( Case 65)
Cognition—the process of perceiving and knowing—underlies all of human problem solving and decision making. In recent years, attempts by scientists in several fields to understand cognitive processes have converged within the discipline of cognitive science. Cognitive science views the mind as an information processor that receives, transforms, retrieves, and transmits information. The discipline seeks to learn how information is stored and how the processes that interpret this information operate. In this chapter, we describe briefly some of the theories about the storage of information, the characteristics of memory, the nature of strategies for searching material stored in memory, the nature of expertise, and the attainment of expertise.
Studying Mental Processes
( Case 62 , Case 65)
Scientists and philosophers have struggled for centuries to understand the structure and function of the mind. For many years, theories of cognition were formulated by learned people based on their personal, introspective theories of their own mental processes. In recent years, however, the theories derived from such introspective approaches have come under considerable question because of inconsistencies between them and experimental observations.[ 88],[ 89] Because introspection is not considered trustworthy, other approaches to understanding the function of the mind have evolved. Prominent among these approaches is the detailed analysis of transcripts of recordings of individuals who were “thinking aloud” as they solved problems, including clinical problems (the process of protocol analysis, or transcript analysis).[ 18],[ 19],[ 36],[ 47],[ 61],[ 89],[ 90]
Typically a problem is presented to a subject; the subject describes what he or she is doing while solving the problem; the session is recorded and transcribed verbatim; and a domain expert then analyzes the transcript. This so-called descriptive approach to the study of reasoning assumes that speaking while thinking is not dissimilar to thinking without speaking. Although investigators who use transcript analysis acknowledge that not all mental strategies are captured by this technique because some mental processes may not be verbalized, such as the short-cut heuristics described earlier,[ 88],[ 91] they infer that the transcripts provide selective glimpses of intermediate points and illuminate states that people pass through as they solve problems.[ 90],[ 92 ], [ 93]–[ 94]
They assume that the analysis provides a running series of responses of behavior from which one can infer the sequence of mental states and reasoning processes that operate in solving problems. In many instances, data from transcript studies have been implemented as a working computer program. This implementation provides evidence that the information obtained in such studies is sufficient to perform the task at hand, although few would claim that the computer program directly models the function of the mind.
The Structure of Memory
( Case 63, Case 64, Case 66)
The mechanisms by which knowledge is stored are avidly debated. One hypothesis, the physical symbol system hypothesis, argues that information is stored in the form of symbols that represent objects, events, and relations between these elements. The form in which symbols might exist is actively debated. One such form is a structure known as production rules. A production rule (also known as a condition–action pair) is a compiled form of categorical knowledge in the form of an “IF-THEN” statement, with the IF part of the statement representing some semantically meaningful condition (e.g., a symptom cluster such as dyspnea on exertion and orthopnea) and the THEN part of the statement representing some action to be implemented whenever the IF condition is satisfied. In the case of this particular IF example, the THEN part might be “generate the hypothesis left ventricular failure.”
Another form is a structure called a frame. A frame is a list of declarative (factual) and procedural (processing) aspects for dealing with a given entity. A frame for a disease entity would contain some hierarchical structure into which the entity fits, findings necessary and sufficient to define the entity, factors that cause the disorder, complications of the disorder, approaches to distinguish it from other entities, and some mechanism to score the relative importance of expected findings. Although frames have been implemented as computer-based diagnostic decision support, they have not added much to our understanding of human cognition.
A third symbolic form has been named a script, a complex description of a particular kind of experiential episode, such as a patient encounter. According to this hypothesis, our memory does not contain abstract descriptions or models of diseases but instead comprises individual specific “training cases,” and we interpret a new case by recalling a similar specific instance or example (called an exemplar) for comparison.[ 95], [ 96]–[ 97] A discipline known as case-based reasoning exploits the notion of exemplars as an approach to understanding reasoning.[ 98], [ 99], [ 100]–[ 101]
Case-based reasoning holds that the storage of specific cases is important in diagnosis. The concept proposes that routine diagnosis is done by reference to knowledge structures that contain case-specific information about the context in which the disease develops, the clinical features, a description of the malfunction, and the disease’s consequences.[ 13],[ 97], [ 98], [ 99], [ 100], [ 101]–[ 102 ] Such knowledge is thought to be tied together with causal links and organized in a temporal sequence that integrates the events as a cohesive story. This story is thought to be the content of a script.
According to this hypothesis, diagnosis involves identifying the information obtained for a patient, searching for an appropriate script by some process of pattern recognition, selecting the script, and verifying the script. Scripts can be prototypes of disease (the most general) or they can be exemplars, that is, descriptions of individual patients (the most specific).
This concept holds that knowledge of clinical medicine may exist at various levels and that this knowledge changes as expertise develops. The first and most elementary level contains extensive pathophysiologic details in some kind of network.[ 47] After experience with more cases, these causal models become simplified, compressed, and compiled.[ 47],[ 103] The second level consists of such compiled knowledge constructed into general diagnostic skeletons that describe either a category of disease or a specific disease entity. The third level consists of multiple exemplars, that is, idiosyncratic scripts based on actual experience with a specific patient (instance scripts).
