Pitt-CMU Medical Informatics Colloquium

Natural products are highly popular and often perceived as a “safe” alternative that people want to use to supplement their treatment regimens. There is a concern that in some cases concomitant use of the natural product and conventional medications may lead to adverse events due to natural product-drug interaction (NPDI). This proposal is focused on two important barriers toward building effective decision support for NPDI: (1) the synthesis of evidence for NPDI is challenging due to inconsistency of the NPDI knowledge bases, lack of connection to real-life clinical outcomes, and the overall sparsity of the NPDI studies, and (2) clinical decision support for NPDI at a point of care does not exist. To address the first barrier, we propose to test a novel framework for evaluating the quality of evidence and providing recommendations about potential NPDI, designed to assist pharmacology research experts (i.e., compendia editors) as they develop content for NPDI knowledge bases (Aim 1). To address the second barrier, we propose to build the first shared clinical decision-making support tool focused on NPDIs that will take into an account a patient’s preference and explicitly accounts for clinical outcomes and patient-specific risk factors (Aim2).

Complementary health approaches involving the use of botanical and other natural products (NPs) such as green tea, kratom, and cannabis have increased in the US in the past few decades, with up to 18% of adults reporting regular consumption of NPs. Older adults are the largest consumers of both prescription drugs and NPs, with up to 88% adults reporting concomitant use. However, co-consumption of NPs with pharmaceutical drugs can lead to pharmacokinetic herb-drug interactions or NP-drug interactions (NPDIs), which can further lead to unwanted drug response. We develop a framework combining an ontology-based biomedical knowledge graph with broad-scope machine reading of NP-related literature to generate mechanistic explanations for potential pharmacokinetic NPDIs. The project includes construction of the knowledge graph, integration of machine reading, and ontology extensions to address the gap in knowledge related to pharmacokinetic NPDIs in biomedical ontologies. The knowledge graph framework is evaluated using known NPDIs involving green tea and kratom to show its potential for generating hypotheses to guide scientific research in future.

Dr. Malec will discuss previous work (related to his NIH/NLM-funded K99/R00 project, "Using the literature to build causal models of retrospective observational data") using representations of causal knowledge to advance causal inference from real-world data such as electronic health records (EHR). Dr. Malec will illustrate how causal knowledge can be useful in drug safety and Alzheimer’s disease research. The methods discussed are general and can be applied in many settings. The presenter encourages active discussion during the session.

The harms of implicit bias in clinical settings are acknowledged but poorly understood and difficult to overcome. We discuss how structural components of electronic medical record (EMR) user interfaces may contribute to sex and gender-based discrimination against patients via constant, duplicative presentation of stigmatizing sexually transmitted infection (STI) data irrespective of clinical significance. Via comparison with symbolism and representative quotes in Hawthorne’s 1850 novel "The Scarlet Letter," we propose a metaphor to examine how EMRs function as a platform for moral judgement. We grapple with the tension between benefits of clinical utility and potential risks of stigmatization, focusing our analysis on the exemplary use case regarding STI history in the setting of pregnancy. We conclude with recommendations for how to address these challenges to improve ethical stewardship of sensitive sexual/reproductive health data.

Rapid advances in machine learning (ML) have made it applicable to healthcare practices. However, the deployment of these ML models remains a challenge due to the lack of user-centered designs and model interpretability and adaptability. In this talk, I will present the findings from iterative engagements with therapists and post-stroke patients to design, develop, and evaluate two interactive systems to improve practices of physical stroke rehabilitation. Specifically, I will first introduce an interactive hybrid approach that combines an ML model with a rule-based model from experts to support transparent and personalized interactions with a user. Then, I will describe case studies of this approach to improve the therapist’s rehabilitation assessment and patient’s engagement in rehabilitation therapy.

AI-based decision support tools (ADS) are increasingly used to augment human decision-making in high-stakes social contexts. In this talk, I consider the case of child welfare, a domain in which ADS has been deployed with the intention of reducing biases and guiding decisions about which families receive services. In particular, I will present findings from Allegheny County’s Children, Youth and Families department, whose use of the Allegheny Family Screening Tool (AFST) has sparked controversy surrounding the use of ADS since its deployment in 2016. We conducted a series of interviews and contextual inquiries with thirteen call screeners and supervisors to understand how they currently make algorithm-assisted child maltreatment screening decisions. I will also present preliminary feedback on design prototypes that probed workers’ attitudes towards different notions of model interpretability, such as feature explanations, uncertainty, and historical outcomes. Finally, I will discuss algorithmic challenges and design opportunities at the interface, model, and organizational levels that may support more effective human-AI decision-making.

This talk will provide insights into the emerging science of human profiling from voice. Due to the enormous number of parameters that play a role in voice production, no two voices in the world are alike. This opens up the possibility of voice being both, an identifier (a biometric) and an information source about the factors that influence the speaker. As an identifier, voice is in fact more unique than either DNA or fingerprints. As a descriptor, voice is more revealing than DNA and fingerprints. It carries information that can be linked to the current (referring to the time of production) physical, physiological, demographic, medical, environmental and myriad other bio-relevant characteristics of the speaker. The science of profiling builds on the hypothesis that if any factor whatsoever influences the human mind or body, and if a biological pathway exists between that influence and the voice production mechanism, then there must exist an effect on voice. The challenge lies in discovering and quantifying these effects.

