Latent subtypes of manic and/or irritable episode symptoms in two population-based cohorts

Abstract

Background

Mood disorders are characterised by pronounced symptom heterogeneity, which presents a substantial challenge both to clinical practice and research. Identification of subgroups of individuals with homogeneous symptom profiles that cut across current diagnostic categories could provide insights in to the transdiagnostic relevance of individual symptoms, which current categorical diagnostic systems cannot impart.

Aims

To identify groups of people with homogeneous clinical characteristics using symptoms of manic and/or irritable mood and explore differences between groups in diagnoses, functional outcomes. and genetic liability.

Methods

We used latent class analysis (LCA) on eight binary self-reported symptoms of manic and irritable mood in UK Biobank and PROTECT studies to investigate how individuals formed latent subgroups. We tested associations between the latent classes and diagnoses of psychiatric disorders, sociodemographic characteristics, and polygenic risk scores (PRS).

Results

Five latent classes were derived in UK Biobank. (N=42,183) and were replicated in the independent PROTECT cohort (N=4,445), including ‘minimally affected’, ‘inactive restless’, active restless’, ‘focused creative’, and ‘extensively affected’ individuals. These classes differed in disorder risk, PRS, and functional outcomes. One class that experienced disruptive episodes of mostly irritable mood was largely comprised of cases of depression/anxiety, and a class of individuals with increased confidence/creativity reported comparatively lower disruptiveness and functional impairment.

Conclusion

Findings suggest that data-driven investigations of psychopathological symptoms that include sub-diagnostic threshold conditions, can complement research of clinical diagnoses. Improved classification systems of psychopathology could investigate a weighted approach to symptoms, towards a more dimensional classification of mood disorders. 245/250

Introduction

Background

Mood disorders are common in the general population (1,2) and lead to significant impairment in the individual, as well as direct and indirect costs to society (3). The episodic nature and intra-individual symptom heterogeneity of these conditions can make diagnosis based on subjective symptom reports challenging in early phases of the disorder (4). DSM-5 (5) diagnostic criteria specify that bipolar disorder diagnosis requires a distinct period of abnormally and persistently elevated, euphoric, or irritable mood the presence of a specified number of additional concurrent symptoms and usually some degree of impairment. The additional symptoms in DSM-5 encompass: 1)_inflated self-esteem or grandiosity, 2)_decreased need for sleep, 3)_increased talkativeness, 4)_racing thoughts, 5)_being easily distracted, 6)_increased goal-directed activity or psychomotor agitation, and 7)_engagement in activities that hold the potential for painful consequences (5). Bipolar disorder type I and type II are differentiated by the presence of mania in type I, compared to hypomania (a condition less disruptive to life than mania) in type II.

Data-driven classifications

Epidemiological studies of bipolar spectrum disorders use questionnaires to ascertain symptoms, with various approaches proposed (6–8). In the UK Biobank (9), questions based on DSM-IV criteria were used to assess presence and severity of symptoms (10)(11), and responses can be used to determine potential current or past disease. Whereas both diagnostic and epidemiological classifications reflect common clinical understanding of mood disorders, the use of data-driven approaches could justify or optimise such classifications. Further explorations of mental health definitions could aid epidemiological studies to refine the cases into more homogenous groups for investigation_(12). Precise phenotypes (or disease endotypes) will be instrumental in the shift to precision medicine and patient-specific tailored treatments, based on a more data-centric approach to disease taxonomy, with various frameworks and solutions already proposed (13–16).

Latent class analysis (LCA) is a model-based probabilistic method of identifying homogenous subgroups of individuals (termed “classes”) based on patterns in a set of categorical indicator variables. Previous studies have used this data-driven approach to identify subtypes of disease based on symptoms data. A general-population study of both manic/irritable and psychotic episode symptoms (n=1846) identified five classes differentiated by the presence of irritability and psychotic experiences, as well as differential associations with sociodemographic and clinical characteristics current or past disease. Whereas both (17). Other clustering methods have also been employed to inform data-driven distinctions between mood disorders, such as with longitudinal patterns of mood to identify individuals with bipolar disorder type I (18). Previous studies conducting LCA of symptoms have often lacked replication in external datasets or been performed in small samples.

Aims

In this study, we conducted a data-driven exploratory analysis of latent structure in reported symptoms experienced during manic and/or irritable episodes. Our aims were two-fold: 1) to identify latent classes with homogeneous clinical characteristics and functional outcomes that may have clinical or biological relevance independent of diagnostic categories; 2) to investigate the correspondence of latent classes with reported psychiatric diagnoses and genetic liability to those, in order to aid in refining commonly used epidemiological definitions of probable bipolar disorder.

