Does Reinforcement Learning Improve Outcomes for Critically Ill Patients? A Systematic Review and Level-of-Readiness Assessment.

OBJECTIVE: Reinforcement learning (RL) is a machine learning technique uniquely effective at sequential decision-making, which makes it potentially relevant to ICU treatment challenges. We set out to systematically review, assess level-of-readiness and meta-analyze the effect of RL on outcomes for critically ill patients. DATA SOURCES: A systematic search was performed in PubMed, Embase.com, Clarivate Analytics/Web of Science Core Collection, Elsevier/SCOPUS and the Institute of Electrical and Electronics Engineers Xplore Digital Library from inception to March 25, 2022, with subsequent citation tracking. DATA EXTRACTION: Journal articles that used an RL technique in an ICU population and reported on patient health-related outcomes were included for full analysis. Conference papers were included for level-of-readiness assessment only. Descriptive statistics, characteristics of the models, outcome compared with clinician's policy and level-of-readiness were collected. RL-health risk of bias and applicability assessment was performed. DATA SYNTHESIS: A total of 1,033 articles were screened, of which 18 journal articles and 18 conference papers, were included. Thirty of those were prototyping or modeling articles and six were validation articles. All articles reported RL algorithms to outperform clinical decision-making by ICU professionals, but only in retrospective data. The modeling techniques for the state-space, action-space, reward function, RL model training, and evaluation varied widely. The risk of bias was high in all articles, mainly due to the evaluation procedure. CONCLUSION: In this first systematic review on the application of RL in intensive care medicine we found no studies that demonstrated improved patient outcomes from RL-based technologies. All studies reported that RL-agent policies outperformed clinician policies, but such assessments were all based on retrospective off-policy evaluation.

Predicting Readmission or Death After Discharge From the ICU: External Validation and Retraining of a Machine Learning Model.

OBJECTIVES: Many machine learning (ML) models have been developed for application in the ICU, but few models have been subjected to external validation. The performance of these models in new settings therefore remains unknown. The objective of this study was to assess the performance of an existing decision support tool based on a ML model predicting readmission or death within 7 days after ICU discharge before, during, and after retraining and recalibration. DESIGN: A gradient boosted ML model was developed and validated on electronic health record data from 2004 to 2021. We performed an independent validation of this model on electronic health record data from 2011 to 2019 from a different tertiary care center. SETTING: Two ICUs in tertiary care centers in The Netherlands. PATIENTS: Adult patients who were admitted to the ICU and stayed for longer than 12 hours. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: We assessed discrimination by area under the receiver operating characteristic curve (AUC) and calibration (slope and intercept). We retrained and recalibrated the original model and assessed performance via a temporal validation design. The final retrained model was cross-validated on all data from the new site. Readmission or death within 7 days after ICU discharge occurred in 577 of 10,052 ICU admissions (5.7%) at the new site. External validation revealed moderate discrimination with an AUC of 0.72 (95% CI 0.67-0.76). Retrained models showed improved discrimination with AUC 0.79 (95% CI 0.75-0.82) for the final validation model. Calibration was poor initially and good after recalibration via isotonic regression. CONCLUSIONS: In this era of expanding availability of ML models, external validation and retraining are key steps to consider before applying ML models to new settings. Clinicians and decision-makers should take this into account when considering applying new ML models to their local settings.

Oxygenation thresholds for invasive ventilation in hypoxemic respiratory failure: a target trial emulation in two cohorts.

BACKGROUND: The optimal thresholds for the initiation of invasive ventilation in patients with hypoxemic respiratory failure are unknown. Using the saturation-to-inspired oxygen ratio (SF), we compared lower versus higher hypoxemia severity thresholds for initiating invasive ventilation. METHODS: This target trial emulation included patients from the Medical Information Mart for Intensive Care (MIMIC-IV, 2008-2019) and the Amsterdam University Medical Centers (AmsterdamUMCdb, 2003-2016) databases admitted to intensive care and receiving inspired oxygen fraction ≥ 0.4 via non-rebreather mask, noninvasive ventilation, or high-flow nasal cannula. We compared the effect of using invasive ventilation initiation thresholds of SF < 110, < 98, and < 88 on 28-day mortality. MIMIC-IV was used for the primary analysis and AmsterdamUMCdb for the secondary analysis. We obtained posterior means and 95% credible intervals (CrI) with nonparametric Bayesian G-computation. RESULTS: We studied 3,357 patients in the primary analysis. For invasive ventilation initiation thresholds SF < 110, SF < 98, and SF < 88, the predicted 28-day probabilities of invasive ventilation were 72%, 47%, and 19%. Predicted 28-day mortality was lowest with threshold SF < 110 (22.2%, CrI 19.2 to 25.0), compared to SF < 98 (absolute risk increase 1.6%, CrI 0.6 to 2.6) or SF < 88 (absolute risk increase 3.5%, CrI 1.4 to 5.4). In the secondary analysis (1,279 patients), the predicted 28-day probability of invasive ventilation was 50% for initiation threshold SF < 110, 28% for SF < 98, and 19% for SF < 88. In contrast with the primary analysis, predicted mortality was highest with threshold SF < 110 (14.6%, CrI 7.7 to 22.3), compared to SF < 98 (absolute risk decrease 0.5%, CrI 0.0 to 0.9) or SF < 88 (absolute risk decrease 1.9%, CrI 0.9 to 2.8). CONCLUSION: Initiating invasive ventilation at lower hypoxemia severity will increase the rate of invasive ventilation, but this can either increase or decrease the expected mortality, with the direction of effect likely depending on baseline mortality risk and clinical context.

