Medical Device & Diagnostic Industry Magazine
MDDI Article Index

Adding Intelligence to Medical Devices

Ralph J. Begley, Mark Riege, John Rosenblum, and Daniel Tseng

The medical device industry is seeing an emergence of computer-based intelligent decision support systems (DSSs) and expert systems, the current success of which reflects a maturation of artificial intelligence (AI) technology. The addition of intelligence to a medical device can be extremely successful. Consider, for example, the Agilent Acute Cardiac Ischemia Time-Insensitive Predictive Instrument (Agilent Technologies; Andover, MA), an intelligent ECG device that predicts the probability of acute cardiac ischemia (ACI), a common form of heart attack. A study conducted by the Agency for Healthcare Policy and Research concluded that more than 200,000 unnecessary hospitalizations and 100,000 unnecessary cardiac care unit admissions could be prevented each year if this device were used in U.S. emergency rooms. Approximately $100 would be saved for each of the 7 million annual U.S. emergency room visits for chest pain, translating into a savings of roughly $700 million yearly in hospital costs.

The Agilent Acute Cardiac Ischemia Time-Insensitive Predictive Instrument from Agilent Technologies (Andover, MA) increases the accuracy of diagnosing acute cardiac ischemia.

Defining "intelligence" in a computer system can be problematic since intelligence can be defined as something that humans possess and computers do not. In AI, however, an "intelligent system" is one that exhibits behaviors normally associated with human intelligence. Systems that perform complex diagnostic procedures and pattern matching are considered intelligent.

This article will focus on intelligent DSSs and not the lower-level DSSs that simply organize data to assist a decision maker. An expert system, which can be considered a subset of the DSS technology, is software that contains expert knowledge and attempts to solve problems at a level equivalent to or better than human experts. In general, expert systems provide decisions, whereas other DSSs support the decision-making process of clinical experts.

Until recently only a few intelligent systems achieved common clinical usage. This article illustrates how focused, specialized intelligent systems can be used in the healthcare environment whenever their knowledge areas are well bounded.

TECHNOLOGY OVERVIEW

Beginning in the mid-1950s, a frequent goal of clinical expert systems was to virtually replace the physician with a Greek oracle model of clinical decision making. The object of this line of thinking was to create a "doctor in a box" capable of querying the physician or medical technician regarding a patient's symptoms and generating a diagnosis. Early expert systems sparked considerable excitement and resulted in a high level of expectation in the late 1960s and 1970s. However, the quantity of information required and the need for rigorous organization and informed decision criteria for reliable output bogged down these programs. They also failed to provide the kind of support that physicians wanted: an assistant rather than a replacement.

Early examples of expert systems used Bayesian probabilities (see sidebar, below) and heuristic reasoning, which can be described as rule-of-thumb techniques. The 1970s saw the introduction of rule-based expert systems (see Table I), such as Mycin, which used its rule base to collect information for the identification of organisms causing bacteremia and meningitis. Many rule-based systems have been developed over the years, however—due to the extreme complexity of maintaining rule sets with more than a few thousand rules—rule-based systems have historically been devoted to narrow application areas.

Table I. Currently, numerous technologies are used for the creation of DSSs and expert systems.

Neural networks and genetic algorithms (see Table I) form one of the most recent trends in the development of computer-assisted diagnosis. While the previously mentioned DSSs are knowledge-based systems drawing on existing bodies of encoded medical knowledge, neural networks and genetic algorithms must "learn" their knowledge interactively from the user. These types of applications have been used in the treatment of back pain, the diagnosis of breast cancer, and the classification of giant-cell arthritis and acute myocardial infarction.

CURRENT APPLICATIONS

As DSS technology continues to develop, devices not employing such systems may become obsolete. In general, physicians are more willing to accept systems focused on very specific areas than those attempting to solve more generalized problems such as diagnosis (see Table II). While teaching institutions and universities were early adopters of DSS technology, its acceptance in the general healthcare market is rapidly increasing. Today, DSSs are being used successfully in many areas of the medical device industry, including cardiac monitoring and automated ECG, medical imaging, clinical laboratory analysis, respiratory monitoring, electroencephalography, and anesthesia.

Table II. General types of DSS applications.

