bs_bs_banner Journal of Evaluation in Clinical Practice ISSN 1365-2753 Selecting clinical diagnoses: logical strategies informed by experience Donald Edward Stanley FCAP1 and Daniel G. Campos PhD2 1 Pathologist, Department of Pathology, Maine Medical Center, Portland, Maine, USA Associate Professor, Department of Philosophy, Brooklyn College, The City University of New York, New York, NY, USA 2 Keywords abduction, Bayesian probability, diagnosis, hypothesis, induction, Peirce Correspondence Dr Donald Edward Stanley Associates in Pathology 500 West Neck Road Nobleboro, Maine 04555 USA E-mail: dstanley@tidewater.net Accepted for publication: 8 June 2015 Abstract This article describes reasoning strategies used by clinicians in different diagnostic circumstances and how these modes of inquiry may allow further insight into the evaluation and treatment of patients. Specifically, it aims to make explicit the implicit logical considerations that guide a variety of strategies in the diagnostic process, as exemplified in specific clinical cases. It focuses, in particular, in strategies that clinicians use to move from a large set of possible diagnoses initially suggested by abductive inferences – the process of hypothesis generation that creates a diagnostic space – to a narrower set or even to a single ‘best’ diagnosis, where the criteria to determine what is ‘best’ may differ according to different strategies. Experienced clinicians should have a diversified kit of strategies – for example, Bayesian probability or inference to a lovely explanation – to select from among previously generated hypotheses, rather than rely on any one approach every time. doi:10.1111/jep.12417 Introduction In this article, we shall describe how reasoning strategies identified by the American pragmatists and other philosophers are used by clinicians under different diagnostic circumstances and how these modes of inquiry may allow further insight into the evaluation and treatment of patients. Specifically, we aim to describe and make explicit the implicit logical considerations that guide a variety of strategies in the diagnostic process, as exemplified in specific clinical cases. We are interested, in particular, in strategies that clinicians use to move from a large set of possible diagnoses initially suggested by abductive inferences – the diagnostic space – to a narrower set of testable diagnoses or even to a single ‘best’ diagnosis, where the criteria to determine what is ‘best’ may differ according to different strategies. Experienced clinicians should have a diversified kit of strategies rather than relying on any one approach every time. Our starting point is our previous thesis that the diagnostic process begins with an abductive search for the differential diagnosis. Clinicians begin by knowing nothing definitive about the origin of the signs and symptoms and search for their explanation through abductive conjecturing [1]. By abduction, we mean the logical process of forming an explanatory hypothesis and, following Charles Peirce, we hold that even though abduction only Journal of Evaluation in Clinical Practice (2015) © 2015 John Wiley & Sons, Ltd. asserts its conclusion ‘problematically or conjecturally’, it is an inference that has a definite logical form, namely The surprising fact, C, is observed; But if A were true, C would be a matter of course. Hence, there is reason to suspect that A is true [2]. For example, a doctor may auscultate a dyspneic patient’s chest for crackles. She reasons that if the patient had pulmonary congestion, then crackles would be a predictable consequence. Therefore, she conjectures that the patient may have pulmonary congestion. For Peirce, ‘A cannot be abductively inferred, or . . . cannot be abductively conjectured, until its entire contents is already present in the premiss, “If A were true, C would be a matter of course”’ [2]. This premise amounts to stating that A would explain C, so Peirce is arguing that we cannot infer A (pulmonary congestion) unless our inference is to an explanatory hypothesis. In short, a condition for the admissibility of a hypothesis is that the hypothesis would account for the facts (crackles), and on those explanatory grounds we hold the hypothesis to be plausible. However, for any given set of symptoms, doctors often may conjecture many possible diagnoses; thus, they also attempt to limit the practical error of trying to know everything, and thereby generating too many hypotheses for testing. We have proposed that the elegance of practical diagnosis lies in striking a balance between the generation of hypotheses and the selection of the most 1 Selecting clinical diagnoses appropriate – the highest practical ranking – for testing [1]. In this article, we are focused on the latter, selective stage of diagnostic reasoning.1 Through detailed cases, we will describe several experimental strategies used by clinicians to select plausible diagnoses for further testing or treatment, including 1 Selection of hypotheses by deducing consequences from them and testing experimentally for whether the consequences are realized (Peircean logical inquiry); 2 Assessment of the probability of hypotheses by way of Bayesian and/or frequentist statistical approaches; 3 Selection of hypotheses by way of inference to the loveliest explanation (a version of inference to the best explanation); 4 Selection of hypotheses by criteria such as uberty or fecundity with potential truth, simplicity, explanatory power, theoretical unification and so on, and 5 Combinations of these – for example, estimating the likelihood of diagnostic hypotheses given the evidence through surrogate criteria such as uberty or loveliness, or estimating prior Bayesian probabilities through clinical acumen resulting from previous experiences. According to the Peircean account, after generating a plausible explanatory hypothesis, the scientific inquirer deduces the observable consequences that should follow from that hypothesis. Then, he or she tests experimentally whether those consequences are in fact realized. This is the abductive-deductive-inductive pattern of logical inquiry [2]. Besides being able to deduce testable consequences, however, there are further considerations as to which hypotheses are worthy of testing. At the outset of inquiry, we may abductively generate and tentatively uphold for further scrutiny fertile hypotheses, even if they may seem unlikely or a priori improbable. Peirce distinguishes between ‘security’ and ‘uberty’ in inferential reasoning. The security of an inference is its degree of certainty; the uberty of an inference – especially the inference to a plausible hypothesis – is its potential to lead to undiscovered truth. The uberty of an abductive hypothesis is different from its a priori likeliness, and even from its immediate fruitfulness in solving a problem. However, there tends to be a tension between security and uberty intrinsic to scientific reasoning. Secure reasonings, such as deductions and well-tested inductions, do not open new paths of investigation, new hypotheses, that may be ‘gravid’ with potential truth, even if these paths seem risky and a priori unlikely. The upshot is that in conflating the scientific generation of gravid hypotheses with an evaluation of their likeliness, we neglect the 1 Ross Upshur has argued previously that the ‘dispute between the proponents of clinical common sense and evidence-based medicine can likely find common ground in the philosophy of C.S. Peirce’ [13]. He advised that ‘circumspection must accompany the use of statistical models in clinical reasoning’ and claimed that Peirce’s logic provides ‘a promising framework in which to develop a theory of clinical reasoning that is both rigorous and probabilistic [while also being] able to recognize the uncertainties and particularities of day-to-day clinical practice’ [13]. He did not develop, however, a detailed account of this theory of clinical reasoning. Here we aim to develop one important aspect of it, by expounding the theory and illustrating the practice of diagnosis selection. In this diagnostic process, clinicians must have recourse to a variety of logical strategies in which clinical acumen, insight and experience play a central role to guide statistical, economic, and other considerations. 