Applications of Mosaic in Health Care Delivery

K. Srinivas, K. Gopinath **, V. Jagannathan, R. Karinthi, Matthew Fuchs, Y.V. Reddy, George Almasi, Tad Davis

Department of Computer Science and Concurrent Engineering Research Center

West Virginia University

Morgantown, WV 26505

Email: {sriniv, gopi, juggy, raghu,fuchs, rar, almasi, tad}@cerc.wvu.edu

** Dept of Computer Science and Automation

Indian Insitute of Science, Bangalore, India

• physicians treating patients using patient records and knowledge (such as cost information and research findings) from distributed sources;
• primary care physicians consulting with remote specialists using enhancements to desktop conferencing technologies in the areas of perinatology and radiology, facilitated with computer support for X-rays, Ultrasound, voice-annotations and other multimedia information; and
• collection of primary care and specialized care providers collaborating to meet a community's health care needs through technologies such as multimedia mail
• clinical administrators tracking and evaluating patient outcomes

Introduction

Our technology is based on the belief that information sharing, communication, and coordination are the crucial elements of any collaborative endeavor. In the health care domain, collaboration is the cooperative activities of health care providers aiming to deliver total and real-time care for their patients. Communication among providers and managed access to distributed patient records should enable health care providers to make informed decisions about their patients in a timely manner. Patients should have universal access to the services of any convenient provider, and providers should be able to use relevant information from even the last episode of care in the patient record. Such a patient-centered perspective is in keeping with the real mission of health care providers.

Since information sharing, communication and coordination crucial aspects of collaboration, we need support for doing these in a transparent manner. We not only need facilities that respond to a user request on a one-time basis (such as data base query response), but also facilities that scour the network for information or keep track of certain data over a period of time or provide automatic notification. The agent technologies are designed to provide these services and in this paper we outline specific agents relevant to patient-centered health care. With this mission, we have integrated various technology elements that facilitate collaboration with the goal of fielding it in several clinics of Valley Health Systems Inc. and the Cabell-Huntington Hospital.

ARTEMIS System Implementation

• Adoption of the `HTTP' protocol for client-end interoperability.
• Adoption of the `CORBA'[5] specifications for server-end interoperability using Orbix.
• Gateways to a commercial relational database (Oracle).
• Adoption of the `Kerberos' standard for authentication.
• Model-based wide-area access to patient records and update capabilities to structured and unstructured information.
• Federated access control mechanisms (information provider decides who can access information).
• Adoption of hyper-text document metaphor (Mosaic) to support ease of use.
• Desktop conferencing among healthcare providers using the MONET[8] system.
• Synchronous information sharing for patient information and images such as X-rays
• Notification and asynchronous communication based on MIME-compliant multi-media mail for ordering laboratory tests and referrals.

Specialist Collaboration

The system we are building has the following capabilities:

• It enables a primary care physician to order an X-ray or an ultrasound scan via multimedia mail by attaching the appropriate forms as needed by the specialist. For example, in ordering an ultrasound scan, the primary care physician typically includes the prenatal flow sheets and the POPRAS form.
• It enables a specialist to respond to a test the primary care physician ordered via multimedia mail, by including his/her evaluation and the test results. For example, in the case of an X-ray, the radiologist typically responds with the X-ray image itself and his/her interpretation of the X-ray.
• It enables the specialist and the primary care physician to discuss a case in real time via desktop conferencing. The MONET system has been customized for the health care scenario. In this system the physicians will be able to see each other, talk to each other as well as be able to type from the keyboard or include portions of the patient record when needed. The physicians will be able to share an application such as an X-ray viewer and jointly discuss the data, such as an X-ray image. In the case of X-rays they can mark up specific portions of the X-ray in order to point to a region while discussing with the other physician. The conference minutes can also be archived.

Enabling Technologies for ARTEMIS

Information Sharing System

Architecture for information sharing

Figure above highlights the various components associated with our information server for our health care application. The various components associated with this figure are explained below.

