5. Limitations of Current Accounts of Medical Diagnosis
Throughout the preceding case study, we have discussed a number of decisions made by the physicians regarding the generation and pursuit of diagnostic hypotheses. In the follow two sections we, first, highlight some limitations of the two primary frameworks used for discussing diagnostic reasoning in the medical literature: (i) the normative, probabilistic framework associated with the threshold approach and (ii) descriptive frameworks based on cognitive psychology. Second, we then offer our constructive proposal: to conceptualize the process of diagnosis in terms of strategic reasoning.
The probabilistic framework is the most popular normative framework employed in the methodological discussions of diagnosis in the medical literature, especially among proponents of evidence-based medicine. This approach is typically summarized as follows (e.g. Richardson and Wilson 2015): First, the physician identifies a plausible differential diagnosis for the patient and assigns an initial prior (or “pretest”) probability to each of the hypotheses in the differential diagnosis. Second, the clinician compares the initial probabilities of the hypotheses to the probability thresholds, as determined by the decision-theoretic models of threshold approach, in order to decide whether to test or treat for the disease. Third, as test-results become available, the clinician should use Bayes’ Theorem together with information about test reliability to update their probabilities.
While this probabilistic framework can highlight important lessons for clinical reasoning,12 it does not provide a general framework for explicating clinical reasoning; the probabilistic approach presents an idealized, simplified picture of clinical decision-making which leaves out many important aspects of the process of diagnosis. In our case study, factors that eventually led to successful diagnoses included: (i) decisions by the emergency room clinician and later by the cardiologist about when to generate more diagnoses for consideration; (ii) choosing effective and efficient strategies for generating relevant hypotheses; (iii) recognizing whether the generated hypotheses can explain the salient symptoms; (iv) recognizing the importance of subtle clues, such as the dilated aortic root or the diastolic murmur, that may initially appear puzzling or unimportant, as well as knowing which features (most of them unmentioned in our description of the case) to ignore; (v) the strategic choice of test (the CT-scan) which could reveal important information for further inquiry even if it failed to confirm the hypothesis tested.
This last point is crucial. The decision-theoretic models of the threshold approach are limited to considering the direct benefits and harms of testing or treating. They do not take into account the kinds of downstream consequences highlighted in Section 3. However, these considerations proved crucial to the successful resolution of our case: the CT-scan produced the crucial clue that eventually led the cardiologist on the right track. Considerations of this type are difficult to represent directly in the probabilistic framework, since it is difficult to assign meaningful probabilities or utilities to these unknown unknowns. What is the probability that a given test will produce a valuable clue for a diagnosis we have not thought of yet? What is the utility of treating this as-yet-unknown disease? Successful diagnosis depends, in part, on recognizing and considering these possibilities. Of course, one can always add a term into the decision-theoretic calculus to represent the weight these considerations are given relative to the direct consequences of testing/treating. But this does not represent the reasoning that leads physicians to give them that weight.
Finally, probabilistic models start from the assumption that one has already formulated a diagnostic hypothesis. In its current form, it only addresses the question of whether the hypotheses generated satisfy the goal of being pursuit-worthy. To the extent that it succeeds in the latter, it at best represents the aim of generative reasoning, rather than describing this reasoning in itself.
When hypothesis generation is discussed in the medical literature, it is done primarily within the framework of cognitive psychology. For instance, while Kassirer, Wong and Kopelman (2010, Ch. 13), discuss hypothesis generation in several case studies, their focus is on which structures of memory allow (or prevent) physicians from recalling the correct diagnosis—e.g. perhaps the physician’s memory is structured in condition-action pairs, one of which state (say) that IF an adult has a high serum cholesterol value, THEN consider the possibility of hypothyroidism (ibid. , 75)? While much can no doubt be learned about generative reasoning from cognitive psychology, these analyzes currently lack a guiding normative framework. Correct diagnosis of course requires physicians to have structures of memory which allow them to recall the correct diagnosis, but as a normative account it amounts to asking “how can we recall the right diagnosis?” As we have argued, the relevant question is rather: which strategies for hypothesis generation allow physicians to generate a manageable set of hypotheses that are most important to consider at the given stage of inquiry?