Shyam Visweswaran
Has grant
Chair & Professor
Make your next move.
Prepare your application
Review drafts before you send them.
Explore related searches
Recent Grants
Grant: Close
Precision Approach to healthCARE enrollment Site (PA CARES)
Open Date: 2016-07-06
Close Date: 2021-06-30
Grant: Close
Development and Evaluation of a Learning Electronic Medical Record System
Open Date: 2015-09-15
Close Date: 2019-06-30
Articles (12)
Objective We developed and validated a detection-guided artifact removal framework for clinical electroencephalography (EEG). The framework applies artifact correction only to the contaminated segments and preserves artifact-free data without modification. Approach The framework uses convolutional neural network (CNN) detectors trained on the Temple University Hospital (TUH) Artifact Corpus (150 recordings, 105 patients). For eye movement artifacts (20 s windows), the framework applies independent component analysis (ICA) and canonical correlation analysis (CCA). For muscle artifacts (5 s windows), the framework applies wavelet thresholding and empirical mode decomposition (EMD). For non-physiological artifacts (1 s windows), the framework applies spherical spline interpolation and artifact subspace reconstruction (ASR). Removal methods were applied only to detector-flagged artifact windows. For windows that are not flagged as artifacts, the output signal is equal to the input signal. In a test set held-out (21 patients, 30 recordings), we compared detection-guided selective removal with global removal, which applies the same method to all windows. We evaluated the preservation of artifact-free windows using correlation, root mean squared error (RMSE), and peak signal-to-noise ratio (PSNR). Main results Selective removal outperformed global removal in all six methods and 18 metric comparisons ( p < 10 −105 ). Selective processing maintained a clean-segment correlation above 0.987 for all methods. Global removal reduced the correlation to values as low as 0.39 (CCA) and 0.47 (ASR). The framework achieved artifact suppression in detector-flagged windows. CCA removed 74.6% of the amplitude of the eye movement artifact. EMD removed 99.8% of high-frequency (30–40 Hz) muscle contamination. ASR reduced the non-physiological artifact amplitude by 37.1% in detector-flagged windows. Preservation of artifact-free windows remained high in all methods, indicating minimal distortion of clean EEG. Significance Detection-guided selective removal addresses a key limitation of pipelines that apply correction globally and that can remove clean electroencephalography (EEG) signals. The framework enables automated artifact removal without manual review and preserves signal fidelity for clinical interpretation. The modular design supports integration into real-time monitoring systems for acute and perioperative care applications.
Year:
2026
Objectives To explore the potential effects of recent restrictions on discussions regarding diversity, equity, and inclusion (DEI) in the field of biomedical informatics. Materials and Methods Executive orders issued by the U.S. federal government regarding diversity and gender issues are discussed in the context of implications for biomedical informatics research. Results Restrictions on specific terminology can hinder research into critical topics such as bias and fairness in clinical artificial intelligence and machine learning algorithms. Additionally, these limitations may narrow the scope of questions that informatics research can address and obstruct efforts to enhance the diversity of perspectives within the field. Discussion Responding to these threats requires a community response. The American Medical Informatics Association (AMIA) can help the informatics community present a united front in support of DEI research in multiple ways. Conclusion The informatics community should take a strong and unambiguous response to support diversity, equity, and inclusion of underrepresented perspectives in the field.
Year:
2025
Long COVID poses a significant disease burden globally, but its heterogeneous presentation and unreliable coding practices render it difficult to study. Developing efficient phenotyping algorithms is crucial to enabling risk prediction and effective management of Long COVID. We introduce the LAbel-efficienT Long COVID pHenotyping (LATCH) algorithm, which synthesizes a small number of gold-standard labels and a large, unlabeled dataset with many electronic health record (EHR) features. Both internal validation and external validation demonstrated the superior performance of LATCH over methods using the U09.9 Long COVID EHR code alone. Our downstream analysis revealed a pattern of elevated healthcare utilization due to Long COVID, peaking at and continuing beyond the fourth month following COVID infection. LATCH enhances the classification of Long COVID by fully utilizing both labeled and unlabeled data, providing vital insights into healthcare utilization trends, informing clinical and public health responses to the enduring consequences of COVID-19.
