Deviation from the intended perioperative care pathway, particularly unplanned admission after day-case surgery, affects patient safety and the efficient organisation of care. We aimed to develop and compare interpretable machine-learning models predicting such deviations from information available before surgery. In this retrospective study, 51,112 consecutive ambulatory anaesthesia procedures from a Belgian network were analysed, with unplanned admission as the outcome. Data were partitioned into training, validation, and test sets; class imbalance was addressed with SMOTENC on the training data alone. Seven algorithms: logistic regression, decision tree, random forest, XGBoost, LightGBM, histogram-based gradient boosting, and explainable boosting machine (EBM) were compared using discrimination, precision–recall, and calibration metrics at two operating points (fixed sensitivity and fixed specificity). Unplanned admission occurred in 9.08
To demonstrate a method for the identification of safe cerebral perfusion pressure (CPP) in patients with severe traumatic brain injury using brain tissue oxygen (PbtO₂) readings as a surrogate marker of cerebral blood flow. Retrospective analysis of prospectively collected data from patients with severe TBI (sTBI). A total of 9 patients with sTBI enrolled in the BONANZA-GT trial with continuous monitoring of mean arterial blood pressure (MAP), intracranial pressure (ICP), cerebral perfusion pressure (CPP) and PbtO₂. MAP, ICP, CPP and high frequency PbtO₂ signals were continuously recorded. stoCPP was calculated as the coefficient of determination (r²) of short-term changes in PbtO₂ regressed on short-term changes in CPP within non-overlapping 144-second correlation windows, assigned to 5-mmHg CPP bins using a window-median binning approach. Algorithm parameters were optimised using Bayesian optimisation with a composite scoring function benchmarked against CPPopt. Per-subject r² thresholds were derived by block-shuffle calibration (median 0.12, range 0.06–0.26). stoCPP was successfully generated for all 9 patients from raw unfiltered data. The lower limit of the stoCPP range (LLstoCPP) was defined in approximately 22
Early detection of clinical deterioration on surgical wards is vital to prevent adverse outcomes. Conventional ward monitoring using the Modified Early Warning Score (MEWS) relies on intermittent nurse-measured vital signs, but its role may change when continuous wearable monitoring is already available. This study evaluated the transition from mandatory to discretionary nurse-measured MEWS assessment in a continuously monitored surgical ward and compared the diagnostic performance of two wearable-derived Early Warning Scores (EWS) with conventional MEWS. In this single-center cohort study conducted between October 2024 and June 2025, postoperative patients were continuously monitored using a wearable sensor measuring heart rate, respiratory rate, and oxygen saturation. Continuous wearable monitoring was available throughout the entire study period. During Phase 1, nurse-measured MEWS assessments were mandatory according to local protocol. During Phase 2, nurse-measured MEWS assessments were performed at staff discretion. Two wearable-derived algorithms were evaluated: the Continuous Remote EWS (CREWS), based on a previously developed algorithm, and Remote 3 EWS (R3EWS), a modified version of the hospital’s standard MEWS using three parameters. The primary outcome was diagnostic accuracy for clinical deterioration, defined as postoperative complications graded Clavien–Dindo ≥ II. A total of 544 admissions (516 patients) were included, of which 98 admissions (18
Automated pupillometry provides a standardised, quantitative, highly reproducible measurement of the pupillary light reactivity and other pupillary variables. The Neurologic pupillary index (NPi) has shown good prognostic value in patients with acute brain injury and cardiac arrest. However, few data on the comparison of NPi with a recently introduced index (quantitative pupillary index, QPI) are available. To compare the performance of two commercially available automated pupillometry devices, and in particular QPI and NPi, in a cohort of neuro-critically ill patients. Single-center observational study, including adult (> 18 years) patients admitted to a neuro-intensive care unit over a 6-month period. Pupillary reactivity was assessed in both eyes of each patient using the two pupillometers currently available, the Neuroptics NPi-300 (Neuroptics, Irvine, CA, USA) and the NeuroLight (ID-Med, Marseilles, France). A total of 73 patients were included in the final analysis, with a median age of 65 years (interquartile range [IQR] 54–77). The mean bias between the two pupillometers was 0.44 mm (95
