Steven E. Hyman is professor of Stem Cell and Regenerative Biology at Harvard University. He holds degrees from Yale University, the University of Cambridge, and the Harvard Medical School, where he was professor of Psychiatry from 1998 to 2001. He served as director of the National Institute of Mental Health in the United States (1996-2001), where he emphasized investment in neuroscience and genetic technologies. From 2001 to 2011 he was provost of Harvard University, the university’s chief academic officer. At present he is director of the Stanley Center for Psychiatric Research, which is part of the Broad Institute of Harvard and the Massachusetts Institute of Technology.
Hyman is president of the Society for Neuroscience, which has about 40.000 members; from 2008 to 2014 he was the founding president of the International Neoroethics Society. He is a Fellow of the American Academy of Arts and Sciences, Fellow of the American Association for the Advancement of Science, Fellow of the American College of Neuropsychopharmacology, and a Distinguished Life Fellow of the American Psychiatric Association. He is especially interested in the genetics of neuropsychiatric disorders, aiming to develop models that will help to discover new therapeutics.
By now, most mental health researchers would attest that DSM-5 (American Psychiatric Association 2013), the widely used diagnostic classification of mental disorders, does not stand up to scientific scrutiny. Useful alternatives are not, however, readily at hand. Classifications like the DSM-5 are, in a formal sense, cognitive schemata applied to data with the purpose of rendering them understandable and useful. Given their reductive nature, medical and biological classifications gain utility by failing to capture all of nature’s complexity. For example scientists revising the ‘tree of life’, which attempts to show the relationships among all living organisms, must often handle situations in which genetic and morphological data yield divergent or indeterminate conclusions. However, the most serious problems with such classifications arise not from their recognizable limitations, but when they become reified and unconsciously applied in a manner that blinds scientists to important observations or leads them to see falsifying data as an uninformative outlier. While issues of ‘cognitive blinding’ have been well documented by psychologists, philosophers, and historians of science, any potential warnings have failed to inoculate organized psychiatry against its perils.
Even as problems with the application of the Diagnostic and Statistical Manual of Mental Disorders (DSM) have continued to mount, many grant-making bodies and journal editors across the world continued to require that diagnoses conform to definitions in DSM-5 (American Psychiatric association 2013) or the similar ICD-10 (World Health Organization 1992) in the belief—mistaken in my view—that shared DSM definitions would ensure the comparability of research across different settings. The insistence on use of DSM (or ICD) diagnostic criteria in research, along with their central role in the training and practice of clinicians has facilitated their dominance of psychiatric thinking and has enshrined research on ‘unnatural’ population groupings originally devised in the 1970’s by expert committees who lacked validating scientific data (Hyman 2007, 2010). As a result, little illness-related research has been performed that could yield much beyond incremental tinkering with DSM diagnoses—at least until the findings of modern genetics, which I will briefly describe, began to erode the core scientific framework of the DSM system: the concept of mental disorders as discontinuous categories that define homogeneous groups of patients.
Recognizing the damage to the research enterprise arising from the hegemonic position of the DSM system, the US National Institute of Mental Health (NIMH) has recently eschewed its use as a basis for grant applications. Instead, NIMH has begun to develop a new approach to the classification of psychopathology called the Research Domain Criteria (RDoC; Cuthbert and Insel 2013) centered on the relationship of symptoms to underlying neurobiology. An immediate benefit of this stance is to liberate investigators from use of the DSM. In order to make replication of their studies possible, scientists must carefully detail their sampling frame instead of reporting a DSM diagnosis. NIMH does not claim, however, that RDoC is intended as the basis of a broad diagnostic classification meant to replace the DSM, nor is such an alternative likely in the near term. To achieve better diagnostic standards, much work would first need to be done that connects clinical observations (not DSM disorders) with genetic, neurobiological, and cognitive mechanisms—followed by a worldwide process that demonstrated validity and clinical utility, at least for the most common and severe psychiatric conditions. For the foreseeable future, however, clinicians are left to employ DSM-5 or ICD-10—indeed they are often required to use these systems for administrative purposes even when they recognize that many of their patients do not actually fit the criteria by which they have been diagnosed.
