Major depressive disorder subtypes to predict long-term course

Abstract

Background

Variation in course of major depressive disorder (MDD) is not strongly predicted by existing subtype distinctions. A new subtyping approach is considered here.

Methods

Two data mining techniques, ensemble recursive partitioning and Lasso generalized linear models (GLMs) followed by k-means cluster analysis, are used to search for subtypes based on index episode symptoms predicting subsequent MDD course in the World Mental Health (WMH) Surveys. The WMH surveys are community surveys in 16 countries. Lifetime DSM-IV MDD was reported by 8,261 respondents. Retrospectively reported outcomes included measures of persistence (number of years with an episode; number of with an episode lasting most of the year) and severity (hospitalization for MDD; disability due to MDD).

Results

Recursive partitioning found significant clusters defined by the conjunctions of early onset, suicidality, and anxiety (irritability, panic, nervousness-worry-anxiety) during the index episode. GLMs found additional associations involving a number of individual symptoms. Predicted values of the four outcomes were strongly correlated. Cluster analysis of these predicted values found three clusters having consistently high, intermediate, or low predicted scores across all outcomes. The high-risk cluster (30.0% of respondents) accounted for 52.9-69.7% of high persistence and severity and was most strongly predicted by index episode severe dysphoria, suicidality, anxiety, and early onset. A total symptom count, in comparison, was not a significant predictor.

Conclusions

Despite being based on retrospective reports, results suggest that useful MDD subtyping distinctions can be made using data mining methods. Further studies are needed to test and expand these results with prospective data.

INTRODUCTION

Patients with major depressive disorder (MDD) vary substantially in treatment response and illness course. Recognition of this variation has led researchers to search for depression subtypes defined either by presumed causes (e.g., postnatal depression),^[1,2] clinical presentation (e.g., atypical or melancholic depression,^[3,4]) or empirically-derived symptom profiles using cluster analysis,^[5] factor analysis,^[6] or latent class analysis,^[7] in hopes that patients in subtypes would be sufficiently similar in psychopathological processes to help identify underlying molecular etiologies or predict treatment response.^[7-9] However, subtyping distinctions up to now have not lived up to these expectations,^[8,10] although some commentators suggest that subtyping using endophenotypes or intermediate phenotypes might hold more promise.^[11,12]

Another potentially useful approach to subtyping, given the goal of prediction, would be to define subtypes using recursive partitioning^[13,14] and related data mining methods^[15,16] that search for synergistic associations of predictors with illness course. Such methods have been used in other areas of medicine^[17,18] and relatively simple applications have been used in psychiatry to predict depression treatment response^[19-23] and suicidality.^[24-26]

The current report presents results of preliminary analyses designed to find symptom-based subtypes predicting course of major depressive disorder using more complex data mining methods than in previous studies. The analysis is preliminary because it uses retrospective data on depression course collected in cross-sectional population epidemiological surveys rather than longitudinal clinical studies. Results are nonetheless useful in providing a proof of concept of the approach in a large and diverse sample of subjects who were asked about potentially important subtyping variables in their index episodes and assessed for multiple indicators of subsequent depression persistence and severity.

MATERIALS AND METHODS

Sample

Data come from the World Health Organization World Mental Health (WMH) surveys (www.hcp.med.harvard.edu/wmh), a series of well-characterized community epidemiological surveys^[27-30] administered in six countries classified by the World Bank as high income (Israel, Japan, New Zealand, Northern Ireland, Portugal, United States,), five upper-middle income (Brazil, Bulgaria, Lebanon, Mexico, Romania), and five low/lower-middle income (Colombia, Iraq, Nigeria, Peoples Republic of China, Ukraine).^[31] Most surveys feature nationally representative household samples, while two (Colombia, Mexico) represent all urban areas in the country, one selected states (Nigeria), and three selected Metropolitan Areas (Brazil, Japan, Peoples Republic of China). (Table 1) A total of 93,167 adults (age 18+) participated, 8,261 of whom met lifetime DSM-IV criteria for MDD. Sample sizes range from 2,357 (Romania) to 12,790 (New Zealand). The average weighted response rate was 73.7% (range: 55.1-95.2%). Weights adjusted for differential probabilities of selection and discrepancies with population socio-demographic/geographic distributions. Further details about WMH sampling and weighting are available elsewhere.^[32]

Table 1.

