Accurate and interpretable prediction of ICU-acquired AKI.

PURPOSE: We developed and validated two parsimonious algorithms to predict the time of diagnosis of any stage of acute kidney injury (any-AKI) or moderate-to-severe AKI in clinically actionable prediction windows. MATERIALS AND METHODS: In this retrospective single-center cohort of adult ICU admissions, we trained two gradient-boosting models: 1) any-AKI model, predicting the risk of any-AKI at least 6 h before diagnosis (50,342 admissions), and 2) moderate-to-severe AKI model, predicting the risk of moderate-to-severe AKI at least 12 h before diagnosis (39,087 admissions). Performance was assessed before disease diagnosis and validated prospectively. RESULTS: The models achieved an area under the receiver operating characteristic curve (AUROC) of 0.756 at six hours (any-AKI) and 0.721 at 12 h (moderate-to-severe AKI) prior. Prospectively, both models had high positive predictive values (0.796 and 0.546 for any-AKI and moderate-to-severe AKI models, respectively) and triggered more in patients who developed AKI vs. those who did not (median of 1.82 [IQR 0-4.71] vs. 0 [IQR 0-0.73] and 2.35 [IQR 0.14-4.96] vs. 0 [IQR 0-0.8] triggers per 8 h for any-AKI and moderate-to-severe AKI models, respectively). CONCLUSIONS: The two AKI prediction models have good discriminative performance using common features, which can aid in accurately and informatively monitoring AKI risk in ICU patients.

Estimation of Baseline Serum Creatinine with Machine Learning.

INTRODUCTION: Comparing current to baseline serum creatinine is important in detecting acute kidney injury. In this study, we report a regression-based machine learning model to predict baseline serum creatinine. METHODS: We developed and internally validated a gradient boosting model on patients admitted in Mayo Clinic intensive care units from 2005 to 2017 to predict baseline creatinine. The model was externally validated on the Medical Information Mart for Intensive Care III (MIMIC III) cohort in all ICU admissions from 2001 to 2012. The predicted baseline creatinine from the model was compared with measured serum creatinine levels. We compared the performance of our model with that of the backcalculated estimated serum creatinine from the Modification of Diet in Renal Disease (MDRD) equation. RESULTS: Following ascertainment of eligibility criteria, 44,370 patients from the Mayo Clinic and 6,112 individuals from the MIMIC III cohort were enrolled. Our model used 6 features from the Mayo Clinic and MIMIC III datasets, including the presence of chronic kidney disease, weight, height, and age. Our model had significantly lower error than the MDRD backcalculation (mean absolute error [MAE] of 0.248 vs. 0.374 in the Mayo Clinic test data; MAE of 0.387 vs. 0.465 in the MIMIC III cohort) and higher correlation (intraclass correlation coefficient [ICC] of 0.559 vs. 0.050 in the Mayo Clinic test data; ICC of 0.357 vs. 0.030 in the MIMIC III cohort). DISCUSSION/CONCLUSION: Using machine learning models, baseline serum creatinine could be estimated with higher accuracy than the backcalculated estimated serum creatinine level.

Predicting acute kidney injury in critically ill patients using comorbid conditions utilizing machine learning.

BACKGROUND: Acute kidney injury (AKI) carries a poor prognosis. Its incidence is increasing in the intensive care unit (ICU). Our purpose in this study is to develop and externally validate a model for predicting AKI in the ICU using patient data present prior to ICU admission. METHODS: We used data of 98 472 adult ICU admissions at Mayo Clinic between 1 January 2005 and 31 December 2017 and 51 801 encounters from Medical Information Mart for Intensive Care III (MIMIC-III) cohort. A gradient-boosting model was trained on 80% of the Mayo Clinic cohort using a set of features to predict AKI acquired in the ICU. RESULTS: AKI was identified in 39 307 (39.9%) encounters in the Mayo Clinic cohort. Patients who developed AKI in the ICU were older and had higher ICU and in-hospital mortality compared to patients without AKI. A 30-feature model yielded an area under the receiver operating curve of 0.690 [95% confidence interval (CI) 0.682-0.697] in the Mayo Clinic cohort set and 0.656 (95% CI 0.648-0.664) in the MIMIC-III cohort. CONCLUSIONS: Using machine learning, AKI among ICU patients can be predicted using information available prior to admission. This model is independent of ICU information, making it valuable for stratifying patients at admission.

