Philosophers of medicine have concentrated much of their attention on treatment. By contrast, diagnosis is understudied by philosophers. The distinctive issues posed in diagnostics are sidelined as analysis is lavished on the evaluation of treatments and the causal mechanisms of treatment and disease. But diagnostics is a vital, fundamental element of clinical practice deserving of significant philosophical attention.
The issues which diagnosis poses for philosophers of science are a mixture of the novel and the familiar. Diagnostics deeply embeds venerable questions of health and disease, of evidence and causal inference, and of the art, science and politics of medicine. Philosophers cannot analyse diagnostics without engaging with these deep philosophical questions. But nor can a full account of the philosophy of medicine be developed without treating diagnosis independently of these abstract debates, because it raises issues which are distinct and throws new challenging nuances into the pathway of old questions.
The existing philosophical literature on diagnosis is underdeveloped and marginal. Perhaps the most influential treatments of diagnosis have come from Kazem Sadegh-Zadeh, whose analytic philosophy of medicine draws strongly on fuzzy logic to underpin his approach, particularly in his Fundamentals of Clinical Methodology (2000) and The Logic of Diagnosis (2011). Sadegh-Zadeh’s analysis in terms of fuzzy logic has been so influential that the Stanford Encyclopedia of Philosophy reports: “In recent years there is a relative consensus among medical professionals and those involved in medical informatics that medical diagnosis almost certainly relies on some form of “fuzzy logic”” (Reiss & Ankeny, 2016).
However, there are further philosophical issues to explore in diagnostics, and Sadegh-Zadeh’s account faces some serious challenges and indeed at least one fatal flaw. In some places, there are ways to work around these limitations and rehabilitate elements of his account. In others, new approaches will be required. This series of articles will unpack six interconnected philosophical problems of diagnosis. By making this fine-grained distinction between different problems, the hope is that we will be better equipped to articulate specific solutions, repurpose elements of the work of Sadegh-Zadeh and others where appropriate, and expose gaps in existing accounts to facilitate new work.
This first paper introduces the philosophy of diagnosis by attempting three tasks: to outline the roles of diagnosis within medical practice; to distinguish six different salient problems of diagnostics which philosophers can engage with; and, to map the interrelationship between these six problems.
The Functions of Diagnosis
To motivate and focus a philosophical analysis of diagnostics, it is useful to understand the functions or roles which diagnosis plays within medicine. Why diagnose patients? There are six broad purposes:
- Prognostic (or ‘Predictive’)
- Therapeutic (or ‘Interventional’)
The first three functions are perhaps the best appreciated roles of diagnosis, while the epidemiological and academic roles are primarily adopted in public health and clinical research, and the palliative role is a less accepted, marginal element of diagnostics which some may view as a side-effect rather than an intentional role.
In its explanatory function, diagnosis provides for patients and practitioners an explanation of the patients’ symptoms. This explanation is usually an appeal to a series of causal relationships and mechanisms between the symptoms and some malady or maladies, often accompanied by a series of statistical relationships which state risk-factors. The causal mechanisms generally do most of the work of explaining how the symptoms are related and how some underlying cause(s) brought them about. Did one symptom lead to another? Are they caused in common by an underlying disease or injury? Meanwhile, the explanatory role of statistical relationships is usually primarily as part of an explanation of why this patient or patients developed these symptoms or this condition. Why are they afflicted when others are not? Why did the condition manifest this way? The statistical relationships are often risk-factors which made the patient more susceptible to the conditions involved, or to the particular manifestation in terms of symptoms.
For instance, a patient with a persistent worsening cough and breathlessness might be diagnosed by a series of causal mechanisms which explain that the patient has an underlying condition – lung cancer – which causes both the cough and the breathlessness due to its physiological effects on the lungs. But the diagnostic explanation might equally note that the patient is a heavy smoker and that smoking is a major risk-factor for developing lung cancer. Thus, the explanatory diagnosis invokes both causal mechanisms and statistical dependencies, attempting an explanation both of the patient’s symptoms in terms of causal mechanisms linking the symptoms to an underlying disease, and of the patient’s contracting the disease in terms of statistical risks.
The explanatory role is about providing understanding of what is happening and has happened to the patient, often creating a form of causal ordering amongst symptoms and conditions. Having an explanation of their symptoms may provide a degree of psychological comfort for some patients. Understanding the causal relationships and mechanisms of the disease may be vital to subsequent steps, particularly prognosis and therapy. However, the explanatory function is meaningfully distinct from these further uses of the explanation. In the most extreme case, explanatory diagnosis can characterise the work of forensic autopsy, in which explaining the death of the patient is the end function of the diagnostic process, divorced entirely from prognostic or therapeutic concerns.