Evolution of Clinical Phenotypes of COVID-19 Patients During Intensive Care Treatment: An Unsupervised Machine Learning Analysis.

BACKGROUND: Identification of clinical phenotypes in critically ill COVID-19 patients could improve understanding of the disease heterogeneity and enable prognostic and predictive enrichment. However, previous attempts did not take into account temporal dynamics with high granularity. By including the dimension of time, we aim to gain further insights into the heterogeneity of COVID-19. METHODS: We used granular data from 3202 adult COVID patients in the Dutch Data Warehouse that were admitted to one of 25 Dutch ICUs between February 2020 and March 2021. Parameters including demographics, clinical observations, medications, laboratory values, vital signs, and data from life support devices were selected. Twenty-one datasets were created that each covered 24 h of ICU data for each day of ICU treatment. Clinical phenotypes in each dataset were identified by performing cluster analyses. Both evolution of the clinical phenotypes over time and patient allocation to these clusters over time were tracked. RESULTS: The final patient cohort consisted of 2438 COVID-19 patients with a ICU mortality outcome. Forty-one parameters were chosen for cluster analysis. On admission, both a mild and a severe clinical phenotype were found. After day 4, the severe phenotype split into an intermediate and a severe phenotype for 11 consecutive days. Heterogeneity between phenotypes appears to be driven by inflammation and dead space ventilation. During the 21-day period, only 8.2% and 4.6% of patients in the initial mild and severe clusters remained assigned to the same phenotype respectively. The clinical phenotype half-life was between 5 and 6 days for the mild and severe phenotypes, and about 3 days for the medium severe phenotype. CONCLUSIONS: Patients typically do not remain in the same cluster throughout intensive care treatment. This may have important implications for prognostic or predictive enrichment. Prominent dissimilarities between clinical phenotypes are predominantly driven by inflammation and dead space ventilation.

Systematic Review and Comparison of Publicly Available ICU Data Sets-A Decision Guide for Clinicians and Data Scientists.

OBJECTIVE: As data science and artificial intelligence continue to rapidly gain traction, the publication of freely available ICU datasets has become invaluable to propel data-driven clinical research. In this guide for clinicians and researchers, we aim to: 1) systematically search and identify all publicly available adult clinical ICU datasets, 2) compare their characteristics, data quality, and richness and critically appraise their strengths and weaknesses, and 3) provide researchers with suggestions, which datasets are appropriate for answering their clinical question. DATA SOURCES: A systematic search was performed in Pubmed, ArXiv, MedRxiv, and BioRxiv. STUDY SELECTION: We selected all studies that reported on publicly available adult patient-level intensive care datasets. DATA EXTRACTION: A total of four publicly available, adult, critical care, patient-level databases were included (Amsterdam University Medical Center data base [AmsterdamUMCdb], eICU Collaborative Research Database eICU CRD], High time-resolution intensive care unit dataset [HiRID], and Medical Information Mart for Intensive Care-IV). Databases were compared using a priori defined categories, including demographics, patient characteristics, and data richness. The study protocol and search strategy were prospectively registered. DATA SYNTHESIS: Four ICU databases fulfilled all criteria for inclusion and were queried using SQL (PostgreSQL version 12; PostgreSQL Global Development Group) and analyzed using R (R Foundation for Statistical Computing, Vienna, Austria). The number of unique patient admissions varied between 23,106 (AmsterdamUMCdb) and 200,859 (eICU-CRD). Frequency of laboratory values and vital signs was highest in HiRID, for example, 5.2 (±3.4) lactate values per day and 29.7 (±10.2) systolic blood pressure values per hour. Treatment intensity varied with vasopressor and ventilatory support in 69.0% and 83.0% of patients in AmsterdamUMCdb versus 12.0% and 21.0% in eICU-CRD, respectively. ICU mortality ranged from 5.5% in eICU-CRD to 9.9% in AmsterdamUMCdb. CONCLUSIONS: We identified four publicly available adult clinical ICU datasets. Sample size, severity of illness, treatment intensity, and frequency of reported parameters differ markedly between the databases. This should guide clinicians and researchers which databases to best answer their clinical questions.

