Search

Machine learning (ML) is increasingly applied to predict adverse postoperative outcomes in cardiac surgery. Commonly used ML models fail to translate to clinical practice due to absent model explainability, limited uncertainty quantification, and no flexibility to missing data. We aimed to develop and benchmark a novel ML approach, the uncertainty-aware attention network (UAN), to overcome these common limitations. Two Bayesian uncertainty quantification methods were tested, generalized variational inference (GVI) or a posterior network (PN). The UAN models were compared with an ensemble of XGBoost models and a Bayesian logistic regression model (LR) with imputation. The derivation datasets consisted of 153,932 surgery events from the Australian and New Zealand Society of Cardiac and Thoracic Surgeons (ANZSCTS) Cardiac Surgery Database. An external validation consisted of 7343 surgery events which were extracted from the Medical Information Mart for Intensive Care (MIMIC) III critical care dataset. The highest performing model on the external validation dataset was a UAN-GVI with an area under the receiver operating characteristic curve (AUC) of 0.78 (0.01). Model performance improved on high confidence samples with an AUC of 0.81 (0.01). Confidence calibration for aleatoric uncertainty was excellent for all models. Calibration for epistemic uncertainty was more variable, with an ensemble of XGBoost models performing the best with an AUC of 0.84 (0.08). Epistemic uncertainty was improved using the PN approach, compared to GVI. UAN is able to use an interpretable and flexible deep learning approach to provide estimates of model uncertainty alongside state-of-the-art predictions. The model has been made freely available as an easy-to-use web application demonstrating that by designing uncertainty-aware models with innately explainable predictions deep learning may become more suitable for routine clinical use. [ABSTRACT FROM AUTHOR]

Objective: To compare the characteristics, use of invasive ventilation and outcomes of patients admitted with critical asthma syndrome (CAS) to ICUs in Australia and New Zealand (ANZ), and a large cohort of ICUs in the United States (US). Methods: We examined two large databases of ICU for patients admitted with CAS in 2014 and 2015. We obtained, analyzed, and compared information on demographic and physiological characteristics, use of invasive mechanical ventilation, and clinical outcome and derived predictive models. Results: Overall, 2202 and 762 patients were admitted with a primary diagnosis of CAS in the ANZ and US databases respectively (0.73% vs. 0.46% of all ICU admissions, P < 0.001). A similar percentage of patients received invasive mechanical ventilation in the first 24 h (24.7% vs. 24.4%, P = 0.87) but ANZ patients had lower respiratory rates and higher PaCO2 levels. Overall mortality was low (1.23 for ANZ and 1.71 for USA; P = 0.36) and even among invasively ventilated patients (2.4% for ANZ vs. 1.1% for USA; P = 0.38). However, ANZ patients also had longer length of stay in ICU (43 vs. 37 h, P = 0.001) and hospital (105 vs. 78 h, P = 0.003). Conclusions: Patients admitted to ANZ and USA ICU with CAS are broadly similar and have a low and similar rate of invasive ventilation and mortality. However, ANZ patients made up a greater proportion of ICU patients and had longer ICU and hospital stays. These findings provide a modern invasive ventilation and mortality rates benchmark for future studies of CAS. [ABSTRACT FROM AUTHOR]

Purpose: Comparisons between institutions of intensive care unit (ICU) length of stay (LOS) are significantly confounded by individual patient characteristics, and currently there is a paucity of methods available to calculate risk-adjusted metrics. Methods: We extracted de-identified data from the Australian and New Zealand Intensive Care Society (ANZICS) Adult Patient Database for admissions between January 1 2011 and December 31 2015. We used a mixed-effects log-normal regression model to predict LOS using patient and admission characteristics. We calculated a risk-adjusted LOS ratio (RALOSR) by dividing the geometric mean observed LOS by the exponent of the expected Ln-LOS for each site and year. The RALOSR is scaled such that values <1 indicate a LOS shorter than expected, while values >1 indicate a LOS longer than expected. Secondary mixed effects regression modelling was used to assess the stability of the estimate in units over time. Results: During the study there were a total of 662,525 admissions to 168 units (median annual admissions = 767, IQR:426–1121). The mean observed LOS was 3.21 days (median = 1.79 IQR = 0.92–3.52) over the entire period, and declined on average 1.97 hours per year (95%CI:1.76–2.18) from 2011 to 2015. The RALOSR varied considerably between units, ranging from 0.35 to 2.34 indicating large differences after accounting for case-mix. Conclusions: There are large disparities in risk-adjusted LOS among Australian and New Zealand ICUs which may reflect differences in resource utilization. [ABSTRACT FROM AUTHOR]

