University of Pennsylvania Health System: Inside an Imaging-informatics Incubator

Twitter icon
Facebook icon
LinkedIn icon
e-mail icon
Google icon
By observing the quantity and quality of informatics innovation emerging from a radiology department, it is possible to identify those institutions that are nurturing the next wave of informaticists in radiology. One beacon is the University of Pennsylvania Health System (UPHS) in Philadelphia, where R. Nick Bryan, MD, PhD, is chair of radiology. Bryan is quick to attribute the department’s success in this arena both to his predecessors and to others in the department, beginning with former chair Ronald Arenson, MD, (instrumental in the development of the first RIS, in the 1980s) and including Steven Horii, MD; Reuben Mezrich, MD, (during his time at UPHS); and, most recently, Curtis Langlotz, MD, PhD, vice chair for informatics. Bryan reports that focus, organization, and culture have made imaging informatics a hallmark of the radiology department at UPHS, and that immediate health-care exigencies are influencing current activities and priorities. Reinforcing Tradition A continuous line of faculty with expertise and interest in informatics has made informatics an innate part of the department, Bryan says, attracting—in turn—a consistent flow of young people interested in the field. “Having people with those levels of skill and knowledge just reinforces informatics within a department,” he says. “They impress and attract young people who want to follow their path.” To maintain the informatics edge, Bryan has had to dedicate appropriate resources, in the form of recruiting both clinical and nonclinical staff. “Informatics and health-services research, a near neighbor, are viewed as one of a number of main academic fields of interest of the department,” Bryan explains. A strong medical-informatics group, headed by Dan Morton, PhD, provides critical support for researchers. “That group has had strong intellectual leadership and good on-the-ground people to support the faculty,” Bryan says. More recently, with the recognition of the need for increased efficiency, Bryan has strongly encouraged some of the younger staff members to invest their time in informatics. “Informatics is the key to gaining the efficiency we will need and, at the same time, the means to provide the quality of care and document it,” he says. He continues, “Staff members are rewarded on the basis of their presentations, publications, and patents—all of the incentives of the traditional academic environment. In some cases, they gain additionally if they link that to commercial activities, but most of them are doing this because it is what they like, this is what they are interested in, and that is the academic part of their career.” Three years ago, Bryan formed a new committee on departmental efficiency and appointed Woojin Kim, MD, as the chair. Under Bryan, the already sizable informatics group has grown 10% to 20%, he estimates. Traditionally, the department has two or three senior faculty for whom informatics is the main focus. A combination of natural interest and encouragement by leaders has resulted in an increase in the number of trainees and junior faculty interested in this area. “We have three to four junior faculty members who are focusing on this area, and that is more than we have traditionally had,” Bryan notes. The Hit Parade Among the recent activities of the department, Bryan highlights involvement in the CMS Medicare Imaging Demonstration for electronic decision support as part of a consortium with Brigham and Women’s Hospital (Boston, Massachusetts); Geisinger Health System (Danville, Pennsylvania); and Weill Cornell Medical College/New York–Presbyterian Hospital (New York, New York). “Curtis Langlotz is leading the effort on our campus, and we actually have that turned on and running,” he reports. He also cites the RADIANCE project of Tessa Cook, MD. “In terms of quality issues, I think Tessa’s RADIANCE is a very good example,” he says. “We now have an automatic system for retrieving our patients’ radiation exposures, and for documenting and reporting them in a variety of fashions, so that people can make use of them in their patient care.” A third project that Bryan cites is Presto, the UPHS name for the management tool developed by Kim; William Boonn, MD; and colleagues. “It indexes data from multiple information systems and provides a very effective user interface to extract the data and create reports online that we use every day in the management of our department,” Bryan reports. Moving forward, Bryan predicts that his informatics team will make further contributions in the areas of quality and efficiency that will resonate throughout the health system. “All large health systems are looking for programs and projects that increase efficiency, while maintaining the quality of patient care,” Bryan says. “Our department has been looked at by the system as something of a leader in using informatics to deal with these issues. Decision support and electronic order entry are things the health system recognizes, from the regulatory and reimbursement perspectives, as critical for our institution, and we have a mission of providing leadership in that area.” Bryan continues, “In general—given the unknowns in health-care delivery and (in particular) with the possibility of more managed-care or population-accountability components of health care—having a very sophisticated informatics system is critical for being able to deliver care in that fashion.” The Next Productivity Leap Informatics is an important strategic tool for Bryan, and he intends to use it to increase clinical efficiency by the 20% to 30% necessary just to counterbalance the decrease in clinical volume anticipated as a result