Yelena Yesha’s research while affiliated with University of Miami and other places

Background Conflicting reports from varying stakeholders related to prognosis and outcomes following placement of temporomandibular joint (TMJ) implants gave rise to the development of the TMJ Patient-Led RoundTable initiative. Following an assessment of the current availability of data, the RoundTable concluded that a strategically Coordinated Registry Network (CRN) is needed to collect and generate accessible data on temporomandibular disorder (TMD) and its care. The aim of this study was therefore to advance the clinical understanding, usage, and adoption of a core minimum dataset for TMD patients as the first foundational step toward building the CRN. Methods Candidate data elements were extracted from existing data sources and included in a Delphi survey administered to 92 participants. Data elements receiving less than 75% consensus were dropped. A purposive multi-stakeholder sub-group triangulated the items across patient and clinician-based experience to remove redundancies or duplicate items and reduce the response burden for both patients and clinicians. To reliably collect the identified data elements, the identified core minimum data elements were defined in the context of technical implementation within High-performance Integrated Virtual Environment (HIVE) web-application framework. HIVE was integrated with CHIOS™, an innovative permissioned blockchain platform, to strengthen the provenance of data captured in the registry and drive metadata to record all registry transaction and create a robust consent network. Results A total of 59 multi-stakeholder participants responded to the Delphi survey. The completion of the Delphi surveys followed by the application of the required group consensus threshold resulted in the selection of 397 data elements (254 for patient-generated data elements and 143 for clinician generated data elements). The infrastructure development and integration of HIVE and CHIOS™ was completed showing the maintenance of all data transaction information in blockchain, flexible recording of patient consent, data cataloging, and consent validation through smart contracts. Conclusion The identified data elements and development of the technological platform establishes a data infrastructure that facilitates the standardization and harmonization of data as well as perform high performance analytics needed to fully leverage the captured patient-generated data, clinical evidence, and other healthcare ecosystem data within the TMJ/TMD-CRN.

Background Venous thromboembolism (VTE) is a preventable complication of hospitalization. Risk-stratification is the cornerstone of prevention. The Caprini and Padua are the most commonly used risk-assessment models to quantify VTE risk. Both models perform well in select, high-risk cohorts. While VTE risk-stratification is recommended for all hospital admissions, few studies have evaluated the models in a large, unselected cohort of patients. Methods We analyzed consecutive first hospital admissions of 1,252,460 unique surgical and non-surgical patients to 1,298 VA facilities nationwide between January 2016 and December 2021. Caprini and Padua scores were generated using the VA’s national data repository. We first assessed the ability of the two RAMs to predict VTE within 90 days of admission. In secondary analyses, we evaluated prediction at 30 and 60 days, in surgical versus non-surgical patients, after excluding patients with upper extremity DVT, in patients hospitalized ≥72 hours, after including all-cause mortality in the composite outcome, and after accounting for prophylaxis in the predictive model. We used area under the receiver-operating characteristic curves (AUC) as the metric of prediction. Results A total of 330,388 (26.4%) surgical and 922,072 (73.6%) non-surgical consecutively hospitalized patients (total n=1,252,460) were analyzed. Caprini scores ranged from 0-28 (median, interquartile range: 4, 3-6); Padua scores ranged from 0-13 (1, 1-3). The RAMs showed good calibration and higher scores were associated with higher VTE rates. VTE developed in 35,557 patients (2.8%) within 90 days of admission. The ability of both models to predict 90-day VTE was low (AUCs: Caprini 0.56 [95% CI 0.56-0.56], Padua 0.59 [0.58-0.59]). Prediction remained low for surgical (Caprini 0.54 [0.53-0.54], Padua 0.56 [0.56-0.57]) and non-surgical patients (Caprini 0.59 [0.58-0.59], Padua 0.59 [0.59-0.60]). There was no clinically meaningful change in predictive performance in patients admitted for ≥72 hours, after excluding upper extremity DVT from the outcome, after including all-cause mortality in the outcome, or after accounting for ongoing VTE prophylaxis. Conclusions Caprini and Padua risk-assessment model scores have low ability to predict VTE events in a cohort of unselected consecutive hospitalizations. Improved VTE risk-assessment models must be developed before they can be applied to a general hospital population.