2 D.E. Stanley and D.G. Campos need for generating, pondering and pursuing hypotheses that are rich in possibilities, even if they are a priori improbable. On the other hand, the Peircean emphasis on uberty seems most appropriate for theoretical scientific inquiry. However, we have suggested that the situation in medical diagnosis is different, due to the immediacy of its practical aims: we must find, as economically and quickly as possible, the explanation for the patient’s condition. Therefore, probabilistic considerations may enter more quickly into the assessment and selection of hypotheses for testing. In other words, once explanatory hypotheses are generated, considerations as to their likeliness to be true soon become important [1]. Questions arise, then, regarding how to assess the likeliness of the candidate explanatory hypotheses. Doctors may, for instance, adopt a Bayesian approach and try to ascertain the probability of the diagnosis (hypothesis) given the evidence – symptoms, preliminary test results and so on. We thus will pay particular attention to the Bayesian model of hypothesis selection and its possible variants or modifications. By Bayesian selection, we mean the process through which, once diagnoses have been generated abductively, the clinician ranks them probabilistically or quasi-probabilistically in order to select the best for testing or treatment. Recall the formulation of Bayes’ rule to estimate conditional probabilities: P ( H E ) = [ P ( E H ) × P ( H )] P ( E ) . This is often interpreted as meaning that the conditional probability that a hypothesis H is true given the evidence E can be estimated from: (a) the probability of the evidence given the supposition that the hypothesis is true – P(E/H), that is, the probability that E will occur if H is true – ; (b) the probability of the evidence – P(E), that is, the statistical frequency with which the facts counted as evidence happen in general – , and the probability that the hypothesis is true, P(H).2 A drawback of with this strategy is the difficulty of establishing the prior probability of the hypothesis, P(H). This is often considered to be a purely subjective guess on the part of an expert. Presumably, however, the assessment of prior probabilities relies on the doctor’s experience and knowledge. We will look at strategies through which experienced doctors may find some reasonable basis for assigning a prior probability P(H) to plausible diagnoses. We will also propose that ‘imaginative deliberation’ enables clinicians to take actionable steps in diagnosis and treatment. By imaginative deliberation we mean a process of reasoning through which doctors ask questions and seek answers in order to select possible diagnoses and discard others. This process involves inter- 2 Notice, however, that in moving from the right-hand to the left-hand side of the mathematical equation, an epistemological shift occurs regarding the facts under study. The patient’s condition, which originally was a fact in need of explanation – or an explanandum – , becomes evidence to assess the probability of a hypothesis – that is, that the diagnosis is correct. We could mark this epistemological distinction by rewriting the equation as P(H/E’) = [P(E/H)*P(H)]/P(E), where E is a fact regarded as an explanandum and E’ is the same fact regarded as evidence. To simplify our discussion, however, in this paper, we will not focus on this epistemological shift from facts regarded as explananda, E, to facts functioning as evidence, E’. © 2015 John Wiley & Sons, Ltd. Selecting clinical diagnoses D.E. Stanley and D.G. Campos play between memory and imagination [3].3 Experts in clinical diagnosis use several strategies to process signs and symptoms, refine the differential diagnosis – that is, the list of possible or probable diagnoses that ought to be considered – in order to take action – for example, further testing, additional family information, biopsy. Medical experience is the link between theoretical knowing and doing. All attempts to formalize medical thinking (knowing) have held our interest in developing expert programmes – for example, DxPlain, Knowledge Coupler, Isabel – but we maintain that clinical diagnosis and decisions (doing) are too complex to reduce to artificial intelligence or computer programs because diagnostic acumen requires the imagination.4 Experienced diagnosis consists in, first, abductive generation and examination of plausible diagnoses, and then the selection of diagnoses for further refinement by testing. This is the experimental method of scientific medicine. This selection often starts from the prior probabilities, P(H), refined by clinical experience so that an expert’s priors are conditionalized as P(H’) = P(H/E) and become her ‘informed priors’. That is, her experience guides and refines ‘naïve or uninformed priors’ – for example, by performing experiments in clinical training to form informed priors. This is what experience means. In order to arrive at informed priors, though, we need something that Bayes does not offer. We need to start with our clinical acumen. We start with our observational assessments (as was Darwin’s case observing the classification barnacles over an 8-year period or Mendel with the constant differentiating characteristics of peas over a 9-year period) from training in medical decisions; through these visual assessments, we take our priors, or they are handed down to us by our teachers, medical journals, mentors. The Bayesian process is an attempt to update our experience after we have chosen which evidence E supports which hypothesis H’ and how prevalence – a measure of the rate of occurrence of a disease in the general population – intersects with H’. Medical diagnosis also depends on the transformation of the patients’ history, physical and laboratory examinations from an indeterminate array into an actionable strategy: that is, how to probe the family history, to reassess the physical examination, to refine testing driven by imaginative deliberation – for example, could this be a manifestation of a common disease presenting in an uncommon fashion? Might it even be a disease not yet described? 3 Colapietro writes: ‘Deliberation understood as the transformative interplay between memory and imagination, is the key to understanding reason in Peirce’s sense; human rationality is, at bottom, the emergent capacity of imaginative deliberations’ [3]. 4 As an anonymous reviewer of this paper noted, in modern medical education students are instructed in the use of diagnostic algorithms and even computer-assisted diagnostics, and these are likely to play an important role in clinical practice of these soon-to-be doctors. A discussion of the interplay of experience and imagination with increasingly powerful automated diagnostic tools would not only desirable but salutary for the future of clinical practice. However, this issue requires a discussion in itself that is beyond the scope of this paper. Our conjecture, based on several decades of experience with the development and attempted application of these programmes by one of the authors, is that these programmes cannot in principle and in practice capture the complexity and variety of manifestations of disease. In the end, there will be no substitute for the clinician’s experienced hypothesis making when confronted with difficult cases. © 2015 John Wiley & Sons, Ltd. Having introduced our general thesis, we now proceed to examine several medical scenarios that exemplify a variety of strategies doctors use to reduce a large set of possible diagnoses – the diagnostic space of observation – to a narrower set or even to a single ‘best’ diagnosis, thus facilitating the choice of a course of action for treatment. Let us emphasize again that in what follows, we reconstruct