WMH sample characteristics by World Bank income categories^a

						Sample size
Country by income category	Surveyb	Sample characteristicsc	Field dates	Age range	Response rated	Total	With MDD
I. High-income countries
Israel	NHS	Nationally representative.	2002-4	21-98	72.6	4,859	284
Japan	WMHJ2002-2006	Eleven metropolitan areas.	2002-6	20-98	55.1	4,129	219
New Zealandf	NZMHS	Nationally representative.	2003-4	18-98	73.3	12,790	1,908
N. Ireland	NISHS	Nationally representative.	2004-7	18-97	68.4	4,340	423
Portugal	NMHS	Nationally representative.	2008-9	18-81	57.3	3,849	379
United States	NCS-R	Nationally representative.	2002-3	18-99	70.9	9,282	1,562
Total					67.9	(39,249)	(4,775)
II. Upper-middle income countries
Brazil – São Paulo	São Paulo Megacity	São Paulo metropolitan area.	2005-7	18-93	81.3	5,037	408
Bulgaria	NSHS	Nationally representative.	2003-7	18-98	72.0	5,318	283
Lebanon	L.E.B.A.N.O.N	Nationally representative.	2002-3	18-94	70.0	2,857	267
Mexico	M-NCS	All urban areas of the country (approximately 75% of the total national population).	2001-2	18-65	76.6	5,782	397
Romania	RMHS	Nationally representative.	2005-6	18-96	70.9	2,357	54
Total					74.8	(21,351)	(1,409)
III. Low and lower-middle income countries
Colombia	NSMH	All urban areas of the country (approximately 73% of the total national population)	2003	18-65	87.7	4,426	476
Iraq	IMHS	Nationally representative.	2006-7	18-96	95.2	4,332	193
Nigeria	NSMHW	21 of the 36 states in the country, representing 57% of the national population. The surveys were conducted in Yoruba, Igbo, Hausa and Efik languages.	2002-3	18-100	79.3	6,752	176
PRCe – Beijing/Shanghai	B-WMH/S-WMH	Beijing and Shanghai metropolitan areas.	2002-3	18-70	74.7	5,201	151
PRCe – Shenzhenf	Shenzhen	Shenzhen metropolitan area. Included temporary residents as well as household residents.	2006-7	18-88	80.0	7,132	452
Ukrainef	CMDPSD	Nationally representative.	2002	18-91	78.3	4,724	629
Total					81.4	(32,567)	(2,077)

^aThe World Bank. (2012). Data. Accessed June 5, 2012 at: http://data.worldbank.org/country.

^bNSMH (The Colombian National Study of Mental Health); IMHS (Iraq Mental Health Survey); NSMHW (The Nigerian Survey of Mental Health and Wellbeing); B-WMH (The Beijing World Mental Health Survey); S-WMH (The Shanghai World Mental Health Survey); CMDPSD (Comorbid Mental Disorders during Periods of Social Disruption); NSHS (Bulgaria National Survey of Health and Stress); LEBANON (Lebanese Evaluation of the Burden of Ailments and Needs of the Nation); M-NCS (The Mexico National Comorbidity Survey); RMHS (Romania Mental Health Survey); NHS (Israel National Health Survey); WMHJ2002-2006 (World Mental Health Japan Survey); NZMHS (New Zealand Mental Health Survey); NISHS (Northern Ireland Study of Health and Stress); NMHS (Portugal National Mental Health Survey); NCS-R (The US National Comorbidity Survey Replication).

^cMost WMH surveys are based on stratified multistage clustered area probability household samples in which samples of areas equivalent to counties or municipalities in the US were selected in the first stage followed by one or more subsequent stages of geographic sampling (e.g., towns within counties, blocks within towns, households within blocks) to arrive at a sample of households, in each of which a listing of household members was created and one or two people were selected from this listing to be interviewed. No substitution was allowed when the originally sampled household resident could not be interviewed. These household samples were selected from Census area data in all countries other than France (where telephone directories were used to select households) and the Netherlands (where postal registries were used to select households). Several WMH surveys (Belgium, Germany, Italy, Poland) used municipal resident registries to select respondents without listing households. The Japanese sample is the only totally un-clustered sample, with households randomly selected in each of the 11 metropolitan areas and one random respondent selected in each sample household. 19 of the 26 surveys are based on nationally representative household samples.

^dThe response rate is calculated as the ratio of the number of households in which an interview was completed to the number of households originally sampled, excluding from the denominator households known not to be eligible either because of being vacant at the time of initial contact or because the residents were unable to speak the designated languages of the survey. The weighted average response rate is 73.7%.

^ePeople's Republic of China

^fFor the purposes of cross-national comparisons, we limit the sample to those 18+.

Measures

Interview procedures

Translation, back-translation, and harmonization of the interview schedule used standardized procedures.^[33] Interviews were fully-structured and administered face-to-face in the homes of respondents by trained lay interviewers. Rigorous interviewer training and quality control procedures were employed.^[34] The research presented here is in compliance with the Code of Ethics of the World Medical Association (Declaration of Helsinki). The institutional review board of the organization that coordinated the survey in each country approved and monitored compliance with procedures for obtaining informed consent and protecting human subjects.