The Potential Cost-Effectiveness of a Machine Learning Tool That Can Prevent Untimely Intensive Care Unit Discharge.

OBJECTIVES: The machine learning prediction model Pacmed Critical (PC), currently under development, may guide intensivists in their decision-making process on the most appropriate time to discharge a patient from the intensive care unit (ICU). Given the financial pressure on healthcare budgets, this study assessed whether PC has the potential to be cost-effective compared with standard care, without the use of PC, for Dutch patients in the ICU from a societal perspective. METHODS: A 1-year, 7-state Markov model reflecting the ICU care pathway and incorporating the PC decision tool was developed. A hypothetical cohort of 1000 adult Dutch patients admitted in the ICU was entered in the model. We used the literature, expert opinion, and data from Amsterdam University Medical Center for model parameters. The uncertainty surrounding the incremental cost-effectiveness ratio was assessed using deterministic and probabilistic sensitivity analyses and scenario analyses. RESULTS: PC was a cost-effective strategy with an incremental cost-effectiveness ratio of €18 507 per quality-adjusted life-year. PC remained cost-effective over standard care in multiple scenarios and sensitivity analyses. The likelihood that PC will be cost-effective was 71% at a willingness-to-pay threshold of €30 000 per quality-adjusted life-year. The key driver of the results was the parameter "reduction in ICU length of stay." CONCLUSIONS: We showed that PC has the potential to be cost-effective for Dutch ICUs in a time horizon of 1 year. This study is one of the first cost-effectiveness analyses of a machine learning device. Further research is needed to validate the effectiveness of PC, thereby focusing on the key parameter "reduction in ICU length of stay" and potential spill-over effects.

A novel Vascular Leak Index identifies sepsis patients with a higher risk for in-hospital death and fluid accumulation.

PURPOSE: Sepsis is a leading cause of morbidity and mortality worldwide and is characterized by vascular leak. Treatment for sepsis, specifically intravenous fluids, may worsen deterioration in the context of vascular leak. We therefore sought to quantify vascular leak in sepsis patients to guide fluid resuscitation. METHODS: We performed a retrospective cohort study of sepsis patients in four ICU databases in North America, Europe, and Asia. We developed an intuitive vascular leak index (VLI) and explored the relationship between VLI and in-hospital death and fluid balance using generalized additive models (GAM). RESULTS: Using a GAM, we found that increased VLI is associated with an increased risk of in-hospital death. Patients with a VLI in the highest quartile (Q4), across the four datasets, had a 1.61-2.31 times increased odds of dying in the hospital compared to patients with a VLI in the lowest quartile (Q1). VLI Q2 and Q3 were also associated with increased odds of dying. The relationship between VLI, treated as a continuous variable, and in-hospital death and fluid balance was statistically significant in the three datasets with large sample sizes. Specifically, we observed that as VLI increased, there was increase in the risk for in-hospital death and 36-84 h fluid balance. CONCLUSIONS: Our VLI identifies groups of patients who may be at higher risk for in-hospital death or for fluid accumulation. This relationship persisted in models developed to control for severity of illness and chronic comorbidities.

Right dose, right now: bedside, real-time, data-driven, and personalised antibiotic dosing in critically ill patients with sepsis or septic shock-a two-centre randomised clinical trial.

BACKGROUND: Adequate antibiotic dosing may improve outcomes in critically ill patients but is challenging due to altered and variable pharmacokinetics. To address this challenge, AutoKinetics was developed, a decision support system for bedside, real-time, data-driven and personalised antibiotic dosing. This study evaluates the feasibility, safety and efficacy of its clinical implementation. METHODS: In this two-centre randomised clinical trial, critically ill patients with sepsis or septic shock were randomised to AutoKinetics dosing or standard dosing for four antibiotics: vancomycin, ciprofloxacin, meropenem, and ceftriaxone. Adult patients with a confirmed or suspected infection and either lactate > 2 mmol/L or vasopressor requirement were eligible for inclusion. The primary outcome was pharmacokinetic target attainment in the first 24 h after randomisation. Clinical endpoints included mortality, ICU length of stay and incidence of acute kidney injury. RESULTS: After inclusion of 252 patients, the study was stopped early due to the COVID-19 pandemic. In the ciprofloxacin intervention group, the primary outcome was obtained in 69% compared to 3% in the control group (OR 62.5, CI 11.4-1173.78, p < 0.001). Furthermore, target attainment was faster (26 h, CI 18-42 h, p < 0.001) and better (65% increase, CI 49-84%, p < 0.001). For the other antibiotics, AutoKinetics dosing did not improve target attainment. Clinical endpoints were not significantly different. Importantly, higher dosing did not lead to increased mortality or renal failure. CONCLUSIONS: In critically ill patients, personalised dosing was feasible, safe and significantly improved target attainment for ciprofloxacin. TRIAL REGISTRATION: The trial was prospectively registered at Netherlands Trial Register (NTR), NL6501/NTR6689 on 25 August 2017 and at the European Clinical Trials Database (EudraCT), 2017-002478-37 on 6 November 2017.

