Full title

Computer-aided decision making and human-machine diversity: The Mammography Case Study

Keywords

diverse redundancy, decision making, decision support, computer-aided detection, mammography, breast cancer screening, probabilistic modelling, ethnography, automation bias

Summary

Computer aided decision making involves human operators supported by automated devices that offer advice, filter or enhance information, issue alerts and prompts. It thus involves a human-machine system whose dependability is affected by the design of the decision support tools, the procedures for their deployment and use, the training of operators and the definition of their roles.

The intended role of automation is often purely auxiliary: the human user retains the authority and responsibility for decisions. The design intention is to exploit protective redundancy with diversity: the machine protects against some human errors, and vice versa.

The assumption that "computer support may improve things but not make them worse" would be very helpful in decisions about deploying a system. So, it is important to ask whether the assumption holds in a specific case; in which cases it is likely not to hold; and how to judge e.g. how unreliable a support tool must be for it to stop helping and start damaging human decision making.

As an example of such systems, DIRC studied the dependability of Computer Aided Detection (CAD) in mammography. This is a decision system formed by clinicians using computer support (CAD tools) in reading X-ray images (mammograms) for breast cancer screening.

The main features of this case study are the following:

• we used an interdisciplinary approach which combined insights from: reliability engineering, psychology, human factors, ethnography;

• we obtained intriguing, hitherto unreported, findings about potentially detrimental effects of CAD on human decisions for some categories of cases and users (which we summarise below);

• our findings are unexpected and potentially important and have been acknowledged as such by mammography practitioners: one of our papers [Alberdi et al., 2004] won the 2005 Best Clinical Paper award from the Association of University Radiologists;

• these results have practical implications for the design, deployment and evaluation of CAD for mammography and computer aided decision making in general.

CAD for mammography:

• CAD tools are designed to help mammogram readers not to overlook details that they ought to examine to reach their decisions

• The starting point for DIRC's involvement with this problem was an earlier study carried out by University College London (UCL) for the NHS Health Technology Assessment (HTA) programme, which compared mammogram readers' performance  with and without CAD

Interdisciplinary approach:

• We collected new data specifically to investigate how clinicians react to incorrect output from the computer aid;

• We conducted more detailed statistical analyses than used in previous statistical studies, highlighting how the machine's effect varies between mammograms and between clinicians and as a function of whether the machine provides correct advice or not;

• We used insight from work on software systems with diverse redundancy to drive probabilistic models; these underline that: variation and co-variation of "difficulty" of input cases for different system components affect the dependability of the overall system substantially; focusing on the average probabilities of the failures of the components, or assuming statistical independence among their failures, can be misleading;

• We exploited direct ethnographic observations of users at work in real and experimental settings, to provide checks on the realism of experiments and to help identify limits to the generalisability of results, as well as suggesting possible phenomena and causal mechanisms.

Main outcomes:

Fig. 1. Impact of CAD on different categories of cases.

Fig. 2. Impact of CAD on different categories of readers.

• Even though the HTA study showed no significant overall effect of CAD on human performance, our analyses suggest that CAD did affect decisions of the clinicians who participated in the trial: CAD was useful for some classes of cases and readers (mostly when it provided correct advice); but also had a detrimental effect for other categories of cases and readers (mostly when it provided incorrect advice). Figure 1 & Figure 2 illustrate the kind of effects observed;

• Our new experiments on expert readers also indicated that incorrect output from CAD may have detrimental effects on the decisions of some of its human users, even if well-experienced and confident, for some categories of mammograms;

• Ethnographic observation highlighted how readers reasoned about the machine's behaviour through the explanations they gave of and for its patterns of error and successes: it showed how readers anticipated the system's responses, made inferences about how it worked, and formed judgements as to its effectiveness. Ethnography also identified ways that the natural behaviour of readers may violate the intended use of the CAD tool.

Practical implications

• The way the effect of CAD varies over case-reader pairs may affect its overall effect on a patient population more than its average rates of false negative and false positive errors [Strigini et al., 2003];this complicates the assessment of a new technology, but may open cost-effective ways of achieving better overall quality of decisions [Alberdi et al., 2005(b)].

• Our interdisciplinary approach has helped in conjecturing specific cognitive processes that account for the detected behavioural patterns; this, in turn, may further help designers to avoid undesired effects [Alberdi et al., 2005(a)].

• We expect these results and methods to be useful for research, design practice and policy making in many other areas of use of advisory systems.

Links

The Mammography Case Study at CSR, City

Diversity Theme

Probabilistic Models

Ethnography

Papers

Author

Eugenio Alberdi, Andrey Povyakalo, Lorenzo Strigini and Mark Hartswood