Counterfactual Sepsis Outcome Prediction Under Dynamic and Time-Varying Treatment Regimes.

Sepsis is a life-threatening condition that occurs when the body's normal response to an infection is out of balance. A key part of managing sepsis involves the administration of intravenous fluids and vasopressors. In this work, we explore the application of G-Net, a deep sequential modeling framework for g-computation, to predict outcomes under counterfactual fluid treatment strategies in a real-world cohort of sepsis patients. Utilizing observational data collected from the intensive care unit (ICU), we evaluate the performance of multiple deep learning implementations of G-Net and compare their predictive performance with linear models in forecasting patient outcomes and trajectories over time under the observational treatment regime. We then demonstrate that G-Net can generate counterfactual prediction of covariate trajectories that align with clinical expectations across various fluid limiting regimes. Our study demonstrates the potential clinical utility of G-Net in predicting counterfactual treatment outcomes, aiding clinicians in informed decision-making for sepsis patients in the ICU.

The impact of aggressive and conservative propensity for initiation of neuromuscular blockade in mechanically ventilated patients with hypoxemic respiratory failure.

INTRODUCTION: Neuromuscular blockade (NMB) in ventilated patients may cause benefit or harm. We applied "incremental interventions" to determine the impact of altering NMB initiation aggressiveness. METHODS: Retrospective cohort study of ventilated patients with PaO2/FiO2 ratio < 150 mmHg and PEEP≥ 8cmH2O from the Medical Information Mart of Intensive Care IV database (MIMIC-IV version 1.0) estimating the effect of incremental interventions on in-hospital mortality and ventilator-free days, modifying hourly propensity for NMB initiation to be aggressive or conservative relative to usual care, adjusting for confounding with inverse probability weighting. RESULTS: 5221 patients were included (13.3% initiated on NMB). Incremental interventions estimated a strong effect on NMB usage: 5-fold higher hourly odds of initiation increased usage to 36.5% (CI = [34.3%,38.7%]) and 5-fold lower odds decreased usage to 3.8% (CI = [3.3%,4.3%]). Aggressive and conservative strategies demonstrated a U-shaped mortality relationship. 5-fold higher or lower propensity increased in-hospital mortality by 2.6% (0.95 CI = [1.5%,3.7%]) or 1.3% (0.95 CI = [0.1%,2.5%]) respectively. In secondary analysis of a healthier patient cohort, results were similar, however conservative strategies also improved ventilator-free days. INTERPRETATION: Aggressive or conservative initiation of NMB may worsen mortality. In healthier populations, marginally conservative NMB initiation strategies may lead to increased ventilator free days with minimal impact on mortality.

Effects of aggressive and conservative strategies for mechanical ventilation liberation.

BACKGROUND: The optimal approach for transitioning from strict lung protective ventilation to support modes of ventilation when patients determine their own respiratory rate and tidal volume remains unclear. While aggressive liberation from lung protective settings could expedite extubation and prevent harm from prolonged ventilation and sedation, conservative liberation could prevent lung injury from spontaneous breathing. RESEARCH QUESTION: Should physicians take a more aggressive or conservative approach to liberation? METHODS: Retrospective cohort study of mechanically ventilated patients from the Medical Information Mart for Intensive Care IV database (MIMIC-IV version 1.0) estimating effects of incremental interventions modifying the propensity for liberation to be more aggressive or conservative relative to usual care, with adjustment for confounding via inverse probability weighting. Outcomes included in-hospital mortality, ventilator free days, and ICU free days. Analysis was performed on the entire cohort as well as subgroups differentiated by PaO2/FiO2 ratio, and SOFA. RESULTS: 7433 patients were included. Strategies multiplying the odds of a first liberation relative to usual care at each hour had a large impact on time to first liberation attempt (43 h under usual care, 24 h (0.95 CI = [23,25]) with an aggressive strategy doubling liberation odds, and 74 h (0.95 CI = [69,78]) under a conservative strategy halving liberation odds). In the full cohort, we estimated aggressive liberation increased ICU-free days by 0.9 days (0.95 CI = [0.8,1.0]) and ventilator free days by 0.82 days (0.95 CI = [0.67,0.97]), but had minimal effect on mortality (only a 0.3% (0.95 CI = [-0.2%,0.8%]) difference between minimum and maximum rates). With baseline SOFA≥ 12 (n = 1355), aggressive liberation moderately increased mortality (58.5% [0.95 CI = (55.7%,61.2%)]) compared with conservative liberation (55.1% [0.95 CI = (51.6%,58.6%)]). INTERPRETATION: Aggressive liberation may improve ventilator free and ICU free days with little impact on mortality in patients with SOFA score < 12. Trials are needed.