Cardiac Diagnosis and Monitoring; Automated ECG Analysis. Cardiac applications—in particular, ECG analysis—constitute a major area where diagnostic decision support has taken hold. The ACI TIPI mentioned in the first paragraph of this article provides real-time guidance in the diagnosis of ACI and improves the accuracy of triage decisions. It has been tested in controlled clinical trials at a number of hospital emergency medical departments within the United States. The input consists of a series of questions to be answered by the physician, as well as data taken from the patient's ECG. The output is an assessment of the probability that the patient has ACI. This is an example of a statistical, probability-based DSS, which allows emergency-care physicians to make better-informed decisions in critical situations where speed of diagnosis is important.

GE Marquette (Milwaukee) produces automated ECG analysis systems that are in widespread use. The CardioSys Exercise Testing System enables the physician to monitor and analyze data from a patient undergoing exercise testing procedures. This device, as well as the MAC 5000 Resting Test System, incorporates the Marquette 12SL ECG analysis program, an integrated DSS that uses newly developed digital processing methods and diagnostic program algorithms to interpret and classify ECG waveforms.

A series of diagnostic ultrasound systems has been developed and marketed by ATL Ultrasound (Bothell, WA) for the purpose of imaging and monitoring cardiac tissue structure and activity. The system uses an adaptive intelligence algorithm to examine specific tissues by actually optimizing several thousand parameters during a patient examination, thus eliminating irrelevant frequencies in returned signals. Over 10,000 systems are in use in clinics and hospitals worldwide.

Perfex, a rule-based expert system developed at Georgia Tech, aids in the diagnosis of heart disease and is currently undergoing clinical evaluation. The system infers the extent and severity of coronary artery disease from myocardial perfusion imaging and produces a report that summarizes the condition of the three main arteries. Perfex was developed using Blaze Software's (Mountain View, CA) Nexpert, an object-oriented development environment for rule-based expert systems. Ambulatory blood-pressure monitors produced by many firms, such as DynaPulse by Pulse Metric (San Diego) and Omron by Omron Healthcare Inc. (Vernon Hills, IL), use pattern recognition and fuzzy-logic algorithms to increase measurement accuracy.

A system to aid in early diagnosis of bacterial sepsis in newborn premature infants is under development by Medical Automation Systems (Charlottesville, VA) in collaboration with researchers at the University of Virginia Medical Center. A statistical analysis of heart rate variability in ECG data finds abnormal patterns that may help diagnose the disease 12 to 24 hours earlier than is currently possible.

Medical Imaging and Microscopy. Image matching is one of the major areas of application for artificial intelligence algorithms that access pattern databases and attempt to match them to patient data to determine specific medical conditions.

The Micro21 Microscopy Workstation, developed by Intelligent Medical Imaging (Palm Beach Gardens, FL), performs automated white blood cell differential tests, red blood cell morphology analysis, platelet estimates, and white blood cell estimates using a neural-network-based algorithm to locate, preclassify, and display both white and red blood cells.

IRIS Inc. (Chatsworth, CA) has developed diagnostic imaging systems that incorporate automated intelligent microscopy technology for use in hospitals around the world. White Iris, a leukocyte differential analyzer, is a fully automated microscope system that detects and classifies leukocytes according to size, color, and image profile using a digital processing algorithm. Yellow Iris is a workstation for urinalysis imaging that performs automatic detection and classification of particles and cell types found in urine samples.

Papnet, by Neuromedical Systems Inc. (Upper Saddle River, NJ), uses neural network technology to scan Pap smears and identify suspicious cells for review by cytotechnologists in cancer screening. Clinical data have shown that significantly more abnormalities are detected when using the Papnet system than by manual screening with a microscope.

Respiratory and Vital-Signs Monitoring. Various systems for monitoring respiratory conditions have been developed and tested in intensive- and emergency-care units. Using the feedback from patient monitors, these systems control respiratory ventilators, which provide breathing assistance to patients.

A closed-loop ventilator system called NéoGanesh, manufactured by the National Institut for Health and Medical Research (INSERM) and the department of physiology and the ICU department of the Henri Mondor Hospital (Créteil, France), incorporates an explicit representation of time and a knowledge base representing a physician's expertise. The system interprets clinical data in real time and controls the mechanical assistance for a patient suffering from lung disease.

The University of Pennsylvania Medical Center (Philadelphia) has developed a "smart" ICU system that improves the vital-signs monitoring of critically ill patients. A combination of neural-network and fuzzy-logic technology is used to convert a patient's vital-sign measurements into easy-to-follow visual models to assist physicians and nurses in monitoring patients' physiological parameters.

An example of DSS system work flow. The DSS is connected to a testing device for the purpose of presenting real-time evaluations of results to assist the healthcare provider.