Rapid growth in adoption of electronic health records (EHRs) has led to an unprecedented expansion in the availability of large longitudinal datasets. Large initiatives such as the Electronic Medical Records and Genomics (eMERGE) Network, the Patient-Centered Outcomes Research Network (PCORNet), and the Observational Health Data Science and Informatics (OHDSI) consortium, have been established and have reported successful applications of secondary use of EHRs in clinical research and practice. In these applications, natural language processing (NLP) technologies have played a crucial role as much of detailed patient information in EHRs is embedded in narrative clinical documents. Meanwhile, a number of clinical NLP systems, such as MedLEE, MetaMap/MetaMap Lite, cTAKES, MedTagger, and i2b2 have been developed and utilized to extract useful information from diverse types of clinical text, such as clinical notes, radiology reports, and pathology reports. This talk will walk through some successful applications of NLP techniques in the clinical domain with potential opportunities and challenges.

The experience of disability is nearly universal, whether through long-term disability, temporary disablement, or the experience of friends or family. However, disability and human functioning have largely been left out of advancements in medical data science. This proposal will advance natural language processing (NLP) technologies to target two key barriers to collecting and analyzing better information on function and disability: (1) the function gap in Electronic Health Records (EHRs) and (2) the exclusion of the patient’s voice from the EHR. The proposed K99/R00 project will accomplish three significant steps in addressing these barriers: (Aim 1) Developing and evaluating enriched NLP models for analyzing functional status information in inpatient notes; (Aim 2) Expanding NLP methods for analyzing functional limitations to outpatient records and daily living domains; and (Aim 3) Developing and evaluating NLP models to bridge patient and provider perspectives on the experience of disability.

Postoperative venous thromboembolic events (VTE) are considered preventable causes of morbidity and mortality. These serious adverse events are responsible for 10% of hospital mortality and account for $1billion of hospital cost/year [cite]. Existing VTE risk scoring systems, which can require 30+ variables many of which are physician-dependent, can be cumbersome and difficult to translate to patient-specific clinical decision making. Recent efforts have been made to develop machine learning algorithms for VTE risk prediction and stratification in specialized surgeries, however no comparable effort has been made in the setting of general surgery. In this study, we aim to evaluate several machine learning models in their ability to generalize VTE prediction from specialized surgery to the broader umbrella of general surgery settings. The goal of this study is to 1) predict the VTE risk for a surgical patient post-discharge within 30 days 2) Stratify the VTE risk of the patient into high and low risk categories. 3) Determine which variables are most strongly associated with postoperative VTE. We analyzed the structured EHR data from post-operative patients who presented at one of several hospitals in our multi-center system and used machine learning algorithms including Linear regression with L1/L2 regularization, random forest, and XGBoost to create predictive models to analyze the risk of VTE. XGBoost had the highest performance and was able to stratify the patients into high and low risk categories using the output probabilities. Our results show that machine learning models can predict and stratify the VTE risk score of post-operative patients with high accuracy, sensitivity and specificity in addition to identifying the most predictive risk factors of VTE. The risk score and the categories predicted by the model can assist the physician in understanding the risk of a patient developing VTE and help in creating a personalized VTE prophylaxis regimen and monitoring plan.

A "new normal" is emerging in medicine integrating caring for COVID-19 patients, which includes the inpatient post-acute (or inpatient rehabilitation) setting. Dr. Levy will take this opportunity to share reflections of some impacts on physician workflow and general factors affecting burnout, share accessible clinical resources for a rapidly evolving novel condition, and review several use-cases of telehealth applications. What initially appeared to be short-term changes in workflow to accommodate enhanced respiratory precautions and a novel condition, are now persisting at the one-year mark. As the cycle of change continues to impress, collaboration and bridging of multidisciplinary teams can facilitate ongoing innovation.

As patients increasingly play more active roles in their health care, the Internet has become a prominent source of health information to guide their decision-making and self-management activities. Despite the great potential of the Internet, many patients who sought health information on the web reported feeling overwhelmed by the vast amount of unfiltered information and unqualified to determine the quality, veracity, and relevance of the information. The overall goal of this proposal is to build and implement a “Health E-Librarian with Personalized Recommendations (HELPeR)” - a personalized information access system with a hybrid recommender engine that adapts to different aspects of the patient. This would be the first implementation of a patient-centered system that can serve as a virtual health librarian. The HELPeR recommender engine is innovative in its capacity to integrate three dimensions of an individual patient (i.e., information needs based on the user’s profile, the user’s unique expressed information interests, and the level of user’s disease-related knowledge) to direct patients to highly personalized sets of information, that are high quality, trustworthy, and appropriate for each patient’s knowledge level. We have selected ovarian cancer (OvCa) as our initial population as it represents a complex disease with multiple tumor types and a range of prognoses, requiring personalized treatments and supportive care needs that evolve over time. HELPeR will be housed on a standalone website linked to the online health community (OHC) of the National Ovarian Cancer Coalition, a national OvCa advocacy organization. In order to attain our goal of implementing HELPeR, the aims of this proposal are: (1) Define user needs, preferences, and expectations for personalized health information, (2) Develop and evaluate the HELPeR system that is able to adapt to three types of individual user characteristics across the disease trajectory evolving information needs, personal information preferences, and progressive cancer-related knowledge, and (3) Conduct a field trial with OvCa patients to determine the acceptability and value of HELPeR in a real-world setting. HELPeR can be easily adapted to filter information for other cancers and chronic conditions, making it highly transferable to any chronic disease where patients repeatedly seek online information to better manage their condition.