The purpose was to describe patient monitor alarm burden in a cardiac surgical intensive care unit and explore whether routinely available patient characteristics are associated with alarm frequency, type, and duration. Secondary objectives were to assess indirect signs of alarm fatigue and compare alarm rates with two previous studies using identical monitors in different clinical units. We retrospectively collected patient-monitor alarm data and patient characteristics over four consecutive weeks in the cardiac surgical intensive care unit of Helsinki University Hospital. The study population comprised 81 patients. Alarm data were analysed descriptively and modelled using negative-binomial regression. Patients produced 45,044 alarms, with a median alarm rate of 6.8 [4.8 to 10.8] alarms per patient per hour. Older age was associated with fewer alarms (IRR = 0.90, 95
Routinely collected real time data from ICU bedside monitors is a valuable source of information which can assist the clinical team caring for a patient. This data almost always contains artifacts caused by clinical staff attending to the patient throughout the day, e.g. blood samples being taken, body washing, administration of IV medication. We present a small (10 patient) study but with accurately annotated clinical event data constructed by observing standard clinical procedures and merging this information with the physiologic data (ECG, ABP, CVP) from bedside monitors. In addition, to assess the feasibility for using accelerometers in the ICU environment, we received ethical approval to place an accelerometer [1] on the chest of all patients and the X, Y Z accelerometer plane data was merged with the physiological and event annotation data. The non-trivial methods required to capture the data and initial characterisation of results are detailed and discussed. A baseline model cross-correlating ECG and ABP waveform signals over a 10 s window was developed and features derived from this model were used in both GLM and Random Forest (RF) models for predicting events that routinely cause movement related artifact in waveform quality time-series data. RF model performance improved when accelerometer data was added to the model for most movement related artifact types. For the GLM models, only selected models (ANY Type, Turns, Bed Movement and Blood Sample) showed an improvement in model fit with the addition of the accelerometer data (P < 0.01). Implications for observational ICU research and potential “Closed Loop” clinical modelling studies are discussed.
To evaluate whether jugular venous NIRS (rSjvO₂) and baseline-calibrated cerebral NIRS (cal-rScO₂) can serve as surrogates for mixed venous oxygen saturation (SvO₂), and whether these inputs can support continuous cardiac output (CO) estimation during off-pump coronary artery bypass grafting (OPCAB). In this prospective observational study, 25 adults undergoing elective OPCAB were enrolled at a single tertiary center. NIRS sensors were placed over the left internal jugular vein (IJV) under ultrasound guidance and over the bilateral forehead. Pulmonary artery catheter measurements served as the reference for SvO₂ and thermodilution CO (COTD). Agreement was assessed using Bland–Altman analysis with linear mixed-effects modeling, and trending ability by error grid analysis. CO was estimated using a previously derived modified Fick equation. rSjvO₂ showed modest bias versus SvO₂ (2.0
To investigate the relationship between hemoglobin concentration and peripheral perfusion index (PI) in normovolemic, hemodynamically stable patients. This prospective observational study included 90 adult emergency department patients, classified as anemic (n = 45) and non-anemic (n = 45) according to World Health Organization criteria. PI was measured using standardized pulse oximetry conditions. Group comparisons were performed using appropriate statistical tests. The association between hemoglobin and PI was evaluated using Spearman correlation. A multivariable linear regression model with log-transformed PI was used to identify independent predictors, adjusting for clinically relevant variables. Model assumptions and influential observations were assessed. PI values were significantly lower in anemic patients compared with non-anemic patients (median 2.1 vs. 3.5, p < 0.001). Hemoglobin concentration showed a moderate positive correlation with PI (r = 0.425, p < 0.001). In multivariable analysis, hemoglobin remained independently associated with PI (β = 0.306, p = 0.021). The model demonstrated modest explanatory power (adjusted R²=0.069). Sensitivity analysis excluding influential observations yielded consistent findings. Hemoglobin level is independently associated with PI in normovolemic patients. Although PI primarily reflects flow-related peripheral perfusion, hemoglobin contributes to its variability. These findings suggest that anemia and peripheral perfusion are interrelated but distinct physiological parameters that should be considered together in clinical interpretation.
