Thoughts - User interaction models for clinical decision support| The Informatics Review Thoughts - User interaction models for clinical decision support |
Informaticians have been developing knowledge-based clinical decision support systems for over 30 years with many notable successes [Reisman, 1996]. Once the arduous tasks of collecting the medical knowledge required and developing an accurate computer understandable representation scheme are complete, the informatician must determine the appropriate user interaction model. Several different interaction models have been developed over the years. The following sections briefly describe each of the five main user interaction models and provide examples of successful systems that have been developed using each of them.
Interpretation systems present information to clinicians passively. These systems can create relatively simple, yet nicely formatted clinical laboratory reports and graphs, or sophisticated interpretations of such things as electrocardiograms (EKGs) [Klingeman, 1967]. Interpretation systems work best when interfaced directly to the data generated by laboratory instruments that produce numeric data. In addition, there should be a well understood physiologic model in existence which unambiguously interprets the data. These systems have found their greatest clinical utility in areas such as arterial blood gas interpretation [Gardner, 1975], spirometry [Ostler, 1984], and automated PAP smear analysis [Wilbur, 1998], to name just a few.
Consultation systems carry on an interactive dialogue with clinicians in an attempt to help them arrive at a correct diagnoses or therapeutic decisions. These systems can be used by physicians to determine the test or procedure that will be most likely to help them confirm or rule out a specific diagnosis. One of the earliest consultation systems, MYCIN, was developed by Shortliffe in the early 1970s [Shortliffe, 1975]. Clinicians interacted with MYCIN through a long series of questions designed to elicit the patient's clinical state and then help the clinician select the appropriate antibiotic. More recently informaticians have utilized the consultation model for the implementation of clinical practice guidelines. Very few of these systems have met with extensive or even continued use in real-life clinical situations.
Monitoring systems "watch" the clinical database for the storage of particular data items or the passage of a predetermined amount of time. Once such an item is stored in the database, a program is called which "decides" whether the particular data value (or combination of data values) warrants notifying a clinician. Monitoring systems work best on problem areas in which the medical knowledge can be represented in one or more if-then-else type constructs. These systems have met with considerable success in areas as simple as the detection of abnormal laboratory results [Bradshaw, 1989] and adverse drug events [Classen, 1992] or as complicated as ventilator monitoring in ICU patients [Sittig, 1989]. Most clinicians find the "safety net" effect of such systems reassuring and more often than not are happy to comply with the computer's suggestion.
Critiquing systems require that all the patient's clinical data, as well as the clinician's anticipated action be available in the computer. The critiquing system then generates a review of this decision based on its "understanding" of the patient's underlying patho-physiological condition and the risks associated with the planned therapeutic alternative chosen. Such systems have met with their greatest success when incorporated into physician order entry systems. In such cases they can inform physicians of potential drug-drug, drug-lab, or drug-allergy interactions, as well as suggest less expensive alternatives [Teich, 1997]. Given the diagnosis the physician is attempting to confirm or rule-out, a critiquing system might mention a more appropriate radiological exam [Harpole, 1997].
Any of the above mentioned interaction models can be enhanced by offering a "teaching mode" to the user. Such a mode would allow the system to "explain" its reasoning to the clinician. In a landmark article, Teach and Shortliffe stated that the ability of a system to "explain" its reasoning was one of the key factors in clinician acceptance of decision support systems [Teach, 1981]. Since that time many systems have been successfully deployed without this capability, although system developers are still encouraged to provide it when possible. Many developers skirt this issue by citing a scientific journal article or displaying the actual rules (along with the patient's data values) the system used to reach the conclusion.
Choosing the appropriate user interaction model is one of the most difficult and important decisions the informatician must make. The key item that must be considered is the nature of the medical decision being made. Many medical decisions are based on numerous simple and widely agreed upon, rules that all clinicians know but have difficulty bringing to bear with 100% accuracy. Examples of such decisions might include: Does this infant need an MMR vaccination today? Do these particular arterial blood gas values represent a metabolic or respiratory acidosis? Has the patient's sodium value fallen more than 25% over the last 12 hours? These determinations are best implemented as interpretation or monitoring systems. Other decisions are fraught with complicated risk assessments and competing alternatives; they have no clear-cut "best" solutions. Such decisions are best implemented as critiquing systems. The management of a patient's hypertension [Miller, 1984] or the selection of the anesthetic technique to be employed during surgery [Miller, 1983] are two excellent examples of medical decisions which beg for a critiquing system. The consultation mode, on the other hand, has not met with much success in the clinical realm for the simple reason that clinicians are reluctant to spend extended periods of time entering data into a computer in order to receive advice.
Building knowledge-based clinical decision support systems that are intended for routine use in the clinical setting is a difficult, but rewarding challenge. Finding the appropriate user interaction model is one of the most important, but often overlooked, tasks.
References:
Ó 1998 Dean F. Sittig
| The Informatics Review Thoughts - User interaction models for clinical decision support |