Year:
2025
Objectives The primary objective was to compile a comprehensive list of Logical Observation Identifiers Names and Codes (LOINC) terms that may be associated with patient, healthcare provider, and healthcare facility identifying information. Materials and Methods We developed a 2-step procedure for identifying LOINC terms, which consists of a keyword search of Long Common Names and filtering on selected property values, followed by expert physician review to confirm and categorize the terms. Results The final list comprises 1309 LOINC terms potentially associated with identifying information of patients, providers, and facilities. This list is publicly available on GitHub. Discussion Compared with electronic health record data coded with other terminologies, LOINC-coded data present unique challenges for deidentification, and a resource of LOINC terms that may be associated with identifying information will be helpful for this purpose. Conclusion This resource is valuable for deidentifying LOINC-coded data, ensuring compliance with the Health Insurance Portability and Accountability Act (HIPAA), and preserving the privacy of patients, providers, and facilities.
Year:
2025
Background Intraoperative hypotension burden not equally distributed during various periods of a general anesthetic. Post-induction hypotension usually has an iatrogenic cause and is related to the combination of drug doses used for induction of anesthesia. Predicting post-induction blood pressures, prior to induction, may inform anesthesiologists on the most judicious selection of anesthetic drugs and dosages. Our objective was to explore patterns of post-induction hypotension, apply machine learning (ML) to predict the mean blood pressure (MBP) immediately after induction, and develop a prototype of a clinical tool that can implement the ML model. Materials and Methods We extracted data from preoperative and intraoperative anesthetic episodes over an 8-year period at the University of Pittsburgh Medical Center. The dataset consisted of 93,037 anesthetic episodes, which was partitioned into training, validation, and test sets. Preoperative and pre-induction predictor variables included demographics, comorbidities, laboratory test values, pre-induction MBP, and administrations of induction drugs. The target consisted of 15 minutes of MBP values immediately after induction. Results The best-performing model was extreme-gradient boosted trees (XGB), with an R-squared value of 0.31 and a mean absolute error of 11.96, which is a moderately good performance. The performance of the model decreased over each minute post-induction. Conclusion ML modeling to predict post-induction MBP is feasible, and a clinical tool that incorporates ML can potentially aid in preventing hypotension in the operating room.
Year:
2025
Objective Federated research networks, like Evolve to Next-Gen Accrual of patients to Clinical Trials (ENACT), aim to facilitate medical research by exchanging electronic health record (EHR) data. However, poor data quality can hinder this goal. While networks typically set guidelines and standards to address this problem, we developed an organically evolving, data-centric method using patient counts to identify data quality issues, applicable even to sites not yet in the network. Materials and Methods We distribute high-performance patient counting scripts as part of Integrating Biology at the Bedside (i2b2), which all ENACT sites operate. They produce counts of patients associated with ENACT ontology terms for each site. At the ENACT Hub, our pipeline aggregates site-contributed counts to produce network statistics, which our self-service web application, Data Quality Explorer (DQE), ingests to help sites conduct data quality investigation relative to the network. Results Thirteen ENACT sites have contributed their patient counts, and currently seven sites have signed up to use DQE to analyze data quality issues. We announced a call to all ENACT sites to contribute additional patient counts. Discussion Identifying site data quality problems relative to the network is novel. Using a metric based on evolving network statistics complements rigid data quality checks. It is adaptable to any network and has low barriers of entry, with patient counting being the sole requirement. Conclusion We implemented a metric for conducting data quality investigation in ENACT using patient counting and network statistics. Our end-to-end pipeline is privacy-preserving and the underlying design is generalizable.