Rapid Evaluation of Coronavirus Illness Severity (RECOILS) in intensive care: Development and validation of a prognostic tool for in-hospital mortality.

BACKGROUND: The prediction of in-hospital mortality for ICU patients with COVID-19 is fundamental to treatment and resource allocation. The main purpose was to develop an easily implemented score for such prediction. METHODS: This was an observational, multicenter, development, and validation study on a national critical care dataset of COVID-19 patients. A systematic literature review was performed to determine variables possibly important for COVID-19 mortality prediction. Using a logistic multivariable model with a LASSO penalty, we developed the Rapid Evaluation of Coronavirus Illness Severity (RECOILS) score and compared its performance against published scores. RESULTS: Our development (validation) cohort consisted of 1480 (937) adult patients from 14 (11) Dutch ICUs admitted between March 2020 and April 2021. Median age was 65 (65) years, 31% (26%) died in hospital, 74% (72%) were males, average length of ICU stay was 7.83 (10.25) days and average length of hospital stay was 15.90 (19.92) days. Age, platelets, PaO2/FiO2 ratio, pH, blood urea nitrogen, temperature, PaCO2, Glasgow Coma Scale (GCS) score measured within +/-24 h of ICU admission were used to develop the score. The AUROC of RECOILS score was 0.75 (CI 0.71-0.78) which was higher than that of any previously reported predictive scores (0.68 [CI 0.64-0.71], 0.61 [CI 0.58-0.66], 0.67 [CI 0.63-0.70], 0.70 [CI 0.67-0.74] for ISARIC 4C Mortality Score, SOFA, SAPS-III, and age, respectively). CONCLUSIONS: Using a large dataset from multiple Dutch ICUs, we developed a predictive score for mortality of COVID-19 patients admitted to ICU, which outperformed other predictive scores reported so far.

Sharing ICU Patient Data Responsibly Under the Society of Critical Care Medicine/European Society of Intensive Care Medicine Joint Data Science Collaboration: The Amsterdam University Medical Centers Database (AmsterdamUMCdb) Example.

OBJECTIVES: Critical care medicine is a natural environment for machine learning approaches to improve outcomes for critically ill patients as admissions to ICUs generate vast amounts of data. However, technical, legal, ethical, and privacy concerns have so far limited the critical care medicine community from making these data readily available. The Society of Critical Care Medicine and the European Society of Intensive Care Medicine have identified ICU patient data sharing as one of the priorities under their Joint Data Science Collaboration. To encourage ICUs worldwide to share their patient data responsibly, we now describe the development and release of Amsterdam University Medical Centers Database (AmsterdamUMCdb), the first freely available critical care database in full compliance with privacy laws from both the United States and Europe, as an example of the feasibility of sharing complex critical care data. SETTING: University hospital ICU. SUBJECTS: Data from ICU patients admitted between 2003 and 2016. INTERVENTIONS: We used a risk-based deidentification strategy to maintain data utility while preserving privacy. In addition, we implemented contractual and governance processes, and a communication strategy. Patient organizations, supporting hospitals, and experts on ethics and privacy audited these processes and the database. MEASUREMENTS AND MAIN RESULTS: AmsterdamUMCdb contains approximately 1 billion clinical data points from 23,106 admissions of 20,109 patients. The privacy audit concluded that reidentification is not reasonably likely, and AmsterdamUMCdb can therefore be considered as anonymous information, both in the context of the U.S. Health Insurance Portability and Accountability Act and the European General Data Protection Regulation. The ethics audit concluded that responsible data sharing imposes minimal burden, whereas the potential benefit is tremendous. CONCLUSIONS: Technical, legal, ethical, and privacy challenges related to responsible data sharing can be addressed using a multidisciplinary approach. A risk-based deidentification strategy, that complies with both U.S. and European privacy regulations, should be the preferred approach to releasing ICU patient data. This supports the shared Society of Critical Care Medicine and European Society of Intensive Care Medicine vision to improve critical care outcomes through scientific inquiry of vast and combined ICU datasets.

External Evaluation of Population Pharmacokinetic Models of Vancomycin in Large Cohorts of Intensive Care Unit Patients.

