Subtyping of psychiatric patients by cluster analysis of QEEG

Download Subtyping of psychiatric patients by cluster analysis of QEEG

Post on 13-Aug-2016




2 download

Embed Size (px)


<ul><li><p>Brain Topography, Volume 4, Number 4, 1992 321 </p><p>Subtyping of Psychiatric Patients by Cluster Analysis of QEEG </p><p>E.R. John *t, L.S. Prichep *t, and M. Almas* </p><p>Summary: We have previously reported successful classification of patients with a variety of psychiatric disorders, using multiple discriminant functions based upon selected neurometric QEEG variables. In independent replications, these functions accurately separate patients with different DSM-III-R diagnoses from one another and from normals. This capability demonstrates that distinctive and replicable patterns of neurometric abnormalities are correlated with the clinical symptom clusters upon which DSM-III-R diagnostic criteria are based. However, patients with the same clinical diagnoses often respond very differently to the same treatments. Similar symptoms may arise from different pathophysiology. This study explored the 'natural structure' of a population of psychiatric patients in 8 diagnostic categories, using uninformed cluster analysis based upon the same set of neurometric variables found useful in separating each of these categories from normal. This preliminary numerical taxonomic approach reveals that groups of patients in each of these DSM-III-R categories contain subtypes with markedly different pathophysiology; further, patients in different DSM-III-R categories were aggregated together within each cluster, displaying similar pathophysiological profiles. Objective classification based on such physiological measurements may add information useful to improve treatment outcomes. </p><p>Key words: Cluster analysis; Heterogeneity within disease; Subtyping; Numerical taxonomy; Pathophysiological profile; Neurometrics. </p><p>Introduction </p><p>The variables extracted by QEEG procedures have a variety of physical dimensions; absolute power measures in microvolts squared, relative power in percent, mean frequency in cycles per second, etc. The diversity of these units of univariate measurement offers a major obstacle to the concise description and quantification of the mul- tivariate abnormalities which characterize most brain disorders. Comparison of such data to normative values to obtain a Z-score, which has become almost uniform practice since it was introduced by ourselves (John et al. 1977; John et al. 1980) and Duffy (1981), replaces these disparate physical units by the common metric of prob- ability. </p><p>The results of a QEEG examination on an individual </p><p>*Dept. of Psychiatry, New York University Medical Center, New York, N.Y. </p><p>tNathan S. Kline Psychiatric Research Institute, Orangeburg, N.Y. Accepted for publication: April 14, 1992. Acknowledgment: The technological assistance of MeeLee Tom, </p><p>Henry Merkin and Bryant Howard is acknowledged. In addition, we acknowledge Heather Stein for assistance in preparation of this manuscript. This work was supported in part by Cadwell Laboratories, Kennewick, Washington. </p><p>Correspondence and reprint requests should be addressed to Dr. E.R. John, New York University, Brain Research Laboratories, Depart- ment of Psychiatry, NYU Medical Center @ Old Bellevue, 27th Street and 1st Avenue, 8th Floor, New York, NY 10016, USA. </p><p>Copyright 1992 Human Sciences Press, Inc. </p><p>can now be represented by an abnormality vector in a brain signal space, the dimensions of which are scaled in Z- scores or standard deviations from the normative mean value, as in figure 1. After correction for intercorrelations to achieve orthogonal dimensions for the signal space, the length of the abnormality vector can be calculated accurately and should ideally be proportional to the over- all severity of the brain dysfunction (John and Prichep 1990). </p><p>This is in fact the case in many conditions (John and Prichep 1990; Prichep et al. 1990b). Figure 2 illustrates the high correlation between some QEEG features and the Global Deterioration Score, a multidisciplinary scale which measures the ability of elderly individuals to per- form the tasks of daily life. </p><p>Sensitivity of neurometric features in a wide variety of clinical conditions has been confirmed and extended by other investigators (Thatcher et al. 1989; Jonkman et al. 1985; Veering et al. 1986; de Weerd et al. 1989; Har- mony et al. 1987; Mas et al. 1991b; Alper et al. 1991; Struve et al. 1989; Struve et al., 1990; Senf 1988; Fitz-Gerald and Patrick 1991). The direction of the abnormality vector is indicative of the nature or quality of the illness. The pattern of abnormal QEEG/EP findings should differ for different diseases. We have shown this to be true for both QEEG and EP features (John et al. 1988a; John et al. 1988b; John et al. 1989). </p><p>A neurometric exam yields an abnormality matrix, in which the columns are local or composite brain regions, </p></li><li><p>322 John et al, </p><p>Z; ABNORMALITY VECTOR </p><p>/ / . </p><p>Zk </p><p>zj </p><p>Figure 11 Abnormality vector in brain signal space with dimensions scaled in standard deviations from the norma- tive mean, proportional to