MDD

DSM-IV MDD was assessed with the Composite International Diagnostic Interview (CIDI), Version 3.0,^[35] a fully-structured diagnostic interview designed for administration by trained lay interviewers. The CIDI translation, back-translation, and harmonization protocol required culturally competent bilingual clinicians to review, modify, and approve key phrases describing symptoms. Clinical reappraisal studies conducted in several WMH countries found good concordance between lifetime DSM-IV/CIDI diagnoses of major depression and independent diagnoses based on blinded SCID clinical reappraisal interviews,^[36] with area under the ROC curve (AUC) averaging .75 and LR+ averaging 8.8 (a level close to the threshold considered definitive for ruling in a clinical diagnosis from a screen).^[37]

Respondents with lifetime DSM-IV/CIDI MDD were asked retrospective questions about age-of-onset (AOO), whether their first lifetime depressive episode “was brought on by some stressful experience” or happened “out of the blue,” all DSM-IV Criterion A-D symptoms of MDE for the index episode (including separate questions about weight loss and weight gain, insomnia and hypersomnia, psychomotor agitation and retardation, and thoughts of death, suicide ideation, suicide plans, and suicide gestures-attempts), ICD-10 severity specifiers, questions to operationalize diagnostic hierarchy rule exclusions, and questions about symptoms during the index episode that might be markers of (i) dysthymia (inability to cope; social withdrawal), (ii) mixed episodes (sleep much less than usual and still not feel tired; racing thoughts), and (iii) anxious depression (feeling irritable; nervous-anxious-worried; having sudden attacks of intense fear or panic).

Four retrospective questions were asked about subsequent lifetime MDD course: number of years since AOO when the respondent had an episode (i) lasting two weeks or longer or (ii) lasting most days throughout the year; (iii) a dichotomous measure of whether the depression was ever so severe that the respondent was hospitalized overnight (and, if so, age of first hospitalization); and (iv) a dichotomous measure of whether the respondent was currently disabled (at least 50% limitation in ability to perform paid work) because of depression. These are the four outcomes considered here. The two measures of years in episode were divided by number of years between age-at-interview (AAI) and AOO+1 to create continuous outcomes in the range 0-100%.

Other predictors

In addition to the information described above about the index episode, additional predictors included discretized information about the respondent's AOO in eight nested age categories selected for sensitivity in the age range with most onsets (less than or equal to ages 12, 15, 19, 24, 29, 34, 39, 59), similarly nested and discretized information about AAI-AOO, and a binary variable for respondent Family History Research Diagnostic Criteria Interview^[38] reports for whether respondents’ parents had a history of major depression.

Analysis methods

Analysis of the de-identified WMH master dataset was approved by the Institutional Review Board of Harvard Medical School, the site of the WMH Data Coordination Center. An ensemble of 100 classification trees was used to find important interactions among predictors of the outcomes. The ensemble approach (i.e., combining results across a large number of replicates, each replicate estimated in a different simulated pseudo-sample) was used to reduce risk of over-fitting.^[13-15] The recursive partitioning R package rpart^[39] was used for this purpose. The minimum number of observations in a node for further splitting was set at 20 and the threshold complexity parameter (cp) at 0.01. The models to predict years in episode, which used a Poisson link function, were estimated among respondents where AAI-AOO was either 10+ years (years with episodes lasting most of the year) or 15+ years (years with any episode) based on preliminary inspection showing that outcome scores stabilized after these cut-points. Proportional hazards survival models were used to predict age at first hospitalization for depression among respondents who were not hospitalized for depression at AOO. Logistic regression models were used to predict current disability in the total sample.

Each tree in the ensemble was built in a randomly selected bootstrap sample drawn without replacement from the sample and cross-validated among the remaining respondents to determine appropriate tree depth. Inspection of summary frequencies of unique terminal nodes (i.e., subgroups of respondents defined by the conjunction of the dichotomous predictors selected to optimize prediction of the outcome) across the 100 trees was used to select the interactions to retain in a second step of analysis. This second step fitted a separate generalized linear model (GLM) for the multivariate associations of all predictors with each outcome. Included here were additive associations of the individual predictors with the outcome, the interactions found to occur repeatedly in the tree models, and nested dichotomies to describe the total number of symptoms endorsed. The inclusion of the latter predictors was important to distinguish differential predictive effects of especially important symptoms from predictive effects of an overall symptom count.

As some of the predictors in the GLM models were highly correlated, conventional regression methods yielded unstable results. Stepwise regression,^[40] which is often used to address this problem, over-fits and performs poorly in new samples.^[41] A number of data mining methods have been developed to improve on stepwise regression. We used one such method, the Lasso,^[42] to address this problem. The Lasso is one of several penalized regression methods that trades off bias to increase the efficiency of estimation by constraining the sum of variance of nonzero values of standardized regression coefficients with coefficient shrinkage parameters We selected Lasso instead of alternatives, as this penalty handles high correlations among predictors by yielding a sparse model (i.e., forces coefficients of weak predictors to zero).^[43] The R-package glmnet^[44] was used to estimate the Lasso GLMs using the same link functions as in the regression tree models. Coefficients from the Lasso models were exponentiated to create incidence density ratios (IDRs) to predict proportion of years in episode, hazard ratios (HRs) to predict hospitalization, and odds-ratios (ORs) to predict disability. No confidence intervals were generated, as standard errors in such models are biased.