MIMIC-IV, a freely accessible electronic health record dataset.

Digital data collection during routine clinical practice is now ubiquitous within hospitals. The data contains valuable information on the care of patients and their response to treatments, offering exciting opportunities for research. Typically, data are stored within archival systems that are not intended to support research. These systems are often inaccessible to researchers and structured for optimal storage, rather than interpretability and analysis. Here we present MIMIC-IV, a publicly available database sourced from the electronic health record of the Beth Israel Deaconess Medical Center. Information available includes patient measurements, orders, diagnoses, procedures, treatments, and deidentified free-text clinical notes. MIMIC-IV is intended to support a wide array of research studies and educational material, helping to reduce barriers to conducting clinical research.

Knowledge Distillation via Constrained Variational Inference.

Knowledge distillation has been used to capture the knowledge of a teacher model and distill it into a student model with some desirable characteristics such as being smaller, more efficient, or more generalizable. In this paper, we propose a framework for distilling the knowledge of a powerful discriminative model such as a neural network into commonly used graphical models known to be more interpretable (e.g., topic models, autoregressive Hidden Markov Models). Posterior of latent variables in these graphical models (e.g., topic proportions in topic models) is often used as feature representation for predictive tasks. However, these posterior-derived features are known to have poor predictive performance compared to the features learned via purely discriminative approaches. Our framework constrains variational inference for posterior variables in graphical models with a similarity preserving constraint. This constraint distills the knowledge of the discriminative model into the graphical model by ensuring that input pairs with (dis)similar representation in the teacher model also have (dis)similar representation in the student model. By adding this constraint to the variational inference scheme, we guide the graphical model to be a reasonable density model for the data while having predictive features which are as close as possible to those of a discriminative model. To make our framework applicable to a wide range of graphical models, we build upon the Automatic Differentiation Variational Inference (ADVI), a black-box inference framework for graphical models. We demonstrate the effectiveness of our framework on two real-world tasks of disease subtyping and disease trajectory modeling.

Titration of Ventilator Settings to Target Driving Pressure and Mechanical Power.

PURPOSE: Driving pressure (ΔP) and mechanical power (MP) may be important mediators of lung injury in acute respiratory distress syndrome (ARDS) however there is little evidence for strategies directed at lowering these parameters. We applied predictive modeling to estimate the effects of modifying ventilator parameters on ΔP and MP. METHODS: 2,622 ARDS patients (Berlin criteria) from the Medical Information Mart for Intensive Care IV database (MIMIC-IV version1.0) admitted to the intensive care unit (ICU) at Beth Israel Deaconess Medical Center between 2008 and 2019 were included. Flexible confounding-adjusted regression models for time varying data were fit to estimate the effects of adjusting PEEP and tidal volume (VT) on ΔP, and adjusting VT and respiratory rate (f) on MP. RESULTS: Reduction in VT reduced ΔP and MP, with more pronounced effect on MP with lower compliance. Strategies reducing f, consistently increased MP (when VT was adjusted to maintain consistent minute ventilation). Adjustment of PEEP yielded a U-shaped effect on ΔP. CONCLUSIONS: This novel conditional modeling confirmed expected response patterns for ΔP, with the response to adjustments depending on patients' lung mechanics. Furthermore a VT -driven approach should be favored over a f -driven approach when aiming to reduce MP.

A contrastive learning approach for ICU false arrhythmia alarm reduction.

The high rate of false arrhythmia alarms in Intensive Care Units (ICUs) can lead to disruption of care, negatively impacting patients' health through noise disturbances, and slow staff response time due to alarm fatigue. Prior false-alarm reduction approaches are often rule-based and require hand-crafted features from physiological waveforms as inputs to machine learning classifiers. Despite considerable prior efforts to address the problem, false alarms are a continuing problem in the ICUs. In this work, we present a deep learning framework to automatically learn feature representations of physiological waveforms using convolutional neural networks (CNNs) to discriminate between true vs. false arrhythmia alarms. We use Contrastive Learning to simultaneously minimize a binary cross entropy classification loss and a proposed similarity loss from pair-wise comparisons of waveform segments over time as a discriminative constraint. Furthermore, we augment our deep models with learned embeddings from a rule-based method to leverage prior domain knowledge for each alarm type. We evaluate our method using the dataset from the 2015 PhysioNet Computing in Cardiology Challenge. Ablation analysis demonstrates that Contrastive Learning significantly improves the performance of a combined deep learning and rule-based-embedding approach. Our results indicate that the final proposed deep learning framework achieves superior performance in comparison to the winning entries of the Challenge.

