Subtyping of psychiatric disorders: implications for drug development.

Psychiatric diagnosis suffers from being based on phenomenology and not on pathophysiology. Data are presented showing that psychiatric patients reveal consistent quantitative electroencephalographic abnormalities, such that they can be separated from normals and from each other. Clustering these pathophysiological groupings reveals an underlying variability, which permits useful subtyping. Data are presented relating subtyping to pharmacological treatment.

The problem of pathophysiological diagnosis in psychiatry is unmet, with the possible exception of Alzheimer's disease. Diagnostic efforts including International Classification of Diseases (ICD)[ and Diagnostic and Statistical Manual of Mental Disorders (DSM),[ are descriptive in nature and based on phenomenology. Virtually all of the phenomenological “markers” can be arrived at through different gene-environment interactions and via totally different pathways. The result is a diagnosis based on phenomenological similarity and a diagnostic category that is heterogeneous and unclear regarding etiopathogenesis. The individuals so labeled may resemble each other at a given moment in time, but they are not classified on the basis of etiopathogenesis. For the last 100 years, diagnosis in medicine has moved away from phenomenology and toward etiopathogenesis. It is that movement that has made for a truly scientific medicine. Psychiatry must follow this path. The quest for pathophysiological markers goes back to Emil Kraepelin and continued for many years thereafter. With the advent of psychodynamic thinking, the search for pathophysiology diminished and was replaced by the search for internalized conflicts. Part of the reason for the failure of that pathophysiological quest included limitations in the scientific methods available to investigators. The development of imaging technology has brought a dramatic change in the power available to investigators.

Discriminates

In an article published in Science, [3] it was demonstrated that data derived from quantitative electroencephalography (EEG) were strongly correlated with DSM diagnoses. The data were age-corrected and Z-transformed, so as to make it possible to use appropriately powerful statistical techniques (“neurometric analysis”) ( Discriminate equations could then be written, which, on the basis of EEG findings, could reliably separate psychiatric patients from normals and classify patients along the lines of the DSM nomenclature. The importance of this finding was initially not fully recognized and brushed aside as “merely correlational in nature.” Nevertheless, there was a consistent and replicable demonstration of abnormal brain activity as a function of diagnostic category. A major limitation of the methodology was that the signal is derived from the scalp, and the source of the signal was not localized three-dimensionally (

Power distribution in different diagnostic categories.

Reproduced from reference 3: John ER, Prichep LS, Friedman J, Easton P. Neurometrics: computer-assisted differential diagnosis of brain dysfunctions. Science. 1988:293:162-169. Copyright © 1988, American Association for the Advancement of Science.

It became clear over time that some features of the abnormal signal did not change with treatment or even with clinical improvement. It can only be concluded that the signal was a mixture of state and trait variables. Nevertheless, it was clear that the patients who improved clinically tended to move toward the normal space and were less abnormal statistically than they had been prior to successful treatment.

Cluster analysis

An interesting question then arose. While it is possible to group patients according to their abnormal quantitative EEG (qEEG) findings, does this mean that the groups were homogeneous within themselves? The technique of discriminate analysis cannot address this question. On the other hand, the use of a cluster analysis technique will assist in resolving this issue.[4]

As can be seen in a perfect discriminate will separate a group into variable sets, but it does not identify where they are located along the vector that separates those variable sets. The cluster analysis will permit an examination of which person identified as belonging to a discriminate group most resembles his or her neighbor. In other words, once we have separated a group via the qEEG methodology into a diagnostic category, we can ask which members of that category look most like their neighbors and which do not.

A cluster analysis on obsessive-compulsive disorder (OCD) revealed two distinct clusters ( [5,6] While the patients could be identified by qEEG as OCD, they clustered into two groupings. Being able to cluster individuals has no meaning if the cluster is not related to something useful. The question was, do these clusters differ in some clinically meaningful fashion? It turned out that members of cluster 1 were predominantly nonresponders to selective serotonin reuptake inhibitors (SSRIs), while members of cluster 2 were predominantly responders to SSRIs. These rates of response and nonresponse of approximately 80% are astonishing, especially given the fact that the data were derived from the scalp and not from the actual source of the abnormality. Three-dimensional source localization via variable resolution electrical tomography (VARETA) or magnetoencephalography would undoubtedly yield results that are more refined.