The best-fitting Lasso coefficients were then used to generate predicted values of each outcome for all respondents. Based on evidence of strong correlations among these predicted values across outcomes, k-means cluster analysis was used to partition the sample into subtypes with similar multivariate profiles of predicted scores across the four outcomes using the R-package stats^[45] and using 100 random starts for each number of clusters. Inspection of observed (as opposed to predicted) mean dichotomized outcome scores (percentages of respondents with high persistence and chronicity, hospitalization, and disability) and calculation of AUC (adjusted appropriately for the survival outcome,^[46] were used to select an optimal number of clusters. Associations of cluster membership with dichotomized versions of outcomes were then examined by calculating relative-risk of the adverse outcomes in the high-risk versus other clusters, positive predictive value (PPV; the proportion of high-risk cluster respondents that experienced the adverse outcomes), and sensitivity (SN; the proportion of all adverse outcomes that occurred in the high-risk cluster).

RESULTS

Distributions of the outcomes

The mean, median, and inter-quartile range (25^th-75^th percentiles) percentages of years after AOO when respondents in the analysis sample reported having a depressive episode lasting two weeks or longer were 25.8%, 13.0%, and 6.2-29.4%, respectively. The comparable percentages for years having a depressive episode lasting most days throughout the year were 9.5%, 0.0%, and 0.0-9.3%. Lifetime hospitalization for a depressive episode was reported by 4.3% of respondents and current disability due to depression was reported by 1.6% of respondents.

Recursive partitioning

The terminal nodes repeatedly predicting outcomes in recursive partitioning all involved two-way or three-way interactions between child-adolescent (before age 19) AOO, suicidality, and anxiety (nervous-anxious-worried, irritable, attacks of fear-panic) during index depressive episodes. The conjunction of later AOO (age 35+) with anxiety and suicidality also predicted chronicity. The cells defined by the conjunction of early onset, suicidality, and anxiety had either the highest or, in one case (disability), second highest scores on all outcomes across cells of the table defined by these predictors. (Detailed results are available on request.) Based on these results, all two-way and three-way interactions among AOO, anxiety, and suicidality were included in the Lasso GLMs.

Lasso generalized linear models

Four predictors of persistence, eight of chronicity, and 11 each of hospitalization and disability were retained in the GLMs with Lasso coefficients meaningfully different from zero. (Table 2) The vast majority (85%) of these coefficients were positive. The positive IDRs for years in episode were in the range 1.1-1.4. The positive HRs for hospitalization and ORs for disability were in the range 1.1-1.9. Only one predictor, severe dysphoria, was retained in all four models. Severe dysphoria was also the strongest predictor of chronicity (IDR=1.4) and one of the strongest predictors of hospitalization (OR=1.7). Four other predictors with consistently positive coefficients retained in three of the four models included suicidality (1.1-1.6), panic attacks (1.1-1.5), the multivariate profile of pediatric onset and anxiety (either nervousness-anxiety-worry or panic) (1.1-1.3), and parental history of major depression (1.2). One of these four, suicidality, was also among the strongest predictors of hospitalization (HR=1.6) and disability (OR=1.5), while panic was one of the strongest predictors of disability (OR=1.5). Other strong predictors of hospitalization included inability to cope (HR=1.9) and hypersomnia (HR=1.5), while inability to cope was also one of the strongest predictors of disability (OR=1.4). Early-AOO-suicidality also predicted disability, while later-AOO (older than age 34)-suicidality predicted chronicity. The latter represented a nonlinearity in the effect of the multivariate AOO-anxiety-suicidality profile.

Table 2.

Lasso GLM coefficients to predict subsequent course of DSM-IV major depressive disorder based on characteristics of the incident episode^a

	Percent of years in episode
	Any episode IDRb	Episode lasting most of year IDRb	Hospitalized HRb	Disabled ORb
I. Criterion A symptoms of major depression
Severe dysphoriac (ICD-10 severity specifier)	1.1	1.4	1.7	1.2
Anhedonia				1.1
Weight loss	0.9
Weight gain		1.1		0.8
Insomnia			1.3
Hypersomnia			1.5
Psychomotor agitation			1.2
Psychomotor retardation				1.2
Suicidality		1.1	1.6	1.5
II. Symptoms of dysthymia
Inability to cope			1.9	1.4
III. Sym ptoms of anxiety
Irritability	1.1		0.8	1.2
Panic		1.1	1.3	1.5
IV. Symptoms of mixed episode
Racing thoughts		0.8
High energy			1.2
V. Multivariate symptom profiles
AOO < 19 and suicidality				1.3
AOO < 19 and anxiety	1.1	1.3	1.2
AOO ≥ 35 and suicidality and anxiety		1.2
VI. Other predictorsd
Endogenous				0.7
Parental history of depression		1.2	1.2	1.2
N	(2,869)	(3,958)	(6,465)	(8,261)

^aBased on Lasso GLM penalized regression models, with the size of penalty determined by 10-fold cross-validation to select the penalty yielding cross-validating results with minimum mean squared prediction error. No Confidence intervals are reported because standard errors of such simulated models are biased. See the text for a discussion of differences in link functions and sample sizes.