diagnostic reasoning in a variety of cases according to logical theories (Peircean inquiry, Bayesianism, inference to the best explanation and so on), rather than provide a fully descriptive account of how clinicians reason. We acknowledge that clinicians often do not know the philosophical literature on abduction, inference to the best explanation and so on, and they only sometimes explicitly use Bayes’ rule. But we hold that these reasoning patterns and strategies are implicit in their reasoning. Making them explicit will help to enrich the diagnostic training of future clinicians. Case no. 1: abductive diagnosis, deductive consequences and inductive testing: trochanteric pain syndrome Clinical details An 18-year-old female high school student, who plays field hockey. She complains of pain in the right hip during the early season of training. Pain is overlying greater trochanter and involves tensor fascia. Pressure elicits tenderness along lateral thigh. No drugs, no falls, usual preseason training and observed normal gait with slight hesitation on right forward step. Clinician: ‘Initially this falls into the category of a very common problem: trochanteric pain syndrome/IT band syndrome. There are no ‘red flags’. but the clinician may inquire about: history of osteoporosis, abnormal menarche, history of eating disorder, night pain, hip snapping, and perform a careful physical exam to rule out asymmetry of joint motion or joint pain. Other possible diagnoses include femoral stress fracture, referred lumbogenic pain, and coxa saltans (snapping hip).If the physical examination – testing for full motion, rotation, strength against resistance and elicitation of any pain on passive movement – and history were negative at this point the patient would need instruction about trochanteric pain syndrome/IT band syndrome, and initiate a short course of therapy and NSAIDS with sports activity modification. Athletic trainer might be advised of diagnosis to plan continuing preseason training. If there were any suspicion of intraarticular pain, then an AP pelvis and lateral right hip radiograph to rule out developmental dysplasia of the hip and a femoral neck stress fracture. Further imaging could be obtained based on the results of the plain films (MRI arthrogram and/or bone scan).’ Logical discussion The focus of investigation is on the top differential diagnosis: a syndrome called trochanteric pain. It is the most common cause of lateral hip discomfort in this age group. The logical form of this initial abductive diagnosis is Premise 1: The patient, an 18-year-old female field-hockey player, presents pain in her right hip, overlying greater trochanter. 3 Selecting clinical diagnoses Premise 2: If the patient were suffering from trochanteric pain syndrome, these symptoms would be a matter of course. Conclusion: There is reason to conjecture that she is suffering from trochanteric pain syndrome. Other possible diagnoses, such as femoral stress fracture, are generated in a similar way and could be entertained. However, the clinician deduces what other symptoms could be observed as a result of such a fracture, and observing during testing that none of them occur (an implicit inductive test), quickly discards such diagnoses. Note that trochanteric pain syndrome is selected from the initial list of plausible diagnoses because it explains all the observed symptoms and frequency statistics suggest that it is the most likely. As we will see in further cases, this coincidence of explanatory plausibility and statistical probability is not always present. The clinician, therefore, here perceives a set of associated symptoms and interprets it as being the effect of a known syndrome. This interpretation involves an explanation that relies on his medical knowledge: a general, well-known condition would explain all of the observed facts. This is called a habitual abduction, that is, one in which the explanation of the observed facts is already known to the inquiring doctor. It might be likened to pattern recognition. There is no need for a creative abduction, that is, one in which the plausible explanation is not yet known and must be originated by the creative medical scientist [4]. The clinician also identifies tests that would eliminate or confirm alternative diagnoses if additional symptoms – for example, intraarticular pain – were to surface. This exemplifies, then, the canonical scenario of a habitual Peircean inquiry: an abductive hypothesis-diagnosis, well-known to the doctor, would explain all the observed facts-symptoms; the procedure for observing, or testing, this hypothesis-diagnosis is already standardized; and treatment follows the standardized course, unless new, unexpected facts or symptoms were to emerge. Case no. 2: progressive elimination (testing) of abductive hypotheses by interrogation and imaginative deliberation: vertebral artery dissection Clinical details Patient: Clinician: Patient: 4 ‘The room was spinning. I was sweating profusely. I sat up briefly. I went back to sleep on my back for a few hours and then woke up. The spinning was less severe, but it was still there’ [5]. ‘When this event started, according to the history, you had rolled over in bed. Can you describe to me exactly what happened to you earlier; before the event began?’ ‘I had center court seats for the New York Open and I was there every day. I generally practice serving and volleying 5 days a week; last month it was every day. The day before the dizziness, I played six sets of tennis before my coach had me practice serving for 90 minutes. I needed to increase my serve velocity by reaching higher and higher and arching my neck even further to get above the ball.’ D.E. Stanley and D.G. Campos ‘This could be due to vestibular hydrops (fluid in Clinician (to herself): inner ear), Meniere’s disease, otolithic (inner ear) disturbance, toxins.’ Clinician: ‘I understand that you do a lot of vigorous exercise, including competitive tennis at the “semipro level.” Had you done anything unusual before this episode?’ Logical discussion In many situations, the abductive generation and selection of diagnoses for further testing and eventual treatment occurs sequentially. As the clinician conjectures and then discards possible causes to explain the patient’s symptoms, she seeks further information from the patient and deliberates with herself, until finding a precise hypothesis that explains all the symptoms. This case illustrates that sequential diagnostic process in the form of a dialogue between clinician and patient and the clinician’s own deliberation. At this point, the clinician has begun to generate plausible hypotheses, such as fluid in the inner ear or Meniere’s disease. Any of these would explain the symptoms. But the list of possibilities is too long already; it would be uneconomical and impractical to test them all, and there is not enough information to eliminate any of them. So she proffers further questions to narrow the diagnostic space, eliminating some hypotheses while refining others. Further clinical details Clinician (to herself reviewing the patient’s history): ‘The patient reports lying in bed, noted vertigo, diaphoresis (sweating), and the development of progressive and fixed symptoms, including tilting to the left, nausea, vomiting, headache, oscillopsia (the visual perception of objects moving when they are actually stationary), and changes in facial sensation. The first consideration, the most common, is vestibular origin secondary to a viral inflammation. But there were no reported upper respiratory symptoms; this might eliminate the most common etiology. I must construct a chronology that might account for the patient’s reports. Neurologic symptoms can be localized to the central nervous system or the peripheral nervous system; most of the signs and symptoms in this case are of central origin. But I must also consider rare diseases, e.g. an inherited late manifesting storage disease, or late onset of muscular dystrophy. Further history and testing might eliminate a rare cause; but these entities must be kept in the diagnostic space until a commoner, more plausible explanation arises.’ Logical discussion The clinician, by interviewing the patient, searches for historical clues in order to discard some of the initial possible diagnoses under consideration or perhaps generate a new plausible diagnosis not yet considered. In this case, the absence of upper respiratory symptoms discards – or at least demotes in the order of possible hypotheses – the most frequent diagnosis. She is trying. Thus, she tries to identify ‘clinical pearls’ and ‘red flags’ – that is, facts that © 2015 John Wiley & Sons, Ltd. Selecting clinical diagnoses D.E. Stanley and D.G. Campos may suggest a salient hypothesis worth investigating. She rehearses the salient reports from the patient. At this point, she confronts a set of facts that need to be explained, economically, by a single diagnosis. The clinician asks whether any one explanatory hypothesis can explain all of these symptoms. She is seeking for a hypothesis that captures and amplifies the symptomatology even though the hypothesis may at first consideration seem implausible. If true, however, it would elegantly explain the clinical picture. While she would like to have one unifying diagnosis, she knows that sometimes this is not the case. There may be two or three different diagnoses, one hiding another. The image of a Russian doll may illustrate this situation. Next, the doctor’s reasoning strategy will be further refined by imagining the results of the physiological and anatomic descriptions. She will ponder the basic anatomic structures that may account for a richer hypothesis. Though it is infrequent, the problem may be a dissection of a vertebral artery. Even if it is an a priori unlikely hypothesis, it would be consistent with the patient’s reported activities and it would explain all of the symptoms. In this scenario, then, we see that the generation and selection of plausible hypotheses is often sequential and many are kept in the space of possible differential diagnoses. It is important to rank them. Frequency, or a priori likeliness, is an initial criterion but it is often trumped by the facts of the case. The doctor then must appeal to her imagination and observation of the crucial facts to link them in a way that may lead to an a priori unlikely but simplifying diagnosis. Case no. 3: Peircean inquiry – abductive-deductive-inductive reasoning, with modelling and economic considerations: skin lesions that look alike Further clinical details Clinician: ‘First, I will try to identify the anatomical localization and cause of the present illness from the history provided.’ (summarizing for herself): ‘The patient is a 49-year-old, left-handed avid semi-pro tennis player who, after rolling onto his stomach . . . ’ She thinks constructively. If this were a stroke then the common precursors e.g. hypertension, diabetes, atrial fibrillation should be present. If not, still what if an ischemic event accounts for this congeries? The clinician narrows the diagnostic list by considering less common etiologies; she ranks ischemia over inflammation even though this is a young patient. She ventures it may be a vascular anomaly, a ruptured aneurysm, even an arterial dissection brought on by the vigorous exercise especially the exaggerated hyperextended neck movements during tennis match observations and by practicing serving the ball by ‘getting above the toss to change the trajectory across the net.’ Diagnosis: dissection of vertebral artery. Logical discussion Here, the strategy is to put hypotheses in an ordinal fashion in order to proceed efficiently. The experienced clinician will invoke Ockham’s razor – a maxim to seek parsimony in explanatory hypotheses by eliminating unnecessary elements from them – as often as she can. For instance, she demotes, in the ordinal rank of plausible diagnoses, the hypothesis of a stroke because the precursors are not present in the young athlete. In this case, the discussant is searching the clinical landscape for markers to identify the most efficient strategy to pursue in order to make a diagnosis. She is interrogating her own experience, imagination and clinical acumen to link parts of the history, tests and infer a tentative explanation. She considers a couple of possible hypotheses and ranks them according to prior probability – ischaemia over inflammation – but as she imagines the athlete’s physical activities before the onset of symptoms, she focuses her attention on an important fact, namely, the ‘exaggerated hyperextended neck movements’. This leads to insight. © 2015 John Wiley & Sons, Ltd. Clinical details Clinician: ‘This woman presented with recurring painful erosions and with mucosal involvement during one episode. Her condition responded quickly to prednisone. Laboratory testing was positive for multiple markers of rheumatologic diseases as well as history of NSAID use for pain. Three skin biopsies were performed, and examination of each specimen was suggestive of a different diagnosis – erythema multiforme, toxic epidermal necrolysis, and Stevens-Johnson syndrome’ [6]. Logical discussion The first stage in this inquiry is abductive – at least three possible differential diagnoses are suggested in order to explain the observed signs and symptoms. These are all generated from the doctor’s habitual background knowledge. He knows, for example, that if the patient had erythema multiforme, her observed symptoms would result. But at least two other plausible causes must also be considered. His reasoning could continue by invoking each hypothetical cause, deducing observable and testable consequences if the hypothesis were true, and testing the hypothetical consequences against the clinical history and laboratory findings and therapy. In terms of the Peircean triadic account of inquiry, this is the abduction-deduction-induction sequence. To simplify the scenario, she needs a strategy to narrow down the diagnostic space, one which relies on her clinical acumen. For this purpose, the presence of markers of rheumatologic disease and ingested drugs is foremost in the clinician’s mind. These are important clues to a unifying and simplifying hypothesis. Further clinical details The next step is to have another review of these three biopsies. This would be standard procedure when controversial interpretations are in the medical chart. This appeal to retrospection is very important. As the nosology of painful papulo-nodular skin lesions is extensive, a narrowing criterion – outside of, and supporting, the histo5 Selecting clinical diagnoses logic diagnosis – is required, while at the same time, an assessment of the threat to the patient is driving a swift hypothetical approach. So the common denominator would be, most probably, a result of an autoimmune phenomenon; namely, the presence of rheumatologic markers in the patient’s serum. First, treat with steroids to prevent life-threatening disease such as toxic epidermal necrolysis, Staphylococcus scalded skin syndrome, etc. This is the primary move, as the latter two diagnoses are possibly life threatening and require emergent therapy, even before a firm histopathologic (microscopic skin biopsy) diagnosis can be established. If the emergency therapy does not show marked improvement, then a biopsy and serologic repeat tests for an immunemediate skin ulcer would be indicated (the first tests have been falsely positive). Clinician (to herself): ‘I already have a hunch that this is an immunologic response ( the underlying etiology) to anti-self antibodies to skin cells. I should be prepared to keep the positive immunologic serum markers as the possible “trump card.” I might ask for the biopsies to be stained for: immunoglobulins IgG, to objectively show that immunoglobulins are the cause of this skin disease. Done expeditiously, the diagnosis would be to treat the systemic skin disease (autoimmune, e.g. lupus erythrmatosus), predicting that the lesions will resolve.’ Logical discussion The clinician, relying on her knowledge and experience, identifies a fact (pre-existing immunologic disease) that helps to narrow the diagnostic space. This fact would most frequently be explained by an autoimmune disease. This moves her strongly in the direction of diagnosing and treating immunologic manifestation in the skin of a systemic immunologic disease. She has in fact arrived at an abductive model, informed by medical theory, of the patient’s overall condition. This means that the clinician now has created an overarching schematic or diagrammatic understanding of the relations between the patient’s clinical history (pre-existing condition) and current signs and symptoms, on the one hand, and the conjectured causes on the other.5 Specifically, the reasoning to correctly diagnose skin biopsies and the patient’s presentation of signs and symptoms involves: 1 A histologic theory of bullous skin lesions – where the separation of the skin layers occurs, for example, intra-corneal, intraepidermal, sub-basalar, subepidermal; 5 We will not delve into the nature and function of these schematic diagrams or models created by the doctor’s imagination. Let us briefly say that the ‘imagination’ is the reasoning ability that allows her to transform complex facts and relations into simplifying signs or ‘diagrams’. such as in schemata and figures, for further investigation. A ‘diagram’ is a sign, whether drawn or only imagined, that presents all the relevant facts and embodies their relations [14,15]. The clinical diagram suggests explanatory relations between observed effects-symptoms and possible causes (aetiologies) of diseases. The clinician investigates whether the schematic picture adequately represents the incorporated elements, that is, facts and relations such as symptoms, diseases and causal relations. Are the elements consistent, in medical terms, with one another and with the suspect disease? Does consistency with the imagined schema of symptoms and disease affect the incorporation of pathophysiology and the speaker’s words, especially those that are selected by the hearer/medical person? For further discussion of the semiotics of medical practice, see Silveira [16]. 6 D.E. Stanley and D.G. Campos 2 Observation – the clinical examination, attention to the distribution and character of these lesions may allow a single pathologic diagnosis; 3 Creating a model (the probable natural history) of the disease while using additional clinical data (immunologic markers) to inform and refine the model, while maintaining an urgency to avoid serious complications – for example, treating an infectious disease with steroids, or not giving steroids where indicated, and 4 Basing further definitive testing or even treatment on the theoretical model that fits the presentation. Note, however, that beyond theoretical modelling that would explain abductively the patient’s condition, the clinician must keep other criteria for diagnosis and treatment in mind. The foremost additional consideration is ‘economy’ in diagnosis – in this case, economy of time. Two of the possible diagnoses are lifethreatening, so the doctor must be swift to prescribe treatment while searching for the correct diagnosis. This is in line with Peirce’s procedural logic of abduction for narrowing down a vast space of generated plausible hypotheses. One criterion is to ascertain which hypotheses are recommended for testing by the economy of research, that is, economy of money, time, thought and energy [7]. In this respect, this scenario also provides a curtailed example of what Neurath suggested last century: we can make adjustments to the theory when inconsistencies arise, or we can accommodate the theory by adding more clinical data, but we must proceed to the remedy without exhaustive and extensive testing – this may be an emergency! There are more than 25 dermatologic conditions that may be confused with one another! They are look-alikes and can masquerade. They are identical for the viewer. We do not have the time to start with parts, we must treat the whole, and expediently: the ship is on sea! [8].6 Clinical conclusion: ‘She is therefore required to take into account the previous skin biopsies interpretations and, since the histologic features overlap in multiple skin diseases, to use additional criteria – e.g. constitutional symptoms, systematic findings, history of drug ingestion – in order to specifically identify which of the three histologic findings is the correct one, or better fits the larger clinical picture; and this must be done rapidly. Look-alikes are common and test the experience of the clinician to recognize an emergency. There may be no time to biopsy. She must know more about the disease in order to sift out the most likely disease. She must understand and incorporate which etiologies (causes) would eventuate in these skin manifestations; she would need to exclude infectious, malignant, parasitic, and allergic (except to her own skin constituents!) ones. But she is using her clinical experience to decide this is indeed an emergency.’ 6 Neurath’s point is primarily about the language of science, but we are adapting it to our needs in accounting for the logic of diagnosis. In his words: ‘There is no way to establish fully secured, neat protocol statements as starting points of the sciences. There is no tabula rasa. We are like sailors who have to rebuild their ship on the open sea, without ever being able to dismantle it in dry-dock and reconstructed from its best components. Only metaphysics can disappear without a trace. Imprecise ‘verbal clusters’ [Ballungen] are somehow always part of the ship. If imprecision is diminished at one place, it may well reappear at another place to a stronger degree’ [8]. © 2015 John Wiley & Sons, Ltd. Selecting clinical diagnoses D.E. Stanley and D.G. Campos Case no. 4: red flags, experience and informed Bayesian probabilities: viral lymphadenopathy Clinical details A 36-year-old-man, previously in good health, experiences fatigue, mild weight loss and palpable lymph nodes in the groin, the axilla and the anterior and posterior neck. He had returned from a business trip to Asia 3 weeks earlier. The lymph nodes are firm, but movable. There is no node larger than about 1.0 cm. A chest X-ray shows multiple enlarged lymph nodes of similar size in the mediastinum and retroperitoneal area. He has low-grade fever of 38°C. He does not use illicit drugs, does not smoke and is a graduate student in philosophy. Before he travelled on business, he received vaccinations against several tropical diseases as well as a skin test for tuberculosis [9]. Clinician building a diagnostic strategy: ‘Whenever we listen to the history of an illness and the case presentation, whether they are being described by another physician or the patient, certain words and phrases seem to carry more weight than others, and experts are trained to listen for these words pregnant with potential truth: words of uberty. Why is this choice the case for some clinicians and not for others? In this case, the initial impression was dominated by the patient’s young age and previous good health, the brief duration of his present illness, his recent travel to Indonesia, the low grade fever and enlarged lymph nodes.’ The task is to construct a differential diagnosis around the development of lymphadenopathy and then narrow the list on the basis of what we know about this patient. Logical observation Initially, a doctor could abduce several possible causes of lymph node enlargement, such as early HIV infection, reaction to a medication, exposure to an allergen, early lymphoma, systemic lupus erythematosus, infectious mononucleosis, Castleman’s disease, Kimura disease or Rosai–Dorfman disease. There are numerous possible abductive diagnoses that flash into a clinician’s mind, so it is necessary to narrow the list down by looking for critical facts or ‘red flags’, that is, facts that may eliminate some possible hypotheses while making others salient for further examination. Further clinical details In this case, the patient is a young man, with few risk factors, who has a monotonous pattern of small, firm, but movable lymph nodes. This would point to a systemic rather than to a local disease. His mild symptoms also rule against metastatic cancer, lymphoma, and is suggestive of a reactive process. Therefore, the clinician chooses a research pathway that would put reactive lymph node enlargement high on the list of plausible diagnoses worthy of further testing. Regarding other possible initial abductive hypotheses, reactive lymph nodes of this size and distribution would not suggest some lymph node diseases, for example, Kimura, Rosai– Dorfman or Castleman’s. The symptoms are understood as being compatible with some possible hypotheses but not with others. © 2015 John Wiley & Sons, Ltd. Logical discussion The clinician identified some facts that quickly eliminated a series of possible hypotheses from the space of differential diagnoses deserving further scrutiny. This is an important step, but the inquiry requires further strategic moves. There may be associated probabilistic considerations to assess whether a diagnosis ought to be selected to inform a future course of action. Recall the formulation of Bayes’ rule to estimate conditional probabilities: P ( H E ) = [ P ( E H ) × P ( H )] P ( E ) If doctors reason along Bayesian lines as they look for red flags, they have learned from clinical experience to posit the prior probabilities of diagnoses-hypotheses, P(H), that clinical training has conditionalized by experience. The beginner with little experience in the practice of medicine posits priors from her training. As she gains experience, the priors, having been tested, are refined and become ‘informed priors’ that more likely fit the diagnostic space and are then utilized as priors. These priors are empirically based rather than merely subjective in part because past frequencies are reliable for common diseases. This is the contribution of observed frequency data; the education of the clinician is modelled on her exposure to frequencies of occurrence. For illustrative purposes, we may say that an inexperienced clinician posits a P(H) informed by her training, but an experienced clinician posits a P(H)’ informed not only by training but by her developed acumen. In other words, from exposure to history, signs and symptoms similar to the case encountered, experienced clinicians are able to recall how strongly, if at all, their prior suspicion H, plus the evidence E, lead to a high P(H/E). And these conditionalized priors, P(H)’ = P(H/E), are used when iteration of the signs and symptoms, E, and the prevalence of the conjectured disease that causes them, P(H), are considered in new cases. In clinical practice, P(H) is estimated by the prevalence of a disease. Prevalence is the rate for a disease that is equal to the number of patients per 100 000 persons per year in the population who have the disease at the time of the study; in other words, it measures the rate of occurrence of the disease in the general population.7 The experienced doctor is able to estimate this initial assessment of P(H) in the specific clinical situation, geography and to consider the environmental exposure under investigation. We claim this ability is what experienced clinicians possess from exposure to the congeries of signs and symptoms, or evidence, inserted into the probability estimate. Travel to Indonesia may change the likelihood of the considered priors because the evidence is strengthened and complemented by the foreign exposure; thus, the prior is shifted from that without this exposure, P(H), to that with exposure, P(H)’. In terms of this specific scenario, the initial red flag is the homogeneous size of the lymph nodes. But the clinician has further considerations. She pauses for reflection and allows for re-estimation of the prevalence and re-evaluation of exposure to foreign travel as part of the evidence. The pause from closure is an index of the reflec7 Prevalence should be distinguished from incidence; the difference is time dependency. Incidence is the number of patients with the disease per 100 000 people per year. So the prevalence of a disease like the flu may be high during the ‘flu season’, but the incidence may be low during the ensuing year. 7 Selecting clinical diagnoses tive mode of generating and selecting hypotheses. The clinician asks herself, reflexively, what could I be missing? What other patterns would fit these congeries? Exposure during foreign travel becomes relevant. By attending to this fact, she is able to adjust the probabilities she assigns to possible diagnoses, giving higher priority for further investigation to one or a few hypotheses, and discarding others as improbable given the two ‘red flags’ – size of nodes and foreign travel. This is part of learning from experience. In this case, the clinical discussant structures the discussion on signs and history – lymphadenopathy, exposure to tropical diseases – and from recall is able to simplify the discussion because her experience has shown her that lymph nodes smaller than 1.0 cm are probably reactive rather than neoplastic – that is, superficial and deep seated lymph nodes that are similar in size and even consistency suggest a reactive pattern rather than a malignant disease. She finally suggests a diagnosis – either Kikuchi-Fujimoto disease or miliary tuberculosis – based on two red flags: size of lymph nodes and travel to Indonesia. Here, the foreign travel narrows the differential diagnosis as it shifts the prevalence, at least on the first examination; yet biopsy may be indicated if the doctor is not satisfied to proceed to treatment. She must also tolerate confounders and keep her vision open to other inputs, since some cases of infectious disease – for example, Epstein–Barr virus – may cause lymphadenopathy larger than 1 cm. Case no. 5: statistics (frequentist and Bayesian) and economy: man with mid-back pain Clinical details A 57-year-old man with 3-month history of mid-back pain, nonradiating. Location: T12-L3, tender to palpation. He does not perform any daily physical exercises. The condition occurred once in his early 40s. He describes it as aching and worse standing. NSAIDs are of little help. He is a non-smoker and occasional alcohol drinker. He works as a mid-level manager in an insurance company. Routine visits to primary care doctor: normal CBC, plain radiograph of lumbosacral area shows no bony pathology. This falls into the category of back pain that has not resolved in the usual 4–6 weeks as would be expected if this were a case of non-specific lower back pain. In the scenario above, there are no ‘red flags’, but the clinician may inquire about area of spine represented in radiograph – especially since many lumbar spine radiographs do not contain thoracic vertebrae (metastatic pancreas cancer does involve lower thoracic vertebrae). He may also inquire about spine trauma, history of osteoporosis, history of cancer, fever, chills, weight loss, illicit drug use, immunosuppression, night pain and do a careful neurological exam. If all of these were normal, only then would he consider screening with bone scan (to rule out metastases and insufficiency fracture), laboratory tests (CRP and ESR to rule out occult infection, rheumatologic disorder and serum protein electrophoresis to rule out multiple myeloma). In the background of his clinical experience, lower back pain of unknown aetiology is the most common diagnosis; yet, he maintains an open question, as the history is somewhat longer than he would expect. The patient is slightly older but his inactive physical state is still consistent. His differential diagnosis space remains open. 8 D.E. Stanley and D.G. Campos If the first examination were normal, additional tests could be obtained serially to minimize cost. In this case, the clinician wishes to perform the test that rapidly yields the most information about a condition that is very dangerous, namely, he wants to rule out cancer or myeloma. At this age, both conditions must be ruled out. Logical discussion The clinician is