Search

Background: Reinforcement learning (RL) holds great promise for intensive care medicine given the abundant availability of data and frequent sequential decision-making. But despite the emergence of promising algorithms, RL driven bedside clinical decision support is still far from reality. Major challenges include trust and safety. To help address these issues, we introduce cross off-policy evaluation and policy restriction and show how detailed policy analysis may increase clinical interpretability. As an example, we apply these in the setting of RL to optimise ventilator settings in intubated covid-19 patients. Methods: With data from the Dutch ICU Data Warehouse and using an exhaustive hyperparameter grid search, we identified an optimal set of Dueling Double-Deep Q Network RL models. The state space comprised ventilator, medication, and clinical data. The action space focused on positive end-expiratory pressure (peep) and fraction of inspired oxygen (FiO2) concentration. We used gas exchange indices as interim rewards, and mortality and state duration as final rewards. We designed a novel evaluation method called cross off-policy evaluation (OPE) to assess the efficacy of models under varying weightings between the interim and terminal reward components. In addition, we implemented policy restriction to prevent potentially hazardous model actions. We introduce delta-Q to compare physician versus policy action quality and in-depth policy inspection using visualisations. Results: We created trajectories for 1118 intensive care unit (ICU) admissions and trained 69,120 models using 8 model architectures with 128 hyperparameter combinations. For each model, policy restrictions were applied. In the first evaluation step, 17,182/138,240 policies had good performance, but cross-OPE revealed suboptimal performance for 44% of those by varying the reward function used for evaluation. Clinical policy inspection facilitated assessment of action decisions for individual patients, including identification of action space regions that may benefit most from optimisation. Conclusion: Cross-OPE can serve as a robust evaluation framework for safe RL model implementation by identifying policies with good generalisability. Policy restriction helps prevent potentially unsafe model recommendations. Finally, the novel delta-Q metric can be used to operationalise RL models in clinical practice. Our findings offer a promising pathway towards application of RL in intensive care medicine and beyond. [ABSTRACT FROM AUTHOR]

Author(s): Lucas M. Fleuren [sup.1] [sup.2], Patrick Thoral [sup.1], Duncan Shillan [sup.3], Ari Ercole [sup.4] [sup.5], Paul W. G. Elbers [sup.1] [sup.5], Mark Hoogendoorn, Ben Gibbison, Thomas L. T. Klausch, [...]

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. [ABSTRACT FROM AUTHOR]

In population pharmacokinetic (PK) models, interindividual variability is explained by implementation of covariates in the model. The widely used forward stepwise selection method is sensitive to bias, which may lead to an incorrect inclusion of covariates. Alternatives, such as the full fixed effects model, reduce this bias but are dependent on the chosen implementation of each covariate. As the correct functional forms are unknown, this may still lead to an inaccurate selection of covariates. Machine learning (ML) techniques can potentially be used to learn the optimal functional forms for implementing covariates directly from data. A recent study suggested that using ML resulted in an improved selection of influential covariates. However, how do we select the appropriate functional form for including these covariates? In this work, we use SHapley Additive exPlanations (SHAP) to infer the relationship between covariates and PK parameters from ML models. As a case‐study, we use data from 119 patients with hemophilia A receiving clotting factor VIII concentrate peri‐operatively. We fit both a random forest and a XGBoost model to predict empirical Bayes estimated clearance and central volume from a base nonlinear mixed effects model. Next, we show that SHAP reveals covariate relationships which match previous findings. In addition, we can reveal subtle effects arising from combinations of covariates difficult to obtain using other methods of covariate analysis. We conclude that the proposed method can be used to extend ML‐based covariate selection, and holds potential as a complete full model alternative to classical covariate analyses. [ABSTRACT FROM AUTHOR]