A novel artificial intelligence based intensive care unit monitoring system: using physiological waveforms to identify sepsis.

A massive amount of multimodal data are continuously collected in the intensive care unit (ICU) along each patient stay, offering a great opportunity for the development of smart monitoring devices based on artificial intelligence (AI). The two main sources of relevant information collected in the ICU are the electronic health records (EHRs) and vital sign waveforms continuously recorded at the bedside. While EHRs are already widely processed by AI algorithms for prompt diagnosis and prognosis, AI-based assessments of the patients' pathophysiological state using waveforms are less developed, and their use is still limited to real-time monitoring for basic visual vital sign feedback at the bedside. This study uses data from the MIMIC-III database (PhysioNet) to propose a novel AI approach in ICU patient monitoring that incorporates features estimated by a closed-loop cardiovascular model, with the specific goal of identifying sepsis within the first hour of admission. Our top benchmark results (AUROC = 0.92, AUPRC = 0.90) suggest that features derived by cardiovascular control models may play a key role in identifying sepsis, by continuous monitoring performed through advanced multivariate modelling of vital sign waveforms. This work lays foundations for a deeper data integration paradigm which will help clinicians in their decision-making processes. This article is part of the theme issue 'Advanced computation in cardiovascular physiology: new challenges and opportunities'.

Multi-View Spatial-Temporal Graph Convolutional Networks With Domain Generalization for Sleep Stage Classification.

Sleep stage classification is essential for sleep assessment and disease diagnosis. Although previous attempts to classify sleep stages have achieved high classification performance, several challenges remain open: 1) How to effectively utilize time-varying spatial and temporal features from multi-channel brain signals remains challenging. Prior works have not been able to fully utilize the spatial topological information among brain regions. 2) Due to the many differences found in individual biological signals, how to overcome the differences of subjects and improve the generalization of deep neural networks is important. 3) Most deep learning methods ignore the interpretability of the model to the brain. To address the above challenges, we propose a multi-view spatial-temporal graph convolutional networks (MSTGCN) with domain generalization for sleep stage classification. Specifically, we construct two brain view graphs for MSTGCN based on the functional connectivity and physical distance proximity of the brain regions. The MSTGCN consists of graph convolutions for extracting spatial features and temporal convolutions for capturing the transition rules among sleep stages. In addition, attention mechanism is employed for capturing the most relevant spatial-temporal information for sleep stage classification. Finally, domain generalization and MSTGCN are integrated into a unified framework to extract subject-invariant sleep features. Experiments on two public datasets demonstrate that the proposed model outperforms the state-of-the-art baselines.

Delaying initiation of diuretics in critically ill patients with recent vasopressor use and high positive fluid balance.

BACKGROUND: Fluid overload is associated with poor outcomes. Clinicians might be reluctant to initiate diuretic therapy for patients with recent vasopressor use. We estimated the effect on 30-day mortality of withholding or delaying diuretics after vasopressor use in patients with probable fluid overload. METHODS: This was a retrospective cohort study of adults admitted to ICUs of an academic medical centre between 2008 and 2012. Using a database of time-stamped patient records, we followed individuals from the time they first required vasopressor support and had >5 L cumulative positive fluid balance (plus additional inclusion/exclusion criteria). We compared mortality under usual care (the mix of care actually delivered in the cohort) and treatment strategies restricting diuretic initiation during and for various durations after vasopressor use. We adjusted for baseline and time-varying confounding via inverse probability weighting. RESULTS: The study included 1501 patients, and the observed 30-day mortality rate was 11%. After adjusting for observed confounders, withholding diuretics for at least 24 h after stopping most recent vasopressor use was estimated to increase 30-day mortality rate by 2.2% (95% confidence interval [CI], 0.9-3.6%) compared with usual care. Data were consistent with moderate harm or slight benefit from withholding diuretic initiation only during concomitant vasopressor use; the estimated mortality rate increased by 0.5% (95% CI, -0.2% to 1.1%). CONCLUSIONS: Withholding diuretic initiation after vasopressor use in patients with high cumulative positive balance (>5 L) was estimated to increase 30-day mortality. These findings are hypothesis generating and should be tested in a clinical trial.

