Leveraging Eye Tracking to Prioritize Relevant Medical Record Data: Comparative Machine Learning Study.

BACKGROUND: Electronic medical record (EMR) systems capture large amounts of data per patient and present that data to physicians with little prioritization. Without prioritization, physicians must mentally identify and collate relevant data, an activity that can lead to cognitive overload. To mitigate cognitive overload, a Learning EMR (LEMR) system prioritizes the display of relevant medical record data. Relevant data are those that are pertinent to a context-defined as the combination of the user, clinical task, and patient case. To determine which data are relevant in a specific context, a LEMR system uses supervised machine learning models of physician information-seeking behavior. Since obtaining information-seeking behavior data via manual annotation is slow and expensive, automatic methods for capturing such data are needed. OBJECTIVE: The goal of the research was to propose and evaluate eye tracking as a high-throughput method to automatically acquire physician information-seeking behavior useful for training models for a LEMR system. METHODS: Critical care medicine physicians reviewed intensive care unit patient cases in an EMR interface developed for the study. Participants manually identified patient data that were relevant in the context of a clinical task: preparing a patient summary to present at morning rounds. We used eye tracking to capture each physician's gaze dwell time on each data item (eg, blood glucose measurements). Manual annotations and gaze dwell times were used to define target variables for developing supervised machine learning models of physician information-seeking behavior. We compared the performance of manual selection and gaze-derived models on an independent set of patient cases. RESULTS: A total of 68 pairs of manual selection and gaze-derived machine learning models were developed from training data and evaluated on an independent evaluation data set. A paired Wilcoxon signed-rank test showed similar performance of manual selection and gaze-derived models on area under the receiver operating characteristic curve (P=.40). CONCLUSIONS: We used eye tracking to automatically capture physician information-seeking behavior and used it to train models for a LEMR system. The models that were trained using eye tracking performed like models that were trained using manual annotations. These results support further development of eye tracking as a high-throughput method for training clinical decision support systems that prioritize the display of relevant medical record data.

Using machine learning to selectively highlight patient information.

BACKGROUND: Electronic medical record (EMR) systems need functionality that decreases cognitive overload by drawing the clinician's attention to the right data, at the right time. We developed a Learning EMR (LEMR) system that learns statistical models of clinician information-seeking behavior and applies those models to direct the display of data in future patients. We evaluated the performance of the system in identifying relevant patient data in intensive care unit (ICU) patient cases. METHODS: To capture information-seeking behavior, we enlisted critical care medicine physicians who reviewed a set of patient cases and selected data items relevant to the task of presenting at morning rounds. Using patient EMR data as predictors, we built machine learning models to predict their relevancy. We prospectively evaluated the predictions of a set of high performing models. RESULTS: On an independent evaluation data set, 25 models achieved precision of 0.52, 95% CI [0.49, 0.54] and recall of 0.77, 95% CI [0.75, 0.80] in identifying relevant patient data items. For data items missed by the system, the reviewers rated the effect of not seeing those data from no impact to minor impact on patient care in about 82% of the cases. CONCLUSION: Data-driven approaches for adaptively displaying data in EMR systems, like the LEMR system, show promise in using information-seeking behavior of clinicians to identify and highlight relevant patient data.

Outlier-based detection of unusual patient-management actions: An ICU study.

Medical errors remain a significant problem in healthcare. This paper investigates a data-driven outlier-based monitoring and alerting framework that uses data in the Electronic Medical Records (EMRs) repositories of past patient cases to identify any unusual clinical actions in the EMR of a current patient. Our conjecture is that these unusual clinical actions correspond to medical errors often enough to justify their detection and alerting. Our approach works by using EMR repositories to learn statistical models that relate patient states to patient-management actions. We evaluated this approach on the EMR data for 24,658 intensive care unit (ICU) patient cases. A total of 16,500 cases were used to train statistical models for ordering medications and laboratory tests given the patient state summarizing the patient's clinical history. The models were applied to a separate test set of 8158 ICU patient cases and used to generate alerts. A subset of 240 alerts generated by the models were evaluated and assessed by eighteen ICU clinicians. The overall true positive rates for the alerts (TPARs) ranged from 0.44 to 0.71. The TPAR for medication order alerts specifically ranged from 0.31 to 0.61 and for laboratory order alerts from 0.44 to 0.75. These results support outlier-based alerting as a promising new approach to data-driven clinical alerting that is generated automatically based on past EMR data.