Energy disaggregation, a.k.a. Non-Intrusive Load Monitoring, aims to separate the energy consumption of individual appliances from the readings of a mains power meter measuring the total energy consumption of, e.g., a whole house. Energy consumption of individual appliances can be useful in many applications, e.g., providing appliance-level feedback to the end users to help them understand their energy consumption and ultimately save energy. Recently, with the availability of large-scale energy consumption datasets, various neural network models such as convolutional neural networks and recurrent neural networks have been investigated to solve the energy disaggregation problem. Neural network models can learn complex patterns from large amounts of data and have been shown to outperform the traditional machine learning methods such as variants of hiddenMarkov models. However, current neural network methods for energy disaggregation are either computational expensive or are not capable of handling long-term dependencies. In this article, we investigate the application of the recently developed WaveNet models for the task of energy disaggregation. Based on a real-world energy dataset collected from 20 households over 2 years, we show thatWaveNet models outperforms the state-of-the-art deep learning methods proposed in the literature for energy disaggregation in terms of both error measures and computational cost. On the basis of energy disaggregation, we then investigate the performance of two deep-learning based frameworks for the task of on/off detection which aims at estimating whether an appliance is in operation or not. The first framework obtains the on/off states of an appliance by binarising the predictions of a regression model trained for energy disaggregation, while the second framework obtains the on/off states of an appliance by directly training a binary classifier with binarised energy readings of the appliance serving as the target values. Based on the same dataset, we show that for the task of on/off detection the second framework, i.e., directly training a binary classifier, achieves better performance in terms of F1 score. [ABSTRACT FROM AUTHOR]

Background Self-reported client assessments during online treatments enable the development of statistical models for the prediction of client improvement and symptom development. Evaluation of these models is mandatory to ensure their validity. Methods For this purpose, we suggest besides a model evaluation based on study data the use of a simulation analysis. The simulation analysis provides insight into the model performance and enables to analyse reasons for a low predictive accuracy. In this study, we evaluate a temporal causal model (TCM) and show that it does not provide reliable predictions of clients' future mood levels. Results Based on the simulation analysis we investigate the potential reasons for the low predictive performance, for example, noisy measurements and sampling frequency. We conclude that the analysed TCM in its current form is not sufficient to describe the underlying psychological processes. Conclusions The results demonstrate the importance of model evaluation and the benefit of a simulation analysis. The current manuscript provides practical guidance for conducting model evaluation including simulation analysis. [ABSTRACT FROM AUTHOR]

Predicting therapeutic outcome in the mental health domain is of utmost importance to enable therapists to provide the most effective treatment to a patient. Using information from the writings of a patient can potentially be a valuable source of information, especially now that more and more treatments involve computer-based exercises or electronic conversations between patient and therapist. In this paper, we study predictive modeling using writings of patients under treatment for a social anxiety disorder. We extract a wealth of information from the text written by patients including their usage of words, the topics they talk about, the sentiment of the messages, and the style of writing. In addition, we study trends over time with respect to those measures. We then apply machine learning algorithms to generate the predictive models. Based on a dataset of 69 patients, we are able to show that we can predict therapy outcome with an area under the curve of 0.83 halfway through the therapy and with a precision of 0.78 when using the full data (i.e., the entire treatment period). Due to the limited number of participants, it is hard to generalize the results, but they do show great potential in this type of information. [ABSTRACT FROM AUTHOR]

More than a decade after the first extensive overview on parameter control, we revisit the field and present a survey of the state-of-the-art. We briefly summarize the development of the field and discuss existing work related to each major parameter or component of an evolutionary algorithm. Based on this overview, we observe trends in the area, identify some (methodological) shortcomings, and give recommendations for future research. [ABSTRACT FROM AUTHOR]

The goals of this study were to examine whether machine-learning algorithms outperform multivariable logistic regression in the prediction of insufficient response to methotrexate (MTX); secondly, to examine which features are essential for correct prediction; and finally, to investigate whether the best performing model specifically identifies insufficient responders to MTX (combination) therapy. The prediction of insufficient response (3-month Disease Activity Score 28-Erythrocyte-sedimentation rate (DAS28-ESR) > 3.2) was assessed using logistic regression, least absolute shrinkage and selection operator (LASSO), random forest, and extreme gradient boosting (XGBoost). The baseline features of 355 rheumatoid arthritis (RA) patients from the "treatment in the Rotterdam Early Arthritis CoHort" (tREACH) and the U-Act-Early trial were combined for analyses. The model performances were compared using area under the curve (AUC) of receiver operating characteristic (ROC) curves, 95% confidence intervals (95% CI), and sensitivity and specificity. Finally, the best performing model following feature selection was tested on 101 RA patients starting tocilizumab (TCZ)-monotherapy. Logistic regression (AUC = 0.77 95% CI: 0.68–0.86) performed as well as LASSO (AUC = 0.76, 95% CI: 0.67–0.85), random forest (AUC = 0.71, 95% CI: 0.61 = 0.81), and XGBoost (AUC = 0.70, 95% CI: 0.61–0.81), yet logistic regression reached the highest sensitivity (81%). The most important features were baseline DAS28 (components). For all algorithms, models with six features performed similarly to those with 16. When applied to the TCZ-monotherapy group, logistic regression's sensitivity significantly dropped from 83% to 69% (p = 0.03). In the current dataset, logistic regression performed equally well compared to machine-learning algorithms in the prediction of insufficient response to MTX. Models could be reduced to six features, which are more conducive for clinical implementation. Interestingly, the prediction model was specific to MTX (combination) therapy response. [ABSTRACT FROM AUTHOR]