of utilization control and reimbursement reductions. He estimates that electronic workflow and hard work by the faculty have allowed the department to increase productivity by 30% over the past 10 years. “We are looking for 20% on top of that,” he says, “and the sooner, the better.” In retrospect, what RIS and PACS have done is keep radiologists busy, according to Bryan. “When we are on clinical assignment, we are working all the time,” he says. “There is always a study in front of us; there’s no downtime. We’ve gained efficiency by RIS and PACS allowing us to work harder at more or less the same thing that we used to do.” While some aspects of electronic workflow (especially voice recognition) might have slowed radiologists, overall, electronic workflow has enabled radiologists to increase productivity by working more quickly, Bryan says. “We probably can’t pedal any faster,” he adds. “Now, we are going to have to get informatics to provide us with some intelligent tools that will actually make us more efficient—on an individual, case-by-case basis. That’s a big challenge, and that’s where we are really pushing informaticists right now.”

Google-like Access to Hospital Data By Woojin Kim, MD

By Woojin Kim, MD Just before joining the Hospital of the University of Pennsylvania—part of the University of Pennsylvania Health System in Philadelphia—as a full-time faculty member, I found myself asking why it was so difficult to find a case that I had dictated only several weeks ago. Without knowing the patient’s name, medical-record number, or study-accession number, looking for a case that I had dictated a while back was a very difficult task. When searching for a case, one typically remembers the type of study and certain words or phrases used in the original report. For example, it was not an easy task to do a search for all cases of Morton neuroma seen on MRI on which I had reported within the past five years—yet we are living in a world where we can find all kinds of information online instantaneously. There are a number of reasons that radiologists would look for past cases. They might want to find cases for retrospective research, case reports or series, lectures, quality-improvement projects, or teaching files. They might be looking at a case and want to refer to similar cases in the past for help with clinical decisions. An attending physician might be reading a case with a resident, might come across a case of extraperitoneal bladder rupture, and might want to show similar cases to supplement the resident’s education. A resident might simply want to find out how a particular attending physician likes to dictate reports and might look up prior reports by that attending physician. The possibilities are endless; in fact, naming them is like asking someone why and how he or she uses Google. The reference to Google also applies to the design and functionality of the application that was created as a result of unmet case-finding needs. A purely Web-based search and data-mining tool was created to meet these needs, but the software application was designed, from the beginning, to go beyond searching just radiology reports. Pathology reports also were incorporated, and the tool was designed to combine different report databases on demand, to allow searching for cases with both radiology and pathology reports—all in a matter of seconds (instead of weeks or months). Soon after the development of the search tool, an automated pathology follow-up module was created. Radiologists often look at an image and wonder what the anomaly that they are seeing will turn out to be. Using an intelligent report-matching algorithm, the automated pathology follow-up module notifies the user of pathology reports pertinent to the radiology reports that he or she has previously dictated. For example, a patient undergoes surgery to remove his renal cell carcinoma, and the pathology report becomes available today. The application then notifies the radiologist of that pathology result, based on the MRI exam that he or she read a month ago. Features of this application were presented at national meetings, and after each presentation, many attendees expressed interest in having a similar tool installed at their institutions. This led to commercialization of the software application. Software Enhancements Since that commercialization, additional uses of data mining have been developed. By mining the metadata (including exam code, exam type, modality, reporting physician, and ordering physician), we were able to create business-intelligence and analytics tools, allowing measurement of productivity and analysis of the referral base. Intelligent automining of the reports allowed the creation of quality dashboards, automatically evaluating reports for critical test results, and finding cases with dictation errors (such as mistaking right for left body sides or male for female patients).
image
Figure 1. The search and data-mining system developed at the University of Pennsylvania can retrieve and analyze needed information in less than a second, offering both clinical and administrative staff the power of the search function through an intuitive user interface. It is a zero-download, browser- based application that makes searching and data mining possible from any computer or mobile device.
image
Figure 2. The application provides multiple options for exploring search results, including this bird’s-eye view. It combines radiology and pathology search functions, enabling users to find pathology- proven cases easily.
image
Figure 3. The analytics package optimizes practice by providing tools for measuring key performance indicators. Interactive tools display turnaround time and RVUs to help users tailor physicians’ schedules and optimize productivity.
