The expanding influence of computers on society is being felt in medicine as well. Essentially all hospitals and clinics depend on computers for administrative and financial functions and for providing access to clinical data. Most physicians have been exposed to the powerful available systems for searching the biomedical literature by computer. Modern imaging techniques depend on computers for image generation, small computers have become mandatory elements in the research laboratory, and information systems are becoming vital topics for medical education. The clinical community has long anticipated the day when computers would be able to assist with diagnosis and with making decisions about patient management. Examples show that technology increasingly will provide physicians with clinically useful decision-support tools. This chapter describes some of the issues in building such systems and in making them easily available, clinically useful, and broadly acceptable.
About 30 years ago, the first research programs for medical diagnosis were shown to be highly accurate in their diagnostic predictions. Why have such systems resisted widespread implementation and acceptance in the intervening years? The barriers reflect the subtleties of the medical practice environment and the increasing recognition that good advice from such programs cannot ensure their use and acceptance. Confounding the implementation efforts have been logistical constraints, the sometimes awkward mechanics of computer use, confusion about whether such programs are intended to be tools for rather than competitors with clinicians, and a pervasive belief that computers do not really “understand” medicine and therefore cannot be expected to reach complex expert decisions. Early successes largely reflect an increased understanding of the importance of such issues and a gradual change in physician attitudes regarding computers and their potential for a beneficial clinical role.
Efforts to construct effective medical decision-support tools are closely tied to research into the nature of medical knowledge and its use for problem solving. Psychologic studies have taught us that medical expertise involves the application of factual knowledge and skills in using hypotheses to guide data collection. This important distinction between what is true and how to do things has had a profound effect on the development of computer-based advice tools. The research has also affected medical education, demonstrating that it is as important to teach decision-making skills and techniques for knowledge use as it is to expose students to the factual basis of medical practice. The Need for Assistance With Clinical Decision Making
Knowing how to find the information needed for clinical decision making is as important as trying to memorize it. Innovations such as medline are a tribute to the way in which computers can be used to help physicians and others to find the information that they need to make good decisions. Computers have superb memories, but until recently, there has been little they could do with pieces of text information other than display them. Modern programs contain detailed clinical information and have the knowledge to help users to determine how that information should be applied. As computer-based tools offer this kind of expertise, physicians have begun to consult programs to obtain reasoned advice about diagnosis and management of specific cases, while the ultimate decision-making roles are maintained for provider and patient. Obstacles to the Effective Introduction of Decision-Support Systems
Experience in applying computers in medicine over two decades has shown that logistical issues are the principal causes of system failures, especially for decision-support systems that rely on physician interaction for optimal use. Any medical computing system will fail to be accepted if it is unduly time consuming or if the cost is not clearly justified by the benefit gained. There is also an issue of inertia-the disinclination of busy people to use a computer if it requires an interruption in the normal routine. This implies the need for an integrated model whereby computers are used routinely for traditional data-management tasks, perhaps in lieu of traditional pen and paper data-recording techniques, and from which the physician obtains advice as a byproduct of this ongoing interaction. Decision-support tools will be more feasible as improved medical record and information systems begin to appear in hospitals and offices.
Equally important is the design of decision-making tools that are sensitive to the traditional independence and skill of physicians. There is a need for system transparency (ie, ability for the program to provide explanations) and for tactful presentation of advice. Such features make it clear that the unique skills of trained physicians are respected and that the system should be viewed as a knowledge-management tool rather than as the decision maker itself.
Medical informatics researchers have long sought to design interactive techniques that avoid clumsiness, typing, or the need for prolonged training of the intended users. Programs have tended to use a pointing device, such as a light pen or a mouse, as the means for manual selection of items and navigation through a program’s features. Options for physician-computer interaction have been greatly expanded by the introduction of hand-held mobile computers, controlled by the use of a pen or stylus and capable of communicating with other computers through wireless networks. Considering the clinical settings in which physicians work, moving from room to room or bed to bed, mobile computing is especially attractive.
The introduction of the “information superhighway,” a concept based on the evolving Internet, has made it reasonable to expect that physicians will soon use portable devices and wireless communications to access clinically pertinent information and to garner advice from remarkably diverse and distant resources. It is already possible to find a large amount of medical information by searching the Internet for topics of interest. A wide variety of resources are available, such as online versions of the clinical practice guidelines that have been developed and distributed by the Agency for Health Care Policy and Research . We will soon see hand-held machines, controlled by the use of a pen and sufficiently portable to be carried like a clipboard, which provide wireless access to all the information on the Internet from the patient’s bedside or the clinic examining room.