An Efficient Pattern Mining Approach for Event Detection in Multivariate Temporal Data.

This work proposes a pattern mining approach to learn event detection models from complex multivariate temporal data, such as electronic health records. We present Recent Temporal Pattern mining, a novel approach for efficiently finding predictive patterns for event detection problems. This approach first converts the time series data into time-interval sequences of temporal abstractions. It then constructs more complex time-interval patterns backward in time using temporal operators. We also present the Minimal Predictive Recent Temporal Patterns framework for selecting a small set of predictive and non-spurious patterns. We apply our methods for predicting adverse medical events in real-world clinical data. The results demonstrate the benefits of our methods in learning accurate event detection models, which is a key step for developing intelligent patient monitoring and decision support systems.

The influence of microbial colonization on inflammatory versus pro-healing trajectories in combat extremity wounds.

A combination of improved body armor, medical transportation, and treatment has led to the increased survival of warfighters from combat extremity injuries predominantly caused by blasts in modern conflicts. Despite advances, a high rate of complications such as wound infections, wound failure, amputations, and a decreased quality of life exist. To study the molecular underpinnings of wound failure, wound tissue biopsies from combat extremity injuries had RNA extracted and sequenced. Wounds were classified by colonization (colonized vs. non-colonized) and outcome (healed vs. failed) status. Differences in gene expression were investigated between timepoints at a gene level, and longitudinally by multi-gene networks, inferred proportions of immune cells, and expression of healing-related functions. Differences between wound outcomes in colonized wounds were more apparent than in non-colonized wounds. Colonized/healed wounds appeared able to mount an adaptive immune response to infection and progress beyond the inflammatory stage of healing, while colonized/failed wounds did not. Although, both colonized and non-colonized failed wounds showed increasing inferred immune and inflammatory programs, non-colonized/failed wounds progressed beyond the inflammatory stage, suggesting different mechanisms of failure dependent on colonization status. Overall, these data reveal gene expression profile differences in healing wounds that may be utilized to improve clinical treatment paradigms.

Learning classification models from multiple experts.

Building classification models from clinical data using machine learning methods often relies on labeling of patient examples by human experts. Standard machine learning framework assumes the labels are assigned by a homogeneous process. However, in reality the labels may come from multiple experts and it may be difficult to obtain a set of class labels everybody agrees on; it is not uncommon that different experts have different subjective opinions on how a specific patient example should be classified. In this work we propose and study a new multi-expert learning framework that assumes the class labels are provided by multiple experts and that these experts may differ in their class label assessments. The framework explicitly models different sources of disagreements and lets us naturally combine labels from different human experts to obtain: (1) a consensus classification model representing the model the group of experts converge to, as well as, and (2) individual expert models. We test the proposed framework by building a model for the problem of detection of the Heparin Induced Thrombocytopenia (HIT) where examples are labeled by three experts. We show that our framework is superior to multiple baselines (including standard machine learning framework in which expert differences are ignored) and that our framework leads to both improved consensus and individual expert models.

Outlier detection for patient monitoring and alerting.

We develop and evaluate a data-driven approach for detecting unusual (anomalous) patient-management decisions using past patient cases stored in electronic health records (EHRs). Our hypothesis is that a patient-management decision that is unusual with respect to past patient care may be due to an error and that it is worthwhile to generate an alert if such a decision is encountered. We evaluate this hypothesis using data obtained from EHRs of 4486 post-cardiac surgical patients and a subset of 222 alerts generated from the data. We base the evaluation on the opinions of a panel of experts. The results of the study support our hypothesis that the outlier-based alerting can lead to promising true alert rates. We observed true alert rates that ranged from 25% to 66% for a variety of patient-management actions, with 66% corresponding to the strongest outliers.

Learning EKG Diagnostic Models with Hierarchical Class Label Dependencies.

