Early prediction of hemodynamic interventions in the intensive care unit using machine learning.

BACKGROUND: Timely recognition of hemodynamic instability in critically ill patients enables increased vigilance and early treatment opportunities. We develop the Hemodynamic Stability Index (HSI), which highlights situational awareness of possible hemodynamic instability occurring at the bedside and to prompt assessment for potential hemodynamic interventions. METHODS: We used an ensemble of decision trees to obtain a real-time risk score that predicts the initiation of hemodynamic interventions an hour into the future. We developed the model using the eICU Research Institute (eRI) database, based on adult ICU admissions from 2012 to 2016. A total of 208,375 ICU stays met the inclusion criteria, with 32,896 patients (prevalence = 18%) experiencing at least one instability event where they received one of the interventions during their stay. Predictors included vital signs, laboratory measurements, and ventilation settings. RESULTS: HSI showed significantly better performance compared to single parameters like systolic blood pressure and shock index (heart rate/systolic blood pressure) and showed good generalization across patient subgroups. HSI AUC was 0.82 and predicted 52% of all hemodynamic interventions with a lead time of 1-h with a specificity of 92%. In addition to predicting future hemodynamic interventions, our model provides confidence intervals and a ranked list of clinical features that contribute to each prediction. Importantly, HSI can use a sparse set of physiologic variables and abstains from making a prediction when the confidence is below an acceptable threshold. CONCLUSIONS: The HSI algorithm provides a single score that summarizes hemodynamic status in real time using multiple physiologic parameters in patient monitors and electronic medical records (EMR). Importantly, HSI is designed for real-world deployment, demonstrating generalizability, strong performance under different data availability conditions, and providing model explanation in the form of feature importance and prediction confidence.

A multimodal encoding model applied to imaging decision-related neural cascades in the human brain.

Perception and cognition in the brain are naturally characterized as spatiotemporal processes. Decision-making, for example, depends on coordinated patterns of neural activity cascading across the brain, running in time from stimulus to response and in space from primary sensory regions to the frontal lobe. Measuring this cascade is key to developing an understanding of brain function. Here we report on a novel methodology that employs multi-modal imaging for inferring this cascade in humans at unprecedented spatiotemporal resolution. Specifically, we develop an encoding model to link simultaneously measured electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) signals to infer high-resolution spatiotemporal brain dynamics during a perceptual decision. After demonstrating replication of results from the literature, we report previously unobserved sequential reactivation of a substantial fraction of the pre-response network whose magnitude correlates with a proxy for decision confidence. Our encoding model, which temporally tags BOLD activations using time localized EEG variability, identifies a coordinated and spatially distributed neural cascade that is associated with a perceptual decision. In general the methodology illuminates complex brain dynamics that would otherwise be unobservable using fMRI or EEG acquired separately.

Densely connected convolutional networks for detection of atrial fibrillation from short single-lead ECG recordings.

The development of new technology such as wearables that record high-quality single channel ECG, provides an opportunity for ECG screening in a larger population, especially for atrial fibrillation screening. The main goal of this study is to develop an automatic classification algorithm for normal sinus rhythm (NSR), atrial fibrillation (AF), other rhythms (O), and noise from a single channel short ECG segment (9-60 s). For this purpose, we combined a signal quality index (SQI) algorithm, to assess noisy instances, and trained densely connected convolutional neural networks to classify ECG recordings. Two convolutional neural network (CNN) models (a main model that accepts 15 s ECG segments and a secondary model that processes shorter 9 s segments) were trained using the training data set. If the recording is determined to be of low quality by SQI, it is immediately classified as noisy. Otherwise, it is transformed to a time-frequency representation and classified with the CNN as NSR, AF, O, or noise. The results achieved on the 2017 PhysioNet/Computing in Cardiology challenge test dataset were an overall F1 score of 0.82 (F1 for NSR, AF, and O were 0.91, 0.83, and 0.72, respectively). Compared with 80 challenge entries, this was the third best overall score achieved on the evaluation dataset.

A clinical prediction model to identify patients at high risk of hemodynamic instability in the pediatric intensive care unit.

