OBJECTIVE: To develop an electronic medical record (EMR) data processing tool that confers clinical context to machine learning (ML) algorithms for error handling, bias mitigation, and interpretability. MATERIALS AND METHODS: We present Trust-MAPS, an algorithm that translates clinical domain knowledge into high-dimensional, mixed-integer programming models that capture physiological and biological constraints on clinical measurements. EMR data are projected onto this constrained space, effectively bringing outliers to fall within a physiologically feasible range. We then compute the distance of each data point from the constrained space modeling healthy physiology to quantify deviation from the norm. These distances, termed "trust-scores," are integrated into the feature space for downstream ML applications. We demonstrate the utility of Trust-MAPS by training a binary classifier for early sepsis prediction on data from the 2019 PhysioNet Computing in Cardiology Challenge, using the XGBoost algorithm and applying SMOTE for overcoming class-imbalance. RESULTS: The Trust-MAPS framework shows desirable behavior in handling potential errors and boosting predictive performance. We achieve an area under the receiver operating characteristic curve of 0.91 (95% CI, 0.89-0.92) for predicting sepsis 6 hours before onset-a marked 15% improvement over a baseline model trained without Trust-MAPS. DISCUSSIONS: Downstream classification performance improves after Trust-MAPS preprocessing, highlighting the bias reducing capabilities of the error-handling projections. Trust-scores emerge as clinically meaningful features that not only boost predictive performance for clinical decision support tasks but also lend interpretability to ML models. CONCLUSION: This work is the first to translate clinical domain knowledge into mathematical constraints, model cross-vital dependencies, and identify aberrations in high-dimensional medical data. Our method allows for error handling in EMR and confers interpretability and superior predictive power to models trained for clinical decision support.

Segmenting lung regions in ICU Chest X-rays (CXR's) is vital for diagnosing lung-related disorders, but existing methods require extensive annotations or training on large datasets. We present LuGSAM, a novel framework that integrates text prompts with the Segment Anything Model (SAM) for segmentation tasks, enhancing precision and adaptability in clinical settings. Our approach combines Grounding DINO, a zero-shot object detector using textual prompts (e.g., "right lobe"), and Meta AI's SAM. Grounding DINO generates bounding boxes based on word-level prompts. These bounding boxes serve as an input to SAM, to generate precise segmentation masks. To further improve accuracy, we propose an iterative bounding box adjustment algorithm that refines object detections through multiple iterations. The Vision Transformer huge (Vit-h) variant of SAM achieved the highest overlap score (IoU = 0.95) for right lung segmentation. Grounding DINO demonstrated high detection accuracy for prompts like "right lung" with a confidence score of 0.58. The Binarized Predicted IoU (BPIoU) metric showed significant improvements in segmentation quality, making this framework a promising tool for clinical applications.

Acute respiratory distress syndrome (ARDS) is a severe organ dysfunction associated with significant mortality and morbidity among critically ill patients admitted to the Intensive Care Unit (ICU). The etiology related to ARDS can be highly heterogeneous, with infection or trauma as the most common associations. The Berlin criteria, the current gold standard for ARDS diagnosis, often necessitates manual adjudication of chest radiographs, limiting automation tools. ARDS diagnosis relies on the presence of bilateral infiltrates on radiographs, which is often not available in Electronic Medical Records (EMRs). Automated identification of radiological evidence would facilitate a comprehensive study of the syndrome, eliminating the need for costly individual image inspections by physicians. Radiological reports enable Natural Language Processing (NLP) to assess lung status and evaluate imaging criteria. We developed a NLP pipeline to analyze radiology notes of 362 patients satisfying sepsis-3 criteria from the EMR for possible ARDS diagnosis using BERT model for classification. These classification models showed F1-score of 74.5% and 64.22% for Emory and Grady dataset respectively.

Modern health care systems are conducting continuous, automated surveillance of the electronic medical record (EMR) to identify adverse events with increasing frequency; however, many events such as sepsis do not have elucidated prodromes (i.e., event chains) that can be used to identify and intercept the adverse event early in its course. Clinically relevant and interpretable results require a framework that can (i) infer temporal interactions across multiple patient features found in EMR data (e.g., Labs, vital signs, etc.) and (ii) identify patterns that precede and are specific to an impending adverse event (e.g., sepsis). In this work, we propose a linear multivariate Hawkes process model, coupled with ReLU link function, to recover a Granger Causal (GC) graph with both exciting and inhibiting effects. We develop a scalable two-phase gradient-based method to obtain a maximum surrogate-likelihood estimator, which is shown to be effective via extensive numerical simulation. Our method is subsequently extended to a data set of patients admitted to Grady hospital system in Atlanta, GA, USA, where the estimated GC graph identifies several highly interpretable GC chains that precede sepsis. The code is available at https://github.com/SongWei-GT/two-phase-MHP.