Electrocardiogram (EKG/ECG) is a key diagnostic tool to assess patient's cardiac condition and is widely used in clinical applications such as patient monitoring, surgery support, and heart medicine research. With recent advances in machine learning (ML) technology there has been a growing interest in the development of models supporting automatic EKG interpretation and diagnosis based on past EKG data. The problem can be modeled as multi-label classification (MLC), where the objective is to learn a function that maps each EKG reading to a vector of diagnostic class labels reflecting the underlying patient condition at different levels of abstraction. In this paper, we propose and investigate an ML model that considers class-label dependency embedded in the hierarchical organization of EKG diagnoses to improve the EKG classification performance. Our model first transforms the EKG signals into a low-dimensional vector, and after that uses the vector to predict different class labels with the help of the conditional tree structured Bayesian network (CTBN) that is able to capture hierarchical dependencies among class variables. We evaluate our model on the publicly available PTB-XL dataset. Our experiments demonstrate that modeling of hierarchical dependencies among class variables improves the diagnostic model performance under multiple classification performance metrics as compared to classification models that predict each class label independently.

Personalized event prediction for Electronic Health Records.

Clinical event sequences consist of hundreds of clinical events that represent records of patient care in time. Developing accurate predictive models of such sequences is of a great importance for supporting a variety of models for interpreting/classifying the current patient condition, or predicting adverse clinical events and outcomes, all aimed to improve patient care. One important challenge of learning predictive models of clinical sequences is their patient-specific variability. Based on underlying clinical conditions, each patient's sequence may consist of different sets of clinical events (observations, lab results, medications, procedures). Hence, simple population-wide models learned from event sequences for many different patients may not accurately predict patient-specific dynamics of event sequences and their differences. To address the problem, we propose and investigate multiple new event sequence prediction models and methods that let us better adjust the prediction for individual patients and their specific conditions. The methods developed in this work pursue refinement of population-wide models to subpopulations, self-adaptation, and a meta-level model switching that is able to adaptively select the model with the best chance to support the immediate prediction. We analyze and test the performance of these models on clinical event sequences of patients in MIMIC-III database.

Hierarchical Active Learning With Qualitative Feedback on Regions.

Learning classification models in practice usually requires numerous labeled data for training. However, instance-based annotation can be inefficient for humans to perform. In this article, we propose and study a new type of human supervision that is fast to perform and useful for model learning. Instead of labeling individual instances, humans provide supervision to data regions, which are subspaces of the input data space, representing subpopulations of data. Since labeling now is performed on a region level, 0/1 labeling becomes imprecise. Thus, we design the region label to be a qualitative assessment of the class proportion, which coarsely preserves the labeling precision but is also easy for humans to do. To identify informative regions for labeling and learning, we further devise a hierarchical active learning process that recursively constructs a region hierarchy. This process is semisupervised in the sense that it is driven by both active learning strategies and human expertise, where humans can provide discriminative features. To evaluate our framework, we conducted extensive experiments on nine datasets as well as a real user study on a survival analysis of colorectal cancer patients. The results have clearly demonstrated the superiority of our region-based active learning framework against many instance-based active learning methods.

Transcriptomic, Proteomic, and Morphologic Characterization of Healing in Volumetric Muscle Loss.

Skeletal muscle has a robust, inherent ability to regenerate in response to injury from acute to chronic. In severe trauma, however, complete regeneration is not possible, resulting in a permanent loss of skeletal muscle tissue referred to as volumetric muscle loss (VML). There are few consistently reliable therapeutic or surgical options to address VML. A major limitation in investigation of possible therapies is the absence of a well-characterized large animal model. In this study, we present results of a comprehensive transcriptomic, proteomic, and morphologic characterization of wound healing following VML in a novel canine model of VML which we compare to a nine-patient cohort of combat-associated VML. The canine model is translationally relevant as it provides both a regional (spatial) and temporal map of the wound healing processes that occur in human VML. Collectively, these data show the spatiotemporal transcriptomic, proteomic, and morphologic properties of canine VML healing as a framework and model system applicable to future studies investigating novel therapies for human VML. Impact Statement The spatiotemporal transcriptomic, proteomic, and morphologic properties of canine volumetric muscle loss (VML) healing is a translational framework and model system applicable to future studies investigating novel therapies for human VML.

Event Outlier Detection in Continuous Time.