This hypothesis supposes that learning proceeds through a series of transitory stages, starting with pathophysiology, proceeding with a compiled version, and ending at the highest level with exemplars.[ 13],[ 34] The attractiveness of the hypothesis is in the capacity of these multiple stored exemplars not only to represent disease polymorphism, but also to explain expertise. Expert diagnostic performance, according to proponents of this hypothesis, is achieved after accretion of a myriad of exemplars in the form of instance scripts.
Presumably experts use pathophysiologic knowledge only when the problem is difficult and other methods fail (i.e., when script knowledge does not apply or is not available).[ 47],[ 103] This concept is consistent with experiments on expertise in the field of physics: Such studies show that expertise is a function of knowledge structures available in several different forms.[ 96]
If there are symbolic knowledge structures in memory, scripts are not the only ones. Where no specific script exists (e.g., when an individual encounters a new situation), presumably a set of general rules exists to solve the problem. We suppose that many different knowledge structures could be accessed to solve such a problem. Such structures could include items, goals, themes, and plans.[ 102 ] In medicine, certain forms of knowledge that cut across disease entities might be stored in nonscript form, possibly as rules (perhaps the IF-THEN rules described earlier). Forms of knowledge that might be coded in this fashion include prevalence of disease and characteristics of tests and treatments. It is difficult to imagine, for example, that we index the efficacy of computed tomography scans or the complications of various drugs according to specific disease entities or individual exemplars.
It seems more likely that we store the characteristics of procedures and therapies in some kind of generalization independent of specific diseases. Furthermore, it seems quite unlikely that only a single script is accessed when searching for a solution to a problem. Given the powerful effect of reminding (i.e., certain concepts remind us of others within the same domain and even in different domains, just as physical objects and events have reminding effects), a given set of circumstances can bring to mind a solution to the problem at hand even if the circumstances and the problem are not related.[ 102 ] Reminding is an essential aspect of understanding a new situation as a function of previously processed situations.[ 102 ]
Finally, some hold that the brain is a parallel computational device, and that representations of the world are held not as symbolic structures in the form of rules, frames, or scripts but as distributed patterns of activity across a network of neurons. This hypothesis, known as connectionism or parallel distributed processing, proposes that meaningful patterns are generated when sets of neurons are activated jointly and that knowledge is stored in the interconnections among a large number of processing units, namely neurons. This concept gains credibility from studies in which large, rapid, parallel processing computers (“neural networks”) have been programmed to simulate a number of functions such as vision, pattern recognition, and cognitive information processing.[ 104], [ 105]–[ 106]
Storage and retrieval of information depend on the functioning of memory. Long-term memory appears to be infinite in capacity, and although information in it is long lasting, retrieval from it is slow.[ 107] Working memory, otherwise known as short-term memory, contains only information under active manipulation. It is widely accepted that working memory is limited in capacity to some 5 to 10 items, and that its contents rapidly change as attention shifts away from the items.[ 23],[ 108]
Retrieval from working memory, however, is rapid. Skilled memory is a special adaptation of long-term memory. It is thought to contain chunks of semantically meaningful material organized into elaborate cognitive structures. In other words, by clumping bits of information into easily remembered salient “chunks,” recall of these items from memory is enhanced. When information is organized in this fashion, long-term memory becomes an effective extension of short-term memory.
The search for a solution to a problem (including a diagnostic problem) involves developing a representation of the problem, making inferences about possible solutions, gathering and interpreting data, wending a path toward a solution, deciding on the “best” solution, and “confirming” the result. In this discussion, we explore the nature of search strategies in solving diagnostic problems.
To set the stage for our discussion of search strategies, we pose this simple problem: Suppose you are looking in your file of 300 papers on pulmonary embolism for a specific reprint. Let us assume that there is no other access to the data. You remember seeing the reprint in your file recently, but it is not in the pulmonary embolism folder. No one else has access to the file. How do you find the reprint? You might consider checking each of the 300 reprints or selecting reprints at random, but these strategies are highly inefficient and time consuming.
Alternatively, you might speculate that you placed the paper in the wrong folder and look in other folders that are related to pulmonary embolism by some semantically meaningful association: for example, anticoagulation, postoperative complications, phlebothrombosis, or membranous nephropathy. In the latter strategy, you are making an educated guess and then testing it.
How is this example relevant to searches for solutions to medical diagnostic problems? Most medical problems are not as simple as this. First, many do not have a straightforward solution, such as finding the one and only reprint. Second, many medical problems have more than one solution: Two diseases might interact—one might cause the major clinical manifestations and another might cause only a few others. Third, manifestations that initiate a search for a solution are sometimes quite specific and other times quite vague.
The search for a solution may be relatively easy when a heavy smoker presents with cough and hemoptysis but far more difficult when a previously healthy person presents with malaise and weakness. In both medical examples, of course, a systematic search through all possible causes of the individual clinical manifestations is neither efficient nor effective. The “review of systems” probably will turn up interesting and important clues in both hypothetical medical examples but will not be likely to give the “answer.”