Research > Research Areas
Clinical Informatics and Decision Support
Clinical informatics and artificial intelligence (AI) are transforming healthcare by using clinical data from electronic health records and clinical monitoring systems to support accurate, safe and ethical decision-making in the outpatient and inpatient settings.
Our history in applying AI to medical problems dates to the mid-1970s, with pioneering efforts in rule-based expert systems, including INTERNIST-1, led by Jack Myers, Harry Pople and Randy Miller. Our current projects are conducted in collaboration with a broad range of clinical partners across critical care medicine, neurology, surgery, psychiatry and other clinical specialties. Key challenges include ensuring that clinical models are fair, robust, and well integrated into clinical care. This work is facilitated by our close partnership with UPMC, one of the largest and most integrated academic health systems in the United States, along with access to our Neptune clinical data warehouse, which contains data on over 5 million UPMC patients.
Projects
Real-time identification of brain ischemia during surgery
Brain ischemia and stroke can be devastating surgical complications, with the likelihood increasing as the number of surgical procedures increases. Currently, intraoperative monitoring of brain activity to detect ischemia and stroke is performed using electroencephalogram (EEG) technology during high-risk surgeries. This process relies on tiring, error-prone and costly visual monitoring by a trained neurophysiologist. We have developed a machine learning system to analyze EEG signals in real time and alert the surgical team if it detects signs of ischemia or stroke during carotid endarterectomy. This human-in-the-loop artificial intelligence monitoring and alerting system aims to enhance safety in high-risk surgeries and, in the future, could enable automated brain monitoring for all surgical procedures. This project has spun off a startup company.
Faculty: Shyam Visweswaran, MD, PhD, Harry Hochheiser, PhD
Collaborators: Parthasarathy D. Thirumala, MD, Kayhan Batmanghelich, PhD
Sample Publication: Mina AI, Espino JU, Bradley AM, Thirumala PD, Batmanghelich K, Visweswaran S. Detecting cerebral ischemia from electroencephalography during carotid endarterectomy using machine learning. AMIA Jt Summits Transl Sci Proc. 2024 May 31;2024:613-622.
Outlier-based monitoring and alerting
Statistical anomalies in patient-management actions may correspond to medical errors. We are using artificial intelligence and machine learning methods to identify patient-management actions that are unusual relative to those used to manage comparable patients in the past, and to raise alerts for actions deemed statistical outliers. This method complements existing knowledge-based detection and alerting methods that are clinically precise but costly to build. The approach is currently being evaluated in critical care settings within UPMC to assess its ability to identify potential medical errors and support safer, more consistent patient management in high-acuity environments.
Faculty: Gregory F. Cooper, MD, PhD, and Shyam Visweswaran, MD, PhD
Collaborators: Gilles Clermont, MD, MS, and Milos Hauskrecht, PhD
Sample Publication: Hauskrecht M, Batal I, Hong C, Nguyen Q, Cooper GF, Visweswaran S, Clermont G. Outlier-based detection of unusual patient-management actions: an ICU study. J Biomed Inform. 2016 Dec;64:211-221.
Predicting postinduction hypotension during anesthesia
General anesthesia consists of three phases: induction, maintenance and emergence. During induction, anesthetic medications produce rapid loss of consciousness while triggering complex physiological responses that can lead to unstable vital signs and post-induction hypotension, a common complication requiring intervention. Because risk depends on multiple patient-specific factors, we are developing machine learning models to predict hypotension using preoperative and pre-induction data and to support anesthesiologists in selecting optimal medication combinations and dosages for safer anesthesia management.
Faculty: Shyam Visweswaran, MD, PhD
Clinical Collaborator: Harikesh Subramanian, MD, MS
Sample Publication Subramanian H, Visweswaran S, Sadhasivam S, Mahajan A. Post-induction hypotension prediction during general anesthesia using machine learning techniques. medRxiv [Preprint]. 19 Apr 2025.
Predicting the risk of bleeding with direct oral anticoagulants
Atrial fibrillation is a growing global health concern and a major cause of stroke, and oral anticoagulants are commonly used to reduce this risk. While newer direct oral anticoagulants (DOACs) are more convenient and predictable than warfarin, they still carry a significant risk of gastrointestinal or intracranial bleeding. Existing risk models were developed primarily for warfarin and do not accurately predict bleeding risk in patients receiving DOACs. We developed an artificial intelligence and machine learning model to identify patients at high risk of bleeding among those who have been prescribed DOACs.