Including urinary output to define AKI enhances the performance of machine learning models to predict AKI at admission.

PURPOSE: Acute kidney injury (AKI) is a prevalent and detrimental condition in intensive care unit patients. Most AKI predictive models only predict creatinine-triggered AKI (AKICr) and might underperform when predicting urine-output-triggered AKI (AKIUO). We aimed to describe how admission AKICr prediction models perform in all AKI patients. MATERIALS AND METHODS: Three types of models were trained: 1) pAKIany, predicting AKI based on creatinine or urine output, 2) pAKIUO, predicting AKI based only on urine output, and 3) pAKICr, predicting AKI based only on creatinine. We compared model performance and predictive features. RESULTS: The pAKIany models had the best overall performance (AUROC 0.673-0.716) and the most consistent performance across three patient cohorts grouped by type of AKI trigger (min AUROC of 0.636). The pAKICr models had fair performance in predicting AKICr (AUROCs 0.702-0.748) but poor performance predicting AKIUO (AUROCs 0.581-0.695). The predictive features for the pAKICr models and pAKIUO models were distinct, while top features for the pAKIany models were consistently a combination of those for the pAKICr and pAKIUO models. CONCLUSION: Ignoring urine output in the outcome during model training resulted in models that are unlikely to predict AKIUO adequately and may miss a substantial proportion of patients in practice.

Descriptive study of differences in acute kidney injury progression patterns in General and Cardiac Intensive Care Units.

BACKGROUND: Acute kidney injury is common in critically ill patients with detrimental effects on mortality, length of stay and post-discharge outcomes. The Acute Kidney Injury Network developed guidelines based on urine output and serum creatinine to classify patients into stages of acute kidney injury. METHODS: In this analysis we utilize the Acute Kidney Injury Network guidelines to evaluate the acute kidney injury stage in patients admitted to general and cardiac intensive care units over a period of 18 months. Acute kidney injury stage was calculated in real time hourly based on the guidelines and using these temporal stage scores calculated for the population; the prevalence and progression of acute kidney injury stage was compared between the two units. We hypothesized that the prevalence and progression of acute kidney injury stage between the two units may be different. RESULTS: More cardiac intensive care unit patients had no acute kidney injury (stage <1) during their intensive care unit stay but more cardiac intensive care unit patients developed acute kidney injury (stage >1), compared to the General Intensive Care Unit. Both at intensive care unit admission and discharge, more General Intensive Care Unit patients had acute kidney injury; however, the number of cardiac intensive care unit patients with acute kidney injury was three times higher at discharge than admission. Acute kidney injury developed in a different pattern in the two intensive care units over five days of intensive care unit stay. In the General Intensive Care Unit, acute kidney injury was most prevalent on second day of intensive care unit stay and in cardiac intensive care unit acute kidney injury was most prevalent on the third day of intensive care unit stay. We observed the biggest increase in new acute kidney injury in the first day of General Intensive Care Unit and second day of the cardiac intensive care unit stay. CONCLUSIONS: The study demonstrates the different trends of acute kidney injury pattern in general and cardiac intensive care unit patient populations highlighting the earlier development of acute kidney injury on General Intensive Care Unit and more prevalence of acute kidney injury on discharge from cardiac intensive care unit.

Automated Continuous Acute Kidney Injury Prediction and Surveillance: A Random Forest Model.