Interface Manager

• to handle logins
• to translate URL requests from Mosaic clients to document pages

	<A NAME=top>
	<PRE>
	<H2>[SELECT firstname, middlename, lastname FROM demographics 
		where patientid =%pid%] Chart</H2>

Session Manager

Gateways

Models

Meeting On the NETwork (MONET)

Future Extensions

• Agent based technologies for patient tracking
• Advanced User Interface technologies based on enhancements to Mosaic

Value-added Agents for ARTEMIS

We define agents[3,4] as (semi-)autonomous, goal-directed software objects. Agents are preferrably programmable by the end user. The agent definition is deliberately broad, so as not to restrict agents only to certain kinds of tasks or to certain implementations. Programs can dispatch their own agents as necessary. The primary difference between agents for humans and agents for software is in the nature of the agents public interface. We restrict agents to software.

Some of the generic agents we have identified include:

• Monitoring and notification agent
• Prioritization agent
• Scheduling agent
• Filing agent
• Information access agents

Monitoring agents

• Referral and order management agents. These agents are entrusted with sending referrals and orders for tests on patients and informing the provider when the results of the order or summaries of the referral consultations become available. Our current implementation of this agent manages orders for ultrasound tests and diagnostic X-rays. The notification is provided as presented as a HTML document when the provider logs into the system.
• Case-worker support agent for prenatal patients. This agent's goal is to determine if a prenatal patient missed a scheduled appointment and notify a case-worker that followup actions are required. Such appointments are currently tracked manually. Missed appointments have to be followed through with the patient, as the providers are legally responsible for doing whatever is necessary to ensure pregnant women follow care-guidelines. In VHS, the followup of such situations is entrusted to a case-worker.
• Home-monitoring agent An agent, on behalf of the provider, checks with the patient at home (or at a nursing home) on parameters such as blood glucose levels, blood pressure, pulse rate, compliance to treatment and general well-being.
• Sign-off monitoring agent This is an agent which monitors whether sign-offs have taken place. All new information (such as a laboratory test result) has to be endorsed by a VHS provider before it can be included in the patient record. If such sign-offs are not done, some corrective action has to be initiated.

Prioritization Agents

• Sign-off prioritization agents. Providers at Valley currently get their new information for all patients in a big stack on their desks. Not all of these new pieces of information are of the same priority. Typically, any test results which have an abnormal value have a higher priority in getting the provider's attention than others. The patients current condition might also dictate whether a new piece of information has higher priority than others.
• Contact prioritization agents Although case-workers must follow up on all missed appointments, prioritizing the calls ensures that urgent cases are handled appropriately.

Scheduling agents

• Provider-provider consultation scheduling agent. ARTEMIS supports synchronous desktop consultations between providers and specialists. This agent helps in scheduling such consultations\cite{edmondsetal}.
• Patient-visit scheduler A home-ward bound agent can present itself in the home computer of a patient and negotiate the next visit with the patient knowing the schedule of the physician.

Filing agents

Information access agents

Agent Implementation

The Common Object Request Broker Architecture (CORBA)is an industry standard for providing location and language independent method invocation for objects. Once an object is registered with an Object Request Broker (ORB), it can be accessed by other objects, even if those objects reside on another node of the network, or are implemented in another language. The Interface Description Language (IDL) provides a language independent means of describing object interfaces.

Dreme [1] is a distributed superset of the Scheme programming language in which all first-class language objects are mobile in the network. Although a Dreme application can dynamically reconfigure itself, or send new pieces to remote sites on the network, correct communication among the parts is maintained through Scheme's lexical scoping rules. Dreme can support a variety of programming paradigms, including agents, client/server and peer-to-peer. In particular, Dreme can support applications that seamlessly combine both agent and other styles. For example, an application (such as a multimedia conference call) can embed parts of itself in smart agents which move around the network locating an appropriate set of resources, after which the distributed elements of the application function on those nodes in a more traditional manner. Mobile Dreme objects in the health-care network can communicate with local resources through IDL interfaces.

A primary function of agents is the intelligent analysis of information so that it can be filtered, manipulated, or reformatted for the end user. Agents need access to the underlying structure of the information; if this is not provided it must be derived by the agent. The SGML standard can be viewed as a meta-language to describe markup languages for specific types of information (normally called documents, but SGML can be applied to a much larger variety of structured bit vectors). HTML and HTML+, used by the World Wide Web (WWW), are examples of SGML-compliant languages.