Methods

Study populations

UK Biobank

Study participants for the discovery analysis were drawn from the UK Biobank. Briefly, the UK Biobank is a prospective cohort study of over 500,000 individuals across the UK. Participants were aged 40-69 years at recruitment in 2006-2010 (9). Genotype data was available for all UK Biobank participants (19). Ethical approval was granted by the NHS North West Research Ethics Committee (REC reference 11/NW/0382). Written informed consent was obtained from all participants. In a follow-up, participants who had agreed to be recontacted were invited to complete an online mental health questionnaire (MHQ) in 2017, resulting in additional phenotypic data in 157,366 UKB participants (11).

Phenotype data

To characterise probable history of mood disorders, UK Biobank worked with experts in mental health epidemiology to devise a self-completed online questionnaire, as clinical interviews would have been unfeasible given scale of the study. Questions were taken from existing tools at the time of the questionnaire’s creation, aiming to maximise international compatibility. Questions on mania/hypomania were adapted from the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I) as described_(10). Participants answered questions on ever having experienced a manic/irritable episode, as described in Box 1.

Box 1. MHQ questions.

Period of manic/hyper mood (field #20501)

“Have you ever had a period of time when you were feeling so good, “high”, “excited”, or “hyper” that other people thought you were not your normal self or you were so “hyper” that you got into trouble?”

Period of irritable mood (field #20502)

“Have you ever had a period of time when you were so irritable that you found yourself shouting at people or starting fights or arguments?”.

Participants that answered positively to either or both of the above questions were subsequently asked if they had experienced any of the eight following symptoms, during these episodes (field #20548), selecting all that might apply

Please try to remember a period when you were in a “high” or “irritable” state and select all of the following that apply:

• I was more talkative than usual,

• I was more restless than usual,

• I needed less sleep than usual,

• My thoughts were racing,

• I was more creative or had more ideas than usual,

• I was easily distracted,

• I was more confident than usual,

• I was more active than usual

Participants who answered positively to above fields were then asked about:

• The longest duration of any such episode (field #20492): brief (<24 hours); moderate (>24 hours but <1 week); or extended (>1 week).

• The disruptiveness of the episode (field #20493): not disruptive or disruptive (if participants reported that the episode required treatment, caused problems with work, relationships, finances, the law or other aspects of life.).

Sociodemographic data on participant sex, age, smoking status, alcohol intake frequency, Townsend deprivation index (TDI, a measure of area-level deprivation as a proxy for socioeconomic status) and education level were extracted from participant responses to the baseline questionnaire (see Supplementary Methods).

In the MHQ, participants reported past diagnoses by a professional (field #20544) of several disorders, which were used to define seven broad diagnostic categories; attention deficit hyperactivity disorder (ADHD), autism spectrum disorder (ASD), generalised anxiety disorder (GAD), depression, schizophrenia/psychosis, mania/bipolar disorder, and personality disorder (see Supplementary Methods). Neuroticism score was derived from responses to the baseline questionnaire(10) (see Supplementary Methods).

Linked electronic health records to Hospital Episode Statistics (HES), which contain hospital diagnoses recorded with the International Classification of Diseases, 10th Revision (ICD-10) up until June 2020, were used to derive cases status for 4 broad disorder definitions; depression, schizophrenia/psychotic disorder, -mania/bipolar disorder and dementia (see Supplementary Methods).

Polygenic risk scores (PRS)

Genetic data pre-processing and sample exclusions are described in Supplementary Methods. We calculated PRS using PRSice v2 (20,21), with clumping (r2< 0.1 and 500kb window) and a p-value threshold of 1 (all SNPs included) for all analyses. PRS were residualized for the first 6 genetic principal components (PCs) and scaled to a mean of zero and standard deviation of 1. Summary results from GWAS of anxiety disorder (23)(22), ADHD (23), ASD (24), major depression (25), bipolar disorder (26) and schizophrenia (27) were used, from studies that did not include UK Biobank (Table S1).

PROTECT – Replication sample

We attempted replication of findings in the Platform for Research Online to Investigate the Genetics and Cognition in Aging (PROTECT) study. Briefly, the PROTECT study is a UK-based online participant registry with continuous, ongoing recruitment beginning in 2015 which tracks the cognitive health of older adults. Study participants must be >50 years old, have no diagnosis of dementia, and must have access to a computer/internet. Beginning in 2015, 14,836 PROTECT study participants were invited to complete the same online MHQ as UK Biobank participants, as a pilot of the questionnaire before roll-out in the UK Biobank. In subsequent PROTECT study enrolment between 2016-2019 21,475 participants in total completed the MHQ. The PROTECT MHQ included the same questions as UK Biobank, on ever experiencing a period of manic and/or irritable mood (see Box 1). Ethical approval was granted by the London Bridge National Research Ethics Committee (13/LO/1578). Written informed consent was obtained from all participants.