Electroencephalography (EEG) Interpretation and Automated Anesthesia Delivery. Aspect Medical Systems (Natick, MA) has developed monitors to assess the depth of the anesthesia state based on the statistically derived Bispectral Index (BIS) reflecting the level of sedation. Community Hospitals Indianapolis is successfully employing the BIS monitor to improve the administration of anesthesia during surgery and has found that this technology contributes to improved patient care and reduced costs. The Automated BIS Controller, in development at the University of Pittsburgh Medical Center, controls the rate of anesthetic drug infusion using the BIS as a feedback control and combining it with a pharmacokinetic control program, such as Stanpump or fuzzy-logic controllers.

Developed at the University of Glasgow and evaluated with several hundred patients, CLAN is yet another system for the closed-loop control of anesthesia. The system, which is capable of automatically controlling anesthesia in a spontaneously breathing patient during surgery, is based on recording and analyzing the auditory evoked response to measure the depth of anesthesia. The system incorporates an EEG amplifier connected to a standard PC and a computer-controlled anesthetic infusion device. Researchers at the Pacific Northwest National Laboratory (Richland, WA) have developed a neural network prototype system to assist anesthesiologists in monitoring the depth of general anesthesia during surgery.

DEVELOPING DSS FOR MEDICAL DEVICES: A HOW-TO GUIDE

This section focuses on the challenges a medical device manufacturer faces when considering the addition of intelligent DSS software to a device. Any development of successful device software has many similarities with DSS development in general. The process is not always separable into clearly defined stages, but it pays to identify the following—sometimes overlapping—phases that explicitly address crucial questions:

Evaluation of Feasibility: Is a DSS Worthwhile?

An initial evaluation should determine whether a DSS will enhance the device's value to the user. For decision support software to be practical and feasible, its complexity must be manageable. The problem should be broken down into manageable pieces that can be described and individually evaluated for feasibility. Since development methods for DSSs and expert systems differ in critical ways from those of more-standard software systems, it is critical to ensure that the developers are prepared to use these methodologies and that experts for the system's knowledge area are available to them.

The medical knowledge used to interpret a device's data needs to be expressed as a logical system of components or objects whose behavior is based on rules or other technologies. To be feasible, the system must also be reconfigurable in order to adapt to the ever-changing requirements of the medical field.

Liability issues also are a major consideration. Systems that make or influence medical decisions can incur significant liability burdens. For instance, though anesthesia-related deaths are likely to decrease with the use of closed-loop anesthesia systems, the manufacturers of those systems may worry that some portion of the liability issues that are usually directed toward the hospitals or anesthesiologists may be transferred to them. It is likely that as intelligent systems become accepted, pressure will mount to make it more practical for manufacturers to develop and market these devices.

Planning: How Should the Development Program Be Structured?

Once an intelligent system is determined to be feasible, the development strategy must be planned. It is crucial for system developers to stay in close contact with the user community and medical experts during the entire development process. To ensure correctness and practicality, the developer's solutions need to be compared to the users' needs and the medical experts' knowledge.

The user community should be defined and its characteristics and abilities understood. Representatives of the user community should be designated to evaluate the system developers' designs.

The device manufacturer needs to determine if the in-house staff is available and has the expertise to undertake this development. If system development is contracted out, in-house resources should be available to provide the developers with information and to review the designs.

Technologies must be evaluated to determine which ones are the most practical and appropriate. For example, neural network technology has been successfully implemented in medical image interpretation, while Bayesian probability systems or belief networks have been used in clinical diagnosis where statistical analysis is involved. Systems using heuristic reasoning based on empirical rules have been most successful with alert or reminder systems where the set of rules remains small and maintainable. The emerging multiple-agent technology allows for the organization of the system as a network of independently acting agents, which can facilitate the creation of complex systems. In practical applications, hybrid systems that combine different technologies have evolved.

A preliminary investigation of various technologies may be useful for finding the best fit. Those who address this preliminary issue should be broadly familiar with the wide range of available technologies. Taking bite-sized chunks of the problem and modeling them from different approaches may point out deficiencies in a tentative approach.

Knowledge Acquisition: How Do System Developers Acquire Information?

The medical knowledge that the intelligent system uses must be thoroughly researched and the decision problems clearly identified at the outset. Developers may find it helpful to conduct detailed and extensive interviews with experts in the field, as well as to consult appropriate Internet sources and medical libraries.

The system designers need to involve themselves with the medical experts' knowledge domain and develop an operational understanding of the field. Not only is such acquisition essential, it is actually quite feasible. Proven methods of knowledge acquisition include protocol analysis techniques using transcripts from experts of varying levels who are asked to think aloud while solving a problem, and cognitive task analysis of video recordings documenting expert solution-finding processes.