Given the early state of the relevant science, any current attempt to define mental disorders would perforce have shortcomings. In contrast to most areas of medicine, for example, there are still no objective tests for any mental disorder. Thus both DSM-5 and ICD-10 have had to base their diagnostic criteria entirely on phenomenology, for the most part on symptoms that can only be reported subjectively. The need to elicit symptoms from patients who may be limited in their ability or willingness to report them partly explains the inability of the DSM system to achieve its central goal of producing high rates of diagnostic agreement concerning any particular patient (inter-rater reliability). In addition, the significance of some reported symptoms may be difficult to classify; for example, discerning the boundary between an overvalued idea and a delusion may not always be possible. Moreover, a diagnosis based on phenomenology is likely to reflect diverse underlying etiologies and disease mechanisms in different individuals. Unsurprisingly, patients given the same DSM-5 diagnosis often have significant differences in treatment response and prognosis, and when entered into research studies, may exhibit heterogeneous results, suggesting different underlying conditions. While the historical (and current) need to rely on phenomenology creates significant challenges for those drawing up a diagnostic system, there were still significant choices to be made in constructing the prototypical DSM-III (1980). Many scientific and clinical problems arising from the use of the DSM reflect those historically contingent decisions: reliance on narrow, highly specified diagnostic categories that were drawn with the primary goal of maximizing inter-rater reliability.
When the DSM system is applied clinically, it yields extremely high rates of co-occurrence, also called comorbidity, far in excess of chance co-occurrence based on prevalence (Krueger and Markon 2006). Thus an individual who receives a single DSM diagnosis has a high likelihood of meeting criteria for multiple additional diagnoses (e.g., multiple anxiety disorders and depression) or of a pattern of shifting diagnoses over the lifespan (e.g., schizophrenia, schizoaffective disorder, and bipolar disorder). In the light of data showing that those DSM disorders that most frequently co-occur share genetic and non-genetic risk factors (Kendler 2011), the most parsimonious explanation for co-occurrence is excessive splitting of single underlying pathologic processes into multiple narrow diagnostic silos.
Another frequent problem with DSM diagnoses arises from the excessive specificity of diagnostic criteria based on the prioritization of reliability. Thus, for example, schizophrenia is well understood to be a chronic disorder. The developers of the DSM saw the term “chronic” as too vague, however; thus DSM-5 requires at least six months of illness to make this diagnosis. In truth, it is very difficult to date the onset of schizophrenia in terms of months, but even if it were readily possible, this pseudo-precision led the DSM writers to invent ‘filler’ categories such as schizophreniform disorder (schizophrenia-like symptoms for less than 6 months), and then, with nary a wink, to treat them as independent diagnostic entities. Even this invention did not solve the problem: may patients with psychotic disorders never fully meet specific criterion sets (6 months or no) and then receive a catch basin ‘not otherwise specified’ (NOS) diagnosis and thus have no real diagnosis at all (Hyman 2007, 2010). Across many diagnostic groupings, NOS diagnoses have, of necessity, been widely used. In employing DSM-IV, for example, eating disorder NOS diagnosis were more widely used than the specific diagnoses of anorexia nervosa or bulimia.
Defenders of the DSM have argued that from the time that DSM-III (1980) created a new prototype diagnostic system grounded in the principles of careful clinical description, each periodic revision has represented a step in a process of successive approximation to true underlying disorders. If indeed the problems of the DSM system were no more than the errors in lumping and splitting, revisions grounded in evidence-based incremental improvements would eventually achieve a scientifically and clinically justifiable classification. I doubt this optimistic scenario. In my view, there is a growing body of evidence to suggest that the DSM classification is grounded in a deep misunderstanding of the nature of psychopathology.
In Procrustean fashion, the DSM defines all mental disorders categorically, i.e., as discontinuous from health and from each other. In contrast, much evidence, most recently from epidemiology and genetics, suggests that mental disorders are better captured dimensionally, where disorder is understood in terms of one or more quantitative deviations from health. Moreover, DSM disorders do not appear to be categorically distinct from each other. Across many areas of psychopathology, including mood and anxiety disorders, personality disorders, neurodevelopmental disorders, and psychotic disorders, patterns of overlapping symptoms and cognitive abnormalities, comorbidities, and familial sharing suggest the conceptual utility of broad spectra instead of narrow diagnostic categories. The spectrum concept is based on the hypothesis that a range of related psychopathology results from partly shared etiological factors and disease mechanisms. It should be added that with clinical utility in mind, the distinction between categories and dimensions in medical diagnosis is not as rigid as may first appear. Dimensional concepts related to severity are often superimposed on categories—thus a particular cancer diagnosis may be further analyzed quantitatively in terms of stage (degree of spread) and grade (malignancy of the cells). Similarly, for purposes of treatment decisions, categories are often superimposed on dimensions; thus progressively higher blood pressures are often binned into the relatively informal categories of borderline, mild, moderate, and severe hypertension.