Objectives: Current definitions of acute kidney injury use a urine output threshold of less than 0.5 mL/kg/hr, which have not been validated in the modern era. We aimed to determine the prognostic importance of urine output within the first 24 hours of admission to the ICU and to evaluate for variance between different admission diagnoses.Design: Retrospective cohort study.Setting: One-hundred eighty-three ICUs throughout Australia and New Zealand from 2006 to 2016.Patients: Patients greater than or equal to 16 years old who were admitted with curative intent who did not regularly receive dialysis. ICU readmissions during the same hospital admission and patients transferred from an external ICU were excluded.Measurements and Main Results: One hundred and sixty-one thousand nine hundred forty patients were included with a mean urine output of 1.05 mL/kg/hr and an overall in-hospital mortality of 7.8%. A urine output less than 0.47 mL/kg/hr was associated with increased unadjusted in-hospital mortality, which varied with admission diagnosis. A machine learning model (extreme gradient boosting) was trained to predict in-hospital mortality and examine interactions between urine output and survival. Low urine output was most strongly associated with mortality in postoperative cardiovascular patients, nonoperative gastrointestinal admissions, nonoperative renal/genitourinary admissions, and patients with sepsis.Conclusions: Consistent with current definitions of acute kidney injury, a urine output threshold of less than 0.5 mL/kg/hr is modestly predictive of mortality in patients admitted to the ICU. The relative importance of urine output for predicting survival varies with admission diagnosis. [ABSTRACT FROM AUTHOR]

Background: Resuscitated cardiac arrest is associated with high mortality; however, the ability to estimate risk of adverse outcomes using existing illness severity scores is limited. Using in-hospital data available within the first 24 hours of admission, we aimed to develop more accurate models of risk prediction using both logistic regression (LR) and machine learning (ML) techniques, with a combination of demographic, physiologic, and biochemical information., Methods and Findings: Patient-level data were extracted from the Australian and New Zealand Intensive Care Society (ANZICS) Adult Patient Database for patients who had experienced a cardiac arrest within 24 hours prior to admission to an intensive care unit (ICU) during the period January 2006 to December 2016. The primary outcome was in-hospital mortality. The models were trained and tested on a dataset (split 90:10) including age, lowest and highest physiologic variables during the first 24 hours, and key past medical history. LR and 5 ML approaches (gradient boosting machine [GBM], support vector classifier [SVC], random forest [RF], artificial neural network [ANN], and an ensemble) were compared to the APACHE III and Australian and New Zealand Risk of Death (ANZROD) predictions. In all, 39,566 patients from 186 ICUs were analysed. Mean (±SD) age was 61 ± 17 years; 65% were male. Overall in-hospital mortality was 45.5%. Models were evaluated in the test set. The APACHE III and ANZROD scores demonstrated good discrimination (area under the receiver operating characteristic curve [AUROC] = 0.80 [95% CI 0.79-0.82] and 0.81 [95% CI 0.8-0.82], respectively) and modest calibration (Brier score 0.19 for both), which was slightly improved by LR (AUROC = 0.82 [95% CI 0.81-0.83], DeLong test, p < 0.001). Discrimination was significantly improved using ML models (ensemble and GBM AUROCs = 0.87 [95% CI 0.86-0.88], DeLong test, p < 0.001), with an improvement in performance (Brier score reduction of 22%). Explainability models were created to assist in identifying the physiologic features that most contributed to an individual patient's survival. Key limitations include the absence of pre-hospital data and absence of external validation., Conclusions: ML approaches significantly enhance predictive discrimination for mortality following cardiac arrest compared to existing illness severity scores and LR, without the use of pre-hospital data. The discriminative ability of these ML models requires validation in external cohorts to establish generalisability., Competing Interests: I have read the journal's policy and the authors of this manuscript have the following competing interests: KR is director of IntelliHQ Pty Ltd, non-profit AI innovation centre for healthcare, connected with Gold Coast University Hospital. KR is owner and Chairman of K. J. Ross & Associates Pty. Ltd. (KJR), professional services firm specialising in IT risk management and assurance. 20% of KJR's work is in healthcare. There is no direct financial stake in the results of the current study.