It is important that the system designers understand the language of the users as well as experts. Quite a few systems have failed because the system developers and users simply did not understand each other, or the system developers were developing the system in isolation from the user community.

Knowledge Mapping: How Should the Information Be Organized?

Once the problem and knowledge domain have been described, this information is mapped to an appropriate "representational entity." That is, the system designers map the knowledge into data structures, sets of rules, or other form of knowledge representation. The experts' knowledge can be expressed, for example, as rules in plain English or as a database, and appropriate data structures can be chosen to express the problem. As an example, if neural networks are used to model the input/output relationship of an ICU monitor, the monitor's input from the patient might naturally correspond to many of the network's input layer nodes, whereas the output layer nodes might be states of the patient.

Research in Medical Informatics has developed extremely useful controlled vocabularies and standards, such as the Systemized Nomenclature of Medicine (SNOMED), the International Classification of Diseases (ICD-9), and the Unified Medical Language System, to represent information from the medical domain and allow for easier interchange and compatibility of medical data between systems.

Prototyping: How Well Does the Initial Set-up Plan Work?

At this point, it is helpful to create prototypes and demonstrations of the system's screens and logical processes to assess both the design's usability and algorithms. Evaluation of prototypes by representatives of the expert and user communities helps to determine if the design will fulfill their needs and if the components, rules, and assumptions are correct.

The design and creation of a system is an iterative process. The experts and representatives of the user communities should be frequently included in the design process. To ensure the creation of a usable system, the designers must repeatedly compare their solutions to the way potential users of the system actually work.

Design: How Will the Final Product Work?

The system design should determine assumptions and maximize reconfigurability. For example, within some rules of a rule-based system there might be variables that the user needs to be able to configure. These variables can be stored in a database while other rules can be coded simply as database queries. When the system's assumptions are clarified, they can be hard-coded (expressed in the source code). In some situations it might be useful to present multiple solutions to a problem and rate them by likelihood or quality.

Designing a system as modular objects will allow the code modules to be reused and often speeds up the development work. Standards such as CORBA for distributed object-oriented technology offer the possibility of distributing platform-independent data over the Internet and standardizing communication between systems. (For information about Common Object Request Broker Architecture by Object Management Group (Needham, MA), visit www.omg.org.)

Several companies are currently attempting to develop knowledge-based tools and methods, such as the Arden Syntax and CORBAmed efforts. A division of OMG, CORBAmed is developing an object-oriented analysis and design methodology to provide the right level of abstractions and definitions of data and processes involved in healthcare.

Implementation: What Software Will Be Used and Developed?

To implement and code the system, developers have to choose whether to customize an off-the-shelf tool or to write the system from scratch. There are reasonably mature off-the-shelf tools for most major technologies: rules, agents, neural networks, genetic algorithms, Bayesian Belief Networks, etc. Expert System Shells—off-the-shelf environments to create expert systems—can be useful in making system development more efficient or for prototyping the system. But complex systems are more likely to be implemented writing custom code, using either a procedural language such as C++ or a logical language such as Prolog.

Another consideration for the long-term success of an application is that the implementation tool should still be available and supported by the vendor for years after the initial development date. For this reason it is sometimes better to use a mainstream development tool rather than a more technically advanced tool which might disappear in a few years.

Quality Assurance: How can Developers Be Sure the System Works and Meets Regulatory Requirements?

Software quality assurance standards for development, testing, and risk management are necessary to assure that the device and the assisting software works well and will be approved. Depending on the criticality of the system and device, the testing as well as the complete development process may have to follow FDA regulations. FDA often considers the software that assists a device as a device in itself. In this case, the software must follow the same regulations as the device. As documented thoroughly in MD&DI, FDA compliance is not a minor matter. In the authors' experience, quality assurance can typically occupy 30% of the resources needed to complete a software development project. Medical DSS developers need expertise in this area in order to succeed. Following quality assurance procedures is often not only required but also speeds up the development process. Clearly written documentation can save the development team time and avoid misunderstandings.

Maintenance and Knowledge Base Tuning: How Can the System Be Kept Up-to-date?

Even when the system is implemented, the development process is not complete. Testing and practical use of a system by qualified users usually shows a multitude of issues and items that need to be adjusted or corrected. Rules may require fine-tuning during alpha and beta test phases. It is of course important that a system's knowledge base continue to reflect the most current medical knowledge. But once a device is on the market, making changes can be an arduous process. Embedded software may demand special consideration since any changes will necessarily require access to the device and such maintenance considerations should be a major part of product planning, especially for an expensive or life-critical device. Validation of changes also is an important issue when FDA approvals are involved.