BACKGROUND: Early recognition and timely intervention are critical steps for the successful management of shock. The objective of this study was to develop a model to predict requirement for hemodynamic intervention in the pediatric intensive care unit (PICU); thus, clinicians can direct their care to patients likely to benefit from interventions to prevent further deterioration. METHODS: The model proposed in this study was trained on a retrospective cohort of all patients admitted to a tertiary PICU at a single center in the United States, and validated on another retrospective cohort of all patients admitted to the PICU at a single center in the United Kingdom. The PICU clinical information system database (Intellivue Clinical Information Portfolio, Philips, UK) was interrogated to collect physiological and laboratory data. The model was trained using a variant of AdaBoost, which learned a set of low-dimensional classifiers, each of which was age adjusted. RESULTS: A total of 7052 patients admitted to the US PICU was used for training the model, and a total of 970 patients admitted to the UK PICU was used for validation. On the training/validation datasets, the model showed better prediction of hemodynamic intervention (area under the receiver operating characteristic (AUROC) = 0.81/0.81) than systolic blood pressure-based (AUCROC = 0.58/0.67) or shock index-based (AUCROC = 0.63/0.65) models. Both of these models were age adjusted using the same classifier. CONCLUSIONS: The proposed model reliably predicted the need for hemodynamic intervention in PICU patients and provides better classification performance when compared to systolic blood pressure-based or shock index-based models alone. This model could readily be built into a clinical information system to identify patients at risk of hemodynamic instability.

Inter-subject alignment of human cortical anatomy using functional connectivity.

Inter-subject alignment of functional MRI (fMRI) data is necessary for group analyses. The standard approach to this problem matches anatomical features of the brain, such as major anatomical landmarks or cortical curvature. Precise alignment of functional cortical topographies, however, cannot be derived using only anatomical features. We propose a new inter-subject registration algorithm that aligns intra-subject patterns of functional connectivity across subjects. We derive functional connectivity patterns by correlating fMRI BOLD time-series, measured during movie viewing, between spatially remote cortical regions. We validate our technique extensively on real fMRI experimental data and compare our method to two state-of-the-art inter-subject registration algorithms. By cross-validating our method on independent datasets, we show that the derived alignment generalizes well to other experimental paradigms.

Machine learning-based clinical decision support for infection risk prediction.

Background: Healthcare-associated infection (HAI) remains a significant risk for hospitalized patients and a challenging burden for the healthcare system. This study presents a clinical decision support tool that can be used in clinical workflows to proactively engage secondary assessments of pre-symptomatic and at-risk infection patients, thereby enabling earlier diagnosis and treatment. Methods: This study applies machine learning, specifically ensemble-based boosted decision trees, on large retrospective hospital datasets to develop an infection risk score that predicts infection before obvious symptoms present. We extracted a stratified machine learning dataset of 36,782 healthcare-associated infection patients. The model leveraged vital signs, laboratory measurements and demographics to predict HAI before clinical suspicion, defined as the order of a microbiology test or administration of antibiotics. Results: Our best performing infection risk model achieves a cross-validated AUC of 0.88 at 1 h before clinical suspicion and maintains an AUC >0.85 for 48 h before suspicion by aggregating information across demographics and a set of 163 vital signs and laboratory measurements. A second model trained on a reduced feature space comprising demographics and the 36 most frequently measured vital signs and laboratory measurements can still achieve an AUC of 0.86 at 1 h before clinical suspicion. These results compare favorably against using temperature alone and clinical rules such as the quick sequential organ failure assessment (qSOFA) score. Along with the performance results, we also provide an analysis of model interpretability via feature importance rankings. Conclusion: The predictive model aggregates information from multiple physiological parameters such as vital signs and laboratory measurements to provide a continuous risk score of infection that can be deployed in hospitals to provide advance warning of patient deterioration.

Continuous noninvasive blood gas estimation in critically ill pediatric patients with respiratory failure.