Objective Hospital-acquired venous thromboembolism (VTE, including pulmonary embolism (PE) and deep vein thrombosis (DVT)) is a preventable cause of hospital death. The Caprini risk assessment model (RAM) is one of the most commonly used tools to assess VTE risk. The RAM is operationalized in clinical practice by grouping several risk scores into VTE risk-categories that drive decisions on prophylaxis. A correlation between increasing Caprini scores and rising VTE risk is well-established. We assessed whether the increasing VTE risk-categories assigned on the basis of recommended score-ranges also correlate with rising VTE risk. Methods We conducted a systematic review of articles that used the Caprini RAM to assign VTE risk-categories and that reported corresponding VTE rates. A Medline and EMBASE search retrieved 895 articles, of which 57 fulfilled inclusion criteria. Results Forty-eight (84%) of the articles were cohort studies, 7 (12%) were case-control studies, and 2 (4%) were cross-sectional studies. The populations varied from post-surgical to medical patients. There was variability in the number of VTE risk-categories assigned by individual studies (6 used 5 risk categories, 37 used 4, 11 used 3, and 3 used 2), and in the cutoff scores defining the risk-categories (scores from 0 alone to 0-10 for the low-risk category; from ≥5 to ≥10 for high-risk). The VTE rates reported for similar risk-categories also varied across studies (0%-12.3% in the low-risk category; 0%-40% for high-risk). The Caprini RAM is designed to assess composite VTE risk, however, 2 studies reported PE or DVT rates alone, and many of the other studies did not specify the types of DVTs analyzed. The Caprini RAM predicts VTE at 30 days post-assessment, however, only 17 studies measured outcomes at 30 days; the remaining studies had either shorter or longer follow-ups (0 to 180 days). Conclusions The utility of the Caprini RAM is limited by heterogeneity in its implementation across centers. The score-derived VTE risk categorization has significant variability in the number of risk-categories being used, the cut-points used to define the risk-categories, the outcome being measured, and the follow-up duration. This leads to similar risk-categories being associated with different VTE rates which impact the clinical and research implications of the results. To enhance generalizability, there is a need for studies that validate the RAM in a broad population of medical and surgical patients, identify standardized risk-categories, define risk of DVT and PE as distinct endpoints, and measure outcomes at standardized follow-up time-points.

Adversarial attacks attempt to disrupt the training, retraining and utilizing of artificial intelligence and machine learning models in large-scale distributed machine learning systems. This causes security risks on its prediction outcome. For example, attackers attempt to poison the model by either presenting inaccurate misrepresentative data or altering the models' parameters. In addition, Byzantine faults including software, hardware, network issues occur in distributed systems which also lead to a negative impact on the prediction outcome. In this paper, we propose a novel distributed training algorithm, partial synchronous stochastic gradient descent (ParSGD), which defends adversarial attacks and/or tolerates Byzantine faults. We demonstrate the effectiveness of our algorithm under three common adversarial attacks again the ML models and a Byzantine fault during the training phase. Our results show that using ParSGD, ML models can still produce accurate predictions as if it is not being attacked nor having failures at all when almost half of the nodes are being compromised or failed. We will report the experimental evaluations of ParSGD in comparison with other algorithms.

Blockchain is the technology used by developers of cryptocurrencies, like Bitcoin, to enable exchange of financial “coins” between participants in the absence of a trusted third party to ensure the transaction, such as is typically done by governments. Blockchain has evolved to become a generic approach to store and process data in a highly decentralized and secure way. In this article, we review blockchain concepts and use cases, and discuss the challenges in using them from a governmental viewpoint. We begin with reviewing the categories of blockchains, the underlying mechanisms, and why blockchains can achieve their security goals. We then review existing known governmental use cases by regions. To show both technical and deployment details of blockchain adoption, we study a few representative use cases in the domains of healthcare and energy infrastructures. Finally, the review of both technical details and use cases helps us summarize the adoption and technical challenges of blockchains.

The analysis of both continuous and categorical attributes generating a heterogeneous mix of attributes poses challenges in data clustering. Traditional clustering techniques like k-means clustering work well when applied to small homogeneous datasets. However, as the data size becomes large, it becomes increasingly difficult to find meaningful and well-formed clusters. In this paper, the authors propose an approach that utilizes a combined similarity function, which looks at similarity across numeric and categorical features and employs this function in a clustering algorithm to identify similarity between data objects. The findings indicate that the proposed approach handles heterogeneous data better by forming well-separated clusters.

COVID-19 is a novel infectious disease responsible for over 800K deaths worldwide as of August 2020. The need for rapid testing is a high priority and alternative testing strategies including X-ray image classification are a promising area of research. However, at present, public datasets for COVID19 x-ray images have low data volumes, making it challenging to develop accurate image classifiers. Several recent papers have made use of Generative Adversarial Networks (GANs) in order to increase the training data volumes. But realistic synthetic COVID19 X-rays remain challenging to generate. We present a novel Mean Teacher + Transfer GAN (MTT-GAN) that generates COVID19 chest X-ray images of high quality. In order to create a more accurate GAN, we employ transfer learning from the Kaggle Pneumonia X-Ray dataset, a highly relevant data source orders of magnitude larger than public COVID19 datasets. Furthermore, we employ the Mean Teacher algorithm as a constraint to improve stability of training. Our qualitative analysis shows that the MTT-GAN generates X-ray images that are greatly superior to a baseline GAN and visually comparable to real X-rays. Although board-certified radiologists can distinguish MTT-GAN fakes from real COVID19 X-rays. Quantitative analysis shows that MTT-GAN greatly improves the accuracy of both a binary COVID19 classifier as well as a multi-class Pneumonia classifier as compared to a baseline GAN. Our classification accuracy is favourable as compared to recently reported results in the literature for similar binary and multi-class COVID19 screening tasks.