OBJECTIVE: To develop and validate a prediction model of acute kidney injury (AKI) of any severity that could be used for AKI surveillance and management to improve clinical outcomes. PATIENTS AND METHODS: This retrospective cohort study was conducted in medical, surgical, and mixed intensive care units (ICUs) at Mayo Clinic in Rochester, Minnesota, including adult (≥18 years of age) ICU-unique patients admitted between October 1, 2004, and April 30, 2011. Our primary objective was prediction of AKI using extant clinical data following ICU admission. We used random forest classification to provide continuous AKI risk score. RESULTS: We included 4572 and 1958 patients in the training and validation mutually exclusive cohorts, respectively. Acute kidney injury occurred in 1355 patients (30%) in the training cohort and 580 (30%) in the validation cohort. We incorporated known AKI risk factors and routinely measured vital characteristics and laboratory results. The model was run throughout ICU admission every 15 minutes and achieved an area under the receiver operating characteristic curve of 0.88 on validation. It was 92% sensitive and 68% specific and detected 30% of AKI cases at least 6 hours before the criterion standard time (AKI stages 1-3). For discrimination of AKI stages 2 to 3, the model had 91% sensitivity, 71% specificity, and 53% detection of AKI cases at least 6 hours before AKI onset. CONCLUSION: We developed and validated an AKI prediction model using random forest for continuous monitoring of ICU patients. This model could be used to identify high-risk patients for preventive measures or identifying patients of prospective interventional trials.

Description of vital signs data measurement frequency in a medical/surgical unit at a community hospital in United States.

Vitals signs are measured at scheduled intervals by nurses in typical general wards. Vital signs may be measured more frequently if the patient condition deteriorates. In many units, the vital signs measurement frequency for some patients is different from the scheduled frequency due to various reasons such as staffing, patient acuity etc. In this article, we describe the actual measurement frequency in patients admitted to general ward in a community hospital in Arizona, US. We present the data in the form of 2 sets of graphs. The first set of graphs are histograms which show the distribution of the number of measurements in a 24 h period for 6 different vital signs. The second set of graphs show the proportion of the patient population who had a measurement of a vital sign for each hour of the last day of patient's general ward stay. The significance of this data on predicting deterioration is discussed in Ghosh et al. (2017) [1].

Early Deterioration Indicator: Data-driven approach to detecting deterioration in general ward.

INTRODUCTION: Early detection of deterioration could facilitate more timely interventions which are instrumental in reducing transfer to higher levels of care such as Intensive Care Unit (ICU) and mortality [1,2]. METHODS AND RESULTS: We developed the Early Deterioration Indicator (EDI) which uses log likelihood risk of vital signs to calculate continuous risk scores. EDI was developed using data from 11,864 general ward admissions. To validate EDI, we calculated EDI scores on an additional 2418 general ward stays and compared it to the Modified Early Warning Score (MEWS) and National Early Warning Score (NEWS). EDI was trained using the most significant variables in predicting deterioration by leveraging the knowledge from a large dataset through data mining. It was implemented electronically for continuous automatic computation. The discriminative performance of EDI, MEWS, and NEWS was calculated before deterioration using the area under the receiver operating characteristic curve (AUROC). Additionally, the performance of the 3 scores for 24h prior to deterioration were computed. EDI was a better discriminator of deterioration than MEWS or NEWS; AUROC values for the validation dataset were: EDI - 0.7655, NEWS - 0.6569, MEWS - 0.6487. EDI also identified more patients likely to deteriorate for the same specificity as NEWS or MEWS. EDI had the best performance among the 3 scores for the last 24h of the patient stay. CONCLUSION: EDI detects more deteriorations for the same specificity as the other two scores. Our results show that EDI performs better at predicting deterioration than commonly used NEWS and MEWS.

Heuristics to determine ventilation times of ICU patients from the MIMIC-II database.