OBJECTIVE: To design and validate a novel deep generative model for seismocardiogram (SCG) dataset augmentation. SCG is a noninvasively acquired cardiomechanical signal used in a wide range of cardivascular monitoring tasks; however, these approaches are limited due to the scarcity of SCG data. METHODS: A deep generative model based on transformer neural networks is proposed to enable SCG dataset augmentation with control over features such as aortic opening (AO), aortic closing (AC), and participant-specific morphology. We compared the generated SCG beats to real human beats using various distribution distance metrics, notably Sliced-Wasserstein Distance (SWD). The benefits of dataset augmentation using the proposed model for other machine learning tasks were also explored. RESULTS: Experimental results showed smaller distribution distances for all metrics between the synthetically generated set of SCG and a test set of human SCG, compared to distances from an animal dataset (1.14× SWD), Gaussian noise (2.5× SWD), or other comparison sets of data. The input and output features also showed minimal error (95% limits of agreement for pre-ejection period [PEP] and left ventricular ejection time [LVET] timings are 0.03 ± 3.81 ms and -0.28 ± 6.08 ms, respectively). Experimental results for data augmentation for a PEP estimation task showed 3.3% accuracy improvement on an average for every 10% augmentation (ratio of synthetic data to real data). CONCLUSION: The model is thus able to generate physiologically diverse, realistic SCG signals with precise control over AO and AC features. This will uniquely enable dataset augmentation for SCG processing and machine learning to overcome data scarcity.

Chronic respiratory diseases affect millions and are leading causes of death in the US and worldwide. Pulmonary auscultation provides clinicians with critical respiratory health information through the study of Lung Sounds (LS) and the context of the breathing-phase and chest location in which they are measured. Existing auscultation technologies, however, do not enable the simultaneous measurement of this context, thereby potentially limiting computerized LS analysis. In this work, LS and Impedance Pneumography (IP) measurements were obtained from 10 healthy volunteers while performing normal and forced-expiratory (FE) breathing maneuvers using our wearable IP and respiratory sounds (WIRS) system. Simultaneous auscultation was performed with the Eko CORE stethoscope (EKO). The breathing-phase context was extracted from the IP signals and used to compute phase-by-phase (Inspiratory (I), expiratory (E), and their ratio (I:E)) and breath-by-breath acoustic features. Their individual and added value was then elucidated through machine learning analysis. We found that the phase-contextualized features effectively captured the underlying acoustic differences between deep and FE breaths, yielding a maximum F1 Score of 84.1 ±11.4% with the phase-by-phase features as the strongest contributors to this performance. Further, the individual phase-contextualized models outperformed the traditional breath-by-breath models in all cases. The validity of the results was demonstrated for the LS obtained with WIRS, EKO, and their combination. These results suggest that incorporating breathing-phase context may enhance computerized LS analysis. Hence, multimodal sensing systems that enable this, such as WIRS, have the potential to advance LS clinical utility beyond traditional manual auscultation and improve patient care.

Machine Learning (ML) research has focused on maximizing the accuracy of predictive tasks. ML models, however, are increasingly more complex, resource intensive, and costlier to deploy in resource-constrained environments. These issues are exacerbated for prediction tasks with sequential classification on progressively transitioned stages with “happens-before” relation between them.We argue that it is possible to “unfold” a monolithic single multi-class classifier, typically trained for all stages using all data, into a series of single-stage classifiers. Each single- stage classifier can be cascaded gradually from cheaper to more expensive binary classifiers that are trained using only the necessary data modalities or features required for that stage. UnfoldML is a cost-aware and uncertainty-based dynamic 2D prediction pipeline for multi-stage classification that enables (1) navigation of the accuracy/cost tradeoff space, (2) reducing the spatio-temporal cost of inference by orders of magnitude, and (3) early prediction on proceeding stages. UnfoldML achieves orders of magnitude better cost in clinical settings, while detecting multi- stage disease development in real time. It achieves within 0.1% accuracy from the highest-performing multi-class baseline, while saving close to 20X on spatio- temporal cost of inference and earlier (3.5hrs) disease onset prediction. We also show that UnfoldML generalizes to image classification, where it can predict different level of labels (from coarse to fine) given different level of abstractions of a image, saving close to 5X cost with as little as 0.4% accuracy reduction.