^bIDR = Incidence density ratio; HR=Hazard ratio; OR = Odds-ratio

^cThis is not the DSM-IV Criterion A symptom of dysphoria but the ICD-10 symptom for somatic depression that the dysphoria is so severe that the patient has a lack of emotional reaction to events or activities that normally produce an emotional response. The DSM-IV symptom of dysphoria, in comparison, was not a significant predictor in any of the models.

^dAn additional 12 predictors were included in the Lasso GLM models that had coefficients of either zero or near zero across all outcomes. These predictors are dysphoria, fatigue/loss of energy, worthlessness or excessive guilt, diminished ability to concentrate or indecisiveness, social withdrawal, nervousness-worry-anxiety, multivariate symptoms profiles of childhood (before age 13) onset with anxiety and/or suicidality, multivariate symptom profiles of AOO before 19 with anxiety and suicidality, other multivariate symptom profiles of AOO either before 13 or before 19 or after 34 with either anxiety and/or suicidality, little need for sleep, total number of symptoms, age of onset, and time between onset and age at interview.

Cluster analysis

Predicted values of each outcome were calculated for each respondent based on the GLM model coefficients. Spearman rank-order correlations among these predicted values were in the range .76-.89. Principal axis exploratory factor analysis showed that the correlations were consistent with the existence of a single underlying factor (factor loadings in the range .89-.94). Based on these results, k-means cluster analysis of transformed (to percentiles) predicted outcome scores searched for multivariate clusters defining differential risk of the outcomes.

Inspection of mean percentile scores for solutions between three and eight clusters showed all solutions defined one class with the highest mean scores on all outcomes, a second class with lowest mean scores on all outcomes, and other classes with consistently intermediate mean scores on all outcomes. (Figure 1a-1f) Based on this observation, alternative three-cluster solutions were constructed from the original four- through eight-cluster solutions by collapsing the intermediate clusters. AUC was then compared across these solutions to predict dichotomous versions of the measures of years in episodes (distinguishing the 5-10 top percentiles of respondents with highest scores), hospitalization, and disability to see if classifications of high-risk or low-risk clusters were refined in solutions with more than three clusters. None of the collapsed solutions had higher AUCs than the original three-cluster solution (.64 for years in episode, .61 for years in episodes lasting more than half the year, .70 for hospitalization, and .72 for disability).

Figure 1.

Mean predicted outcome scores in the three-cluster through eight-cluster k-means¹

*Per = the percentile-transformed predicted score on the persistence outcome variable; Chr = the percentile-transformed predicted score on the chronicity outcome; Hos = the percentile-transformed cumulative predicted probability of hospitalization; Dis = the percentile-transformed predicted probability of disability. ¹k-means cluster analysis of percentile-transformed predicted scores on the four outcomes for all respondents based on the Lasso GLM

The distribution of membership in the three-cluster solution was 30.7% high-risk, 35.6% intermediate-risk, and 33.7% low-risk. Respondents in the high-risk cluster were 2.1-5.1 times as likely as others and 2.5-11.3 times as likely as respondents in the low-risk cluster to have high levels of long-term MDD persistence and severity. (Table 3) Respondents in the high-risk cluster includes 52.9-69.7% of all those with high levels of long-term MDD persistence and severity and 68.4-71.1% of those with two or more such adverse outcomes.

Table 3.

Associations of cluster membership with positive screening characteristics

	Relative-riska in the high-risk cluster vs.
	All othersb		Those in the low-risk cluster		Positive predictive valuec		Sensitivityc
	Est	(95% CI)	Est	(95% CI)	%	(se)	%	(se)
Percent of years in any episode
Top 5 percentile	2.7	(1.7-3.7)	3.3	(1.7-4.9)	7.9	(0.8)	60.0	(4.3)
Top 10 percentile	2.5	(1.9-3.1)	3.1	(1.9-4.2)	16.4	(1.1)	58.0	(3.0)
Percent of years in episodes lasting most of the year
Top 5 percentile	2.7	(1.8-3.5)	4.0	(1.9-6.1)	8.2	(0.8)	59.3	(3.5)
Top 10 percentile	2.1	(1.6-2.5)	2.5	(1.7-3.3)	14.4	(0.9)	52.9	(2.5)
Hospitalized	5.1	(3.4-6.7)	10.4	(4.3-16.6)	9.6	(0.8)	69.7	(3.5)
Disabled	4.6	(2.5-6.7)	11.3	(2.9-19.8)	3.4	(0.4)	67.1	(4.9)
Summary outcomes using top 5 percentile
Anyd	3.2	(2.4-3.9)	4.5	(3.0-6.0)	25.4	(1.6)	64.2	(3.0)
Multiplee	4.3	(1.8-6.9)	6.5	(0.9-12.1)	5.6	(0.8)	71.1	(5.6)
Summary outcomes using top 10 percentile
Anyd	2.6	(2.1-3.1)	3.1	(2.3-4.0)	33.0	(1.9)	59.3	(2.6)
Multiplee	3.8	(2.3-5.3)	4.8	(1.9-7.7)	11.2	(1.1)	68.4	(4.3)