The prognostic function of diagnosis is to support prediction of the future trajectory of the patients’ symptoms. The knowledge that a patient has a worsening hacking cough and breathlessness severely underdetermines the future course of the patient’s symptoms. We need to know more about the causes of the symptoms and their relationship to each other to stand a chance of offering a reliable prognosis. Prognosis usually consists, again, of a mixture of causal mechanistic and statistical reasoning. We create a causal model which explains the patient’s symptoms and use an understanding of this and the progression in this model to forecast the patient’s trajectory. This causal model also gives us information about the reference class to which the patient belongs for the purpose of statistical modelling of the likelihood of future outcomes. Whether the patient’s cough and breathlessness is the result of lung cancer or emphysema will put her into a very different reference class sanctioning different prognoses.
But we also use data about risk-factors for different potential trajectories to identify the probabilistically most likely outcomes, and use risk-factors to further narrow down the reference class as closely as possible in order to deliver a more tailored (and thus hopefully more accurate) prediction. Again, prognosis is closely related to subsequent therapeutic decisions but might be the endpoint of the diagnostic process, particularly where treatment is not possible. Counterfactual prognosis is also often involved in autopsy in the work of coroners and forensic pathologists, asking how long a patient would or could have lived had some event not occurred.
The therapeutic function focuses on supporting specification of the best possible treatment for the patient. Again, a patient’s symptom set alone often underdetermines the best course of action for treatment. While treating symptoms is often possible without diagnosis, this is risky – sometimes the cause of a symptom would contraindicate an otherwise commonplace treatment. Moreover, knowledge of the underlying causal structures at work can facilitate a more lasting, effective and appropriate response. There are interactions between the therapeutic function and both the explanatory and prognostic functions. An accurate causal model allows for better identification of mechanisms which can be interrupted to affect the processes causing the symptoms or underlying malady. Accurate prognosis gives the best estimate of the necessity of treatment, and of how much risk is justifiable in the treatment. So, even where the diagnosis does not affect which treatment is most appropriate, it could affect whether the treatment is administered or not.
The epidemiological role of diagnosis focuses on understanding trends in symptoms and disease at the population level. By determining the underlying causes of patients’ symptoms, diagnosis has a crucial role in monitoring both the spread of a disease, and in charting and analysing the statistical trends of the condition in terms of which patients are more or less likely to be afflicted. This role of diagnosis in epidemiology was dramatically highlighted in the recent Covid-19 pandemic, in which the role of diagnostic testing to assist public health authorities in tracking and tracing the spread of the disease was paramount even independently of attempts to treat patients. This offers a further nuance to the interventional element of diagnostics: here, population-level epidemiological interventions can be mandated such as quarantining affected individuals to preserve the health of the population by slowing or preventing spread.
The closely related academic role of diagnosis describes the function of diagnosis as part of the process of researching a condition. Creating explanatory, prognostic, therapeutic and epidemiologic models of a condition is a central role of academic medical research. Again, this diagnostic function can operate independently of or in concert with the clinical roles described above, and in some cases takes place post mortem.
Finally, the palliative role of diagnosis is closely linked to the explanatory role but merits independent consideration. Receiving a diagnosis may ease a patient’s suffering, both physically and psychologically, to some extent. However, this role can be considered distinct from the explanatory role, as a diagnosis may still offer some palliative relief even when the diagnosis offers no explanation of the symptoms. For instance, it has been claimed that some syndromes are purely extensionally defined (see e.g. Aronowitz, 2001) – that is, such a syndrome is nothing more than a distinctive collection of symptoms, without a clear understanding of any underlying cause or interrelationship between the symptoms. Offering such a syndrome as a diagnosis does not explain the symptoms, but it does put a name to the patient’s suffering and perhaps offer some reassurance that the patient is not alone in their experience. In this sense, palliative effects of diagnosis may be detachable from the explanatory function.
The legitimacy of such diagnoses in the absence of explanatory content is a controversial issue which will be explored later in this series. It is also worth noting that diagnosis may be counterproductive from the palliative perspective: a patient’s reaction to their diagnosis may not be one of relief, but may heighten anxiety, create psychological trauma and even provoke a worsening of symptoms. Throughout our consideration of diagnosis, it should not be viewed as a purely beneficial or neutral endeavour, and the risks of diagnosis should be fully appreciated.