Patients supported by mechanical ventilation require frequent invasive blood gas samples to monitor and adjust the level of support. We developed a transparent and novel blood gas estimation model to provide continuous monitoring of blood pH and arterial CO2 in between gaps of blood draws, using only readily available noninvasive data sources in ventilated patients. The model was trained on a derivation dataset (1,883 patients, 12,344 samples) from a tertiary pediatric intensive care center, and tested on a validation dataset (286 patients, 4030 samples) from the same center obtained at a later time. The model uses pairwise non-linear interactions between predictors and provides point-estimates of blood gas pH and arterial CO2 along with a range of prediction uncertainty. The model predicted within Clinical Laboratory Improvement Amendments of 1988 (CLIA) acceptable blood gas machine equivalent in 74% of pH samples and 80% of PCO2 samples. Prediction uncertainty from the model improved estimation accuracy by 15% by identifying and abstaining on a minority of high-uncertainty samples. The proposed model estimates blood gas pH and CO2 accurately in a large percentage of samples. The model's abstention recommendation coupled with ranked display of top predictors for each estimation lends itself to real-time monitoring of gaps between blood draws, and the model may help users determine when a new blood draw is required and delay blood draws when not needed.

Real-time infection prediction with wearable physiological monitoring and AI to aid military workforce readiness during COVID-19.

Infectious threats, like the COVID-19 pandemic, hinder maintenance of a productive and healthy workforce. If subtle physiological changes precede overt illness, then proactive isolation and testing can reduce labor force impacts. This study hypothesized that an early infection warning service based on wearable physiological monitoring and predictive models created with machine learning could be developed and deployed. We developed a prototype tool, first deployed June 23, 2020, that delivered continuously updated scores of infection risk for SARS-CoV-2 through April 8, 2021. Data were acquired from 9381 United States Department of Defense (US DoD) personnel wearing Garmin and Oura devices, totaling 599,174 user-days of service and 201 million hours of data. There were 491 COVID-19 positive cases. A predictive algorithm identified infection before diagnostic testing with an AUC of 0.82. Barriers to implementation included adequate data capture (at least 48% data was needed) and delays in data transmission. We observe increased risk scores as early as 6 days prior to diagnostic testing (2.3 days average). This study showed feasibility of a real-time risk prediction score to minimize workforce impacts of infection.

Function-based intersubject alignment of human cortical anatomy.

Making conclusions about the functional neuroanatomical organization of the human brain requires methods for relating the functional anatomy of an individual's brain to population variability. We have developed a method for aligning the functional neuroanatomy of individual brains based on the patterns of neural activity that are elicited by viewing a movie. Instead of basing alignment on functionally defined areas, whose location is defined as the center of mass or the local maximum response, the alignment is based on patterns of response as they are distributed spatially both within and across cortical areas. The method is implemented in the two-dimensional manifold of an inflated, spherical cortical surface. The method, although developed using movie data, generalizes successfully to data obtained with another cognitive activation paradigm--viewing static images of objects and faces--and improves group statistics in that experiment as measured by a standard general linear model (GLM) analysis.

Analyzing single-lead short ECG recordings using dense convolutional neural networks and feature-based post-processing to detect atrial fibrillation.

OBJECTIVE: The prevalence of atrial fibrillation (AF) in the general population is 0.5%-1%. As AF is the most common sustained cardiac arrhythmia that is associated with an increased morbidity and mortality, its timely diagnosis is clinically desirable. The main aim of this study as our contribution to the PhysioNet/CinC Challenge 2017 was to develop an automatic algorithm for classification of normal sinus rhythm (NSR), AF, other rhythm (O), and noise using a short single-channel ECG. Furthermore, the impact of changing labels/annotations on performance of the proposed algorithm was studied in this article. APPROACH: The challenge training dataset (8528 ECG recordings) and a complementary dataset (6312 ECG recordings) from other sources were used for algorithm development. Version 3 (v3), which is an updated version of the annotations at the official phase of the challenge (v2), was used in this study. In the proposed algorithm, densely connected convolutional networks were combined with feature-based post-processing after initial signal quality analysis for the classification of ECG recordings. MAIN RESULTS: The F1 scores for classification of NSR, AF, and O were 0.91, 0.83, and 0.72, respectively, which led to a F1 of 0.82. There was a small or no performance difference between the top entries in the official phase of the challenge and our proposed method. An increase of 2.5% in F1 score was observed when the same annotations for training and test was used (using v3 annotations) compared to using different annotations (v2 annotations for training and v3 annotations for the test). SIGNIFICANCE: Our promising results suggest that the availability of more data with improved labeling along with improvement in signal quality analysis make our algorithm suitable for practical clinical applications.