A key insight from the development of SGML is that no single set of markup is sufficient for all information. Information converted to a single markup language, such as HTML, has lost its original semantic structure. SGML provides a standard way for describing the information agents will need to access and a standard way for them to manipulate it, even though that information may be transformed into HTML or Postscript for display. The more the information is structured, the more we can relieve the burden of document analysis from the agent.

An Example

A monitor agent sits and waits for a scheduled visit. As it approaches, it may contact the case worker to schedule a reminder call. Once the scheduled time passes, the agent must examine the sites in the network to determine if the appointment has been kept, as the patient might visit any of the clinics. This can be done by sending sub-agents to each of the sites. If no visit occurred, then a phone call must be arranged. The monitoring agent contacts the case worker's scheduling agent, as well as dispatching another agent to create a patient dossier on the fly. Since the dossier will have a standard structure, the caseworker's scheduler can analyze and prioritize them. Finally, a user interface agent, customized by the case-worker, can take this dossier and turn it into personalized multimedia mail or hypertext. Part of the scheduler's function is to keep track of the case-worker and send him or her the necessary information whether at the hospital, at a clinic, or with a patient.

In this scenario, the agents are all programmed in Dreme; the databases, email systems and user interfaces are all accessed through CORBA interfaces; and the information to be displayed is defined in SGML to facilitate its manipulation by agents.

Enhancements to Mosaic

• High Performance Distributed Web Servers Web servers in the near future will have to service large number of requests. Distributed and multithreaded Web server implementations with I/O optimizations to handle large multimedia objects are being investigated.
• Logical URLs Currently the URL is a specific reference to a particular object at a particular server. This approach has scale-up and fault tolerance problems, particularly for heavily demanded documents. All attempts to access a document are routed to the server named in the URL, forcing delays throughout the network, and making it a single point of failure which may make the links it contains virtually inaccessible (at least to anyone who does not have them on their hotlist). These problems will be particularly onerous to commercial ventures, as they translate into lost business and a low quality of service, which might send customers elsewhere. Document replication is necessary to better balance network traffic and provide continued access in the face of server and network failures, but the current URL protocol provides no means of supporting this. We are looking at two approaches, one short term and one longer term, to resolving this problem:
  • URL-tables to perform server to server translations. It is simple enough to place the same document in several locations, but it is more complex to convey that information to a client. Here, we let the URL designate a primary server which has sent copies of a particular document to some number of mirror servers. The primary server retains the list of secondary servers. When a request comes to a server, it replies with the list of mirror servers, as well as optionally sending the document itself, depending on its own current load. If the document is not returned, then the client has the option of attempting one of the other mirror servers. On the client side, a table of mirror servers is kept for frequently used URLs. If the client wishes to access a mirrored URL, then the servers are contacted in a random fashion until one responds or the request is cancelled. Since all servers return the list of mirror sites, the table can be updated automatically on each request. The size of this table can be controlled by deleting the less frequently accessed URLs. An alternative to the table would be to include the list of mirror servers in the URL, as contained in other documents, but seems rather onerous and difficult to update.
  • Logical (virtual?) URLs. A logical URL names a set of servers which contain the desired document, but does not refer to any particular physical server. When a request is sent to a logical URL, any server in the set may respond. The client is freed from any consideration of the physical server responding to the request, and servers can enter and leave the set without any awareness on the part of clients. This kind of behavior is required in high-availability transaction processing systems {reference ISIS and Teknekron Information Bus}. To implement this on the Internet, we will be using the Reliable Multicast Protocol (RMP) being developed at CERC. RMP creates a virtual token ring in the network which allows members to communicate with each other, and also allows outside processes to send messages to the ring. The set of servers in the logical URL corresponds to the RMP token ring, and the client is an outside process communicating with it.

  Both of these methods require that servers communicate updates to each other "behind the scenes". This is a standard distributed database problem. RMP provides a technology to support this on the network, although there are many alternatives. Due to the growing volume of traffic and the initiation of commercial ventures on the Web, we suspect that there will be a number of methods proposed, not all of which will use the public network.