As these six roles make clear, elements of diagnosis are patient-oriented, practitioner-oriented, population-oriented and knowledge-oriented. Information gained through the diagnostic process can help the patient (directly and indirectly), the practitioner in making decisions and recommendations, public health officials and researchers. This information is explicitly a juncture of statistical and causal mechanistic modelling, and these dual elements will be foregrounded throughout the analysis of diagnostics. Any account of diagnosis which attempts to reduce the process to either purely statistical regularities or purely qualitative causal mechanistic reasoning will fail to capture some critical elements of the endeavour.
Six Problems in the Philosophy of Diagnosis
In The Logic of Diagnosis, Sadegh-Zadeh lays out two problems: specifying the syntactic form of diagnosis statements, and the semantics of diagnoses. He makes an admirable although flawed foray into both, which will be discussed in Part 3 of this series. Unlike Sadegh-Zadeh, the first of the philosophical problems identified here cuts somewhere between his two problems (this move will be defended in Part 5). The remaining problems are commonplace in medical practice but have been seldom defined and often conflated elsewhere. The five problems are:
- The Problem of Diagnosis Structure: What logical form(s) do diagnosis statements take?
- The Problem of Diagnostic Candidacy: Which statement(s) are worth considering as potential diagnoses for a patient P with symptom set S?
- The Problem of Diagnostic Legitimacy: When does offering a diagnosis D for patient P with symptom set S constitute a legitimate attempt to diagnose P?
- The Problem of Diagnostic Quality: How should we assess how good or bad a diagnosis D is of patient P’s symptom set S?
- The Problem of Diagnostic Accuracy: When do we know that a diagnosis of patient P’s symptom set S is correct?
- The Problem of Diagnostic Effect: How can we predict the effects of offering a diagnosis D to patient P, and consequently is it ever ethically defensible to withhold a diagnosis (or some element of a diagnosis) for the good of the patient?
The problem of diagnosis structure is perhaps the least contentious and provocative of the problems, but it underpins the rest of the project. Diagnosis structure is a foundational problem for the logic of diagnosis, which helps us to understand exactly the domain across which the other problems range. Despite seeming comparatively trivial, answering this question fully will embroil us in questions about the nature of disease, abnormality, symptoms and risk. Problems 2-5 are in fact classification problems. They propose four different properties which diagnosis statements could have, lack, or have to some degree: candidacy, legitimacy, quality and accuracy. Insofar as those properties are binary, they ask us to partition the domain of diagnosis statements into the candidates and non-candidates, the legitimate and illegitimate, the good and the bad diagnoses, and the accurate and inaccurate ones. Alternatively, we might think of these as fuzzy properties, following Sadegh-Zadeh’s lead, and see these as questions of defining the degree to which a statement is a candidate, legitimate, good or accurate for a given patient. To begin addressing those problems, we must scope the domain well. This means the problem of diagnosis structure is critical to the rest of the project. The problem is not only of interest to philosophers; it is vital both for clinical researchers and for computer scientists working on the challenges of equipping computer-aided diagnostics systems and machine learning algorithms to attempt to support clinicians in their diagnostic work.
The problem of diagnostic candidacy is similarly crucial to computer science in computer-aided diagnostics. Answering this question helps to break us out of a form of the ‘frame problem’. In computer science, the frame problem is the challenge that an automated system tasked with choosing the best course of action (for instance) must evaluate the consequences of each possible action. But to do this, it must first identify what those consequences might be. Imagine a robot asked to choose whether to move forward or backward by evaluating the consequences of each action. To do this, it needs to determine what effects its movements might have. But there are an infinite variety of possible effects. Will moving forward affect the colour of my t-shirt? What about the shirts in the closet? Their size or shape? Will it change who wrote Wuthering Heights? Will it alter the definition of Planck length? In order to even begin its evaluation, the robot first needs to be able to determine which consequences are relevant without checking whether each possible consequence is relevant. This is a challenging problem. In diagnostics, the analogous problem is to identify which statements which have the logical structure of a diagnosis of the patient are actually worth considering and evaluating. Indeed, as our analysis of diagnosis structure will show, there are infinitely many possible diagnosis statements for any individual. Neither practitioner nor machine can evaluate them all, so for diagnostics to function we must have an approach to diagnostic candidacy, ideally one which does not require it to assess the candidacy of each diagnosis statement individually! Solving that frame problem will be beyond the scope of an initial account of diagnostic candidacy, however.