Arterial pressure waveform analysis has been used to assess hemodynamic status, however, investigations focusing on waveform morphology remain limited. In this study, we examined changes in selected morphological features of the arterial pressure waveform following volume loading versus vasopressor administration in critically ill patients. In this retrospective observational study, we analyzed adult patients admitted to an intensive care unit between April 2022 and June 2023. Participants were included if they exhibited increased mean arterial pressure (MAP) following either blood return during hemodiafiltration discontinuation (Volume group) or intravenous phenylephrine administration (Pressor group). Arterial waveforms were extracted from one minute before to four minutes after the intervention. We compared 21 waveform-derived features and their normalized parameters at peak MAP with baseline values. Pre- and post-intervention differences for each parameter were assessed using the Wilcoxon matched-pairs signed-rank test. The Volume group comprised 26 patients, and the Pressor group included 20 patients. Systolic angle, systolic dP/dt, and systolic area under the curve (AUC) increased in both groups following intervention. After pulse-pressure normalization, however, systolic AUC increased significantly in the Volume group (median paired difference 4.6 [-0.3–10.2] ms, 95
Advanced haemodynamic monitoring of cardiac index is restricted to selected high-risk cases, leaving most patients undergoing major surgery without real-time assessment. We aimed to develop and internally validate a simplified bedside score for continuous, minute-by-minute screening of low cardiac index using routinely available parameters. Retrospective single-centre study of adults undergoing major abdominal surgery with advanced haemodynamic monitoring (January 2023 to February 2026). Low cardiac index was defined as pulse contour cardiac index < 2.2 l.min− 1.m− 2. The dataset was split 80/20 at the patient level. Twelve candidate features were ranked using SHapley Additive exPlanations (SHAP) values; the top five were used to fit a generalised linear mixed model, yielding the integer-weighted Low Cardiac Index (LCI) score. A rule-out threshold (sensitivity ≥ 0.90) was selected, and test-set performance was evaluated with patient-level cluster bootstrap. The analysis cohort comprised 136 patients and 42,346 timepoints (15.5
Early detection of systemic inflammation and sepsis is essential for improving patient outcomes. Heart rate variability (HRV) has been proposed as a non-invasive marker of inflammation; however, its specificity is limited, particularly under anesthesia or sedation. This study introduces the Trend of Sepsis and Inflammation (TSI), a novel index combining HRV and electroencephalogram (EEG)-derived parameters to better isolate inflammation-related autonomic changes.A post-hoc observational analysis was conducted in two cohorts: surgical patients undergoing general anesthesia (N = 41) and septic patients admitted to the intensive care unit (N = 21). Continuous EEG and ECG recordings were obtained using a multimodal monitoring system. HRV parameters from time and frequency domains, together with an EEG-derived Brain Activity Index, were integrated using an Adaptive Neuro-Fuzzy Inference System (ANFIS) to compute the TSI. TSI values were compared across three conditions: awake baseline, post-surgical inflammation, and sepsis. Statistical analysis employed non-parametric tests, while discriminative performance was assessed using prediction probability (Pk) and concordance with predefined interpretation ranges.TSI values increased progressively across conditions. Median values were 2.29 (IQR: 0.00–13.74) at baseline, 24.68 (18.50–32.00) post-surgery, and 65.59 (52.38–76.25) in sepsis, with significant differences (p < 0.0001). The TSI demonstrated strong discriminative ability (Pk = 93.58