Dosing of vancomycin is often guided by therapeutic drug monitoring and population pharmacokinetic models in the intensive care unit (ICU). The validity of these models is crucial, as ICU patients have marked pharmacokinetic variability. Therefore, we set out to evaluate the predictive performance of published population pharmacokinetic models of vancomycin in ICU patients. The PubMed database was used to search for population pharmacokinetic models of vancomycin in adult ICU patients. The identified models were evaluated in two independent data sets which were collected from two large hospitals in the Netherlands (Amsterdam UMC, Location VUmc, and OLVG Oost). We also tested a one-compartment model with fixed values for clearance and volume of distribution, in which a clinical standard dosage regimen (SDR) was mimicked to assess its predictive performance. Prediction error was calculated to assess the predictive performance of the models. Six models plus the SDR model were evaluated. The model of Roberts et al. (J. A. Roberts, F. S. Taccone, A. A. Udy, J.-L. Vincent, F. Jacobs, and J. Lipman, Antimicrob Agents Chemother 55:2704-2709, 2011, https://doi.org/10.1128/AAC.01708-10) performed satisfactorily, with mean and median values of prediction error of 5.1% and -7.5%, respectively, for Amsterdam UMC, Location VUmc, patients, and -12.6% and -17.2% respectively, for OLVG Oost patients. The other models, including the SDR model, yielded high mean values (-49.7% to 87.7%) and median values (-56.1% to 66.1%) for both populations. In conclusion, only the model of Roberts et al. was able to validly predict the concentrations of vancomycin for our data, whereas other models and standard dosing were largely inadequate. Extensive evaluation should precede the adoption of any model in clinical practice for ICU patients.

Amino Acid Loss during Continuous Venovenous Hemofiltration in Critically Ill Patients.

BACKGROUND/OBJECTIVES: During continuous venovenous hemofiltration (CVVH), there is unwanted loss of amino acids (AA) in the ultrafiltrate (UF). Solutes may also be removed by adsorption to the filter membrane. The aim was to quantify the total loss of AA via the CVVH circuit using a high-flux polysulfone membrane and to differentiate between the loss by ultrafiltration and adsorption. METHODS: Prospective observational study in ten critically ill patients, receiving predilution CVVH with a new filter, blood flow 180 mL/min, and predilution flow 2,400 mL/h. Arterial blood, postfilter blood, and UF samples were taken at baseline, and 1, 8, and 24-h after CVVH initiation, to determine AA concentrations and hematocrit. Mass transfer calculations were used to determine AA loss in the filter and by UF, and the difference between these 2. RESULTS: The median AA loss in the filter was 10.4 g/day, the median AA loss by UF was 13.4 g/day, and the median difference was -2.9 g/day (IQR -5.9 to -1.4 g/day). For the individual AA, the difference ranged from -1 g/day to +0.4 g/day, suggesting that some AA were consumed or adsorbed and others were generated. AA losses did not significantly change over the 24-h study period. CONCLUSION: During CVVH with a modern polysulfone membrane, the estimated AA loss was 13.4 g/day, which corresponds to a loss of about 11.2 g of protein per day. Adsorption did not play a major role. However, individual AA behaved differently, suggesting complex interactions and processes at the filter membrane or peripheral AA production.

Application of the Sepsis-3 criteria to describe sepsis epidemiology in the Amsterdam UMCdb intensive care dataset.

INTRODUCTION: Sepsis is a major cause of morbidity and mortality worldwide. In the updated, 2016 Sepsis-3 criteria, sepsis is defined as life-threatening organ dysfunction caused by a dysregulated host response to infection, where organ dysfunction can be represented by an increase in the Sequential Organ Failure Assessment (SOFA) score of 2 points or more. We sought to apply the Sepsis-3 criteria to characterise the septic cohort in the Amsterdam University Medical Centres database (Amsterdam UMCdb). METHODS: We examined adult intensive care unit (ICU) admissions in the Amsterdam UMCdb, which contains de-identified data for patients admitted to a mixed surgical-medical ICU at a tertiary academic medical centre in the Netherlands. We operationalised the Sepsis-3 criteria, defining organ dysfunction as an increase in the SOFA score of 2 points or more, while infection was defined as a new course of antibiotics or an escalation in antibiotic therapy, with at least one antibiotic given intravenously. Patients with sepsis were determined to be in septic shock if they additionally required the use of vasopressors and had a lactate level >2 mmol/L. RESULTS: We identified 18,221 ICU admissions from 16,408 patients in our cohort. There were 6,312 unique sepsis episodes, of which 30.2% met the criteria for septic shock. A total of 4,911/6,312 sepsis (77.8%) episodes occurred on ICU admission. Forty-seven percent of emergency medical admissions and 36.7% of emergency surgical admissions were for sepsis. Overall, there was a 12.5% ICU mortality rate; patients with septic shock had a higher ICU mortality rate (38.4%) than those without shock (11.4%). CONCLUSIONS: We successfully operationalised the Sepsis-3 criteria to the Amsterdam UMCdb, allowing the characterization and comparison of sepsis epidemiology across different centres.