^aRelative-risk is the ratio of the percent of respondents in the high-risk cluster that experienced the adverse outcome compared to the percent in the other clusters or in the low -risk cluster.

^bOthers = Respondents in either the intermediate-risk or low -risk clusters.

^cPositive Predictive Value is the percent of respondents in the high-risk cluster that experienced the adverse outcome; Sensitivity is the percent of observed adverse outcomes that occurred in the high-risk cluster.

^dThese are dichotomous variables that differentiate respondents who had one or more of the following four adverse outcomes: in the top 5 percentile (or 10 percentile) of years with episodes, in the top 5 percentile (or 10 percentile) of years w ith episodes lasting most of the year, hospitalized, or disabled.

^eThese are dichotomous variables that differentiate respondents who had two or more of the four adverse outcomes.

Cluster membership was strongly associated (Cramer's V greater than .50) with only one baseline predictor, suicidality (V=.54), and moderately associated (Cramer's V in the range .30-.50) with eight others, including one Criterion A depressive symptom (worthlessness/excessive guilt, V=.34), the ICD-10 severe dysphoria marker (V=.47), one symptom of dysthymia (inability to cope, V=.50), two of the three symptoms of anxiety (irritability, panic attacks, V=.30-.44), and the early-AOO multivariate symptom profiles retained in the Lasso GLMs (early AOO with either suicidality or anxiety, V=.35-.46). (Table 4) Scores on these variables were consistently higher in the high-risk than intermediate-risk cluster and in the intermediate-risk than the low-risk cluster. However, proportional high-risk versus intermediate-risk differences were relatively modest in most cases (1.1-1.4 risk-ratios) other than for panic (1.7) and the early-AOO multivariate symptoms profiles (2.0-3.2), while proportional intermediate-risk versus low-risk differences were consistently larger, with the highest risk-ratios for panic (2.8), inability to cope (2.5), suicidality (2.0), and the multivariate symptoms profiles (2.4-7.1).

Table 4.