Background: Digital health interventions (DHIs) are poised to reduce target symptoms in a scalable, affordable, and empirically supported way. DHIs that involve coaching or clinical support often collect text data from 2 sources: (1) open correspondence between users and the trained practitioners supporting them through a messaging system and (2) text data recorded during the intervention by users, such as diary entries. Natural language processing (NLP) offers methods for analyzing text, augmenting the understanding of intervention effects, and informing therapeutic decision making.Objective: This study aimed to present a technical framework that supports the automated analysis of both types of text data often present in DHIs. This framework generates text features and helps to build statistical models to predict target variables, including user engagement, symptom change, and therapeutic outcomes.Methods: We first discussed various NLP techniques and demonstrated how they are implemented in the presented framework. We then applied the framework in a case study of the Healthy Body Image Program, a Web-based intervention trial for eating disorders (EDs). A total of 372 participants who screened positive for an ED received a DHI aimed at reducing ED psychopathology (including binge eating and purging behaviors) and improving body image. These users generated 37,228 intervention text snippets and exchanged 4285 user-coach messages, which were analyzed using the proposed model.Results: We applied the framework to predict binge eating behavior, resulting in an area under the curve between 0.57 (when applied to new users) and 0.72 (when applied to new symptom reports of known users). In addition, initial evidence indicated that specific text features predicted the therapeutic outcome of reducing ED symptoms.Conclusions: The case study demonstrates the usefulness of a structured approach to text data analytics. NLP techniques improve the prediction of symptom changes in DHIs. We present a technical framework that can be easily applied in other clinical trials and clinical presentations and encourage other groups to apply the framework in similar contexts. [ABSTRACT FROM AUTHOR]

A variety of effective psychotherapies for depression are available, but patients who suffer from depression vary in their treatment response. Combining face-to-face therapies with internet-based elements in the sense of blended treatment is a new approach to treatment for depression. The goal of this study was to answer the following research questions: (1) What are the most important predictors determining optimal treatment allocation to treatment as usual or blended treatment? and (2) Would model-determined treatment allocation using this predictive information and the personalized advantage index (PAI)-approach result in better treatment outcomes? Bayesian model averaging (BMA) was applied to the data of a randomized controlled trial (RCT) comparing the efficacy of treatment as usual and blended treatment in depressive outpatients. Pre-treatment symptomatology and treatment expectancy predicted outcomes irrespective of treatment condition, whereas different prescriptive predictors were found. A PAI of 2.33 PHQ-9 points was found, meaning that patients who would have received the treatment that is optimal for them would have had a post-treatment PHQ-9 score that is two points lower than if they had received the treatment that is suboptimal for them. For 29% of the sample, the PAI was five or greater, which means that a substantial difference between the two treatments was predicted. The use of the PAI approach for clinical practice must be further confirmed in prospective research; the current study supports the identification of specific interventions favorable for specific patients. [ABSTRACT FROM AUTHOR]

Background: Different treatment alternatives exist for psychological disorders. Both clinical and cost effectiveness of treatment are crucial aspects for policy makers, therapists, and patients and thus play major roles for healthcare decision-making. At the start of an intervention, it is often not clear which specific individuals benefit most from a particular intervention alternative or how costs will be distributed on an individual patient level.Objective: This study aimed at predicting the individual outcome and costs for patients before the start of an internet-based intervention. Based on these predictions, individualized treatment recommendations can be provided. Thus, we expand the discussion of personalized treatment recommendation.Methods: Outcomes and costs were predicted based on baseline data of 350 patients from a two-arm randomized controlled trial that compared treatment as usual and blended therapy for depressive disorders. For this purpose, we evaluated various machine learning techniques, compared the predictive accuracy of these techniques, and revealed features that contributed most to the prediction performance. We then combined these predictions and utilized an incremental cost-effectiveness ratio in order to derive individual treatment recommendations before the start of treatment.Results: Predicting clinical outcomes and costs is a challenging task that comes with high uncertainty when only utilizing baseline information. However, we were able to generate predictions that were more accurate than a predefined reference measure in the shape of mean outcome and cost values. Questionnaires that include anxiety or depression items and questions regarding the mobility of individuals and their energy levels contributed to the prediction performance. We then described how patients can be individually allocated to the most appropriate treatment type. For an incremental cost-effectiveness threshold of 25,000 €/quality-adjusted life year, we demonstrated that our recommendations would have led to slightly worse outcomes (1.98%), but with decreased cost (5.42%).Conclusions: Our results indicate that it was feasible to provide personalized treatment recommendations at baseline and thus allocate patients to the most beneficial treatment type. This could potentially lead to improved decision-making, better outcomes for individuals, and reduced health care costs. [ABSTRACT FROM AUTHOR]

