Curve—AUROC was 0.837 for ICU admission).

Notably, Stieger et al. (2025) suggested that intraoperative data may be essential to determine the ICU triage process, while preoperative variables are more closely related to ICU length of stay.

Immediate postoperative monitoring in dedicated areas (e.g., Post-Anaesthesia Care Unit—PACU or Recovery Room) is essential for determining the patient's condition after surgery and possible subsequent admission to the ICU or general ward. Indeed, the specific hospital organisation plays a crucial role in determining the final clinical decision regarding the level of thoroughness of postoperative monitoring (e.g., ICU bed availability, staff experience and hospital surgical volume, presence of a high dependency unit such as a step-down ICU).

Unnecessary postoperative ICU admission is associated with increased costs and potential patient exposure to complications, such as a higher risk of infection or postoperative delirium. On the other hand, a retrospective study (Pecorelli et al. 2025) that included 1,486 patients undergoing pancreatic surgery demonstrated that patients receiving unplanned late ICU admission showed a significantly increased mortality rate compared to those who underwent direct ICU admission (in-hospital mortality was 14% and 57% in the two groups, respectively, p<0.001). Potential difficulties in analysing this condition are also related to its very low incidence. Wanderer (Wanderer et al. 2013) developed and retrospectively validated a model on 71,996 medical records that included 16 patient-related variables and 12 intraoperative variables, demonstrating accurate prediction of unplanned ICU admissions (AUROC 0.905). In this large cohort, the incidence of this condition was 6.7%.

AI and Prediction of ICU Admission

Li et al. (2022) developed a machine-learning based application to support anaesthesiologists in assessing perioperative risk in patients undergoing hip surgery. The authors analysed data from over 4,000 medical records (3,000 were used to train the algorithm, while the remainder served as a validation cohort). The developed model proved superior to the ASA score in predicting the onset of the primary outcome (a set of adverse events including in-hospital mortality, acute myocardial infarction, stroke, respiratory, hepatic and renal failure, and sepsis, with an AUROC of 0.810 vs 0.629, respectively), ICU admission (AROC 0.835 vs 0.692, respectively) and prolonged length of hospital stay (AUROC 0.832 vs 0.618, respectively). The resulting model was then integrated into a web application used to support the hospital information system.

A retrospective analysis of 4,658 patients undergoing general, vascular and thoracic surgery over a 3-year period, incorporating 16 predictor variables, was used to develop a machine-learning model to predict the risk of unplanned ICU admission (Chiu et al. 2025). In the selected cohort, the incidence of unplanned ICU admission was 2.3%. Long preoperative hospital stays, advanced age and high BMI significantly influenced the model's positive prediction of unplanned ICU admission, while a laparoscopic or robotic surgical approach showed an opposite effect. The developed model proved to be a reliable tool for predicting this rare complication, with an AUROC of 0.80.

In the specific context of non-small cell lung carcinoma surgery, Isik et al. (2025) analysed clinical and laboratory data, radiological tumour characteristics and intraoperative variables from 953 patients who underwent surgery between 2001 and 2023 (Isik et al. 2025). The collected data were divided into training and testing sets. The training data were used to develop the machine learning models, while the testing data were used to evaluate their performance in predicting postoperative ICU admission. The AI algorithm demonstrated an accuracy of 85.3% with an AUROC of 0.83.

A retrospective analysis of 2,268 medical records was conducted to train and test an AI-based natural language processing model. The model proved highly effective in predicting ICU admissions among elective neurosurgical patients at University College London Hospital, achieving a recall of 0.87 for ICU admissions and 0.85 for unplanned recoveries, with an Area Under the Curve (AUC) of 0.99 (Ive et al. 2025).

Langlois et al. (2024) retrospectively analysed data from 7,251 patients undergoing laminectomy, colectomy and thoracic surgery. The resulting machine learning algorithm was used to develop three different clinical phenotypes, each associated with a progressive risk of in-hospital and 30-day mortality, prolonged length of hospital stays and ICU admission. The obtained digital phenotypes achieved a higher AUROC than ASA for each outcome (0.76 vs 0.71 for predicting ICU admission).

The Predictive OpTimal Trees in Emergency Surgery Risk Intensive Care Unit application is another AI-based tool developed to predict the risk of ICU admission after emergency surgery (Gebran et al. 2022). Also in this context, the resulting model proved to be a reliable tool in supporting clinical decisions to triage patients in the ICU with an AUC value of 0.88.

Park developed and validated an AI model to predict prolonged ICU stays and mortality among surgical patients. The authors analysed 6,029 critically ill patients from two medical centres. The model demonstrated good accuracy in predicting prolonged ICU stays, with an AUROC of 0.7376 in the external validation. Emergency surgery, high lactate levels, and low diastolic blood pressure played key roles in the resulting algorithm (Park et al. 2025).

AI-based protocols were also found to be effective in predicting readiness for discharge from the PACU compared to the traditional staff evaluation or Aldrete checklist in a cross-sectional study involving 830 patients undergoing general anaesthesia (Maroufi et al. 2025).

All presented studies include as possible limitations the retrospective nature of the analysis, the single-centre design in most cases, as well as the execution of the analysis and validation process in a specific clinical context and/or in a single hospital with specific organisational structures.

Conclusion and Future Perspectives

Postoperative ICU admission remains a complex clinical decision based on several factors, including specific patient characteristics, the type of surgery, intraoperative variables, vital signs recorded in the immediate postoperative period, staff experience, and specific organisational issues. Unnecessary ICU admission is associated with increased costs, while unplanned late emergency ICU admissions significantly increase mortality and morbidity. AI models, with their ability to process large amounts of data, could provide a fundamental contribution to clinical decisions in this specific area. Preliminary scientific evidence demonstrated a good level of accuracy of AI-based algorithms in predicting the risk of ICU admissions in various surgical contexts. However, all the AI models presented were developed through retrospective analyses, conducted, in most cases, in a few centres, characterised by specific organisational challenges.

Future studies should evaluate the role of AI in supporting physicians in determining the risk of postoperative ICU admission through large, prospective, multicentre studies capable of demonstrating the effectiveness of AI models in different clinical settings and hospital organisations, as well as their potential beneficial effects on clinically relevant outcomes.

Conflict of Interest

None.