WORKSHOP: Exploring Temporal Patterns in Electronic Health Record Data
Thursday, May 29, 2014

overall data transform drug usages and medialc records PVDAS
A workshop of the
31 st  Human-Computer Interaction Lab Symposium
University of Maryland

Overview and Topics

Electronic Health Record (EHR) databases contain millions of patient records including events such as diagnoses, test results or medication prescriptions. The use of EHR databases could be dramatically improved if easy-to-use interfaces allowed clinical researchers and quality improvement analysts to explore complex temporal patterns in order to build and test hypotheses regarding the benefits, risks, comparative effectiveness, and appropriateness of treatments or medication regimens. Novel strategies in interface design and information visualization are needed.

We welcome discussions of the use of visual approaches, statistical methods, machine learning, etc. to study temporal patterns in patient histories, where the goals may be to find common patterns, rare events, or matches to a given patient etc. Topics might include data cleaning, integration from multiple sources, coping with incomplete or conflicting information, and presentation of result sets.

This workshop will include talks (followed by ample time for discussions) from:

  • HCIL researchers (e.g. our latest work on EventFlow and our newer work on cohort comparison)
  • Users of HCIL tools such as EventFlow (who will be invited to report on their case studies), and
  • Other researchers working on similar topics will be selected to present their work (please contact us if you are interested).

TO PARTICIPATE: Come and join us! The workshop is opened to all (space permitting, so register early).


Questions: Please contact Catherine Plaisant (



  • 08:15am - Registration & Breakfast
  • 09:00am - HCIL Symposium Intro + Plenary talks (more information)

    (third plenary talk - starting at about 9:40)
    Interactive Event Sequence Query and Simplification (Megan Monroe, Sana Malik, Chris Imbriano, Fan Du, Catherine Plaisant, Ben Shneiderman)

    • Catherine Plaisant and Ben Shneiderman (HCIL), "Introduction: The Growth of Temporal Event Data Analysis" (slides)
    • Catherine Plaisant (HCIL), "HCIL Partnership with Pulse8" (slides)
    • Chris Imbriano and Megan Monroe (HCIL), "Software Engineering for Consistent Modification Management" (slides)
    • Sana Malik (HCIL), "A Visual Analytics Approach to Comparing Cohorts of Event Sequences" (abstract) (slides)
      A common type of data analysis is finding the differences and similarities between two datasets. With temporal event sequence data, this task is complex because of the variety of ways single events and sequences of events can differ between the two groups (or cohorts) of records: the structure of the event sequences (e.g., event order, co-occurring events, or frequencies of events), the attributes about the events and records (e.g., gender of a patient), or metrics about the timestamps themselves (e.g., duration of an event). Running statistical tests to cover all these cases and determining which results are significant becomes cumbersome. Current visual analytics tools for comparing groups of event sequences emphasize a purely statistical or purely visual approach for comparison. We describe a taxonomy of metrics for comparing cohorts of temporal event sequences, including sequence, time, and attribute metrics. We also present a visual analytics tool, CoCo (for "Cohort Comparison"), which balances automated statistics with user-driven analysis to guide users to significant, distinguishing features between the cohorts.
    • Eberechukwu Onukwugha (University of Maryland, Baltimore), "An Algorithm to Identify Palliative Radiation Therapy in the Metastatic Prostate Cancer Setting: A Proof of Concept Application of Data Visualization Tools" (abstract) (slides)
      Skeletal-related events occurring among prostate cancer (PCa) patients with bone metastasis include palliative radiation to the bone (RttB), pathological fracture (PF), spinal cord compression (SCC), and bone surgery (BS). Reliable measures for identifying the components of SREs are critical to studies investigating the clinical and economic burden of skeletal-related events. Studies using healthcare claims employ various algorithms to identify RttB given that codes available in claims data cannot distinguish RttB from radiation to the prostate gland. We investigate the use of data visualization software to improve the reliability of a claims-based algorithm for RttB.
  • Noon - LUNCH
  • 1:00pm - WORKSHOP: Afternoon Session I
    • Megan Monroe (HCIL), "Temporal Event Analysis of Sports Data" (video)
    • Seth Powsner and Tami Sullivan (Yale University School of Medicine, Dept. of Psychiatry), "A Case Study of over 12000 Daily Reports from Women in Abusive Relationships" (abstract) (slides)
      Detailed daily reports on intimate partner violence, IPV, had been collected from 140 women over 90-day periods. These had been entered in traditional format for statistical processing, one record for each day. Within each record there were fields for event times: arguments, fights, unwanted sex, drinking, and drug use. Each record included fields for four of each type of event; a record for a peaceful day would have many blank fields. The initial conversion to find patterns of violence seemed only to confirm Tolstoy's observation that "All happy families are alike; each unhappy family is unhappy in its own way."