Cluster analysis of quantitative electroencephalography (qEEG) data in obsessive-compulsive disorder (OCD).

Reproduced from reference 6: Prichep LS, Mas F, Holiander E, et al. Quantitative electroencephalographic subtyping of obsessive-compulsive disorder. Psychiatry Res. 1993;50:25-32. Copyright © 1993, Elsevier Science.

shows differences between positron emission tomography (PET) images in OCD responders to SSRI treatment at baseline and after successful treatment with SSRI.[7] The localization of the metabolic changes was consistent with the RRG source localization of the abnormal activity.

Positron emission tomography (PET) in obsessive-compulsive disorder (OCD) responders (n=20) to selective serotonin reuptake inhibitor treatment: comparisons between drug-free baseline and retest. AC, anterior cingulate.

Reproduced from reference 7: Hansen ES, Prichep LS, Bolvvig TG, John ER. Quantitative electroencephalography in OCD patients treated with paroxetine. Clinical EEG. 2003. In press. Copyright © 2003, EEG and Clinical Neuroscience Society

A similar clustering algorithm was utilized for patients suffering from attention-deficit disorder (ADD). The cohort, of ADD cases was divided into two clusters: 76% of cluster 1 responded to methylphenidate, whereas 62% of cluster 2 responded better to dextroamphetamine (Table I). In other words, despite the total similarity of these cases clinically, the differential response to methylphenidate and dextroamphetamine was determined to a large extent by the distinctive pathophysiology revealed by cluster membership. Again, this cluster membership was determined by the scalp signal and not based on three-dimensional source localization (

VARETA images were computed at the qEEG frequencies where the most significant changes occurred. shows VARETA images taken at 6.63 Hz on dextroamphetamine responders before and after medication. One can sec the obvious normalization with medication.

shows VARRTA images at 5.85 Hz of dextroamphetamine nonresponders before and after medication. An examination of this figure shows worsening with medication. It should also be noted that the responders and nonresponders differed according to the VARETA frequency.

The final grouping that will be reported in this paper consists of a group of patients with schizophrenia, which were subtyped into five clusters ( Only members of cluster 1 showed a greater than 25% reduction in Brief Psychiatric Rating Scale (BPRS) with the use of haloperidol. Members of cluster 3 responded best to risperidone. What is apparent is that there were differential responses to medication as a function of cluster membership.

These three clinical examples demonstrate the variability in the pathophysiology within a so-called diagnostic category.

Conclusion

represents an uninformed cluster analysis of a mixed population containing both normal and abnormal individuals. They were clustered without diagnosis and then later grouped by categories ranging from normal through the various diagnostic labels. As is obvious, whether normal or any other clinical category, there was variability of cluster membership. Some members of a particular diagnostic category were in a particular cluster, while others of the same diagnostic category were in other clusters. What, is particularly striking is that while many normals are in cluster 10, patients with a variety of psychiatric disorders can also be found in cluster 10. This raises the question as to whether these are normal people only in the sense that they have not yet, become ill, but in fact have the trait variables that might be manifested in a variety of different diagnostic categories. Belonging to a particular cluster docs not identify whether or not an individual will manifest clinical illness. Ultimately, the task is to use three-dimensional source localization and more refined analysis of the pathophysiology to separate trait from state and thereby identify individuals who are at future risk from those who are not. Finally, developing a better understanding of pathophysiology will lead to more specific and more effective treatment of the subtypes of various psychiatric syndromes.

Table I.

Relationship between quantitative electroencephalography (qEEG) cluster membership and response to treatment in children with attention-deficient disorder.

Responder	Cluster 1		Cluster 2
	n	%	n	%
Methylphenidate	47	76	19	38
Dextroamphetamine	15	24	31	62