Faculty: Shyam Visweswaran, MD, PhD
Collaborator: Rahul Chaudhary, MD, MBA
Sample Publication: Chaudhary R, Nourelahi M, Thoma FW, Gellad WF, Lo-Ciganic WH, Chaudhary R, Dua A, Bliden KP, Gurbel PA, Neal MD, Jain S, Bhonsale A, Mulukutla SR, Wang Y, Harinstein ME, Saba S, Visweswaran S. Machine learning predicts bleeding risk in atrial fibrillation patients on direct oral anticoagulant. Am J Cardiol. 2025 Jun 1;244:58-66.
Predicting the risk of opioid overdose
A critical step in addressing the opioid crisis is the development and implementation of effective risk assessment tools to identify individuals at high risk of opioid overdose. This project is evaluating a machine learning opioid-overdose risk prediction model in UPMC primary care practices. For the evaluation, the team developed a reusable software pipeline that retrieves electronic health record data from our Neptune research data warehouse, applies the machine learning model, and sends the results to UPMC’s EpicCare system.
Faculty: Eugene M. Sadhu, MD, and Shyam Visweswaran, MD, PhD
Collaborators: Walid F. Gellad, MD, MPH, and Wei-Hsuan "Jenny" Lo-Ciganic, PhD, MS
Sample Publication: JGellad WF, Chen YF, Park TW, Yang Q, Arnold JD, Kuza CC, Fedro-Byrom SN, Diiulio J, Militello LG, Whitlock M, Sadhu EM, Visweswaran S, Fine MJ, Abebe KZ, Suda KJ, Lo-Ciganic WH. Machine learning prediction and reducing overdoses with electronic health record nudges (mPROVEN) in the primary care setting: protocol for a cluster randomized controlled trial. JMIR Res Protoc. 2026 May 4;15:e94007.
Concierge clinical evidence service
Clinicians often lack timely, patient-specific evidence at the point of care, particularly when existing guidelines and published studies do not fully address the clinical question at hand. To address this gap, the Evolve to Next-Gen Accrual to Clinical Trials (ENACT) Network launched a pilot of the concierge clinical evidence service (CCES). The ENACT Network is a secure, federated platform that enables queries across more than 142 million patient electronic health records from Clinical and Translational Science Award (CTSA) consortium sites. Clinicians submit clinical questions to the CCES team, which formulates and executes network queries, analyzes the results, and delivers an evidence-based report to support clinical decision-making. Development of the CCES required robust technical infrastructure and close collaboration across disciplines and institutions. Key lessons include the importance of clearly defining patient populations early, engaging clinicians throughout the process, integrating biostatistical expertise from the outset, strengthening communication workflows, sustaining outreach efforts, and systematically evaluating clinician and patient satisfaction.
Faculty: Shyam Visweswaran, MD, PhD, and Eugene M. Sadhu, MD
Collaborators: Olga V. Kravchenko, MS, PhD, and Gary S. Firestein, MD
Bio-digital rapid alert to identify neuromorbidity
Children in the intensive care unit (ICU) face increased risks of developing neurological complications, which may, in some cases, lead to chronic difficulties. The BRAIN-AI project aims to develop predictive models that provide early warnings for children in the ICU who are at high risk of these difficulties. BRAIN-AI models have been developed with an eye towards standardization and streamlining of model creation, using the Fast Health Internet Resources (FHIR) data transport tools and automated machine learning approaches.
Faculty: Harry Hochheiser, PhD
Collaborators: Christopher M. Horvat, MD, MHA, Robert S. B. Clark, MD, and Alicia K. Au, MD
Sample Publication: Horvat CM, Barda AJ, Perez Claudio E, Au AK, Bauman A, Li Q, Li R, Munjal N, Wainwright MS, Boonchalermvichien T, Hochheiser H, Clark RSB. Interoperable models for identifying critically ill children at risk of neurologic morbidity. JAMA Netw Open. 2025 Feb 3;8(2):e2457469.
Predicting postoperative venous thromboembolism with explanations
Postoperative venous thromboembolism (VTE), which includes deep vein thrombosis and pulmonary embolism, is a serious yet preventable complication that can lead to significant morbidity, mortality and increased healthcare costs. Existing risk scoring systems apply only at the time of a patient's discharge from the hospital. This project developed artificial intelligence/machine learning models to predict the daily risk of venous thromboembolism (VTE) in hospitalized patients after surgery, along with explanations. Wet also evaluated the models alongside clinicians to determine whether they could provide accurate, interpretable, and patient-specific risk assessments daily.
Faculty: Gregory F. Cooper, MD, PhD, and Shyam Visweswaran, MD, PhD
Collaborators: Smitha Edakalavan, PhD, and Rafael Ceschin, PhD