Fast bootstrapping and permutation testing for assessing reproducibility and interpretability of multivariate fMRI decoding models.

Multivariate decoding models are increasingly being applied to functional magnetic imaging (fMRI) data to interpret the distributed neural activity in the human brain. These models are typically formulated to optimize an objective function that maximizes decoding accuracy. For decoding models trained on full-brain data, this can result in multiple models that yield the same classification accuracy, though some may be more reproducible than others--i.e. small changes to the training set may result in very different voxels being selected. This issue of reproducibility can be partially controlled by regularizing the decoding model. Regularization, along with the cross-validation used to estimate decoding accuracy, typically requires retraining many (often on the order of thousands) of related decoding models. In this paper we describe an approach that uses a combination of bootstrapping and permutation testing to construct both a measure of cross-validated prediction accuracy and model reproducibility of the learned brain maps. This requires re-training our classification method on many re-sampled versions of the fMRI data. Given the size of fMRI datasets, this is normally a time-consuming process. Our approach leverages an algorithm called fast simultaneous training of generalized linear models (FaSTGLZ) to create a family of classifiers in the space of accuracy vs. reproducibility. The convex hull of this family of classifiers can be used to identify a subset of Pareto optimal classifiers, with a single-optimal classifier selectable based on the relative cost of accuracy vs. reproducibility. We demonstrate our approach using full-brain analysis of elastic-net classifiers trained to discriminate stimulus type in an auditory and visual oddball event-related fMRI design. Our approach and results argue for a computational approach to fMRI decoding models in which the value of the interpretation of the decoding model ultimately depends upon optimizing a joint space of accuracy and reproducibility.

Fast, Exact Model Selection and Permutation Testing for ℓ2-Regularized Logistic Regression.

Regularized logistic regression is a standard classification method used in statistics and machine learning. Unlike regularized least squares problems such as ridge regression, the parameter estimates cannot be computed in closed-form and instead must be estimated using an iterative technique. This paper addresses the computational problem of regularized logistic regression that is commonly encountered in model selection and classifier statistical significance testing, in which a large number of related logistic regression problems must be solved for. Our proposed approach solves the problems simultaneously through an iterative technique, which also garners computational efficiencies by leveraging the redundancies across the related problems. We demonstrate analytically that our method provides a substantial complexity reduction, which is further validated by our results on real-world datasets.

A common, high-dimensional model of the representational space in human ventral temporal cortex.

We present a high-dimensional model of the representational space in human ventral temporal (VT) cortex in which dimensions are response-tuning functions that are common across individuals and patterns of response are modeled as weighted sums of basis patterns associated with these response tunings. We map response-pattern vectors, measured with fMRI, from individual subjects' voxel spaces into this common model space using a new method, "hyperalignment." Hyperalignment parameters based on responses during one experiment--movie viewing--identified 35 common response-tuning functions that captured fine-grained distinctions among a wide range of stimuli in the movie and in two category perception experiments. Between-subject classification (BSC, multivariate pattern classification based on other subjects' data) of response-pattern vectors in common model space greatly exceeded BSC of anatomically aligned responses and matched within-subject classification. Results indicate that population codes for complex visual stimuli in VT cortex are based on response-tuning functions that are common across individuals.

fMRI-Based Inter-Subject Cortical Alignment Using Functional Connectivity.

The inter-subject alignment of functional MRI (fMRI) data is important for improving the statistical power of fMRI group analyses. In contrast to existing anatomically-based methods, we propose a novel multi-subject algorithm that derives a functional correspondence by aligning spatial patterns of functional connectivity across a set of subjects. We test our method on fMRI data collected during a movie viewing experiment. By cross-validating the results of our algorithm, we show that the correspondence successfully generalizes to a secondary movie dataset not used to derive the alignment.