looking first for ‘red flag’ signs or symptoms, which guide the investigation. He finds none beyond the 3-monthlong history, and the patient’s age. He then estimates the prevalence of this condition given the setting of his practice. As defined before, prevalence is the number of patients per 100 000 persons per year in the population who have the disease. The clinician recalls that regardless of the sensitivity and specificity of further testing, prevalence (the tacit hypothesis) is heavily weighted in any Bayesian calculation. For a test Predictive value = (Prevalence )(Sensitivity ) [(Prevalence ) (Sensitivity ) + (1 − Prevalence ) (1 − Specificity )], 8 where the specificity is the true negative rate or the proportion of actual negative cases that are ruled out, and the sensitivity is the true positive rate or the proportion of actual positive cases that are ruled in . These measures are highly dependent on the prevalence of the disease in the population undergoing testing. Alternatively one can calculate the likelihood ratio (LR), which does not depend on prevalence. This is frequently advocated: For a positive test result, LR + = Sensitivity (1-Specificity) . For negative test result, LR − = (1 – Sensitivity ) Specificity . 9 Unfortunately, these values are not readily generalizable.10 Observing no red flags, the doctor proceeds to try to rule in or rule out the pathologies he considers likely, and this is based on the prevalence of this condition in his clinical practice. He stratifies the investigation based on the urgency of the implications of the possible aetiologies. He wishes to obtain data serially in order to minimize cost. In this sense, he follows the criterion of economy identified by Peirce – that is, according to the logic of abduction that scientists apply in practice, one of the criteria for selecting among tentative hypotheses is the cost of testing them (Peirce, CP 5:600; n.d.).11 This dictum is frequently avoided most likely because of medico-legal implications. However, the doctor also stratifies testing by seriousness, for example, myeloma and metastatic cancers are higher in the possible diagnostic list because of history and age and are dangerous. 8 Note how prevalence affects the calculation of predictive value. The prevalence of non-specific lower back pain is high in his office setting. 9 Alternatively, from a 2 × 2 chart of true positives (TP), false positives (FP), true negatives (TN) and false negatives (FN): the predictive value of a positive test is 1[TP/(TP+FP)] × 100, while the predictive value of a negative test is [(TN/TN+FN)] × 100. 10 For a thorough presentation of prevalence and predictive value, see Galen and Bambino [17], pp. 167–264. 11 Following standard practice in Peirce scholarship, references to Peirce’s Collected Papers are abbreviated CP followed by volume and paragraph number and date of drafting or publication. © 2015 John Wiley & Sons, Ltd. Selecting clinical diagnoses D.E. Stanley and D.G. Campos He pauses, perhaps, because the likelihood of these hypotheses is not the highest on an ordinal scale, but these possibilities are the most serious considerations. Case no. 6: inference to the loveliest explanation: Nigerian woman with melena secondary to infectious gastric ulcer Clinical details Discussant: ‘When I first evaluated this 55-year-old Nigerian woman, I thought that identification of the source of melena would quickly lead us to the diagnosis. Melena is typically caused by hemorrhage in the upper gastrointestinal tract, proximal to the ligament of Treitz; however, blood loss in the distal small bowel or proximal colon may also result in melena’ [10]. Logical discussion Here we find the first movement to form a hypothesis, except in this case, the discussant is prompted to entertain a different hypothesis. She has been made aware of the correct diagnosis. The logical situation could be described as follows. The clinician initially generates two possible abductive explanations for melena: haemorrhage in the upper gastrointestinal track or blood loss in lower tract. She also immediately ranks these two possible diagnoses. The typical site is upper gastrointestinal, and so she assigns it a higher prior probability of being the correct diagnosis, but she also keeps in view the possibility of blood loss in the colon, though with a lower prior probability. The clinician, then, has chosen prevalence as the promising path to a hypothesis, but she must be careful about relying exclusively or hastily on prior probabilities. Recall the adage of hoof beats in the office: ‘horse, not zebra’. It is helpful but most be pondered in context. In this case, zebras are still in the running. Closure on diagnostic considerations should not occur too early or too late – this requires a balancing act when efficiency does not jeopardize the well-being of the patients. This adage, very well known in medicine, may be illustrated as follows. In Costa Rica, there is only one form of malaria: Plasmodium vivax, the ‘horse’ so to speak. But a doctor there who was concerned about cerebral malaria of the P. falciparum variety – the ‘zebra’ – would have to be very certain that the patient had not foreign travel outside Costa Rica and no blood transfusions to exculpate the notion of a diagnosis of cerebral malaria. But in Nigeria the presence of P. falciparum exists. Experience in one setting does not allow for easy extrapolation to other settings; this is the limitation of prevalence estimation and thus likelihood. Further clinical details In evaluating this patient, the clinician first considered esophageal causes of melena, including drug-induced esophagitis, reflux esophagitis, a Mallory–Weiss tear and cancer. She also considered causes originating in the stomach and small intestine, including peptic ulcer disease, gastritis. Further details of her history were helpful: this woman had autoimmune hepatitis and cirrhosis treated with steroids, and dia- © 2015 John Wiley & Sons, Ltd. betes mellitus, type II. Medications included prednisone, insulin, furosemide and omeprazole. Thus, knowledge of her national origins, her medical history and condition required further investigation. Logical discussion The discussant is trying to estimate causal prevalence based on her clinical training and experience – the likelihood that origin of melena would lead to a conclusive diagnosis. However, this approach does not take into account the previous history (liver disease, diabetes, national origin, steroid treatment) and does not further identify the site of bleeding. As a result, the clinician has to reassess and perhaps re-interview the patient. Thus, she invokes other considerations and offers a resultant, alternative hypothesis. Final clinical details With prolonged immunosuppression (prednisone) for the liver disease and diabetes, she was exposed to opportunistic infections. The history of melena and probable upper gastrointestinal source of bleeding suggested the need for an upper endoscopy that showed a gastric ulcer – the diagnosis is based on a biopsy showing inflamed gastric wall with fungal hyphae. Reviewing the patient’s history, the clinician recalls she was from Nigeria where parasitic infestations such as Strongyloides, malaria and infectious aetiologies are more common, especially in this immunocompromised woman. Logical conclusion This exercise is a demonstration of the clinician’s ability to investigate one sign of illness (melena) while maintaining the possibility of a zebra. Thinking about the presentation to include her history and country of origin allows her to avoid focusing too early on the initial complaint. The previous medical history of autoimmune liver disease under immunosuppressive treatment brings up a differential diagnosis in which zebras are still on the run. Taking into account the present illness past history and medications, she pursues a different hypothesis because, she suspects, it will lead to a different diagnostic pathway and eliminate excessive, perhaps redundant, investigations. In terms of the logic of