Implementation of the system in the field is likely to change the way personnel work. The expectations of the users may change as the work flow and procedures adjust around the new system. In one example, a system turned a laborious process of several days' work into an automated process of several hours, the users initially seemed amazed at the progress. After several months, the experts who previously performed this function were now free to do other work. However, for the lab techs who were operating the new system and accomplishing the work for which an expert was previously needed, several hours seemed too long. Further research resulted in an update of the system search algorithms so that the process could actually be completed within a few minutes. In general, user-sensitive field tests of a device can demonstrate how a new system might influence the work of the user and allow usability issues to be discovered and addressed before the system goes on the market.

CONCLUSION

DSS and expert system technology is now ready for the medical device industry. The tool set described in this article is varied and well tested, and current examples demonstrate that the technology for developing these systems has been proven in mainstream medical devices. In the near future, many medical device applications must include intelligent software to remain competitive. As the feasibility of DSSs and expert systems is increasingly demonstrated, there will be a rapid proliferation of intelligent systems, which will enhance the performance of medical devices and help users to interpret complex outputs. Intelligent devices will work to improve medical care and at the same time help to contain expanding healthcare costs.

BIBLIOGRAPHY

Berner, Eta S. "Example from S. Andrew Spooner: Mathematical Foundations of Decision Support Systems," in Clinical Decision Support Systems. New York: Springer, 1998.

Bloom, Marc J. [on-line] "Processed EEG to Control Drug Dosing." (University of Pittsburgh Medical Center, 1998).

Coiera, Enrico. [on-line] "Guide to Medical Informatics, the Internet and Telemedicine." (1997).

de Braal, L. et al. "PERFUSE: An Interactive Knowledge-Based System for the Interpretation and Explanation of Cardiac Imagery." Presented to the IEEE Engineering in Medicine and Biology Society (EMBS) Conference, Amsterdam, Netherlands, October 31C`–November 3, 1996.

Dojat, M et al. "NéoGanesh: A Working System for the Automated Control of Assisted ventilation in ICUs." Artificial Intelligence in Medicine 11 (1997): 97–117.

Ezquerra, N et al. "PERFEX: An Expert System for Interpreting Perfusion Images." Expert Systems with Applications 6 (1993): 459–468.

Haddaway, P. "An Overview of Some Recent Developments in Bayesian Problem-Solving Techniques." AAAI Magazine (Summer 1999): 11–19.

Hammad, Tarek A. "Computational Intelligence: Neural Networks Methodology for Health Decision Support," in Health Decision Support Systems, ed. Joseph K. H. Tan. Gaithersburg, MD: Aspen, 1998.

Hedberg, Sara Reese. "Is AI Going Mainstream at Last? A Look Inside Microsoft Research." IEEE Intelligent Systems (March/April 1998): 21–25.

Jensen, FV. An Introduction to Bayesian Networks. New York: Springer, 1996.

Kenny, G and SE Milne. Auditory Evoked Potentials for Closed Loop Anaesthesia (CLAN), State of the Art Technology in Anaesthesia and Intensive Care, ed. K Ikeda, M Doi, and T Kazam. Amsterdam: Elsevier Science B.V., 1998, 105–112.

Manning, William L. Presentation to the Food and Drug Administration & National Library of Medicine Software Policy Workshop, September 3–4, 1996.

Miller, Randolph A. "Medical Diagnostic Decision Support Systems," in Journal of American Medical Informatics Association (Jan/Feb 1994).

Pearl, J. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, revised second printing. San Francisco, CA: Morgan Kaufman.

Selker, Harry P et al. [on-line] Annals of Internal Medicine 129 (1998: 845–855).

Sherpbier, Harm J. [on-line] ASTM Subcommittee E31.15, E1460 (1992).

"Penn Researchers Develop Smart Intensive Care Unit System Using Advanced Computer Intelligence" [on-line] (University of Pennsylvania Health System, 20 October 98).

"Use of Emergency Room Predictive Instrument Could Save More than $700 Million Annually" [on-line] (Rockville, MD: Agency for Health Care Policy and Research, Press Release, 30 November 1998).

Wood, Bill J. "Software Risk Management for Medical Devices." Medical Device & Diagnostic Industry 21, no. 1 (1999): 139–156.