Recruitment maneuvers (RM) can improve oxygenation in patients with acute respiratory distress syndrome (ARDS), but their physiological effects depend on lung recruitability. This secondary analysis of a randomized controlled trial (RCT) evaluated oxygenation, respiratory mechanics, regional ventilation, and cardiorespiratory adverse events responses to a RM followed by electrical impedance tomography (EIT)-guided PEEP titration, using EIT to assess lung recruitability. In this study, fifty patients with moderate-to-severe ARDS underwent a stepwise RM followed by individualized PEEP titration guided by EIT. Lung recruitability was determined using the collapse index at PEEP 6 cmH₂O (CLPEEP6), defined as the proportion of collapsed lung at this PEEP level. Patients were classified into high- and low-recruitability groups based on median CLPEEP6 values. Oxygenation (PaO₂/FiO₂), static compliance (Cstat), driving pressure (Pdriv), regional ventilation distribution, and cardiorespiratory adverse events were compared before and after RM, during subsequent individualized EIT-guided PEEP titration. In patients with high recruitability (CLPEEP6 > 12.5), the PaO₂/FiO₂ ratio and Cstat increased significantly after RM (PaO₂/FiO₂: 100.8 ± 30.3 vs. 125.4 ± 38.3 mmHg, p < 0.05; Cstat: 21.5 ± 5.8 vs. 28.0 ± 7.0 mL/cmH₂O, p < 0.001), Pdriv decreased (19.1 ± 3.5 vs. 15.6 ± 3.2 cmH₂O, p < 0.001). EIT demonstrated a posterior redistribution of ventilation after RM. In contrast, patients with low recruitability (CLPEEP6 ≤ 12.5) showed no significant mechanical or oxygenation improvement and transient hypotension, arrhythmia, and desaturation appeared numerically more common in this group. No barotrauma or cardiac arrest occurred, and ICU mortality was similar between groups. A strategy combining a RM with subsequent individualized EIT-guided PEEP titration was associated with improved oxygenation and lung mechanics in patients with high lung recruitability, whereas patients with low recruitability showed limited physiological benefit, with cardiorespiratory adverse events appearing numerically more frequent. The EIT-derived CLPEEP6 index represents a feasible and clinically applicable https://clinicaltrials.gov/study/NCT06733168.
We developed machine learning (ML) models to perform continuous hourly prediction of arterial blood gas (ABG) and basic metabolic panel (BMP) laboratory values in critically ill patients. We evaluated its impact on prediction of the need for renal replacement therapy (RRT). We compared the performance of the deep learning models using laboratory variables imputed hourly by our ML estimators versus models using laboratory variables imputed by a previous-value baseline. Our ML model incorporated various predictors, including previous laboratory values, administered medications, clinical events, fluid input/output, hemodynamics, and ventilator settings. Accuracy of laboratory imputations were compared using mean absolute error (MAE) and root mean squared error (RMSE). We trained bi-directional long short-term memory models (Bi-LSTM) to predict need for renal replacement therapy (RRT) that differed only in how laboratory variables were imputed over time. We compared performance using area under the receiver operating characteristics curve (AUROC) and area under the precision-recall curve (AUPRC). XGBoost achieved an average error reduction of 33
Maintaining mean arterial pressure (MAP) within a predefined target is central to haemodynamic management in surgical and critically ill adults receiving vasopressors. Closed-loop vasopressor (CLV) systems automate titration to optimise blood pressure control, but their clinical effectiveness remains uncertain. We performed a systematic review and meta-analysis comparing CLV with manual titration. This PRISMA 2020–compliant review was prospectively registered in PROSPERO (CRD420250655697). MEDLINE, Embase, Scopus, Web of Science, CENTRAL, and the Cochrane Library were searched (January 2000–June 2025). Randomised controlled trials enrolling adults receiving vasopressors in perioperative or intensive care settings were included. Primary outcomes were time within the MAP target range and time spent in hypotension or hypertension. Risk of bias was assessed using RoB 2.0 and certainty of evidence using GRADE. Random- or fixed-effects models were selected according to heterogeneity. Six randomized controlled trials (215 patients) were included in the systematic review, whereas five perioperative trials contributed to the meta-analysis of haemodynamic control outcomes, and one ICU-based study was summarized narratively because it did not report comparable MAP control endpoints. CLV increased time within the MAP target range (mean difference [MD] 33.94