severity of abnormality. Length is proportional to multivariate abnormality, direction to quality of illness. </p><p>and the rows are univariate or multivariate features. A topographic map is only a graphic representation of one row in this matrix, but a disease profile is a vector encom- passing entries in several columns and rows. </p><p>Patients with different diseases constitute swarms of points distributed in brain signal space, such that those with one disease will be concentrated in one region while those with another disease are concentrated in another region, as shown in figure 3. </p><p>Discriminantfunctions are mathematical pattern recog- nition algorithms which optimally separate regions of signal space containing patients with different diseases. Using multiple stepwise discriminant analysis, we have succeeded in separating patients in a wide variety of psychiatric categories from one another, with good sen- sitivity and specificity achieved in independent replica- tions (John et al. 1988a; Prichep et al. 1986; Prichep et al. 1990c. For a summary of current discriminant functions and their independently replicated accuracies, see Prichep &amp; John, this issue (Prichep and Jol'm 1992). </p><p>QEEG has a valuable contribution to make as an ad- junct to diagnosis in routine clinical practice, where patients in the early stages of disease or with atypical or multiple disorders pose diagnostic problems. However, this utility is limited to a relatively small percentage of cases confronted by the experienced neurological or psychiatric specialist (Mas et al. 1991a), although it may provide important confirmation in a larger proportion of cases confronting non-specialist practitioners. Our dis- </p><p>Global Deterioration Score (GDS) </p><p>GDS 4-7 (n = 3s) </p><p>GDS 3 (n = 56) </p><p>GDS 2 (n = 76) </p><p>GDS 1 (n = 112) </p><p>%Theta (Left Central) </p><p>Z Score </p><p>Figure 2. Relation between Z-score for % theta in lead C3 and the Global Deterioration Score in 282 elderly subjects. Darkened bar in histogram indicates mean value. ANOVA between Z-score for % theta and GDS was sig- nificant at the 0.0001 level. </p><p>z | </p><p>PA'n ENTS WFrH DISEASE Y </p><p>Zk </p><p>PATIENTS WITH DISEASE X </p><p>oe e </p><p>e </p><p> = </p><p>~q </p><p>e </p><p>*e le ~e e~ o </p><p> n </p><p>Zl </p><p>Patients with different diseases, X and Y, are Figure 3. expected to be swarms of points concentrated in dif- ferent regions of signal space. </p></li><li><p>Subtyping Psychiatric Patients by QEEG Profile 323 </p><p>Discriminant Analysis (Informed) versus Cluster Analysis (Uninformed) </p><p>VARIABLE SET I </p><p>Cluster 1" </p><p>O OO C O00 </p><p>O O ," Cluster 2 o/'~ </p><p>~,:o;...;x 2\ </p><p>VARIABLE SET II </p><p>DISCRIMINANT ..-"" FUNCTION X v O/ " " O~j/"" </p><p>0 O0 ...." , 0 0 ~." </p><p>uster 3,,,,,,"~- N x </p><p>Figure 4. Given a population containing two types of patients previously identified as O's and X's, discriminant analysis is a method to identity a set of variables which will classify that population into those two categories, using pre-labeled O's and X's to build the classification rules. The X versus O discriminant function would cleave the signal space along the dotted line, with O's on one side and X's on the other side. Cluster analysis, classifying individuals in terms of their nearest neighbors and the distance to the center of the nearest point-dense region, would divide the space shown in the example into 3 neighborhoods or 'clusters', each containing both O's and X's. </p><p>criminant studies essentially demonstrate that patients who have received a diagnostic label, based upon the evaluation of clinical symptoms and history by a knowledgeable practitioner in accordance with the standardized procedures stipulated by DSM IIIR conven- tions, have pathophysiological QEEG features sufficient- ly distinctive for those categories to be accurately classified by computer. </p><p>Unfortunately, the demonstration that a patient clas- sifted into a given DSM IIIR category on the basis of his symptoms has recognizable QEEG correlates of that membership is not equivalent to identification of the treatment to which that patient will respond. Similar symptoms may arise from different etiologies. A symptom-based classification system cannot be relied upon to select optimum individualized treatments for each of the multiple causes which may underlie those symptoms. As behaviors are multi-determined, so be- havioral disorders may have multiple causes. </p><p>QEEG studies have already provided preliminary evidence that patients within a homogeneous symptom category may belong to heterogenous subtypes. In our opinion, the crucial contribution of neurometric assess- ment to psychiatry will be not as an adjunct to diagnosis but as a method for identification of subtype membership and the selection of treatment. Patterns of neurometric findings have been reported in the literature demonstrat- ing the separation of the following subtypes of patients: </p><p>1) Obsessive-compulsive disorder (OCD) patients who are responders to serotonin reuptake inhibitors (Prichep et al. 1991; Mas et al. 1991b). </p><p>2) Major affective disorder patients in unipolar and bipolar subtypes (Prichep, 