      To make sense of IPV required looking at shorter time segments, not entire 90 day sampling periods. Splitting subject reports into weeks, then days was more revealing. We review the key steps in conversion and segmentation of the original SPSS dataset into EventFlow: an irregular one to many (or none) relationship between records and events, calendrical calculations, and timeline representation of untimed data (e.g., tried to keep me from leaving house today).
    • Ana Szarfman and Rongjian Lan, "Visualization of Unique Temporal Sequences of Treatments and Events in an Unidentified Clinical Trial Data" (slides)
    • Sean Finan (Harvard Medical School), Piet De Groen (Mayo Clinic College of Medicine), Guergana Savova (Harvard Medical School), "Narrative Event and Temporal Relation Visualization Tool" (abstract) (slides)
      Extensive critical patient information typically is in an unstructured, free text format - the clinical narrative - that can only be accessed by reading the full text. However, the amount of information within the Electronic Medical Record (EMR) of a single patient is expanding beyond the ability of someone to read within a typical appointment slot. New Natural Language Processing (NLP) methods allow automated extraction of medical events and temporal relations among those events from clinical narratives. In order to display the clinically relevant events from a complete life span of a patient, we created a novel visualization tool that allows scrolling and zooming in time while maintaining an overview of the entire timeline within a single frame. We selected four key features of a typical clinical encounter as the main content of a medical timeline: (1) signs and symptoms; (2) tests and procedures; (3) diseases and disorders; and (4) medications. Within these four features, more detailed subset timelines are allowed. We will demonstrate our prototype graphical user interface and discuss some of the challenges unique to the visualization of unstructured clinical narratives as well as our solutions.
    • David Wang (Partners HealthCare), "Sifting through Lines, Events, and Trees: Stories of EHR Visual Exploratory Analysis" (abstract) (slides)
      The Research Information Services and Computing group at Partners HealthCare curates electronic health records from eight Boston area hospitals and provides controlled access for in-house and external research. As a result, we are often involved from the earliest data ETL (extract, transform, load) process to the final analysis result. I will present how a number HCIL technologies - Dynamic Query, Lifelines2, EventFlow, Treemap - are indispensable in our analytical process. I will use examples in our software applications and from recent pharmacovigilance exploratory studies in type 2 diabetes drugs.
  • 3:00pm - BREAK
  • 3:20pm - WORKSHOP: Afternoon Session II
    • Chris Crowley (Commonwealth Informatics, Inc.), "Integrating Event Flow with the Pharmacovigilance Defense Application System (PVDAS)" (abstract) (slides)
      PVDAS is the name given to the primary EHR data warehouse and associated data retrieval application used by epidemiologists, statisticians, and pharmacists at the U.S. Army Pharmacovigilance Center (PVC) to study drug utilization and drug safety issues among the 14-million members of the Military Healthcare System. After gaining some initial exposure to the EventFlow software through several ad-hoc, small-scale experiments, the PVC has funded Commonwealth Informatics to design and implement approaches for directly integrating EventFlow with PVDAS, automating the interchange of data and handling issues of scale-up to potentially very large cohorts of interest. We report on the first fruits of that effort, namely the ability to configure and launch an EffortFlow display based on user selection of key variables (both time-points and intervals) from the output of a PVDAS "descriptive analysis" run.
    • Sophia Wu and Margret Bjarnadottir (Robert H. Smith Business School, University of Maryland), "Exploring Temporal Patterns in Hypertensive Drug Therapy" (abstract) (slides)
      Patients' adherence to medication is of great importance as non-adherence can lead to worsening of conditions and health decline, this is especially true in the case of the management of chronic conditions. A summary statistic, the Medication Possession Ratio (MPR) has often been used to describe how well patients adhere to drug regiment, but the statistic does not adequately capture different adherence patterns of patients, which vary widely.