But not all diagnosis statements which might be candidates for a patient P would count as legitimate attempts at diagnosing the patient. This problem is particularly salient in ethics and medical law. Do medical practitioners have an ethical duty to provide only legitimate diagnoses to their patients? Could legal action be pursued against a clinician who provides an ‘illegitimate’ diagnosis to a patient or fails to offer a legitimate attempt at diagnosing the patient whatsoever? If a doctor knowingly diagnoses a patient with a condition she cannot possibly have, then is the doctor is behaving irresponsibly or even illegally? What of a doctor who diagnoses a condition which is highly unlikely given the patient’s symptoms? Such questions demand clarity about what makes a particular diagnosis legitimate. Intuitively, it seems that candidacy and legitimacy are not always equivalent. Diagnostics often involves considering a potential diagnosis and then excluding it. This suggests that some candidate diagnoses may not be legitimate diagnoses. We will explore this facet of diagnostic logic later in the series.
Assessing diagnostic quality may be the most accessible problem for practitioners. This problem poses the task of formulating an approach or approaches to evaluating how good a diagnosis is for a given patient’s symptom-set. It may be the case that quality is closely linked to legitimacy, and it will be necessary to explore whether there are cases in which offering a patient a poor-quality diagnosis is legitimate, or offering a high-quality one is illegitimate. This question (at least) will require a fuzzy theoretic response reminiscent of Sadegh-Zadeh’s. Quality is a spectrum from the very poorest to the very best diagnoses. The key question here is about the value of diagnoses. How good or how bad a diagnosis is will relate to the value that offering that diagnosis provides to the patient either directly or indirectly through the consequences of the diagnosis for the subsequent process.
Diagnostic accuracy is a challenge of understanding whether, when and how we can be confident that a diagnosis is correct. This will often be easier to tackle through the negative: how we can know that a diagnosis is inaccurate. While this series will build some tools and conceptual apparatus to set up some fledgling accounts of accuracy, this problem will ultimately remain wide open.
The final challenge is that of diagnostic effect. This problem has a different flavour to those before. It focuses upon the consequences of offering a diagnosis, and in some cases could be independent of the details of the diagnosis itself. The problem of diagnostic effect focuses upon the effects on the patient of the diagnosis – though we might also talk about effects on populations, practitioners and researchers. We broached this issue in the discussion of palliative diagnosis, asking whether the act of diagnosing a patient will have beneficial or harmful physical and psychological effects. This question poses the problem of how to anticipate, predict and manage these effects. It treats diagnosis, when communicated to the patient, as a medical intervention in its own right with effects (and unintentional side-effects), which should be evaluated to maximise benefit and minimize risks. With that in mind, we can also ask the ethical question of whether there are circumstances in which it is justifiable for a practitioner to withhold a diagnosis or some detail of that diagnosis from a patient to prevent harm. This vexing question opens provocative territory which is deeply uncomfortable for many clinicians, but is perhaps most familiar in the domain of paediatrics, in which practitioners have never been able to evade the question of what to communicate to a child patient and how to do so.
These problems are not independent. The first five in particular are nested: an account of diagnosis structure is needed to provide an account of diagnostic candidacy, and so on. The problem of diagnosis structure sets the domain for the other problems. Candidate, legitimate, high-quality and accurate diagnoses are all subsets of the domain of statements with the logical form of a diagnosis. So, for that matter, will be non-candidate, illegitimate, low-quality and inaccurate ones.
It seems intuitively clear that all accurate diagnoses should be candidate diagnoses, and that our accounts of each concept will need to do justice to this intuition. Accurate diagnoses are a subset of candidate diagnoses. But there may be (and surely are, in almost every case) inaccurate diagnoses which are still candidates. So, accurate diagnoses are a proper subset of candidate diagnoses. We can establish the following principle governing our accounts of accuracy and candidacy:
(A) If D is an accurate diagnosis for P’s symptom set S, then D is a candidate diagnosis for P.
Similarly, we can assume that giving a correct diagnosis to a patient always constitutes a legitimate attempt to diagnose that patient. But some inaccurate or partially accurate diagnoses (particularly those which are not very inaccurate – we will see that accuracy is also a fuzzy property in this sense) are likely to be legitimate. So, accurate diagnoses are also a proper subset of legitimate diagnoses, and:
(B) If D is an accurate diagnosis for P’s symptom set S, then D is a legitimate diagnosis for P.