The Web’s search engines, such as Google or Bing, work by looking for information on Internet servers (crawling) and then indexing this information. The search engine quickly delivers results for a user’s query by examining its index. Because the usefulness of the search engine largely depends on relevance of the results returned, much attention is paid to algorithms related to improving relevance. Google bases some of its relevance ranking on the PageRank algorithm, for example. Our application functions in a similar fashion, with attention paid to the unique features of medical-report databases and algorithms for understanding medical-report text. The application has minimal hardware requirements for most practices’ study volumes, and it even runs in a virtualized environment. For extremely large operations, it can be horizontally scaled to distribute the load across a number of servers or virtual machines. It has the ability to crawl and index report data residing in any RIS or hospital information system directly, and it accepts standard interchange formats such as HL7. A flexible architecture allows the application to index and search reports outside radiology, including those in cardiology and pathology systems. Great attention has been paid to the overall user experience (Figures 1–3) to ensure the practical and efficient use of the application. Being Web based, the application follows Web standards that ensure cross-browser compatibility, supporting all modern browsers, legacy versions of Internet Explorer, and even mobile browsers such as Android and Mobile Safari. A zero-download application, it provides access from reading rooms, offices, VPN-attached home computers, and mobile devices such as iPhones and iPads. As imaging services are targeted for cost cutting, radiology practices need timely business intelligence to survive. Furthermore, there is growing interest in the improvement of health-care quality, leading to increasing demand for quick access to timely and accurate health-care data. Spending weeks working with a database analyst to produce a static report is no longer a tenable approach to business intelligence in health care. With the ability to extend beyond the radiology-report database to include other specialties in medicine, such as pathology and cardiology, the possibilities for enhanced knowledge discovery and improvement of health care are greatly increased. Woojin Kim, MD, is interim chief of the division of musculoskeletal imaging at the Hospital of the University of Pennsylvania in Philadelphia and is cofounder of a health-care informatics solution; woojin@montagehealthcare.com.

CT Dose Monitoring Using RADIANCE

By Tessa S. Cook, MD, PhD Though no medical procedure is without risk, medical imaging is generally considered safe, so when a number of instances of patient overexposure to imaging-related radiation from CT exams came to light, the radiology and medical-physics communities quickly responded. The ACR® Dose Index Registry (DIR) is actively collecting CT dose parameters from facilities worldwide in order to develop a set of benchmarks and dose-reference levels for diagnostic CT. For pediatric patients, the Society for Pediatric Radiology is sponsoring a similar effort on a smaller scale. In addition, the American Association of Physicists in Medicine has developed size-specific dose estimates that allow normalization of CT dose parameters (based on patient size) for children and small adults. The biggest challenge in the CT dose-monitoring effort has historically been the way that the dose parameters are stored—as pixels on an image-based dose sheet, rather than as structured numeric data. More recently, the major CT-scanner vendors have adopted the DICOM standard for radiation-dose parameters: the Radiation Dose Structured Report (RDSR). Only newly or recently introduced scanner models support RDSR generation, however, leaving a vast repository of existing CT exams worldwide for which no effective means of dose monitoring previously existed.
image
Figure 4. The automated RADIANCE pipeline, which can import dose parameters from either image-based CT dose sheets or the DICOM Radiation Dose Structured Report.