Sepsis is a life-threatening condition, caused by the body’s extreme response to an infection. In the United States, 1.7 million cases of sepsis occur annually, resulting in 265,000 deaths. Delayed diagnosis and treatment are associated with higher mortality rates. An exponential rise in the availability of medical data has allowed for the development of sophisticated machine learning algorithms to predict sepsis earlier than the onset. However, these models often underperform, as the training data are retrospective and do not fully capture the uncertain future. In this study, we develop a novel framework, which we refer to as MLePOMDP, to leverage and combine the underlying, high-level knowledge about sepsis progression and machine learning (ML) for classification. Specifically, we use a hidden Markov model to describe sepsis development at a high level, where the ML model makes the higher-order “observations” from temporal data. Consequently, a partially observable Markov decision process (POMDP) model is developed to make classification decisions. We analytically establish that the optimal policy is of threshold-type, which we exploit to efficiently optimize MLePOMDP. MLePOMDP is calibrated and tested using high-frequency physiological data collected from bedside monitors. Different from past POMDP-based frameworks, MLePOMDP is developed for a prediction task using a very small state definition, produces highly interpretable results, and accounts for a novel and clinically meaningful action space. Our results show that MLePOMDP outperforms machine learning–based benchmarks by up to 8% in precision. Importantly, MLePOMDP is able to reduce false alarms by up to 28%. An additional experiment is conducted to show the generalizability of MLePOMDP to different patient cohorts. Summary of Contribution: This study develops a novel real-time decision support framework for early sepsis prediction by integrating well-known machine learning models (random forest and neural networks) with a well-established sequential decision-making model, namely, a partially observable Markov decision process (POMDP). The structural properties of the optimal policy are further explored and a threshold-type structure is established, which is then leveraged to develop a customized algorithm to solve the problem more efficiently. The resulting framework demonstrates the benefit of applying POMDPs to augment machine learning outputs. Specifically, the framework results in the reduction of false alarms in sepsis predictions where decisions are made in real time, hence improving the overall prediction precision.

Sepsis is a major public concern due to its high mortality, morbidity, and financial cost. There are many existing works of early sepsis prediction using different machine learning models to mitigate the outcomes brought by sepsis. In the practical scenario, the dataset grows dynamically as new patients visit the hospital. Most existing models, being "offline" models and having used retrospective observational data, cannot be updated and improved dynamically using the new observational data. Incorporating the new data to improve the offline models requires retraining the model, which is very computationally expensive. To solve the challenge mentioned above, we propose an Online Artificial Intelligence Experts Competing Framework (OnAI-Comp) for early sepsis detection using an online learning algorithm called Multi-armed Bandit. We selected several machine learning models as the artificial intelligence experts and used average regret to evaluate the performance of our model. The experimental analysis demonstrated that our model would converge to the optimal strategy in the long run. Meanwhile, our model can provide clinically interpretable predictions using existing local interpretable model-agnostic explanation technologies, which can aid clinicians in making decisions and might improve the probability of survival.

Hospital readmission after third-stage palliation for single ventricle physiology (Fontan cavopulmonary anastomosis) approaches 25%. The cause for readmissions is varied, and there is no clear way to predict those at risk for readmission. If an effective predictive model can be used, physicians can preempt readmission in either the inpatient or outpatient setting. In this study, we incorporated electronic health records (EHRs), such as vital signs, administered medications, laboratory results, and demographic information, to determine whether a patient will require readmission after undergoing Fontan surgery. The EHR data were extracted for 338 patients admitted to the Children’s Healthcare of Atlanta from May 2009 to August 2020 for Fontan surgery. 38.8% of the patients were readmitted. Machine learning models, such as extreme gradient boosting, random forest, decision tree, and logistic regression, were employed to make these predictions. The random forest classifier outperformed the rest of the models and achieved 89.6% accuracy [82.1%, 95.5%], 77.8% sensitivity [63.0%, 92.6%], 97.5% specificity [90%, 100%], 95.2% positive predictive value [83.7%, 100%], 86.7% negative predictive value [78.7%, 94.9%], 82.2% area under precision-recall curve [69.8%, 92.7%], 87.0% area under receiver operating characteristic [78.4%, 94.4%], and 85.1% F1 score [72.7%, 94.1%].