Year:
2025
Objectives Statistical and artificial intelligence algorithms are increasingly being developed for use in healthcare. These algorithms may reflect biases that magnify disparities in clinical care, and there is a growing need for understanding how algorithmic biases can be mitigated in pursuit of algorithmic fairness. We conducted a scoping review on algorithmic individual fairness (IF) to understand the current state of research in the metrics and methods developed to achieve IF and their applications in healthcare. Materials and Methods We searched four databases: PubMed, ACM Digital Library, IEEE Xplore, and medRxiv for algorithmic IF metrics, algorithmic bias mitigation, and healthcare applications. Our search was restricted to articles published between January 2013 and November 2024. We identified 2498 articles through database searches and seven additional articles, of which 32 articles were included in the review. Data from the selected articles were extracted, and the findings were synthesized. Results Based on the 32 articles in the review, we identified several themes, including philosophical underpinnings of fairness, IF metrics, mitigation methods for achieving IF, implications of achieving IF on group fairness and vice versa, and applications of IF in healthcare. Discussion We find that research of IF is still in their early stages, particularly in healthcare, as evidenced by the limited number of relevant articles published between 2013 and 2024. While healthcare applications of IF remain sparse, growth has been steady in number of publications since 2012. The limitations of group fairness further emphasize the need for alternative approaches like IF. However, IF itself is not without challenges, including subjective definitions of similarity and potential bias encoding from data-driven methods. These findings, coupled with the limitations of the review process, underscore the need for more comprehensive research on the evolution of IF metrics and definitions to advance this promising field. Conclusion While significant work has been done on algorithmic IF in recent years, the definition, use, and study of IF remain in their infancy, especially in healthcare. Future research is needed to comprehensively apply and evaluate IF in healthcare.
Year:
2024
With generative artificial intelligence (GenAI), particularly large language models (LLMs), continuing to make inroads in healthcare, assessing LLMs with human evaluations is essential to assuring safety and effectiveness. This study reviews existing literature on human evaluation methodologies for LLMs in healthcare across various medical specialties and addresses factors such as evaluation dimensions, sample types and sizes, selection, and recruitment of evaluators, frameworks and metrics, evaluation process, and statistical analysis type. Our literature review of 142 studies shows gaps in reliability, generalizability, and applicability of current human evaluation practices. To overcome such significant obstacles to healthcare LLM developments and deployments, we propose QUEST, a comprehensive and practical framework for human evaluation of LLMs covering three phases of workflow: Planning, Implementation and Adjudication, and Scoring and Review. QUEST is designed with five proposed evaluation principles: Quality of Information, Understanding and Reasoning, Expression Style and Persona, Safety and Harm, and Trust and Confidence.
Year:
2024
BACKGROUND Postoperative venous thromboembolic events (VTE), encompassing deep vein thrombosis (DVT) and pulmonary embolism (PE), are preventable but can cause severe morbidity and mortality in post-surgical patients. Around 1 in 1000 individuals annually experience VTE, making prompt diagnosis and therapy vital, as untreated VTE carries a 30% mortality rate. These events result in $1 billion in hospital costs every year. While prophylaxis for preventing VTE is crucial, some patients may face challenges that limit their ability to take prophylactic measures, such as the risk of bleeding, allergies or adverse reactions, recent surgery or trauma, severe liver or kidney disease, and drug interactions. Of the patients who develop VTE after surgery, around 40% occur during the initial hospital stay , whereas 60% of the patients develop VTE within 30-90 days after discharge from the hospital following surgery. Existing VTE risk scoring systems, such as Caprini and Rogers sub-optimal and can be cumbersome and difficult to translate to patient-specific clinical decision-making. Though machine learning approaches have been used to predict VTE risk in specific surgery types, no validated predictive models are currently used in clinical practice for VTE risk calculation under the broader umbrella of post-operative surgery patients. OBJECTIVE The study seeks to use both the static and the time-varying EHR patient data (aggregated per visit) to 1) predict the VTE risk for a surgical patient within 30 days post-discharge using machine learning (ML); 2) stratify the VTE risk of the patient into high, medium, and low-risk categories; and 3) compare the performance of individual department-specific