Academic journal article Human Factors

Interactive Critiquing as a Form of Decision Support: An Empirical Evaluation

Academic journal article Human Factors

Interactive Critiquing as a Form of Decision Support: An Empirical Evaluation

Article excerpt

INTRODUCTION

In examining the artificial intelligence literature, one may find that there have been many attempts to build decision-support systems that provide the user with a conclusion and an explanation for that conclusion. For example, in the medical domain, systems such as MYCIN (Shortliffe, 1976) and MENINGE (Francois et al., 1993) have been developed to aid the diagnosis of diseases, and systems such as ONCOCIN (Shortliffe et al., 1981) have been developed to aid with the management of a patient's treatment plan. Many expert systems are knowledge-based, but some use mathematical models such as Bayesian reasoning (Sutton, 1989). Almost all of these systems, however, follow the same model of decision support: The computer tries to solve the problem and then gives its results (possibly along with an explanation) to the user for review. Users are then expected to critique that conclusion and decide whether they agree with it.

Traditionally, evaluations of such systems focused on whether or not the computer system was able to generate the gold standard (i.e., the best answer) as either the top answer or a highly rated answer on a range of cases (e.g., Bernelot Moens, 1992; Berner et al., 1994; Francois et al., 1993; Hickam, Shortliffe, Bischoff, Scott, & Jacobs, 1985; Nelson et al., 1985; Plugge, Verhey, & Jolles, 1990; Shamsolmaali, Collinson, Gray, Carson, & Cramp, 1989; Sutton, 1989; Verdaguer, Patak, Sancho, Sierra, & Sanz, 1992; Wellwood, Johannessen, & Spiegelhalter, 1992). Subsequently, however, researchers in medical informatics began to realize that focusing only on the computer's performance was a limited and unrealistic evaluation of a decision-support system if the goal was to successfully incorporate the system into actual practice (e.g., Forsythe & Buchanan, 1992; Miller & Maserie, 1990; Wyatt & Spiegelhalter, 1992). The human interface is almost always cited as a problem (e.g., Berner, Brooks, Miller, Masarie, & Jackson, 1989; Harris & Owens, 1986; Miller, 1984; Shamsolmaali et al., 1989; Shortliffe, 1990), particularly because most expert systems require the practitioner to enter data into the computer so that the computer can have the information necessary to perform its reasoning. Thus, one requirement for a successful medical informatics system is to have the necessary data already on-line (Linnarson, 1993; Miller, 1984; Shortliffe, 1990).

A second usage problem is that these systems may have an incomplete knowledge base or may use simplifying assumptions that make them "brittle," meaning that they can fail on cases that the system was not designed to handle. This leaves the practitioner in the role of having to detect and correct any problems generated by faulty computer reasoning (Aikins, Kunz,& Shortliffe, 1983; Andert, 1992; Bankowitz et al., 1989; Bernard, 1989; Berner et al., 1989; Gregory, 1986; Guerlain et al., 1994; Harris & Owens, 1986; Miller, 1984; Roth, Bennett, & Woods, 1988; Sassen, Buiel, & Hoegee, 1994). With this design model, the human user must decide whether or not to accept the computer's diagnosis or treatment plan. A serious concern, however, is that the user in this role of critiquing the computer's answer may become overreliant on the computer or may be led "down the garden path," failing to adequately evaluate the computer's conclusion.

People may ignore the advice of a system even when it is relevant or may heed the advice of a system even when it is faulty. Complacency may occur when monitoring for automation failures if the automation reliability is unchanging or if the operator is responsible for more than one task (Parasuraman, Molloy, & Singh, 1993; Parasuraman, Mouloua, & Molloy, 1994). In addition, less than obvious failures may cause practitioners to be unduly influenced by an expert system's proposed solutions. This was demonstrated in a recent study in the domain of flight planning (Layton, Smith, & McCoy, 1994). …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.