• Groupware applications The Web currently uses a strictly client/server architecture for object delivery, such as hypertext, with a stateless protocol between clients and servers. The same approach is being used for current commercial applications. Distributed hypertext and on-line catalogs are just part of the potential applications for the Web. The Internet already supports a variety of interactive, multi-user applications, from the Usenet newsgroups through multi-user dungeons (MUDs) to the MBONE multiuser whiteboard. We are looking at ways of using or expanding the current Web architecture to support groupware applications. Although a graphical MUD communicating with Mosaic based users through a Web server will probably be the first significant Web groupware application, other fields, such as health care, can also benefit. In groupware... Mention how we'd like multi-user activity to occur by the same process of discovery as found in hypertext traversal -- people meeting people randomly on the net.
• Smarter Servers, Smarter Clients The development of groupware, commercial services, and other applications to be accessed through the Web represents a fundamental shift in the way the Web will be traversed. The current traversal paradigm, which is hypertext based, assumes that users proceed in a more or less random (or at least unpredictable) walk through the URL graph. The current stateless protocol is perfectly acceptable in this scenario, as there is no reason to retain state which is more likely to be thrown away than kept. With a shift to applications, this will no longer be true. Traversal, if that is still the right term, in an application is both far more predictable and far more stateful. Complex applications, such as groupware, can be implemented using the current architecture through scripts and forking child processes. This starts to become awkward as the applications become more sophisticated. At the same time, the purely fetch/display architecture of the clients severely limits the complexity which can be placed into a single page. We will attack this problem on the server side by placing intelligence directly in the server. We will first wrap the server API in a C++ class library, and then to wrap that in a Dreme interpreter. Dreme is a distributed dialect of Scheme with mobile objects designed for distributed and multi-user applications. Linking this with the server provides either an intelligent server, or applications which use HTML as their GUI. Using a distributed language, such as Dreme, will also simplify implementing the replicated server strategy described above. On the client side we will add the ability to receive sets of forms and pages, as opposed to just a single page at a time. As mentioned above, traversing an application will be significantly more regular than traversing hypertext. We can take advantage of this by downloading working sets of HTML, based on knowledge of the application.
• Prefetching Strategies Since a page is the current focus of attention, all the hot-links visible in the current page are possible candidates for prefetching. We are investigating other strategies to reduce the size of this set.
• HotDirectories In the current implementation management of hotlist may become unwieldy if the hotlist becomes too large (since hot list is a linear structure). We are implementing hierarchical directories that can be organized and managed more easily.

Conclusions

References

1. M.Fuchs, DREME: for Life on the Net , Ph.D. Thesis, New York University, 1995, Preliminary version available from the author.
2. V. Jagannathan, C. Gollapudy, R. Karinthi, K. Srinivas, R. Reddy and S. Reddy, Architectural Alternatives for Communicty Care Networks,Proceedings of the Third Workshop on Enabling Technologies:Infrastructure for Collaborative Enterprises, IEEE Computer Society Press,1994.
3. P. Maes, Agents that Reduce Work and Informations Overload , Communications of the ACM, Vol.37, No. 7, 1994.
4. O. Etzioni and D. Weld, A Softbot-Based Interface to the Internet , Communications of the ACM, Vol. 37, No. 7, 1994.
5. The Common Object Request Broker: Architecture and Specification , Object Management Group, 1991.
6. V. Jagannathan, R. Karinthi, G. Almasi and M. Sobolewski Model Based Information Access, International Journal of Intelligent and Cooperative Information Systems (IJICIS) Volume 3, Number 2, June 1994 pp 107-127.
7. R.Reddy, V. Jagannathan, K. Srinivas, R. Karinthi, S.M.Reddy,C.Gollapudy and S.Friedman, ARTEMIS: A Research Testbed for Collaborative Health Care Informatics, Proceedings of the Second Workshop on Enabling Technologies: Infrastructure for Collaborative Enterprises, April 1993, IEEE Computer Society Press.
8. Srinivas, K., Reddy, Y.V., Chang, L., Babadi, A., Kamana, S., Dai, Z. and Kumar, V MONET: A Multimedia Conferencing System for Collocating People and Programs. Proceedings of the CALS & CE Third National Conference on Concurrent Engineering, Washington, D.C. pp 433-441, 1991.