Introduction: Treatment in the intensive care unit (ICU) generates complex data where machine learning (ML) modelling could be beneficial. Using routine hospital data, we evaluated the ability of multiple ML models to predict inpatient mortality in a paediatric population in a low/middle-income country., Method: We retrospectively analysed hospital record data from 0-59 months old children admitted to the ICU of Dhaka hospital of International Centre for Diarrhoeal Disease Research, Bangladesh. Five commonly used ML models- logistic regression, least absolute shrinkage and selection operator, elastic net, gradient boosting trees (GBT) and random forest (RF), were evaluated using the area under the receiver operating characteristic curve (AUROC). Top predictors were selected using RF mean decrease Gini scores as the feature importance values., Results: Data from 5669 children was used and was reduced to 3505 patients (10% death, 90% survived) following missing data removal. The mean patient age was 10.8 months (SD=10.5). The top performing models based on the validation performance measured by mean 10-fold cross-validation AUROC on the training data set were RF and GBT. Hyperparameters were selected using cross-validation and then tested in an unseen test set. The models developed used demographic, anthropometric, clinical, biochemistry and haematological data for mortality prediction. We found RF consistently outperformed GBT and predicted the mortality with AUROC of ≥0.87 in the test set when three or more laboratory measurements were included. However, after the inclusion of a fourth laboratory measurement, very minor predictive gains (AUROC 0.87 vs 0.88) resulted. The best predictors were the biochemistry and haematological measurements, with the top predictors being total CO 2 , potassium, creatinine and total calcium., Conclusions: Mortality in children admitted to ICU can be predicted with high accuracy using RF ML models in a real-life data set using multiple laboratory measurements with the most important features primarily coming from patient biochemistry and haematology., Competing Interests: Competing interests: No, there are no competing interests., (© Author(s) (or their employer(s)) 2024. Re-use permitted under CC BY-NC. No commercial re-use. See rights and permissions. Published by BMJ.)

Objective: Congenital hypothyroidism (CH) is an inborn thyroid hormone (TH) deficiency mostly caused by thyroidal (primary CH) or hypothalamic/pituitary (central CH) disturbances. Most CH newborn screening (NBS) programs are thyroid-stimulating-hormone (TSH) based, thereby only detecting primary CH. The Dutch NBS is based on measuring total thyroxine (T4) from dried blood spots, aiming to detect primary and central CH at the cost of more false-positive referrals (FPRs) (positive predictive value (PPV) of 21% in 2007-2017). An artificial PPV of 26% was yielded when using a machine learning-based model on the adjusted dataset described based on the Dutch CH NBS. Recently, amino acids (AAs) and acylcarnitines (ACs) have been shown to be associated with TH concentration. We therefore aimed to investigate whether AAs and ACs measured during NBS can contribute to better performance of the CH screening in the Netherlands by using a revised machine learning-based model., Methods: Dutch NBS data between 2007 and 2017 (CH screening results, AAs and ACs) from 1079 FPRs, 515 newborns with primary (431) and central CH (84) and data from 1842 healthy controls were used. A random forest model including these data was developed., Results: The random forest model with an artificial sensitivity of 100% yielded a PPV of 48% and AUROC of 0.99. Besides T4 and TSH, tyrosine, and succinylacetone were the main parameters contributing to the model's performance., Conclusions: The PPV improved significantly (26-48%) by adding several AAs and ACs to our machine learning-based model, suggesting that adding these parameters benefits the current algorithm.