Assessing SOFA score trajectories in sepsis using machine learning: A pragmatic approach to improve the accuracy of mortality prediction.

INTRODUCTION: An increasing amount of longitudinal health data is available on critically ill septic patients in the age of digital medicine, including daily sequential organ failure assessment (SOFA) score measurements. Thus, the assessment in sepsis focuses increasingly on the evaluation of the individual disease's trajectory. Machine learning (ML) algorithms may provide a promising approach here to improve the evaluation of daily SOFA score dynamics. We tested whether ML algorithms can outperform the conventional ΔSOFA score regarding the accuracy of 30-day mortality prediction. METHODS: We used the multicentric SepsisDataNet.NRW study cohort that prospectively enrolled 252 sepsis patients between 03/2018 and 09/2019 for training ML algorithms, i.e. support vector machine (SVM) with polynomial kernel and artificial neural network (aNN). We used the Amsterdam UMC database covering 1,790 sepsis patients for external and independent validation. RESULTS: Both SVM (AUC 0.84; 95% CI: 0.71-0.96) and aNN (AUC 0.82; 95% CI: 0.69-0.95) assessing the SOFA scores of the first seven days led to a more accurate prognosis of 30-day mortality compared to the ΔSOFA score between day 1 and 7 (AUC 0.73; 95% CI: 0.65-0.80; p = 0.02 and p = 0.05, respectively). These differences were even more prominent the shorter the time interval considered. Using the SOFA scores of day 1 to 3 SVM (AUC 0.82; 95% CI: 0.68 0.95) and aNN (AUC 0.80; 95% CI: 0.660.93) led to a more accurate prognosis of 30-day mortality compared to the ΔSOFA score (AUC 0.66; 95% CI: 0.58-0.74; p < 0.01 and p < 0.01, respectively). Strikingly, all these findings could be confirmed in the independent external validation cohort. CONCLUSIONS: The ML-based algorithms using daily SOFA scores markedly improved the accuracy of mortality compared to the conventional ΔSOFA score. Therefore, this approach could provide a promising and automated approach to assess the individual disease trajectory in sepsis. These findings reflect the potential of incorporating ML algorithms as robust and generalizable support tools on intensive care units.

Development of a Reinforcement Learning Algorithm to Optimize Corticosteroid Therapy in Critically Ill Patients with Sepsis.

BACKGROUND: The optimal indication, dose, and timing of corticosteroids in sepsis is controversial. Here, we used reinforcement learning to derive the optimal steroid policy in septic patients based on data on 3051 ICU admissions from the AmsterdamUMCdb intensive care database. METHODS: We identified septic patients according to the 2016 consensus definition. An actor-critic RL algorithm using ICU mortality as a reward signal was developed to determine the optimal treatment policy from time-series data on 277 clinical parameters. We performed off-policy evaluation and testing in independent subsets to assess the algorithm's performance. RESULTS: Agreement between the RL agent's policy and the actual documented treatment reached 59%. Our RL agent's treatment policy was more restrictive compared to the actual clinician behavior: our algorithm suggested withholding corticosteroids in 62% of the patient states, versus 52% according to the physicians' policy. The 95% lower bound of the expected reward was higher for the RL agent than clinicians' historical decisions. ICU mortality after concordant action in the testing dataset was lower both when corticosteroids had been withheld and when corticosteroids had been prescribed by the virtual agent. The most relevant variables were vital parameters and laboratory values, such as blood pressure, heart rate, leucocyte count, and glycemia. CONCLUSIONS: Individualized use of corticosteroids in sepsis may result in a mortality benefit, but optimal treatment policy may be more restrictive than the routine clinical practice. Whilst external validation is needed, our study motivates a 'precision-medicine' approach to future prospective controlled trials and practice.

Determining and assessing characteristics of data element names impacting the performance of annotation using Usagi.