Fluid-limiting treatment strategies among sepsis patients in the ICU: a retrospective causal analysis.

OBJECTIVE: In septic patients, multiple retrospective studies show an association between large volumes of fluids administered in the first 24 h and mortality, suggesting a benefit to fluid restrictive strategies. However, these studies do not directly estimate the causal effects of fluid-restrictive strategies, nor do their analyses properly adjust for time-varying confounding by indication. In this study, we used causal inference techniques to estimate mortality outcomes that would result from imposing a range of arbitrary limits ("caps") on fluid volume administration during the first 24 h of intensive care unit (ICU) care. DESIGN: Retrospective cohort study SETTING: ICUs at the Beth Israel Deaconess Medical Center, 2008-2012 PATIENTS: One thousand six hundred thirty-nine septic patients (defined by Sepsis-3 criteria) 18 years and older, admitted to the ICU from the emergency department (ED), who received less than 4 L fluids administered prior to ICU admission MEASUREMENTS AND MAIN RESULTS: Data were obtained from the Medical Information Mart for Intensive Care III (MIMIC-III). We employed a dynamic Marginal Structural Model fit by inverse probability of treatment weighting to obtain confounding adjusted estimates of mortality rates that would have been observed had fluid resuscitation volume caps between 4 L-12 L been imposed on the population. The 30-day mortality in our cohort was 17%. We estimated that caps between 6 and 10 L on 24 h fluid volume would have reduced 30-day mortality by - 0.6 to - 1.0%, with the greatest reduction at 8 L (- 1.0% mortality, 95% CI [- 1.6%, - 0.3%]). CONCLUSIONS: We found that 30-day mortality would have likely decreased relative to observed mortality under current practice if these patients had been subject to "caps" on the total volume of fluid administered between 6 and 10 L, with the greatest reduction in mortality rate at 8 L.

Is Deep Reinforcement Learning Ready for Practical Applications in Healthcare? A Sensitivity Analysis of Duel-DDQN for Hemodynamic Management in Sepsis Patients.

The potential of Reinforcement Learning (RL) has been demonstrated through successful applications to games such as Go and Atari. However, while it is straightforward to evaluate the performance of an RL algorithm in a game setting by simply using it to play the game, evaluation is a major challenge in clinical settings where it could be unsafe to follow RL policies in practice. Thus, understanding sensitivity of RL policies to the host of decisions made during implementation is an important step toward building the type of trust in RL required for eventual clinical uptake. In this work, we perform a sensitivity analysis on a state-of-the-art RL algorithm (Dueling Double Deep Q-Networks) applied to hemodynamic stabilization treatment strategies for septic patients in the ICU. We consider sensitivity of learned policies to input features, embedding model architecture, time discretization, reward function, and random seeds. We find that varying these settings can significantly impact learned policies, which suggests a need for caution when interpreting RL agent output.

Retaining Privileged Information for Multi-Task Learning.

Knowledge transfer has been of great interest in current machine learning research, as many have speculated its importance in modeling the human ability to rapidly generalize learned models to new scenarios. Particularly in cases where training samples are limited, knowledge transfer shows improvement on both the learning speed and generalization performance of related tasks. Recently, Learning Using Privileged Information (LUPI) has presented a new direction in knowledge transfer by modeling the transfer of prior knowledge as a Teacher-Student interaction process. Under LUPI, a Teacher model uses Privileged Information (PI) that is only available at training time to improve the sample complexity required to train a Student learner for a given task. In this work, we present a LUPI formulation that allows privileged information to be retained in a multi-task learning setting. We propose a novel feature matching algorithm that projects samples from the original feature space and the privilege information space into a joint latent space in a way that informs similarity between training samples. Our experiments show that useful knowledge from PI is maintained in the latent space and greatly improves the sample efficiency of other related learning tasks. We also provide an analysis of sample complexity of the proposed LUPI method, which under some favorable assumptions can achieve a greater sample efficiency than brute force methods.

A Model-Based Machine Learning Approach to Probing Autonomic Regulation From Nonstationary Vital-Sign Time Series.