Mechanical ventilation is an important life support tool for patients in intensive care units (ICU). For various research purposes related to patient hemodynamic and cardiopulmonary monitoring, it is important to know when a patient is on a ventilator. Unfortunately, the widely used MIMIC-II database contains results from user charted data, where the user did not always store ventilation on and off times explicitly and accurately. The resulting ventilation-related data are subject to error. Therefore, there are no simple rules to define ventilation times retrospectively for this dataset. Hence, we designed a simple set of rules to determine the ventilation times using multiple sources of mechanical ventilator-related settings and physiological measurements by expert heuristics. The rules worked well in comparison with nursing notes regarding ventilation events. We conclude that our rule sets for determining ventilation times may be useful in assisting with MIMIC-II database analysis.

Development and evaluation of predictive alerts for hemodynamic instability in ICU patients.

This paper describes an algorithm for identifying ICU patients that are likely to become hemodynamically unstable. The algorithm consists of a set of rules that trigger alerts. Unlike most existing ICU alert mechanisms, it uses data from multiple sources and is often able to identify unstable patients earlier and with more accuracy than alerts based on a single threshold. The rules were generated using a machine learning technique and were tested on retrospective data in the MIMIC II ICU database, yielding a specificity of approximately 0.9 and a sensitivity of 0.6.

Predicting ICU hemodynamic instability using continuous multiparameter trends.

BACKGROUND: Identifying hemodynamically unstable patients in a timely fashion in intensive care units (ICUs) is crucial because it can lead to earlier interventions and thus to potentially better patient outcomes. Current alert algorithms are typically limited to detecting dangerous conditions only after they have occurred and suffer from high false alert rates. Our objective was to predict hemodynamic instability at least two hours before a major clinical intervention (e.g., vasopressor administration), while maintaining a low false alert rate. STUDY POPULATION: From the MIMIC II database, containing ICU minute-by-minute heart rate (HR) and invasive arterial blood pressure (BP) monitoring trend data collected between 2001 and 2005, we identified 132 stable and 104 unstable patients that met our stability-instability criteria and had sufficient data points. METHOD: We first derived additional physiological parameters of shock index, rate pressure product, heart rate variability, and two measures of trending based on HR and BP. Then we developed 220 statistical features and systematically selected a small set to use for classification. We applied multi-variable logistic regression modeling to do classification and implemented validation via bootstrapping. RESULTS: Area under receiver-operating curve (ROC) 0.83+/-0.03, sensitivity 0.75+/-0.06, and specificity 0.80+/-0.07; if the specificity is targeted at 0.90, then the sensitivity is 0.57+/-0.07. Based on our preliminary results, we conclude that the algorithms we developed using HR and BP trend data may provide a promising perspective toward reliable predictive alerts for hemodynamically unstable patients.

Predicting respiratory instability in the ICU.

Acute lung injury (ALI) and acute respiratory distress syndrome (ARDS) contribute to the morbidity and mortality of intensive care patients worldwide, and have large associated human and financial costs. We identified a reference data set of 624 mechanically-ventilated patients in the MIMIC-II intensive care database with and without low PaO(2)/FiO(2) ratios (termed respiratory instability), and developed prediction algorithms for distinguishing these patients prior to the critical event. In the end, we had four rule sets using mean airway pressure, plateau pressure, total respiratory rate and oxygen saturation (SpO(2)), where the specificity/sensitivity rates were either 80%/60% or 90%/50%.

Morphograms: exploiting correlation patterns to efficiently identify clinically significant events in intensive care units.

In this paper, we present a simple technique that utilizes the cross correlations between ECG signals and an arterial blood pressure (ABP) signal for the purpose of assessing signal quality and detecting artifacts in the ABP signal. The technique was tested using cases from a physician-annotated patient monitoring signal database from Beth Israel/Harvard-MIT University data bank. The results were encouraging: 45% of the manually annotated artifacts were correctly classified and 98% of the manually annotated true events were correctly classified.