Vitae

Dr. K. Gopinath obtained his B.Tech. from IIT-Madras and MS in Computer Science from Univ. of Wisconsin. After working at Advanced Micro Devices for two years, he joined Stanford Univ. for a PhD which he obtained in 1988. After a brief stint as a PostDoc at Stanford, he joined Indian Institute of Science as an Asst. Prof. in 1990. He was also a consultant for Sun Microsystems part of the time during 1990-91.

V. "Juggy" Jagannathan , is an Associate Professor of Computer Science and a Senior Member of Technical Staff at CERC. He has directed and participated in research applications of advanced technology in Bell Labs, University of Louisville, Boeing Advanced Technology Center, Cimflex Teknowledge and CERC. He has focussed on the development of collaboration technology in DICE and directed the effort in developing the Information Sharing System, at CERC. He is widely known for his work on blackboard systems and distributed AI and has edited a book on blackboard systems through Academic Press in 1989. He was a guest editor of IEEE Computer for the special issue on Computer Support for Concurrent Engineering which appeared in January 1993. Dr. Jagannathan obtained a Ph.D. in Electrical Engineering from Vanderbilt University in 1981.

R. Karinthi , is an Assistant Professor in the Department of Statistics and Computer Science and the Concurrent Engineering Research Center (CERC) at the West Virginia University, Morgantown, WV. At CERC, he leads a sub task in the ARTEMIS project dealing with modeling of multimedia patient records. For the past two years he has been working on a project relating to integrating heterogeneous information repositories which is part of the DICE collaborative environment. His current research interests include medical informatics, computer graphics, multimedia information systems and collaborative environ- ments. Raghu Karinthi received his bachelors degree in Electrical Engineering in 1984 from the Indian Institute of Technology (IIT), Madras, India. He received his MS and Ph.D. degrees in Computer Science from the University of Maryland at College Park, 1988 and 1990, respectively. He has over 25 papers in refereed conferences and journals.

Matthew Fuchs is a Member of Technical Staff at the Concurrent Enginreering Research Center at West Virginia University. He is expecting his Ph.D. in Computer Science from New York University (NYU) in 1995. Mr. Fuchs' dissertation work extends the fields of mobile distributed computing, distributed multi-user interfaces, and platform independent user interfaces. Mr. Fuchs has developed Dreme, a distributed extension of Scheme in which all first-class language objects are mobile on the Internet, and has demonstrated its utility in developing multi-user applications, distributed systems, and mobile agents. The need for agents to communicate with users across the network led to the development of a platform indpendent user interface description language which agents can use to communicate with local users regardless of host type. The combination of technologies will enable the development of spontaneous collaboration and more sophisticated network browsers.

Ramana Reddy , is a Professor of Computer Science and the Director of the Concurrent Engineering Research Center. He has managed over $50M in sponsored research from both, the Federal Government as well as industry. As Principal Investigator and Chief Designer of the DICE Architecture he has demonstrated leadership skills in dealing with large scale projects spanning multiple organizations. His current principal project, ARTEMIS (Advanced Research Testbed for Medical Informatics, sponsored by ARPA) is directly related to the proposed project. Professor Reddy pioneered the development of Knowledge Based Simulation (while at Carnegie Mellon University), a field which has gained considerable following. Professor Reddy has authored numerous papers on applications of Artificial Intelligence, Concurrent Engineering and computer technologies for collaboration. He was a keynote/plenary speaker at many international conferences.

George Almasi is a Member of Technical Staff at the Concurrent Enginreering Research Center at West Virginia University. His Masters Degree thesis, defended in 1993 at WVU, deals with the Information Sharing System, a heterogeneous distributed database management system. His current area of interest is in interoperability in general and CORBA in particular. He also has a Masters degree in Electrical Engineering, obtained in Cluj, Romania in 1991. Tad Davis is a Member of Technical Staff at the Concurrent Enginreering Research Center at West Virginia University. He obtained his masters' degree in computer science from West Virginia University. Mr. Tad Davis has extensive experience in Object Oriented and User Interface technologies. Prior to joining CERC Mr. Davis was employed at Bell Atlantic Software Systems.