To meet this need for a dose-monitoring solution, my colleagues and I at the Hospital of the University of Pennsylvania in Philadelphia, part of the University of Pennsylvania Health System (UPHS), developed Radiation Dose Intelligent Analytics for CT Exams (RADIANCE). RADIANCE is a freely available dose-monitoring application for CT exams that imports dose parameters from either the image-based dose sheet or the RDSR. Once configured, it can be scheduled to run as an automated pipeline that processes new dose sheets and updates the database without user input (Figure 4). The pipeline can also be triggered manually, as needed. Imported dose parameters include the volumetric CT dose index and dose–length product, as well as other reported parameters that vary by scanner manufacturer. Additional information about the patient, the imaging facility, the type of study performed, and the scanner hardware used is extracted from the DICOM study header. All data are stored in a relational database that resides behind the firewall of the imaging facility. If the facility participates in the DIR, it can elect to submit data to the registry using RADIANCE. Reporting tools built on the RADIANCE database schema can be used to facilitate data analysis and dose reporting. A real-time dashboard provides a snapshot of the database that can be filtered by study type, scanner model, individual patient, or dose-estimate threshold (Figure 5).
image
Figure 5. The RADIANCE dashboard shows individual-study dose estimates for the specified time interval, as well as average- dose and maximum-dose estimates on different scanners within the facility.
Monthly scorecard reports are tailored to the role of the recipient: radiologist, technologist, or radiology administrator (section chief, medical physicist, and so forth). Radiologists and technologists receive a summary of dose estimates for studies interpreted or performed, respectively, during the immediately preceding month. Administrators get a wider view of the data, with 12-month trend information, as well as access to individual radiologists’ or technologists’ scorecards (Figure 6). Comparative data from preceding months for equivalent study types are provided, as are data from equivalent studies that were interpreted or performed by others in the department during the same month.
image
Figure 6. The administrator’s view of the RADIANCE scorecard shows a 12-month trend for average-dose and maximum-dose estimates, in addition to comparative data and summary information for a subset of highest-dose estimates for a particular study type during the preceding month.
A subset of the highest-dose estimates for all patients, as well as for patients under the age of 50, is provided. Scorecard users within UPHS can access the final study interpretation, original study images, or the original image-based dose sheet for the study. Any factors identified that could have contributed to a higher-than-expected dose can be documented directly from the scorecard and attached to the study within the RADIANCE database. Within UPHS, all three of our hospitals use RADIANCE for CT dose tracking. The pipeline is automatically triggered when dose sheets are received by the RADIANCE computer from any of the three hospitals. Radiologists at our main hospital can launch a patient’s CT dose-estimate profile directly from the PACS when viewing an existing CT exam for that patient. The dose profile summarizes individual dose estimates stored in RADIANCE for all CT exams undergone by a patient within our health system, with links to the reports for every exam. Dose estimates for an individual patient can be retrieved from the RADIANCE database during the study-protocol process so that an appropriate choice of modality and study protocol can be made for the patient’s next imaging study. Using RADIANCE data, we have implemented dose-reduction measures for our thoracic and cardiovascular CT exams, as well as for CT urograms. By tightening scan length, decreasing the number of imaging phases (when clinically appropriate), reducing tube current, and customizing protocols to the patient’s body-mass index, we have been able to reduce dose estimates for single-phase thoracic CT exams, pulmonary-embolism CT exams, and high-resolution parenchymal chest CT exams by 38%, 30%, and 65%, respectively. RADIANCE data have also been used to educate radiology trainees about CT-related radiation and how alteration of CT protocols and parameters can reduce dose. Dose-monitoring and dose-reduction efforts continue within our health system; they include education of referring physicians and nonradiology trainees about CT and dose, as well as assessment of the effect of dose-specific decision support on ordering patterns. A successful dose-reduction initiative requires the ability to monitor and analyze dose estimates for specific study types, whether by using an internal dose-monitoring solution, by participating in a dose registry, or both. Combining the facility’s internal dose monitoring with participation in a registry provides the ability to monitor individual-study dose estimates regularly and to determine how well national or regional dose benchmarks are being met. Once an opportunity for dose reduction is identified, a plan must be implemented. The plan might include specific protocol modifications, use of clinical decision-support tools, education of referring physicians regarding imaging-related radiation, or some combination of these interventions. The success of the initiative depends on closing the loop after the plan is put in place and demonstrating a reduction in dose estimates, while ensuring continued dose monitoring via compliance with the systematic improvements. As has been done within UPHS, RADIANCE can be used as the cornerstone of a successful dose-monitoring and dose-reduction effort. For more information and to download RADIANCE, please visit http://www.radiancedose.com. Tessa S. Cook, MD, PhD, is a fourth-year radiology resident at the Hospital of the University of Pennsylvania in Philadelphia.