Continuous-time event sequences represent discrete events occurring in continuous time. Such sequences arise frequently in real-life. Usually we expect the sequences to follow some regular pattern over time. However, sometimes these patterns may be interrupted by unexpected absence or occurrences of events. Identification of these unexpected cases can be very important as they may point to abnormal situations that need human attention. In this work, we study and develop methods for detecting outliers in continuous-time event sequences, including unexpected absence and unexpected occurrences of events. Since the patterns that event sequences tend to follow may change in different contexts, we develop outlier detection methods based on point processes that can take context information into account. Our methods are based on Bayesian decision theory and hypothesis testing with theoretical guarantees. To test the performance of the methods, we conduct experiments on both synthetic data and real-world clinical data and show the effectiveness of the proposed methods.

Neural Clinical Event Sequence Prediction through Personalized Online Adaptive Learning.

Clinical event sequences consist of thousands of clinical events that represent records of patient care in time. Developing accurate prediction models for such sequences is of a great importance for defining representations of a patient state and for improving patient care. One important challenge of learning a good predictive model of clinical sequences is patient-specific variability. Based on underlying clinical complications, each patient's sequence may consist of different sets of clinical events. However, population-based models learned from such sequences may not accurately predict patient-specific dynamics of event sequences. To address the problem, we develop a new adaptive event sequence prediction framework that learns to adjust its prediction for individual patients through an online model update.

Improving Prediction of Low-Prior Clinical Events with Simultaneous General Patient-State Representation Learning.

Low-prior targets are common among many important clinical events, which introduces the challenge of having enough data to support learning of their predictive models. Many prior works have addressed this problem by first building a general patient-state representation model, and then adapting it to a new low-prior prediction target. In this schema, there is potential for the predictive performance to be hindered by the misalignment between the general patient-state model and the target task. To overcome this challenge, we propose a new method that simultaneously optimizes a shared model through multi-task learning of both the low-prior supervised target and general purpose patient-state representation (GPSR). More specifically, our method improves prediction performance of a low-prior task by jointly optimizing a shared model that combines the loss of the target event and a broad range of generic clinical events. We study the approach in the context of Recurrent Neural Networks (RNNs). Through extensive experiments on multiple clinical event targets using MIMIC-III [8] data, we show that the inclusion of general patient-state representation tasks during model training improves the prediction of individual low-prior targets.

Modeling multivariate clinical event time-series with recurrent temporal mechanisms.

In this work, we propose a novel autoregressive event time-series model that can predict future occurrences of multivariate clinical events. Our model represents multivariate event time-series using different temporal mechanisms aimed to fit different temporal characteristics of the time-series. In particular, information about distant past is modeled through the hidden state space defined by an LSTM-based model, information on recently observed clinical events is modeled through discriminative projections, and information about periodic (repeated) events is modeled using a special recurrent mechanism based on probability distributions of inter-event gaps compiled from past data. We evaluate our proposed model on electronic health record (EHRs) data derived from MIMIC-III dataset. We show that our new model equipped with the above temporal mechanisms leads to improved prediction performance compared to multiple baselines.

Feature importance analysis for patient management decisions.

The objective of this paper is to understand what characteristics and features of clinical data influence physician's decision about ordering laboratory tests or prescribing medications the most. We conduct our analysis on data and decisions extracted from electronic health records of 4486 post-surgical cardiac patients. The summary statistics for 335 different lab order decisions and 407 medication decisions are reported. We show that in many cases, physician's lab-order and medication decisions are predicted well by simple patterns such as last value of a single test result, time since a certain lab test was ordered or time since certain procedure was executed.

Hierarchical Active Learning with Overlapping Regions.

Learning of classification models from real-world data often requires substantial human effort devoted to instance annotation. As this process can be very time-consuming and costly, finding effective ways to reduce the annotation cost becomes critical for building such models. To address this problem we explore a new type of human feedback - region-based feedback. Briefly, a region is defined as a hypercubic subspace of the input data space and represents a subpopulation of data instances; the region's label is a human assessment of the class proportion of the data subpopulation. By using learning from label proportions algorithms one can learn instance-based classifiers from such labeled regions. In general, the key challenge is that there can be infinite many regions one can define and query in a given data space. To minimize the number and complexity of region-based queries, we propose and develop a hierarchical active learning solution that aims at incrementally building a concise hierarchy of regions. Furthermore, to avoid building a possibly class-irrelevant region hierarchy, we further propose to grow multiple different hierarchies in parallel and expand those more informative hierarchies. Through experiments on numerous data sets, we demonstrate that methods using region-based feedback can learn very good classifiers from very few and simple queries, and hence are highly effective in reducing human annotation effort needed for building classification models.