Objective: This work explores the perceptions of obstetrical clinicians about artificial intelligence (AI) in order to bridge the gap in uptake of AI between research and medical practice. Identifying potential areas where AI can contribute to clinical practice, enables AI research to align with the needs of clinicians and ultimately patients., Design: Qualitative interview study., Setting: A national study conducted in the Netherlands between November 2022 and February 2023., Participants: Dutch clinicians working in obstetrics with varying relevant work experience, gender and age., Analysis: Thematic analysis of qualitative interview transcripts., Results: Thirteen gynaecologists were interviewed about hypothetical scenarios of an implemented AI model. Thematic analysis identified two major themes: perceived usefulness and trust. Usefulness involved AI extending human brain capacity in complex pattern recognition and information processing, reducing contextual influence and saving time. Trust required validation, explainability and successful personal experience. This result shows two paradoxes: first, AI is expected to provide added value by surpassing human capabilities, yet also a need to understand the parameters and their influence on predictions for trust and adoption was expressed. Second, participants recognised the value of incorporating numerous parameters into a model, but they also believed that certain contextual factors should only be considered by humans, as it would be undesirable for AI models to use that information., Conclusions: Obstetricians' opinions on the potential value of AI highlight the need for clinician-AI researcher collaboration. Trust can be built through conventional means like randomised controlled trials and guidelines. Holistic impact metrics, such as changes in workflow, not just clinical outcomes, should guide AI model development. Further research is needed for evaluating evolving AI systems beyond traditional validation methods., Competing Interests: Competing interests: None declared., (© Author(s) (or their employer(s)) 2023. Re-use permitted under CC BY-NC. No commercial re-use. See rights and permissions. Published by BMJ.)

Introduction: In ageing societies, the number of older adults with complex chronic conditions (CCCs) is rapidly increasing. Care for older persons with CCCs is challenging, due to interactions between multiple conditions and their treatments. In home care and nursing homes, where most older persons with CCCs receive care, professionals often lack appropriate decision support suitable and sufficient to address the medical and functional complexity of persons with CCCs. This EU-funded project aims to develop decision support systems using high-quality, internationally standardised, routine care data to support better prognostication of health trajectories and treatment impact among older persons with CCCs., Methods and Analysis: Real-world data from older persons aged ≥60 years in home care and nursing homes, based on routinely performed comprehensive geriatric assessments using interRAI systems collected in the past 20 years, will be linked with administrative repositories on mortality and care use. These include potentially up to 51 million care recipients from eight countries: Italy, the Netherlands, Finland, Belgium, Canada, USA, Hong Kong and New Zealand. Prognostic algorithms will be developed and validated to better predict various health outcomes. In addition, the modifying impact of pharmacological and non-pharmacological interventions will be examined. A variety of analytical methods will be used, including techniques from the field of artificial intelligence such as machine learning. Based on the results, decision support tools will be developed and pilot tested among health professionals working in home care and nursing homes., Ethics and Dissemination: The study was approved by authorised medical ethical committees in each of the participating countries, and will comply with both local and EU legislation. Study findings will be shared with relevant stakeholders, including publications in peer-reviewed journals and presentations at national and international meetings., Competing Interests: Competing interests: None declared., (© Author(s) (or their employer(s)) 2023. Re-use permitted under CC BY-NC. No commercial re-use. See rights and permissions. Published by BMJ.)

Objectives: Heart failure (HF) is a commonly occurring health problem with high mortality and morbidity. If potential cases could be detected earlier, it may be possible to intervene earlier, which may slow progression in some patients. Preferably, it is desired to reuse already measured data for screening of all persons in an age group, such as general practitioner (GP) data. Furthermore, it is essential to evaluate the number of people needed to screen to find one patient using true incidence rates, as this indicates the generalisability in the true population. Therefore, we aim to create a machine learning model for the prediction of HF using GP data and evaluate the number needed to screen with true incidence rates., Design, Settings and Participants: GP data from 8543 patients (-2 to -1 year before diagnosis) and controls aged 70+ years were obtained retrospectively from 01 January 2012 to 31 December 2019 from the Nivel Primary Care Database. Codes about chronic illness, complaints, diagnostics and medication were obtained. Data were split in a train/test set. Datasets describing demographics, the presence of codes (non-sequential) and upon each other following codes (sequential) were created. Logistic regression, random forest and XGBoost models were trained. Predicted outcome was the presence of HF after 1 year. The ratio case:control in the test set matched true incidence rates (1:45)., Results: Sole demographics performed average (area under the curve (AUC) 0.692, CI 0.677 to 0.706). Adding non-sequential information combined with a logistic regression model performed best and significantly improved performance (AUC 0.772, CI 0.759 to 0.785, p<0.001). Further adding sequential information did not alter performance significantly (AUC 0.767, CI 0.754 to 0.780, p=0.07). The number needed to screen dropped from 14.11 to 5.99 false positives per true positive., Conclusion: This study created a model able to identify patients with pending HF a year before diagnosis., Competing Interests: Competing interests: None declared., (© Author(s) (or their employer(s)) 2022. Re-use permitted under CC BY-NC. No commercial re-use. See rights and permissions. Published by BMJ.)