Phenotype data

Participant responses to the MHQ questions on ever having experienced a manic and/or irritable episode, along with the corresponding response to symptoms and episode duration/disruptiveness in PROTECT were extracted using the same derivation process as UK Biobank. Sociodemographic variables on sex, age, smoking, education level and alcohol consumption frequency were derived from responses to baseline questionnaires.

Genetic data

A subset of PROTECT study participants provided a saliva sample for genotyping. Genetic data pre-processing and sample exclusions are described in Supplementary Methods. The total number of individuals with genetic data following exclusions was 8,272. PRS were calculated as for UK Biobank, residualized for the first 6 genetic PCs and rescaled.

Statistical analysis

Latent class analysis

Latent class analysis (LCA) is a model-based method that estimates the distribution of an underlying unobserved categorical variable, hypothesised to explain the patterns of association between a set of discrete variables. The estimated categorical variable describes subgroups (termed “classes”) of individuals. The method estimates the posterior probabilities of an individual belonging to a particular latent class. LCA was run using the poLCA package(28) in R, which uses the maximum likelihood method. Models with increasing numbers of classes, beginning at 2 and up to 7, were compared for best fit using the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC). The relative entropy (a measure of classification certainty ranging between 0 and 1) was used to assess separation between classes(29). The eight binary symptom responses in participants reporting a manic and/or irritable episode were used as indicators in the LCA (responses to: more talkative, more restless, less sleep, racing thoughts, more creative, easily distracted, more confident, more active).

Multinomial logistic regression

Multinomial logistic regression was used to test for association between class membership, as the outcome (based on most likely class membership probability) and sociodemographic variables, disorder diagnoses (self-reported or hospital) and polygenic risk scores. Posterior probabilities of class membership were used as weights. Relative risk ratios (RR) were estimated for each class, compared to a reference class (the largest class). For categorical variables (education attainment, smoking and alcohol consumption), dummy coding was used for each level, with the reference level of each being all combined remaining levels.

Results

In the UK Biobank MHQ, 42,183 participants responded positively to the questions on a manic and/or irritable episode and completed the episode symptoms questions (Table S2). Characteristics of this analytical subset and all MHQ respondents are shown in Table 1.

Table 1.

Comparison of sociodemographic characteristics in the subset of participants in the latent class analysis (LCA subset) with responses to the stem question on ever having experienced a period of manic and/or irritable mood as well as the subsequent questions of symptoms, and the participants who completed the Mental Health Questionnaire (MHQ - whole sample) in the UK Biobank.

		LCA subset		Full MHQ sample
Total		N	percent	N	percent
		42183	-	151159	-
Sex	Female	24402	58	85557	57
	Male	17779	42	65597	43
	NA/Missing	2	0.005	5	0.003
Education	University degree	18820	45	68467	45
	A levels NVQ HNC or HND	14411	34	49677	33
	O levels or CSE	5901	14	21169	14
	None	2654	6.3	10424	6.9
	NA/ Missing	397	0.94	1422	0.94
Alcohol consumption	Daily	9245	22	35176	23
	Weekly	20973	50	76917	51
	Occasionally	9371	22	30484	20
	Never	2551	6	8459	5.6
	NA/ Missing	43	0.1	123	0.081
Smoking	Current	2877	6.8	7303	4.8
	Past	17124	41	56483	37
	Never	22076	52	87037	58
	NA/ Missing	106	0.25	336	0.22
		Mean	SD	Mean	SD
TDI*		-1.4	3.0	-1.7	2.8
	NA/ Missing	78		196
Age**		53.9	7.8	55.9	7.7
	NA/ Missing	2		5

^*TDI - Townsend Deprivation Index values before rank normalisation (high values = increased deprivation).

^**at baseline

Latent class analysis

LCA was applied to the eight binary symptoms, as indicators, in the subset of participants reporting a manic and/or irritable episode (N=42,183). As the number of classes increased, BIC and AIC both continuously decreased, with no minimum attained (Table S3). Elbow/scree plots (30) (Figure S1) indicated that either a 4 or 5-class model was the optimum model. Plotting the conditional probabilities for each indicator symptom showed that the additional class in the 5-class model was distinct from the other 4 (Figure S2). We therefore selected the 5-class model as the optimum model.