Real-time analysis of streaming physiological data to identify earlier abnormal conditions is an important aspect of precision medicine. However, open-source systems supporting this workflow are lacking. In this paper, we present PhysOnline, a pipeline built on the open-source Apache Spark platform to ingest streaming physiological data for online feature extraction and machine learning. We consider scalability factors for horizontal deployment to support growing analysis requirements. We further integrate real-time feature extraction, including pattern recognition methods as well as descriptive statistical components to identify temporal characteristics of waveform signals. These generated features are then used for machine learning and for real-time classification of abnormal conditions. As a case study, we present the online classification of electrocardiography recordings for screening Paroxysmal Atrial Fibrillation (PAF) and demonstrate that our pipeline can predict persons developing PAF at least 45 min. before an episode of that condition. This pipeline can be applied in domains where pattern matching, temporal abstractions, and morphological characteristics can be used for real-time classification of streaming time-series data.

This paper presents a novel method for hierarchical analysis of machine learning algorithms to improve predictions of at risk patients, thus further enabling prompt therapy. Specifically, we develop a multi-layer machine learning approach to analyze continuous, high-frequency data. We illustrate the capabilities of this approach for early identification of patients at risk of sepsis, a potentially life-threatening complication of an infection, using high-frequency (minute-by-minute) physiological data collected from bedside monitors. In our analysis of a cohort of 586 patients, the model obtained from analyzing the output of a previously developed sepsis prediction model resulted in improved outcomes. Specifically, the original model failed to predict 11.76 ± 4.26% of sepsis patients earlier than Systemic Inflammatory Response Syndrome (SIRS) criteria, commonly used to identify patients at risk for rapid physiological deterioration resulting from sepsis. In contrast, the multi-layer model only failed to predict 3.21 ± 3.11% of sepsis patients earlier than SIRS. In addition, sepsis patients were predicted on average 204.87 ± 7.90 minutes earlier than SIRS criteria using the multi-layer model, which can potentially help reduce mortality and morbidity if implemented in the ICU.

Abstract In this work, we introduce a novel visualization technique, the Temporal Intensity Map, which visually integrates data values over time to reveal the frequency, duration, and timing of significant features in streaming data. We combine the Temporal Intensity Map with several coordinated visualizations of detected events in data streams to create PhysioEx, a visual dashboard for multiple heterogeneous data streams. We have applied PhysioEx in a design study in the field of neonatal medicine, to support clinical researchers exploring physiologic data streams. We evaluated our method through consultations with domain experts. Results show that our tool provides deep insight capabilities, supports hypothesis generation, and can be well integrated into the workflow of clinical researchers.

The critical care environment is a complex and critical environment, containing numerous body sensors attached to critically ill patients and producing continuous streams of physiological data. In conjunction, human-generated clinical data are produced by clinicians providing care for those patients. Currently, physiological information is displayed in a limited univariate method, inherited from conventional practices of biomedical engineering firms who design and manufacture these medical devices. However, the method in which this information is displayed was developed with limited consideration from user-centered design practices, and largely excludes influential relationship to their underlying medical knowledge and experience. Moreover, these univariate displays have limited abilities to project trends. This paper proposes a framework for displaying multi-dimensional and complex data to users in the critical care environment. We present a case study of neonatal intensive care, a form of critical care for premature and ill term infants to illustrate the framework's practical impact.

We present a method for transmission and processing of real-time trans-continental medical data streams. We apply fundamentals of existing network technologies to create a secure tunnel from a remote hospital through an open-network to the Artemis Cloud. We capture and store incoming 1Hz data stream in our real-time event stream processor to allow for online real-time monitoring of the patient status. The contributions of this paper extend the Critical Care as a Service paradigm by incorporating remote monitoring centers. The results establish feasibility of the system to support real-time monitoring. However, existing protocols were required significant optimization to account for variability in throughput and availability of the network.

This paper proposes a framework for mobile nursing documentation enabling the integration of clinical intervention data with both electronic health record systems and real-time intelligent decision support systems for patient monitoring. A brief discussion on the networking and information security concerns is presented in order to provide context for the mobile application design decisions surround data transmission and storage. The framework is demonstrated using an initial case study in a neonatal intensive care unit.