INTRODUCTION: Hospitals generate large amounts of data and this data is generally modeled and labeled in a proprietary way, hampering its exchange and integration. Manually annotating data element names to internationally standardized data element identifiers is a time-consuming effort. Tools can support performing this task automatically. This study aimed to determine what factors influence the quality of automatic annotations. METHODS: Data element names were used from the Dutch COVID-19 ICU Data Warehouse containing data on intensive care patients with COVID-19 from 25 hospitals in the Netherlands. In this data warehouse, the data had been merged using a proprietary terminology system while also storing the original hospital labels (synonymous names). Usagi, an OHDSI annotation tool, was used to perform the annotation for the data. A gold standard was used to determine if Usagi made correct annotations. Logistic regression was used to determine if the number of characters, number of words, match score (Usagi's certainty) and hospital label origin influenced Usagi's performance to annotate correctly. RESULTS: Usagi automatically annotated 30.5% of the data element names correctly and 5.5% of the synonymous names. The match score is the best predictor for Usagi finding the correct annotation. It was determined that the AUC of data element names was 0.651 and 0.752 for the synonymous names respectively. The AUC for the individual hospital label origins varied between 0.460 to 0.905. DISCUSSION: The results show that Usagi performed better to annotate the data element names than the synonymous names. The hospital origin in the synonymous names dataset was associated with the amount of correctly annotated concepts. Hospitals that performed better had shorter synonymous names and fewer words. Using shorter data element names or synonymous names should be considered to optimize the automatic annotating process. Overall, the performance of Usagi is too poor to completely rely on for automatic annotation.

Clinically Interpretable Machine Learning Models for Early Prediction of Mortality in Older Patients with Multiple Organ Dysfunction Syndrome: An International Multicenter Retrospective Study.

BACKGROUND: Multiple organ dysfunction syndrome (MODS) is associated with a high risk of mortality among older patients. Current severity scores are limited in their ability to assist clinicians with triage and management decisions. We aim to develop mortality prediction models for older patients with MODS admitted to the ICU. METHODS: The study analyzed older patients from 197 hospitals in the United States and 1 hospital in the Netherlands. The cohort was divided into the young-old (65-80 years) and old-old (≥80 years), which were separately used to develop and evaluate models including internal, external, and temporal validation. Demographic characteristics, comorbidities, vital signs, laboratory measurements, and treatments were used as predictors. We used the XGBoost algorithm to train models, and the SHapley Additive exPlanations (SHAP) method to interpret predictions. RESULTS: Thirty-four thousand four hundred and ninety-seven young-old (11.3% mortality) and 21 330 old-old (15.7% mortality) patients were analyzed. Discrimination AUROC of internal validation models in 9 046 U.S. patients was as follows: 0.87 and 0.82, respectively; discrimination of external validation models in 1 905 EUR patients was as follows: 0.86 and 0.85, respectively; and discrimination of temporal validation models in 8 690 U.S. patients: 0.85 and 0.78, respectively. These models outperformed standard clinical scores like Sequential Organ Failure Assessment and Acute Physiology Score III. The Glasgow Coma Scale, Charlson Comorbidity Index, and Code Status emerged as top predictors of mortality. CONCLUSIONS: Our models integrate data spanning physiologic and geriatric-relevant variables that outperform existing scores used in older adults with MODS, which represents a proof of concept of how machine learning can streamline data analysis for busy ICU clinicians to potentially optimize prognostication and decision making.

Illness severity assessment of older adults in critical illness using machine learning (ELDER-ICU): an international multicentre study with subgroup bias evaluation.

BACKGROUND: Comorbidity, frailty, and decreased cognitive function lead to a higher risk of death in elderly patients (more than 65 years of age) during acute medical events. Early and accurate illness severity assessment can support appropriate decision making for clinicians caring for these patients. We aimed to develop ELDER-ICU, a machine learning model to assess the illness severity of older adults admitted to the intensive care unit (ICU) with cohort-specific calibration and evaluation for potential model bias. METHODS: In this retrospective, international multicentre study, the ELDER-ICU model was developed using data from 14 US hospitals, and validated in 171 hospitals from the USA and Netherlands. Data were extracted from the Medical Information Mart for Intensive Care database, electronic ICU Collaborative Research Database, and Amsterdam University Medical Centers Database. We used six categories of data as predictors, including demographics and comorbidities, physical frailty, laboratory tests, vital signs, treatments, and urine output. Patient data from the first day of ICU stay were used to predict in-hospital mortality. We used the eXtreme Gradient Boosting algorithm (XGBoost) to develop models and the SHapley Additive exPlanations method to explain model prediction. The trained model was calibrated before internal, external, and temporal validation. The final XGBoost model was compared against three other machine learning algorithms and five clinical scores. We performed subgroup analysis based on age, sex, and race. We assessed the discrimination and calibration of models using the area under receiver operating characteristic (AUROC) and standardised mortality ratio (SMR) with 95% CIs. FINDINGS: Using the development dataset (n=50 366) and predictive model building process, the XGBoost algorithm performed the best in all types of validations compared with other machine learning algorithms and clinical scores (internal validation with 5037 patients from 14 US hospitals, AUROC=0·866 [95% CI 0·851-0·880]; external validation in the US population with 20 541 patients from 169 hospitals, AUROC=0·838 [0·829-0·847]; external validation in European population with 2411 patients from one hospital, AUROC=0·833 [0·812-0·853]; temporal validation with 4311 patients from one hospital, AUROC=0·884 [0·869-0·897]). In the external validation set (US population), the median AUROCs of bias evaluations covering eight subgroups were above 0·81, and the overall SMR was 0·99 (0·96-1·03). The top ten risk predictors were the minimum Glasgow Coma Scale score, total urine output, average respiratory rate, mechanical ventilation use, best state of activity, Charlson Comorbidity Index score, geriatric nutritional risk index, code status, age, and maximum blood urea nitrogen. A simplified model containing only the top 20 features (ELDER-ICU-20) had similar predictive performance to the full model. INTERPRETATION: The ELDER-ICU model reliably predicts the risk of in-hospital mortality using routinely collected clinical features. The predictions could inform clinicians about patients who are at elevated risk of deterioration. Prospective validation of this model in clinical practice and a process for continuous performance monitoring and model recalibration are needed. FUNDING: National Institutes of Health, National Natural Science Foundation of China, National Special Health Science Program, Health Science and Technology Plan of Zhejiang Province, Fundamental Research Funds for the Central Universities, Drug Clinical Evaluate Research of Chinese Pharmaceutical Association, and National Key R&D Program of China.