Symptoms associated with the high-risk, intermediate-risk, and low risk clusters

	High-risk		Intermediate-risk		Low-risk		Total
	%	(se)	%	(se)	%	(se)	%	(se)	χ 2 2	Cramer's V
I. Criterion A symptoms of major depression
Dysphoria	99.6	(0.2)	98.9	(0.2)	96.1	(0.4)	98.2	(0.2)	57.0*	0.11
Severe dysphoria	97.8	(0.4)	91.0	(0.7)	56.0	(1.1)	81.3	(0.5)	713.6*	0.47
Anhedonia	95.0	(0.5)	88.7	(0.7)	79.0	(0.9)	87.3	(0.5)	205.1*	0.20
Weight loss	74.9	(1.0)	74.8	(1.0)	76.4	(1.0)	75.4	(0.6)	1.8	0.02
Weight gain	18.0	(0.8)	14.8	(0.8)	12.0	(0.7)	14.8	(0.5)	28.5*	0.07
Insomnia	85.3	(0.9)	83.1	(0.8)	78.5	(0.9)	82.2	(0.5)	28.4*	0.07
Hypersomnia	10.5	(0.8)	9.0	(0.6)	9.7	(0.6)	9.7	(0.4)	2.4	0.02
Psychomotor agitation	16.9	(0.8)	17.7	(0.9)	14.8	(0.8)	16.5	(0.5)	6.4*	0.03
Psychomotor retardation	67.7	(1.0)	55.0	(1.0)	41.0	(1.1)	54.2	(0.6)	247.9*	0.22
Fatigue/loss of energy	89.4	(0.8)	86.6	(0.8)	81.2	(0.8)	85.7	(0.5)	54.4*	0.10
Worthlessness or excessive guilt	98.2	(0.3)	88.4	(0.7)	68.3	(1.0)	84.6	(0.4)	724.0*	0.34
Diminished ability to concentrate/indecisiveness	96.2	(0.4)	91.7	(0.6)	84.5	(0.7)	90.7	(0.4)	167.9*	0.16
Suicidality	98.8	(0.2)	76.9	(1.0)	38.2	(1.1)	70.6	(0.6)	1449.6*	0.54
II. Symptoms of dysthymia
Inability to cope	86.0	(0.8)	59.5	(1.0)	24.0	(1.0)	55.7	(0.7)	1296.2*	0.50
Social withdrawal/tearfulness	99.2	(0.2)	94.4	(0.5)	88.0	(0.8)	93.7	(0.3)	261.3*	0.19
III. Symptoms of anxiety
Irritability	77.7	(1.0)	62.9	(1.0)	41.1	(1.1)	60.1	(0.6)	541.3*	0.30
Nervousness-worry-anxiety	88.4	(0.9)	76.2	(1.0)	56.6	(1.1)	73.3	(0.6)	369.6*	0.29
Panic	66.4	(1.1)	38.1	(1.1)	13.5	(0.7)	38.5	(0.6)	842.1*	0.44
IV. Symptoms of mixed episode
Little need for sleep	47.1	(1.2)	47.0	(1.2)	41.7	(1.1)	45.2	(0.7)	14.2*	0.05
Racing thoughts	10.6	(0.8)	12.4	(0.7)	15.7	(0.9)	12.9	(0.4)	16.6*	0.06
High energy	3.2	(0.4)	3.0	(0.4)	3.3	(0.4)	3.1	(0.2)	0.3	0.01
V. Multivariate symptom profiles
AOO < 19 and suicidality	48.6	(1.2)	17.7	(0.9)	4.9	(0.5)	22.9	(0.6)	853.0*	0.43
AOO < 19 and anxiety	48.9	(1.2)	24.4	(1.1)	9.9	(0.7)	27.1	(0.6)	658.4*	0.35
AOO < 19 and suicidality and anxiety	47.8	(1.2)	14.9	(0.9)	2.1	(0.3)	20.7	(0.5)	919.8*	0.46
AOO ≥ 35 and suicidality and anxiety	16.4	(0.8)	21.9	(0.9)	9.1	(0.6)	15.9	(0.4)	179.2*	0.15
VI. Other predictors
Endogenous	17.0	(0.9)	16.2	(0.9)	14.7	(0.8)	15.9	(0.6)	4.3	0.03
Parental history of depression	9.9	(0.7)	5.8	(0.6)	3.2	(0.4)	6.2	(0.4)	63.8*	0.11
N	(2,520)		(2,899)		(2,842)		(8,261)

^*Significant at the .05 level, two-sided test

DISCUSSION

The above results are limited by being based on retrospective data collected in fully-structured interviews excluding information on such potentially important predictors as temporally primary comorbid disorders and treatment status. Sample biases could also have been introduced by differential response related to predictors or predictor effects or differential mortality. The limitations involving use of a fully-structured interview and restricted predictors almost certainly led to downward bias in the estimated strength of associations, but the other limitations could have introduced either conservative or anti-conservative biases. Results should be considered only exploratory because of these limitations, although the results have value both as a proof of concept and as a source of ideas about prediction patterns that warrant analysis in future studies.

Within the context of these limitations, three results emerged that could serve as a starting point for future prospective clinical studies. First, the recursive partitioning found an early-onset-anxious-suicidal subtype associated with all four outcomes (persistence, chronicity, hospitalization, disability) and a late-onset-anxious-suicidal subtype associated with chronicity. Second, the GLMs found that a number of index episode symptoms were significant predictors of all outcomes. The most consistent and powerful of these was severe dysphoria, while others included parental history of major depression, suicidality, panic attacks, and multivariate profiles of pediatric onset with anxiety and/or suicidality. Third, strong clustering was found in these predicted values across the outcomes, with the roughly 30% of respondents in the high-risk cluster accounting for more than two-thirds of cases with multiple indicators of high long-term persistence, chronicity, and severity.

Several previous epidemiological studies examined baseline predictors of long-term course either in treatment^[47,48] or community^[49-51] samples, but did not attempt to search for depression subtypes. While these studies found several replicated predictors, including cooccurring anxiety, pain-physical comorbidity, and family history of depression,^[50,52-54] no attempt was made in those studies to examine synergistic effects of predictor clusters other than for summary measures of overall depression symptom number. Importantly, we included a total count of depressive symptoms in our GLMs but this measure was not significant.

As noted in the introduction, subtyping analyses more similar to those reported here have been done to predict treatment response^[19,20] and naturalistic patterns of remission among patients^[23] or in the placebo control group of a depression clinical trial.^[21] A number of recent clinical studies have also used methods similar to ours either to predict suicidality during^[22,25,26] or after termination of^[24] treatment. However, none of those analyses used ensemble methods or combined recursive partitioning with GLM to assess both synergistic and additive predictor effects.