Physiological variables, such as heart rate (HR), blood pressure (BP) and respiration (RESP), are tightly regulated and coupled under healthy conditions, and a break-down in the coupling has been associated with aging and disease. We present an approach that incorporates physiological modeling within a switching linear dynamical systems (SLDS) framework to assess the various functional components of the autonomic regulation through transfer function analysis of nonstationary multivariate time series of vital signs. We validate our proposed SLDS-based transfer function analysis technique in automatically capturing 1) changes in baroreflex gain due to postural changes in a tilt-table study including ten subjects, and 2) the effect of aging on the autonomic control using HR/RESP recordings from 40 healthy adults. Next, using HR/BP time series of more than 450 adult ICU patients, we show that our technique can be used to reveal coupling changes associated with severe sepsis (AUC = 0.74, sensitivity = 0.74, specificity = 0.60). Our findings indicate that reduced HR/BP coupling is significantly associated with severe sepsis even after adjusting for clinical interventions (P  0.001). These results demonstrate the utility of our approach in phenotyping complex vital-sign dynamics, and in providing mechanistic hypotheses in terms of break-down of autoregulatory systems under healthy and disease conditions.

You Snooze, You Win: the PhysioNet/Computing in Cardiology Challenge 2018.

The PhysioNet/Computing in Cardiology Challenge 2018 focused on the use of various physiological signals (EEG, EOG, EMG, ECG, SaO2) collected during polysomnographic sleep studies to detect sources of arousal (non-apnea) during sleep. A total of 1,983 polysomnographic recordings were made available to the entrants. The arousal labels for 994 of the recordings were made available in a public training set while 989 labels were retained in a hidden test set. Challengers were asked to develop an algorithm that could label the presence of arousals within the hidden test set. The performance metric used to assess entrants was the area under the precision-recall curve. A total of twenty-two independent teams entered the Challenge, deploying a variety of methods from generalized linear models to deep neural networks.

Representation Learning Approaches to Detect False Arrhythmia Alarms from ECG Dynamics.

The high rate of intensive care unit false arrhythmia alarms can lead to disruption of care and slow response time due to desensitization of clinical staff. We study the use of machine learning models to detect false ventricular tachycardia (v-tach) alarms using ECG waveform recordings. We propose using a Supervised Denoising Autoencoder (SDAE) to detect false alarms using a low-dimensional representation of ECG dynamics learned by minimizing a combined reconstruction and classification loss. We evaluate our algorithms on the PhysioNet Challenge 2015 dataset, containing over 500 records (over 300 training and 200 testing) with v-tach alarms. Our results indicate that using the SDAE on Fast Fourier Transformed (FFT) ECG at a beat-by-beat level outperforms several competitive baselines on the task of v-tach false alarm classification. We show that it is important to exploit the underlying known physiological structure using beat-by-beat frequency distribution from multiple cardiac cycles of the ECG waveforms to obtain competitive results and improve over previous entries from the 2015 PhysioNet Challenge.

Improving Sepsis Treatment Strategies by Combining Deep and Kernel-Based Reinforcement Learning.

Sepsis is the leading cause of mortality in the ICU. It is challenging to manage because individual patients respond differently to treatment. Thus, tailoring treatment to the individual patient is essential for the best outcomes. In this paper, we take steps toward this goal by applying a mixture-of-experts framework to personalize sepsis treatment. The mixture model selectively alternates between neighbor-based (kernel) and deep reinforcement learning (DRL) experts depending on patient's current history. On a large retrospective cohort, this mixture-based approach outperforms physician, kernel only, and DRL-only experts.

Phenotyping Hypotensive Patients in Critical Care Using Hospital Discharge Summaries.

Among critically-ill patients, hypotension represents a failure in compensatory mechanisms and may lead to organ hypoperfusion and failure. In this work, we adopt a data-driven approach for phenotype discovery and visualization of patient similarity and cohort structure in the intensive care unit (ICU). We used Hierarchical Dirichlet Process (HDP) as a nonparametric topic modeling technique to automatically learn a d-dimensional feature representation of patients that captures the latent "topic" structure of diseases, symptoms, medications, and findings documented in hospital discharge summaries. We then used the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm to convert the d-dimensional latent structure learned from HDP into a matrix of pairwise similarities for visualizing patient similarity and cohort structure. Using discharge summaries of a large patient cohort from the MIMIC II database, we evaluated the clinical utility of the discovered topic structure in phenotyping critically-ill patients who experienced hypotensive episodes. Our results indicate that the approach is able to reveal clinically interpretable clustering structure within our cohort and may potentially provide valuable insights to better understand the association between disease phenotypes and outcomes.