hypothesis selection, in this case, considerations of a priori ‘likeliness’ do not suffice to select the correct diagnosis, given the patient’s origin and past medical history. She sifts the medical history and current illness to search more specifically for a hypothesis that provides a ‘lovely’ explanation that encompasses both past and present illness – that is, an explanation that will take into account all of the patient’s illnesses (Lipton [11], pp. 53–70). Lovely diagnoses are parsimonious – Ockham’s diagnoses – while unifying all observed symptoms and conditions into a powerful, elegant explanation. This view of diagnostic loveliness links simplicity to explanatory power and beauty. In fact, sometimes a clinician will exclaim: ‘What a beautiful diagnosis! It puts together all of the unexplained and unexpected details into a composite picture that portrays exactly what we can now understand and treat’. The sort of loveliness we are suggesting is not a pseudoaesthetic criterion, but rather an epistemological concept. In assessing the logical diagnostic situation, we are following Peter Lipton who distinguishes between the inference to the likeliest and inference to the loveliest explanations. Given the empiri9 Selecting clinical diagnoses cal evidence, the likeliest explanation is ‘the explanation that is most warranted’ by inductive testing, while the loveliest explanation is ‘the one which would, if correct, be the most explanatory or provide the most understanding’ [12]. He writes, ‘likeliness speaks of truth; loveliness of potential understanding’ [12]. In this specific medical scenario, the selected diagnosis was not the most a priori likely in terms of prevalence; however, it yields the best potential understanding of the patient’s condition. It explains all the symptoms and facts of the patient’s medical history in a simple, unified, and causally connected way. It explains clearly which causes led to which effects and why, and it does not leave relevant facts of the patient’s condition or medical history out of account. In short, the diagnosis offers a lovely explanation. Lipton also argues that the loveliness of a hypothesis may function, in the logic of hypothesis selection, as a surrogate for its likelihood [11]. In the medical scenario under analysis here, the loveliness of the more carefully considered diagnosis – melena secondary to an infectious gastric ulcer – indicates that its informed prior probability trumps the prior probability of the first hypothesis – gastrointestinal bleeding – which relied on prevalence (probabilities) alone. This scenario illustrates, then, that diagnostic closure cannot come too early in the process – several simultaneous, potentially conflicting, hypotheses are often under consideration and the experienced clinician requires acumen to choose a diagnosticselection strategy from a wide variety of possible approaches. Summary The foregoing scenarios are illustrative simplifications of the diagnostic process that is messy and does not follow strict rules. Yet, they exemplify a variety of logical strategies for selecting differential diagnoses for further testing and treatment from an abductively generated space of plausible diagnoses. In actual practice, these selective strategies may be used sometimes singly, more often in groups or even seriatim, to choose a diagnosis. The major anchor in establishing a diagnosis is insightful, knowledgeable perception of details that, fitting into a plausible picture, allows for tentative resolution and action. By this, we mean that both experienced perspicacity and scientific knowledge are important to effective diagnosis. These capacities allow the doctor to generate plausible diagnoses and select the most worthy ones for testing and treatment. Sometimes even a minor detail – an inadvertent gesture; a change in voice or speech; the mood of the patient; a fact such as foreign travel mentioned in passing while reviewing the history, or family history, which in the future should receive even more attention as genetics will probably play a larger role in ‘precision medicine’ – can lead the observer in a different direction for diagnosing. A plausible diagnosis that at first may seem more prevalent or likely may become a less viable option to the perspicacious, experienced clinician. Thus, in order to turn insightful, knowledgeable perception into effective diagnosis, the doctor should have recourse to a diverse kit of logical strategies to choose from the space of plausible diagnoses. As the preceding medical cases show, these selective strategies include Peircean triadic inquiry – abduction, deduction and induction – frequentist statistical inference, Bayesian inference relying on prior probabilities informed by experience, inference to the loveliest explanation and so on. Under every strategy, imagi- 10 D.E. Stanley and D.G. Campos native deliberation sharpened by clinical experience is explicitly or implicitly at work. We summarize our discussion, then, by suggesting this dictum: ‘Our best guesses are winnowed by the wind of experience’. Acknowledgement We would like to acknowledge the assistance of Librarians at Maine Medical Center, especial thanks to Amy Moore for editing footnotes. References 1. Stanley, D. E. & Campos, D. G. (2013) The logic of medical diagnosis. Perspectives in Biology and Medicine, 56 (2), 300–315. 2. Peirce, C. S. (1998) The Essential Peirce Selected Philosophical Writings, Vol. 2. Bloomington, IN; London: Indiana University Press. 3. Colapietro, V. (1999) Peirce’s guess at the riddle of rationality: deliberative imagination as the personal locus of human practice. In Classical American Pragmatism: Its Contemporary Vitality (eds S. B. Rosenthal, C. R. Hausman & D. R. Anderson), pp. 15–30. Urbana: University of Illinois Press. 4. Campos, D. G. (2011) On the distinction between Peirce’s abduction and Lipton’s inference to the best explanation. Synthese, 180 (3), 419–442. 5. Engelter, S. T., Grond-Ginsbach, C., Metso, T. M., et al. (2013) Cervical artery dissection: trauma and other potential mechanical trigger events. Neurology, 80 (21), 1950–1957. 6. Roh, E. K., Vleugels, R. A. & Mai, P. H. (2013) Case 39-2013: a 57-year-old woman with painful bullous skin lesions. The New England Journal of Medicine, 369, 2438–2449. 7. Peirce, C. S., Hartshorne, C., Weiss, P. & Burks, A. W. (1960) Collected Papers of Charles Sanders Peirce. Cambridge: Belknap Press of Harvard University Press. 8. Neurath, O., Cohen, R. S. & Neurath, M. (1983) Philosophical Papers 1913–1946 with a Bibliography of Neurath in English. Dordrecht: Springer Netherlands. 9. Ray, A., Muse, V. V. & Boyer, D. F. (2013) Case records of the Massachusetts General Hospital. Case 38-2013. A 30-year-old man with fever and lymphadenopathy. The New England Journal of Medicine, 369 (24), 2333–2343. 10. Corey, K. E., Gupta, N. K., Agarwal, S. & Xiao, H. D. (2013) Case records of the Massachusetts General Hospital. Case 32-2013. A 55-year-old woman with autoimmune hepatitis, cirrhosis, anorexia, and abdominal pain. The New England Journal of Medicine, 369 (16), 1545–1553. 11. Lipton, P. (2004) Inference to the Best Explanation. London; New York: Routledge/Taylor and Francis Group. 12. Lipton, P. (1991) Inference to the Best Explanation. London; New York: Routledge. 13. Upshur, R. (1997) Certainty, probability and abduction: why we should look to C.S. Peirce rather than Gödel for a theory of clinical reasoning. Journal of Evaluation in Clinical Practice, 3 (3), 201–206. 14. Campos, D. G. (2009) Imagination, concentration, and generalization: Peirce on the reasoning abilities of the mathematician. Transactions of the Charles S. Peirce Society, 45 (2), 135–156. 15. Peirce, C. S. (1976) The New Elements of Mathematics. Hague: Mouton Publishers. 16. Silveira, L. (2005) Análise semiótica da diagnose médica. Cognitio: Revista de Filosofia, 6 (1), 94–101. 17. Galen, R. S. & Gambino, S. R. (1975) Beyond Normality: The Predictive Value and Efficiency of Medical Diagnoses. New York: Wiley. © 2015 John Wiley & Sons, Ltd.