We investigated whether pre-induction High Frequency Variability Index (HFVI)/Analgesia Nociception Index (ANI) and Nociception Level (NOL) values could predict hypotension during general anesthesia induction in older patients. This prospective observational study included patients aged ≥ 65 years undergoing elective surgery under general anesthesia. After recording HFVI/ANI and NOL values, remifentanil was administered at 0.2 µg/kg/min for 2 min, followed by a rapid intravenous bolus administration of propofol at 2 mg/kg. Noninvasive blood pressure was measured at 1-min intervals over 5 min. Hypotension was defined as mean arterial pressure (MAP) < 65 mmHg. Predictive performance was evaluated using receiver operating characteristic (ROC) curves. Additionally, the area under MAP < 65 mmHg (AUT-65), a continuous metric integrating the depth and duration of hypotension, was calculated and correlated with nociception index values. We analyzed 51 patients (19 men; 22 in the hypotension group and 29 in the nonhypotension group). The hypotension group demonstrated lower HFVI/ANIi values (58.7 ± 11.1 vs. 69.0 ± 15.1, p = 0.010) and higher NOL values (17.2 ± 9.1 vs. 10.7 ± 8.8, p = 0.013). HFVI/ANIi (AUC = 0.72, cutoff = 62.5) and NOL (AUC = 0.71, cutoff = 13.5) moderately predicted hypotension. AUT-65 was significantly correlated with HFVI/ANIi (Spearman rho = − 0.35, p = 0.013) and NOL (rho = 0.47, p = 0.0005). Pre-induction HFVI/ANI and NOL were associated with the occurrence and cumulative burden of post-induction hypotension in patients aged ≥ 65 years, suggesting their potential utility as pre-induction screening tools. Trial registration This study was approved by the Ethics Committee of Yamagata University Hospital on January 18, 2024 (approval number 2023 − 280). Before commencement, the study was registered as a prospective observational study in the UMIN Clinical Trials Registry (UMIN000053529).
Background Artificial intelligence (AI) can enhance diagnostics, treatment, and workflow efficiency. However, successful integration into clinical practice depends on users’ acceptance. Objective To investigate benefits, barriers, and challenges of AI applications among anaesthesia and intensive care professionals. Design International online survey. Main outcome measures The survey included items on familiarity and experiences with AI applications, perceived benefits, concerns, and demographic variables. Descriptive analyses, fisher exact tests, χ²-tests, odds ratios, and Spearman rank correlations were used to explore associations between responses and demographics. Results The survey was distributed by the European Society of Anaesthesiology and Intensive Care in 2023. A total of 510 respondents completed the entire survey, primarily from Europe (78
Continuous monitoring of vital signs after hospital discharge may support early recognition of deviating vital signs. However, the utility may be challenged by high alert frequencies. This exploratory study aimed to assess the impact of evidence-based augmented filtering algorithms on alert frequency following discharge. Adult patients (≥ 18 years) discharged after acute medical admission were monitored continuously using wearable devices that measured heart rate, respiratory rate, blood pressure, and oxygen saturation. The primary outcome was the number of alerts per patient per day. We compared outcomes across three filtering strategies: (1) no filtering, (2) artefact removal, and (3) filtering with artefact removal and clinical criteria based upon severity and duration. Ninety-eight patients were enrolled; the total vital sign alert frequency was reduced from a median of 74 [IQR 36–125] to 5 [IQR 1–13] alerts/patient/day following application of the clinical criteria filters, corresponding to an 84
To compare the safety and effectiveness of distal radial artery (DRA) versus conventional radial artery (CRA) catheterization for invasive arterial blood pressure monitoring. This meta-analysis followed PRISMA guidelines. Randomized controlled trials published up to December 30, 2025 were systematically searched in PubMed, Embase, Web of Science, the Cochrane Library, CINAHL, CNKI, Wanfang, VIP, and SinoMed. Two reviewers independently screened studies, extracted data, and assessed risk of bias. Meta-analyses were conducted using Review Manager 5.4 and Stata 18.0, and evidence quality was evaluated with the GRADE system. 12 randomized controlled trials (RCTs) involving 1,790 participants were included. For the primary outcomes, compared with CRA, DRA was associated with lower incidences of haematoma (RR = 0.42, 95