In 1960’s and 1970’s, research psychiatrists based at Washington University in St. Louis, revived the careful, long-term observational methods of Emile Kraepelin (1856-1926). These proponents of what has been called ‘descriptive psychiatry’ took on the goal of identifying homogeneous groups of patients for treatment and research. Kraepelinian thinking had been pushed aside during the mid-20th century by then dominant psychoanalytic views that downplayed the significance of discrete diagnoses in favor of more complex formulations. Such formulations were grounded on views of psychopathology as failures to progress successfully along a developmental continuum. The descriptive psychiatrists were motivated, in part, by the need to match patients with newly discovered treatments that had become available with the rapid emergence of psychopharmacology in the 1950’s. There was an additional motivation for psychiatry to get its diagnostic house in order: to answer assertions that its diagnoses were little more than arbitrary labels that society had placed on annoying misfits.
To develop diagnoses that could be applied reliably across settings and that could ultimately be validated, Robins and Guze (1970) focused on identifying symptoms that could be ascertained readily with reasonable precision. Symptoms that typically clustered together formed the basis of diagnostic entities. They further hypothesized that such diagnostic entities that had been based on careful observation would breed true, that long term follow-up studies would be validating by showing stability over the lifespan, and that confirmatory diagnostic laboratory tests would be discovered. Robins and Guze (1970) hypothesized that these different strands of evidence would converge on valid disorders that could be delineated from each other and from health—as Kraepelin had used clinical observation to distinguish dementia praecox (schizophrenia) from manic-depressive illness (bipolar disorder). Kraepelin had further hypothesized that this distinction had a biological basis that would be partly revealed by studies of familial transmission. The prediction that independent lines of evidence would converge was, at the time, a reasonable view of the process of validation. In fact, convergence of lines of evidence has not been observed for disorders as they were defined by the DSM system. Symptoms of DSM disorders often do not breed true, and symptom patterns often change over the lifespan. What the architects of DSM-III had not predicted was the enormous complexity of genetic, developmental, and environmental influences on thought, emotion, behavior, and the illnesses related to them.
Beyond extending the Kraepelinian project of subdividing psychopathology into homogeneous, perhaps biologically based categories, there were several other reasons that the descriptive psychiatrists who set the stage for DSM-III may have favored a categorical classification. Categorical diagnoses seemed to assert the reality of multiple discrete, diagnosable disease entities at a time when the very concept of psychiatric illness had engendered doubt from some quarters. Moreover, this approach stood in sharp contrast to the psychoanalytic concept of psychopathology based on developmental continua. The apparently successful Kraepelinian distinction between dementia praecox and manic-depressive illness may have motivated attempts in the design of the DSM-III to progressively subdivide larger syndromes into narrower, putatively more homogeneous disorders, and even subtypes of disorders. In any case, the DSM-III (1980) emerged with a large number of disorders and subtypes, far outstripping any conceivable empirical justification, with all disorders defined by operationalized criterion sets that specified discontinuous categories. The only dimensional diagnosis in the DSM-III was mental retardation, not generally considered a psychiatric condition at all, based, as it had been historically, on the quantitative IQ scale.
To my mind, the distinction between categorical and dimensional diagnoses speaks to an issue of central importance for psychiatry and society: diagnostic thresholds. There has been a longstanding concern both within psychiatry and more broadly in society that dimensional and spectrum concepts would extend psychiatric diagnoses to individuals with mild symptoms and would thus medicalize ever larger fractions of humanity. Under what circumstances should certain thoughts, emotions and behaviors come under the purview of psychiatry? Which are better understood as expectable responses to stress or loss, as odd or eccentric but not diseased, or as matters for moral or legal rather than medical judgments? In public health discourse and in consideration of the limited resources of healthcare systems it is often debated which conditions, of what severity, deserve coverage by medical insurance, and which conditions should confer the patient role on the sufferer.
In the categorical classification of DSM-5, with its putative discontinuities between disorder and health, diagnostic thresholds are explicitly set by the listed criteria. Major Depressive Disorder (MDD) requires the presence of at least five out of nine DSM-5 criteria for at least two weeks. Like an infectious disease, one has MDD or one does not. In contrast, if a mental disorder is conceptualized as a quantitative deviation from health, the presence of cognitive, emotional, and behavioral symptoms do not, by themselves warrant a diagnosis since they are understood to be experienced also by people who are healthy. The setting of diagnostic thresholds for dimensionally defined disorders makes explicit the need for a policy decision. Such a decision would take into account empirical data concerning such factors as symptom number and severity, levels of distress and impairment, context such as recent bereavement, as well as health outcomes.