ML models to that of a unified model 4) determine which variables are most strongly associated with postoperative VTE METHODS The structured EHR data from post-operative patients from our multi-center hospital system between 2013 and 2019 were used. Various machine learning algorithms, including linear regression with L1/L2 regularization, random forest, and eXtreme Gradient Boosting (XGBoost), were evaluated. Different output probability thresholds were chosen to stratify the patients into low, medium, and high-risk categories. Department-specific ML models were developed, and their performance was compared with the unified model to determine the viability of constructing individual models versus unified models. Feature importance techniques were subsequently applied to identify the most influential features, followed by Rank Biased Overlap (RBO) analysis to quantify the level of overlap among the department-specific models. RESULTS Our findings demonstrate the efficiency of ML models in predicting and categorizing VTE risk among post-operative patients. The top-performing ML model achieved an F1 score of 0.76 with an area under the receiver operating characteristic curve (AUROC) of 0.83 and an area under the precision recall curve (AUPRC) of 0.89. A threshold of less than 0.3 in ML output probability was utilized to designate patients as low risk, while probabilities exceeding 0.7 indicated high risk. Interestingly, our investigation revealed that department-specific models failed to surpass the performance of the unified model. Analysis of feature importance correlation highlighted the length of stay (LOS) as the most influential predictor, followed by various laboratory results such as blood urea nitrogen (BUN), white blood cell count (WBC), vital signs like heart rate and temperature, as well as the patient’s age and body mass index (BMI). Our model integrates the dynamic aspects of a patient’s condition, incorporating changes in laboratory values and vital signs into its assessments. This represents a critical improvement over static models, which fail to account for the fluctuating clinical features of patients. CONCLUSIONS We have demonstrated the utility of machine learning models in predicting and stratifying VTE risk among post-operative
Year:
2024
Objectives To demonstrate that 2 popular cohort discovery tools, Leaf and the Shared Health Research Information Network (SHRINE), are readily interoperable. Specifically, we adapted Leaf to interoperate and function as a node in a federated data network that uses SHRINE and dynamically generate queries for heterogeneous data models. Materials and Methods SHRINE queries are designed to run on the Informatics for Integrating Biology & the Bedside (i2b2) data model. We created functionality in Leaf to interoperate with a SHRINE data network and dynamically translate SHRINE queries to other data models. We randomly selected 500 past queries from the SHRINE-based national Evolve to Next-Gen Accrual to Clinical Trials (ENACT) network for evaluation, and an additional 100 queries to refine and debug Leaf’s translation functionality. We created a script for Leaf to convert the terms in the SHRINE queries into equivalent structured query language (SQL) concepts, which were then executed on 2 other data models. Results and Discussion 91.1% of the generated queries for non-i2b2 models returned counts within 5% (or ±5 patients for counts under 100) of i2b2, with 91.3% recall. Of the 8.9% of queries that exceeded the 5% margin, 77 of 89 (86.5%) were due to errors introduced by the Python script or the extract-transform-load process, which are easily fixed in a production deployment. The remaining errors were due to Leaf’s translation function, which was later fixed. Conclusion Our results support that cohort discovery applications such as Leaf and SHRINE can interoperate in federated data networks with heterogeneous data models.
Year:
2024
Year:
2024
Year:
2024
Collaborators (16)
Katherine Liao
Associate Professor
Harvard Medical School
Zhaohui Geng
Assistant Professor
Ohio State University
Griffin Weber
Associate Professor of Biomedical Informatics
Harvard Medical School
Sunyang Fu
Assistant Professor
The University of Texas Health Science Center at Houston
Rafael Ceschin
Assistant Professor
University of Pittsburgh
Chuan Hong
Duke University
Malarkodi Jebathilagam Samayamuthu
Senior Research Scientist
University of Pittsburgh
Harry Hochheiser
Professor
University of Pittsburgh
Zongqi Xia
Principal Investigator
University of Pittsburgh
Senthilkumar Sadhasivam
University of Pittsburgh Medical Center
Yifan Peng
Associate Professor
Weill Cornell Medicine
Yanshan Wang
Vice Chair of Research and Tenured Associate Professor
University of Pittsburgh
Aman Mahajan
Chair of the Department of Anesthesiology and Perioperative Medicine
University of Pittsburgh
Harikesh Subramanian
Assistant Professor
University of Pittsburgh
Jeffrey Klann
Assistant Professor of Medicine
Harvard University
Parthasarathy Thirumala
Associate Professor
University of Pittsburgh Medical Center
Make your next move.
Prepare your application
Review drafts before you send them.