1987; Prichep et al. 1990a). </p><p>3) Major affective disorder patients who are electrocon- vulsive therapy (ECT) responders (Roemer et al. 1991). </p><p>4) Mildly impaired elderly patients who will show sig- nificant cognitive deterioration within five years (John and Prichep 1990). </p><p>5) Learning disabled children who will respond to methylphenidate (Prichep and John 1990). </p><p>6) Comatose head injured patients who will regain cog- nitive competence (Thatcher et al. 1989). </p><p>7) Schizophrenic patients who are responders to haloperidol (Czobor and Volavka 1991). </p><p>Methods and Subjects Such results indicate the existence of multiple sub- </p><p>types of patients with different therapeutic prognoses within many symptom-defined categories. These find- ings show the importance of developing methods of objective classification, or numerical taxonomy (John et al. 1977), based upon biological measurements rather than clinical descriptions of symptoms. </p><p>As shown by figure 4, discriminant analysis establishes rules for the separation of patients belonging to two or more pre-determined categories on the basis of a priori beliefs and definitions. Such a method is confirmatory of reliable correlations between physiological and clinical data sets. In contrast, cluster analysis divides patients into categories defined by their location inside or outside neighborhoods in signal space defined solely on the basis of "natural structure" such as distances between abnor- mality vectors derived solely from biological measures (John et al. 1977). </p><p>Several major problems arise when one undertakes </p></li><li><p>324 John et al. </p><p>Primary Cluster Structure of the B R L Psychiatr:ic Population </p><p>, oT </p><p>, ,Oo, m mm </p><p>:l m 4O 30 </p><p>' 20 [] </p><p>:I 4O . 30 ,o, m m </p><p>110/ 140 </p><p>54/ 89 </p><p>r </p><p>50 T - - ) 20/' </p><p>:l i 4 3O 2o m t ...... m </p><p>37 </p><p>69/ 99 </p><p>mm s I % iJ '10 </p><p>m '~ 20 ,0ot m m </p><p>:1 ,,,, 40 50 </p><p>" 2O I m </p><p>129/' 197 </p><p>40 103 ~ 3o </p><p>20 , . . [] </p><p>40 31 3o I 2o m l </p><p>I ~J ~ 4 ~ 6 7 6 q l(, l ] 121 </p><p>CLUSTER </p><p>Figure 5. Results of cluster analysis of 764 normals and psychiatric patients in 8 different diagnostic categories, A set of 26 variables previously found useful in discriminant functions which could accurately separate patients in these categories from normals and from one another were used. 12 clusters were constructed. Note that 1) each DSMIII-R category was subdivided into 2 to 4 sub- types; 2) each cluster except 7, 8 and 9 contained patients in 2 or more DSMIII-R categories. </p><p>cluster analysis of a body of data. One major issue is which and how many variables to select. A second major issue is how many clusters to seek. At the present time, there is insufficient experience to provide informed guidance in these decisions. We proceeded pragmatical- ly to carry out a prel iminary cluster analysis of "psychiatric space". </p><p>We assembled a sample of 764 patients and normals </p><p>in 8 major categories based upon DSM IIIR criteria: nor- mal (n--140), unipolar depressed (n=89), bipolar depressed (n=38), schizophrenic (n=136), alcoholic (n=30), mild cognitive impairment (MCI, n=103), primary organic dementia (n=197), and obsessive-com- pulsive disorder (n--31). We selected a set of 26 variables previously found most useful in the construction of mul- tiple discriminant functions to separate patients in these different DSM IIIR categories from normals or from each other. We explored the consequences of constructing 2, 4, 6, 10, 12 and 16 clusters and concluded that 12 clusters yielded a data structure which was sufficiently "consis- tent" and "stable" to provide a first approximation to what this approach might ultimately reveal. </p><p>Results The sample of patients in each DSM IIIR category was </p><p>randomly divided into balanced split-halves, and en- coded for "blind" cluster analysis. After cluster analysis was completed, the samples were decoded. Figure 5 displays the structure which replicated across the split- halves. This replicated structure contained 481 of the 764 patients, or 63%. The remaining patients were scattered in randomly distributed cells which did not replicate. On one hand, the sample of patients in each DSM IIIR category fragmented into several major subtypes, with reasonable split-half replication. This indicated that groups of patients with sufficiently similar clinical symptoms to belong to the same DSM IIIR category displayed markedly different subsets of physiological abnormalities. On the other hand, the patients ag- gregated within each of the clusters with similar physiological profiles displayed membership in diverse DSM IIIR categories. In other words, similar symptoms could arise from different pathophysiology and similar pathophysiology could lead to different symptoms. </p><p>The top half of figure 6 shows selected relative power, coherence and symmetry maps averaged separately across the dementia patients who were members of cluster 1, 8 and 9. The lower half of figure 6 shows the corresponding maps averaged separately across the MCI patients who were members of clusters 1, 6 and 11. The neurometric p...</p></li></ul>