Policy decisions involving expert consensus are needed when scientific observations are applied to society. The problem with the development of DSM-III was not the involvement of experts—although for some health conditions diagnostic boundaries may soon be set computationally—but the lack of data. The use of evidence-based consensus that has taken such matters as long-term outcomes into consideration has produced diagnostic and treatment thresholds for hypertension, type 2 diabetes mellitus, and levels of LDL cholesterol. Similarly there are processes to determine what degree of cellular atypicality (a dimensional trait) might define the threshold for diagnosis of a particular cancer. The categorical nature of the DSM system has mooted research that has, in reality, been much needed, that could inform the setting of diagnostic thresholds (and which, in analogy with hypertension, need not be binary) on conditions such as attention deficit hyperactivity disorder, depression, autism spectrum disorders, and diverse anxiety disorders, where thresholds (and with them medical treatments and provision of other services) are a matter of contention. This is especially true when symptoms and impairments begin early in life that might not conform to a DSM diagnosis.
Evidence from epidemiology and twin studies has long favored an understanding of many psychiatric disorders as extremes on measures of dimensional traits found in the normal population. For example, epidemiological research and twin comparisons (including affected and non-affected co-twins) has identified no natural breakpoints in trait distributions or risk factors between healthy people and individuals diagnosed with disorders—as is also the case for many general medical disorders. For example, symptom number and severity appear continuous across the DSM categorical threshold for depression (Kendler and Gardner 1998). Symptoms of ADHD (Levy et al 1998) and autism spectrum disorders (ASDs; Robinson et al 2011, Lundström et al 2012) are continuous between those affected and healthy populations. The heritability of ASD traits is the same between the extreme (affected) group and the general population with a continuous increase in liability toward the extreme end (Robinson et al 2011).
Advances in genomic technology have permitted geneticists to identify specific DNA sequence variants in human genomes associated with schizophrenia, bipolar disorder, autism, and other disorders. Given the poor existing definitions of these disorders, in addition to the underlying genetic complexity, these studies have required very large numbers of patients and healthy comparison subjects (conferring very great statistical power) in order to succeed. This need has had the effect of motivating the formation of large global consortia that have, by this time in 2015, permitted the genotyping (on microarrays) of nearly 50,000 people with schizophrenia and sequencing of the DNA within the protein coding regions of their genomes in almost 10,000 cases. What has been learned is that with very rare exceptions the genetic contribution to risk of psychiatric disorders is highly polygenic. This means that the genetic component of risk (which acts together with stochastic factors in development—one cannot wire up 100 trillion synapses deterministically—and specific environmental factors) involves genetic variation at many hundreds of loci within the genome. It is hazardous to guess, but perhaps 1,000 of our approximately 20,000 genes might play a role in each of the disorders studied to date. Within these genes, small differences in DNA sequence, exerting modest independent effects, appear to add up in some individuals to produce risk of illness. What was not suspected is the degree to which there are shared risk variants across different psychiatric disorders. Thus there appears to be very significant sharing between schizophrenia and bipolar disorder (70% of common genetic variants), and a lower level of sharing (20%), involving different genes, between schizophrenia and autism (Smoller et al 2013, Lee et al 2013). The large numbers of genes involved in these disorders populate all of our chromosomes and therefore do not segregate together. The mixing and matching of risk associated versions (alleles) of these genes across generations explains the fact that narrowly defined DSM disorders do not breed true, and that within families there may be individuals with different DSM diagnoses, e.g., schizophrenia, bipolar disorder, and schizoaffective disorder, as well as individuals with no diagnosis, but who have some degree of cognitive impairment.
With the emerging ability to study polygenic risk for psychiatric disorders using DNA microarrays, it has become possible to move from twin studies, which make inferences about the genetic basis of phenotypes to large case-control studies in which the burden of risk-associated alleles is identified. Such studies have revealed that polygenic risk for ASDs differs only quantitatively from social impairments in control populations. On a more preliminary basis, the same appears to be true for ADHD and for related behavioral traits in unaffected children (Stergiakouli et al 2015).
It is early days, but findings from genetics would seem to support a substantial reconceptualization of psychaitric disorders in terms of dimensions (e.g., based on the additivity of risk alleles and the lack of an discontinuities between health and illness) and spectra (based on patterns of sharing of risk alleles across currently separate disorders). Genetics will also provide important tools for epidemiologists studying environmental risk factors, which may prove to be as complex. Based on ascertainment of genetic risk alleles, it may be possible to stratify populations studied epidemiologically (and also in cognitive studies, imaging studies, and clinical trials), thus achieving in a more limited and probabilistic fashion, one of the goals of Robins and Guze (1970), i.e., to increase the homogeneity of patient groups.
The implications are significant. Given the cognitive blinders of the DSM system, we have not developed many of the diagnostic instruments necessary to identify dimensions relevant to diagnosis and treatment. It will then be necessary to validate them globally and to determine their utility in both scientific investigation and clinical practice. We must also begin to think explicitly and wisely about thresholds for diagnosis and appropriate treatment interventions, perhaps most importantly in children.