Unexpected ICU readmission is associated with longer length of stay and increased mortality. To prevent ICU readmission and death after ICU discharge, our team of intensivists and data scientists aimed to use AmsterdamUMCdb to develop an explainable machine learning-based real-time bedside decision support tool., Derivation Cohort: Data from patients admitted to a mixed surgical-medical academic medical center ICU from 2004 to 2016., Validation Cohort: Data from 2016 to 2019 from the same center., Prediction Model: Patient characteristics, clinical observations, physiologic measurements, laboratory studies, and treatment data were considered as model features. Different supervised learning algorithms were trained to predict ICU readmission and/or death, both within 7 days from ICU discharge, using 10-fold cross-validation. Feature importance was determined using SHapley Additive exPlanations, and readmission probability-time curves were constructed to identify subgroups. Explainability was established by presenting individualized risk trends and feature importance., Results: Our final derivation dataset included 14,105 admissions. The combined readmission/mortality rate within 7 days of ICU discharge was 5.3%. Using Gradient Boosting, the model achieved an area under the receiver operating characteristic curve of 0.78 (95% CI, 0.75-0.81) and an area under the precision-recall curve of 0.19 on the validation cohort ( n = 3,929). The most predictive features included common physiologic parameters but also less apparent variables like nutritional support. At a 6% risk threshold, the model showed a sensitivity (recall) of 0.72, specificity of 0.70, and a positive predictive value (precision) of 0.15. Impact analysis using probability-time curves and the 6% risk threshold identified specific patient groups at risk and the potential of a change in discharge management to reduce relative risk by 14%., Conclusions: We developed an explainable machine learning model that may aid in identifying patients at high risk for readmission and mortality after ICU discharge using the first freely available European critical care database, AmsterdamUMCdb. Impact analysis showed that a relative risk reduction of 14% could be achievable, which might have significant impact on patients and society. ICU data sharing facilitates collaboration between intensivists and data scientists to accelerate model development., Competing Interests: Amsterdam University Medical Centers is entitled to royalties from Pacmed. Mr. Hovenkamp is cofounder of Pacmed, a Dutch technology company in health care. The remaining authors have disclosed that they do not have any potential conflicts of interest., (Copyright © 2021 The Authors. Published by Wolters Kluwer Health, Inc. on behalf of the Society of Critical Care Medicine.)

Background: The identification of risk factors for adverse outcomes and prolonged intensive care unit (ICU) stay in COVID-19 patients is essential for prognostication, determining treatment intensity, and resource allocation. Previous studies have determined risk factors on admission only, and included a limited number of predictors. Therefore, using data from the highly granular and multicenter Dutch Data Warehouse, we developed machine learning models to identify risk factors for ICU mortality, ventilator-free days and ICU-free days during the course of invasive mechanical ventilation (IMV) in COVID-19 patients., Methods: The DDW is a growing electronic health record database of critically ill COVID-19 patients in the Netherlands. All adult ICU patients on IMV were eligible for inclusion. Transfers, patients admitted for less than 24 h, and patients still admitted at time of data extraction were excluded. Predictors were selected based on the literature, and included medication dosage and fluid balance. Multiple algorithms were trained and validated on up to three sets of observations per patient on day 1, 7, and 14 using fivefold nested cross-validation, keeping observations from an individual patient in the same split., Results: A total of 1152 patients were included in the model. XGBoost models performed best for all outcomes and were used to calculate predictor importance. Using Shapley additive explanations (SHAP), age was the most important demographic risk factor for the outcomes upon start of IMV and throughout its course. The relative probability of death across age values is visualized in Partial Dependence Plots (PDPs), with an increase starting at 54 years. Besides age, acidaemia, low P/F-ratios and high driving pressures demonstrated a higher probability of death. The PDP for driving pressure showed a relative probability increase starting at 12 cmH 2 O., Conclusion: Age is the most important demographic risk factor of ICU mortality, ICU-free days and ventilator-free days throughout the course of invasive mechanical ventilation in critically ill COVID-19 patients. pH, P/F ratio, and driving pressure should be monitored closely over the course of mechanical ventilation as risk factors predictive of these outcomes.