Finally for our account of accuracy, it may be the case that a diagnosis being accurate means that it is also high-quality. But this claim will have to await analysis. There is an intuitively accessible possibility that an accurate diagnosis may be low quality. Insofar as quality relates to the value of the diagnosis, it may be that there are entirely accurate diagnosis statements which do not add anything of substance to the process or provide any other support to the patient. Since this cannot yet be ruled out, it is too early to claim that accurate diagnoses are a subset of high-quality diagnoses or to draw a direct connection between accuracy and quality. We will presume for the moment that, all else being equal, a more accurate diagnosis is higher quality than a less accurate one.
Because they fall into a continuum, high-quality and low-quality diagnoses (and everything in between) have a somewhat less clear intuitive relationship to candidacy and legitimacy. It seems necessary for all diagnoses above a certain quality threshold to be worth considering as candidates. Moreover, a higher quality level would likely render a diagnosis legitimate. But this does not entail that low-quality diagnoses are necessarily not candidates. Legitimacy seems, intuitively, more closely connected with quality. This problem will need disentangling.
Finally, all legitimate diagnoses are diagnostic candidates. If it would be acceptable to offer a given diagnosis to a patient, then it would be reasonable for a clinician (or a machine) to consider and evaluate that diagnosis. But clearly not all candidate diagnoses will be legitimate. So, legitimate diagnoses are a proper subset of candidate diagnoses, and:
(C) If D is a legitimate diagnosis for patient P’s symptom set S, then D is a candidate diagnosis for P.
The problem of diagnostic effect has some relationship to the problems of quality, accuracy and legitimacy. Intuitively, we can anticipate that certain harmful effects are more likely and beneficial ones less likely if an inaccurate diagnosis is given than an accurate one. A particularly evocative class of cases are those in which a patient is inaccurately given a deeply disturbing diagnosis with a morbid prognosis, and adapts their life and behaviour accordingly, only to subsequently discover the error. There are clearly increased risks of harm attendant to inaccurate diagnoses. But accuracy is not a full defence against harm. Intuitively, it seems that diagnostic effects will be closely bound up with legitimacy: premature diagnoses with serious risks of harmful effects upon the patient are perhaps the clearest diagnoses with potential to be candidates but illegitimate diagnoses.
These six problems will form the scaffold for the analysis of diagnosticity in the rest of this series. First, some philosophical work is needed to set our parameters by understanding what is meant by symptoms, signs, diseases, abnormalities and maladies. There is a massive philosophical literature on questions of health and disease, and much of this will need to be bracketed for a philosophy of diagnosis to get off the ground. However, some clear definitions which allow room for constrasting philosophical positions can be articulated. Then, we engage with the analysis of diagnostic structure, unpicking Sadegh-Zadeh’s (2000;2011) account of the logical form of diagnoses and outlining an account of four broad categories of diagnosis statement, each of which will count as a logical structure for diagnosis statements. Next, we will explore causal relevance, which sits at the heart of Sadegh-Zadeh’s account of diagnosticity, and show that there are fundamental flaws to the identification of causal relevance with diagnostic candidacy, legitimacy and quality. Then, we will assess the program of ‘fuzzification’, rendering each of the four classification problems (candidacy, legitimacy, quality and accuracy) in the form of fuzzy predicates, and showing that some, but not all, of these descriptors can be handled well by this account. Then, we will develop a pluralist account of diagnostic legitimacy and quality and suggestions for the development of the analysis of diagnostic accuracy. The problem of diagnostic effects is reserved for a final treatment in its own right, but will never be far from the surface throughout the other analyses.
- Aronowitz, R.A., 2001. When Do Symptoms Become a Disease? Annals of Internal Medicine, 134: 803–808.
- Dennett, D., 1978. Brainstorms, MIT Press.
- Reiss, J. & Ankeny, R.A., 2016. Philosophy of Medicine. The Stanford Encyclopedia of Philosophy (Summer 2016 Edition), Edward N. Zalta (ed.), URL = <https://plato.stanford.edu/archives/sum2016/entries/medicine/>.
- Sadegh-Zadeh, K., 2000. Fundamentals of clinical methodology: 4. Diagnosis. Artificial Intelligence in Medicine, 20(3), pp.227-241.
- Sadegh-Zadeh, K., 2011. The logic of diagnosis. In Philosophy of Medicine, pp. 357-424, North-Holland.