Monitoring ICU Mortality Risk with A Long Short-Term Memory Recurrent Neural Network.

In intensive care units (ICU), mortality prediction is a critical factor not only for effective medical intervention but also for allocation of clinical resources. Structured electronic health records (EHR) contain valuable information for assessing mortality risk in ICU patients, but current mortality prediction models usually require laborious human-engineered features. Furthermore, substantial missing data in EHR is a common problem for both the construction and implementation of a prediction model.Inspired by language-related models, we design a new framework for dynamic monitoring of patients' mortality risk. Our framework uses the bag-of-words representation for all relevant medical events based on most recent history as inputs. By design, it is robust to missing data in EHR and can be easily implemented as an instant scoring system to monitor the medical development of all ICU patients. Specifically, our model uses latent semantic analysis (LSA) to encode the patients' states into low-dimensional embeddings, which are further fed to long short-term memory networks for mortality risk prediction. Our results show that the deep learning based framework performs better than the existing severity scoring system, SAPS-II. We observe that bidirectional long short-term memory demonstrates superior performance, probably due to the successful capture of both forward and backward temporal dependencies.

Modeling physician variability to prioritize relevant medical record information.

OBJECTIVE: Patient information can be retrieved more efficiently in electronic medical record (EMR) systems by using machine learning models that predict which information a physician will seek in a clinical context. However, information-seeking behavior varies across EMR users. To explicitly account for this variability, we derived hierarchical models and compared their performance to nonhierarchical models in identifying relevant patient information in intensive care unit (ICU) cases. MATERIALS AND METHODS: Critical care physicians reviewed ICU patient cases and selected data items relevant for presenting at morning rounds. Using patient EMR data as predictors, we derived hierarchical logistic regression (HLR) and standard logistic regression (LR) models to predict their relevance. RESULTS: In 73 pairs of HLR and LR models, the HLR models achieved an area under the receiver operating characteristic curve of 0.81, 95% confidence interval (CI) [0.80-0.82], which was statistically significantly higher than that of LR models (0.75, 95% CI [0.74-0.76]). Further, the HLR models achieved statistically significantly lower expected calibration error (0.07, 95% CI [0.06-0.08]) than LR models (0.16, 95% CI [0.14-0.17]). DISCUSSION: The physician reviewers demonstrated variability in selecting relevant data. Our results show that HLR models perform significantly better than LR models with respect to both discrimination and calibration. This is likely due to explicitly modeling physician-related variability. CONCLUSION: Hierarchical models can yield better performance when there is physician-related variability as in the case of identifying relevant information in the EMR.

Hierarchical Adaptive Multi-task Learning Framework for Patient Diagnoses and Diagnostic Category Classification.

The problems a patient suffers from can be summarized in terms of a list of patient diagnoses. The diagnoses are typically organized in a hierarchy (or a lattice structure) in which many different low-level diagnoses are covered by one or more diagnostic categories. An interesting machine learning problem is related to learning of a wide range of diagnostic models (at different levels of abstraction) that can automatically assign a diagnosis or a diagnostic category to a specific patient. While one can always approach this problem by learning models for each diagnostic task independently, an interesting open question is how one can leverage the knowledge of a diagnostic hierarchy to improve the classification and outperform independent diagnostic models. In this work, we study this problem by designing a new hierarchical classification learning framework in which multiple diagnostic classification targets are explicitly related via diagnostic hierarchy relations. By conducting experiments on MIMIC-III data and ICD-9 diagnosis hierarchy, we demonstrate that our framework leads to improved classification performance on individual diagnostic tasks when compared to independently learned diagnostic models. This improvement is stronger for diagnoses with a low prior and smaller number of positive training examples.