Intensive Care Unit Physicians' Perspectives on Artificial Intelligence-Based Clinical Decision Support Tools: Preimplementation Survey Study.

BACKGROUND: Artificial intelligence-based clinical decision support (AI-CDS) tools have great potential to benefit intensive care unit (ICU) patients and physicians. There is a gap between the development and implementation of these tools. OBJECTIVE: We aimed to investigate physicians' perspectives and their current decision-making behavior before implementing a discharge AI-CDS tool for predicting readmission and mortality risk after ICU discharge. METHODS: We conducted a survey of physicians involved in decision-making on discharge of patients at two Dutch academic ICUs between July and November 2021. Questions were divided into four domains: (1) physicians' current decision-making behavior with respect to discharging ICU patients, (2) perspectives on the use of AI-CDS tools in general, (3) willingness to incorporate a discharge AI-CDS tool into daily clinical practice, and (4) preferences for using a discharge AI-CDS tool in daily workflows. RESULTS: Most of the 64 respondents (of 93 contacted, 69%) were familiar with AI (62/64, 97%) and had positive expectations of AI, with 55 of 64 (86%) believing that AI could support them in their work as a physician. The respondents disagreed on whether the decision to discharge a patient was complex (23/64, 36% agreed and 22/64, 34% disagreed); nonetheless, most (59/64, 92%) agreed that a discharge AI-CDS tool could be of value. Significant differences were observed between physicians from the 2 academic sites, which may be related to different levels of involvement in the development of the discharge AI-CDS tool. CONCLUSIONS: ICU physicians showed a favorable attitude toward the integration of AI-CDS tools into the ICU setting in general, and in particular toward a tool to predict a patient's risk of readmission and mortality within 7 days after discharge. The findings of this questionnaire will be used to improve the implementation process and training of end users.

Diagnosing acute kidney injury ahead of time in critically ill septic patients using kinetic estimated glomerular filtration rate.

INTRODUCTION: Accurate and actionable diagnosis of Acute Kidney Injury (AKI) ahead of time is important to prevent or mitigate renal insufficiency. The purpose of this study was to evaluate the performance of Kinetic estimated Glomerular Filtration Rate (KeGFR) in timely predicting AKI in critically ill septic patients. METHODS: We conducted a retrospective analysis on septic ICU patients who developed AKI in AmsterdamUMCdb, the first freely available European ICU database. The reference standard for AKI was the Kidney Disease: Improving Global Outcomes (KDIGO) classification based on serum creatinine and urine output (UO). Prediction of AKI was based on stages defined by KeGFR and UO. Classifications were compared by length of ICU stay (LOS), need for renal replacement therapy and 28-day mortality. Predictive performance and time between prediction and diagnosis were calculated. RESULTS: Of 2492 patients in the cohort, 1560 (62.0%) were diagnosed with AKI by KDIGO and 1706 (68.5%) by KeGFR criteria. Disease stages had agreement of kappa = 0.77, with KeGFR sensitivity 93.2%, specificity 73.0% and accuracy 85.7%. Median time to recognition of AKI Stage 1 was 13.2 h faster for KeGFR, and 7.5 h and 5.0 h for Stages 2 and 3. Outcomes revealed a slight difference in LOS and 28-day mortality for Stage 1. CONCLUSIONS: Predictive performance of KeGFR combined with UO criteria for diagnosing AKI is excellent. Compared to KDIGO, deterioration of renal function was identified earlier, most prominently for lower stages of AKI. This may shift the actionable window for preventing and mitigating renal insufficiency.