The use of such network-based information resources will be enhanced if the network serves a wide variety of clinical needs. For this reason, medical computing experts are seeking to ensure that the evolving national network is built with the needs of the health care system in mind. A range of facilities can be provided to practitioners and to patients who seek access to the health care system or to information on disease prevention and health promotion Serious questions exist regarding management of such a complex range of information resources, including quality control and effective means for finding the best information available for a specific clinical problem. >
A clinical decision support system is generally viewed as providing patient-specific advice rather than generic information such as medline or clinical practice guidelines. Many programs have been developed to address such problems. Since the 1950s, researchers have recognized the relevance of Bayes’ rule to the task of diagnosis, with computers being used to manipulate the pertinent conditional probabilities. Many bayesian programs are highly accurate in selecting among competing explanations of a patient’s disease state. The results of such analyses are probabilities, and their defense is based on Bayes’ formula plus the accuracy of the sensitivity, specificity, and prevalence data used. The mathematical orientation of such programs has not always appealed to physicians who are used to discussing a more concept-oriented line of reasoning with their consultants.
Many clinical questions deal more with what actions to take in managing a patient than they do with determining the proper disease classification. Researchers have accordingly developed computer-based tools that draw on the methods of decision analysis. One class of programs permits the specification of decision trees, along with associated probabilities and utilities, and then calculates expected utilities and performs sensitivity analyses. A second class has used prespecified decision-analytic models to advise physicians who are not trained in decision analysis; such programs’ utility is limited to cases that correspond closely to the decision tree provided.
Since the early 1970s, researchers have been applying the techniques of artificial intelligence (AI) to the development of diagnostic and therapy-management consultation programs. AI is the field of computer science that deals with the symbolic (instead of numeric) representation of knowledge and its use in problem solving. The field is closely tied to psychology and to the modeling of logical processes by computer. Of particular pertinence is the subfield of AI research known as expert systems. An expert system is a program that uses knowledge derived from experts in a field to provide the kind of problem analysis and advice that the expert may provide.
Several barriers continue to limit the effective implementation of decision-support tools in clinical settings. However, rapid technologic progress, especially in the development of powerful but inexpensive single-user computers that provide intuitively pleasing interactive environments, has allowed the introduction of the first commercially available diagnostic programs for physicians. Representing Medical Knowledge
An ongoing research challenge is the need to refine methods for encoding the knowledge used by medical experts. Although powerful techniques exist (eg, many systems use frames, rules, or belief networks to express relations among medical concepts), several complex challenges remain. For example, physicians use mental models of the three-dimensional associations among body parts and organs when they are interpreting data or planning therapy. Representing such anatomic knowledge and performing spatial reasoning by computer have proved particularly challenging. Similarly, human beings have a remarkable ability to interpret changes in data over time, assessing temporal trends and developing models of disease progression or the response of disease to past therapies.
Another kind of expertise is the human skill inherent in knowing how to use what is known. In medicine, the skill is called good clinical judgment to properly differentiate it from the memorization of factual knowledge or of data from the literature. Giving computers factual knowledge will not make them expert unless they also are skilled in the proper application of that knowledge. Improved understanding of human problem solving is helping researchers to develop tools that more closely simulate the reasoning of expert clinicians. Acquiring Medical Knowledge
Researchers are working to devise methods that will facilitate the development and maintenance of medical knowledge bases, especially in light of the rapid pace at which medical knowledge evolves. Some have formalized the techniques by which experts should be interviewed to infer how they are solving problems. Efforts have been made to develop computer programs that acquire the knowledge base for a consultant program by interacting directly with the expert, thereby avoiding a computer programmer as intermediary. Physician-Computer Interaction
It is often claimed that physicians will resist computer-based tools until they can talk to them as they would to a human consultant. Although techniques for speech understanding by computer do not yet permit reliable free-form spoken interaction, early commercial systems exist for domains such as radiology and pathology reporting. Innovative graphic environments have demonstrated that computer novices can rapidly learn to use software tools, making the need for speech understanding less of an issue. Integration With Database Systems
The successful introduction of decision-support tools is likely to be tied to their effective integration with routine data-management tasks. This suggests the need for innovative research on how best to tie knowledge-based computer tools with programs designed to store, manipulate, and retrieve patient-specific information. As hospitals and clinics increasingly use multiple small machines optimized for different tasks, the challenges of integration will be inherently tied to issues of networking. It is in the smooth linkage of multiple machines with overlapping functions that distributed but integrated patient data processing will be realized.
With the emergence of medical informatics as an academic discipline and the training of individuals who can work effectively at the interface between medicine and computer science, it is likely that physicians and other health professionals will see rapid development of clinically useful tools that are sensitive to the realities of the health care environment. Among such systems will be decision-support programs that are integrated with patient data-management systems. They will serve as knowledge-access and management tools for physicians who will rely on their support much as they rely on paper-based reference tools today.
Although the scientific and logistical barriers to their successful implementation are great, progress has been rapid, and new technologies have offered especially promising solutions to many of the problems. The first commercial systems are already in use, and their enhanced integration into routine care environments is expected over the next decade.