      In this study, we utilize Eventflow, a novel interactive visualization tool being developed by the University of Maryland Human Computer Interaction Lab (HCIL) to observe and summarize common patterns in hypertensive therapy, using pharmacy claims of 493,022 individuals. Eventflow provides a quick way to detect common use patterns for further evaluation and specific case studies, and possible insights into the development of new summarizing methods that can enable drug therapy comparisons. In particular the visualization clearly demonstrates the limitations of the MPR of capturing true patient experiences.

      In this talk we discuss the analysis of patients taking five hypertensive drug classes,Angiotension-Converting Enzyme-Inhibitors (Ace), Angiotension II Receptor Blockers (ARB), Calcium Channel Blockers (CCBs), Beta blockers(Beta), and Diuretics, We demonstrate how Eventflow was used to visualize general prescription patterns, and led to a focused study of a group of heart failure patients taking CCBs. The case study underlined the large variability in patient adherence and possibly inappropriate prescription patterns. Finally we discuss some future avenues for summarizing prescription patterns of large cohorts.
    • Jessica Lin (George Mason University), "Grammar-Based Medical Time Series Mining and Visualization" (abstract)
      The advancement of computer technology in medicine has enabled more sophisticated patients monitoring, either on-site or remotely. With the massive amount of medical data produced on the daily basis, it has become increasingly evident that efficient methods to search and analyze historic data, and to detect patterns in biomedical time series (e.g. ECG, EEG, respiration rates, blood pressure, etc) are in great demand. In recent work, we developed a grammar-based time series pattern discovery and visualization system called GrammarViz. GrammarViz automatically discovers novel patterns such as recurrent patterns (motifs), rare patterns (anomalies) and periodic patterns by identifying and exploiting the hidden hierarchical structure through grammar induction. We argue that discovering such patterns in biomedical time series has numerous applications including classification, prediction of patient outcome, timely detection of anomalies, and identification of biomarkers or critical patterns associated with patient outcome. Since our approach operates on discretized time series by utilizing the state-of-the-art time series discretization technique, Symbolic Aggregate approXimation (SAX), it can potentially be applied to discrete temporal events sequences commonly seen in Electronic Medical Records.
    • Beth Carter (Children's National Medical Center), "Analysis of task performance during pediatric trauma resuscitation" (abstract) (slides)
      As checklists gain popularity in medical settings, it will be important to understand how they are being used in practice. Checklist compliance is often reported as the percent of checklist boxes checked but this metric does not indicate if the associated task was performed. We introduced a pediatric trauma resuscitation checklist at Children's national with the goal of improving compliance with the Advanced Trauma Life Support (ATLS) protocol. Video review of resuscitations was conducted to determine if tasks checked off on the checklist were performed. EventFlow was used to visualize patterns of checklist use and to measure the difference in time between box checked and task performed.
  • 05:00pm - Symposium Demos & Posters + RECEPTION
  • 06:30pm - Symposium ends

Event Logistics

  • Symposium Registration
  • Accomodations
  • Directions, Maps, and Parking Information
  • All activities will take place in CSIC (Computer Science Instruction Center), near the AV Williams bldg / Computer Science department, i.e. also near Route 1
  • Plan enough time for parking.
  • Coffee and light breakfast will be available in the morning during the demos
  • Lunch will include some vegetarian offerings.

Related HCIL Materials

Past Related Workshops

Web Accessibility