In considering the possibility of future extensions to prospective studies, it is important to note that although we found an early-onset anxious-suicidal depression subtype that predicts all the outcomes (suicidality being the critical element in predicting disability and anxiety in predicting the other outcomes), we failed to find recursive partitioning profiles associated with a larger set of predictors despite the sample being much bigger than in existing prospective studies (i.e., affording good statistical power to detect synergistic symptom profiles if they existed) and the symptoms considered being quite broad. Taken together with the results of a recent secondary analysis that failed to find stable symptom-based MDD subtypes defined by internal consistency,^[10] our failure to find more elaborate subtypes argues against the existence of complex MDD subtypes defined exclusively on the basis of synergistic associations among index depressive episode symptoms other than the early-onset anxious-suicidal subtype.

It is important to note that broader MDD predictive subtypes not defined exclusively by index episode symptoms might be found in either of two other ways. One possibility would involve expanding the search for subtypes beyond symptoms of an index episode. Included here, for example, could be information about temporally primary comorbid mental disorders (e.g., early-onset distress, fear-circuitry, or impulse-control disorders), physical disorders (e.g., metabolic syndrome), socio-demographics, and (neuro)biological factors to define subtypes. We purposefully did not include such expansions here, as we wanted to focus on subtypes defined by index episode symptoms, but future analysis should do so to broaden the search for subtypes to include these other predictors. It would be interesting for future research to examine the possibility that the significant association found here between later-AOO-anxious-suicidal depression in the index episode with later chronicity but no other outcome might reflect the importance of a late-onset depression subtype that might occur in conjunction with a physical comorbidity, such as cardio-metabolic illness^[55] associated with episodes of long duration but not high persistence or severity.

Such a possibility can only be examined by broadening the search for subtypes to include comorbid physical disorders. The potential value of expanding the search for subtypes to include information about biomarkers is illustrated in recent studies showing that the course of atypical and melancholic depression is differentially predicted by HPA-axis, metabolic syndrome and inflammatory parameters^[56] and that inflammatory dysregulation is associated with the onset of ‘mixed state depression’.^[57] Such analyses have the potential to discover clinically meaningful and biologically valid disease clusters across a range of clinically relevant outcomes, an approach consistent with the recent call for what has been referred to as a stratified medicine approach^[58] that bypasses the search for a gold standard and focuses instead on the discovery of subtypes associated with a range of clinically meaningful outcomes.

A second possibility would be to look more closely within the high-risk cluster found in our analysis to search for embedded subtypes. To understand this suggestion it is important to recognize that the clusters we discovered cannot themselves be thought of as subtypes in the classical sense because they were discovered by clustering predicted outcome scores rather than the predictors themselves. A great many different combinations of predictors could yield the same predicted outcome scores. This means that further effort is needed to define subtypes within the high-risk cluster by considering multivariate profiles among the predictors that determine cluster membership so as to take into consideration the differential importance of these predictors within and across outcomes. No attempt was made to do this here, but it is clearly something that warrants future investigation in future studies based on the analysis of a more complete set of predictors.

It would also be useful, finally, if future studies expanded the range of outcomes considered here. The four outcomes in our analysis were selected purely based on availability. Given the discovery that predicted values are strongly correlated across these outcomes, it would be useful to develop an understanding of the range of outcomes over which this consistency occurs. Such an investigation could be carried out informally using the simple correlational methods used here, or a more formal approach might be conceived along the lines of the canonical regression models used to study latent mediators in the development of comorbidity among mental disorders.^[59-61] Or it might be possible to address this issue by adapting the data mining methods developed to discover what have been called master regulators^[62] in molecular genetic studies of physical disorders.^[63-65] Regardless of method, though, the discovery of common predictors of multiple indicators of persistence, chronicity, and severity call out for a more diverse and integrated analysis of clusters and within-cluster subtypes among the predictors of such outcomes.

In thinking of these future developments, it is important to recognize that the recursive partitioning methods used here require a much larger sample size than is likely to exist in prospective clinical samples. This means that the most feasible way to extend the current results in prospective clinical studies would be to evaluate the significance of the synergistic symptom profiles found here rather than to attempt independent data mining exercises, although independent Lasso and cluster analyses using larger sets of predictors (possibly including measures of endophenotypes) and alternative indicators of outcomes would be quite feasible in such studies. Although it is unlikely that clinicians would be willing to collect such data for purposes of making the subtyping distinctions made here, it is conceivable that future studies will document powerful effects of other predictors that could be examined using similar methods and shown to have sufficiently important clinical implications that it would motivate clinicians to collect such information as a routine part of their initial evaluations to guide treatment planning. The technology described here holds great promise in facilitating analyses aimed at documenting such predictors.

CONCLUSION

Despite our analysis being based on retrospective reports, our results suggest that useful symptom-based MDD subtyping distinctions can be made with data mining methods that focus on prediction rather than internal consistency and that the resulting subtypes have meaningful relationships with course of illness. The practical value of this approach, though, can only be judged by replication with prospective data, ideally expanding the analysis to use a wider range of predictors and outcomes.