The conditional probabilities of the eight indicator symptoms in each of the five latent classes are shown in Figure 1A. Individuals in the first class (3.2% of sample) had a high probability of reporting all symptoms and was therefore labelled the ‘extensively affected’ (EA) class. The second class (9.8%) was labelled ‘focused creative’ (FC), as individuals reported being more active, talkative, confident and creative. Individuals in the third class (11.5%), had high probabilities of being more active, talkative, restless, easily distracted and having racing thoughts. This class was labelled the ’active restless’ (AR). Individuals in the fourth class (31.6% of the sample) had a high probability of reporting racing thoughts, feeling more restless and being more easily distracted. This class was labelled the ‘inactive restless’ (IR). The fifth class (43.9%) had low probabilities of reporting all symptoms and was therefore labelled the ‘minimally affected’ (MA) and was used as the reference class in downstream analyses.

Figure 1. Conditional probabilities of each response (symptom) in, (A) the UK Biobank optimum 5-class latent model (N= 42,183), (B) the PROTECT replication study optimum 5-class latent model (N=4,445).

Distributions of responses to the original stem question of ever experiencing a period of manic and/or irritable mood by most likely class membership indicated that the ‘inactive restless’ and the ‘minimally affected’ classes were comprised mostly of individuals reporting an irritable episode. The ‘active restless’ class was comprised of individuals reporting an irritable episode and (to a lesser extent) both a manic and an irritable episode, whereas the ‘focused creative’ class was comprised of individuals reporting an irritable, a manic, or both a manic and an irritable episode. The ‘extensively affected’ class was comprised mostly of individuals reporting both a manic and an irritable episode. (Figure 2(A), Figure S3, Table S4)

Figure 2.

Distributions of responses to (A) the stem question on ever experiencing a period of manic or irritable mood, (B) the subsequent question on episode duration in a subset of N=37,424 participants with available data, and (C) the subsequent question on episode disruptiveness in a subset of N=35,934 participants with available data, by most likely class membership, in the optimum 5-class model. Classes are AR = active restless class, EA = extensively affected class, FC = focused creative class, IR = inactive restless class, MA = minimally affected class.

Associations with episode duration and disruptiveness

For responses to episode duration (N=37,424; brief, moderate or extended duration), individuals in the ‘minimally affected’ class were more likely to report brief duration, while those in the ‘extensively affected’ class mostly reported extended duration (Figure 2(B)). Episode duration patterns did not substantially differ among the remaining three classes. Associations of episode duration with each class when using ‘minimally affected’ as the reference largely reflected the observations from the most likely class membership (Figure S5, Table S5-S6).

Episode disruptiveness (N=35,934) showed a similar pattern to duration, with the highest proportion of reported disruption in the ‘extensively affected’ class (53%) and lowest in the ‘focused creative’ (21%) and ‘minimally affected’ (22%) classes (Figure 2(C) and Table S7). Individuals reporting disruptive episodes were more likely to be in the ‘inactive restless’ and the ‘active restless’ classes, and far more likely to be in the ‘extensively affected’ class (Figure S7, Table S8). Notably, levels of non-response to the questions on episode duration and disruptiveness were high (N=4,759 and N=6,249 respectively, Figures S4, S6).

Associations with sociodemographic characteristics

Associations with sociodemographic characteristics were investigated in a subset of N=41,620 individuals (Tables S9-S17, Figures S8-S16). Being male was associated with an increased risk of being in all other classes when compared to ‘minimally affected’, with a particularly high risk for ‘focused creative’. Higher educational attainment was associated with increased risk of being in the ‘extensively affected’ and the ‘focused creative’ classes. For alcohol intake, individuals in the ‘extensively affected’ and ‘active restless’ classes were less likely to drink alcohol, whereas those in the ‘focused creative’ class were more likely to drink daily. There was an increased risk of current smoking for the ‘extensively affected’ class and a smaller increase for remaining classes. For TDI, there was an increased risk of being in the ‘extensively affected’ class with increasing TDI score (increased deprivation) and smaller but significant increases in risk for the other classes, when compared to ‘minimally affected’.

Associations with self-reported diagnoses of psychiatric disorders

The self-reported diagnoses of six psychiatric disorders differed substantially between the latent classes (N=42,183). Over half of individuals (54.9%) did not report a diagnosis of any of the self-reported disorders studied: ADHD, GAD, ASD, mania/bipolar disorder, depression, and schizophrenia/psychosis. Most individuals that did not report a diagnosis were members of the ‘minimally affected’ (57%) or ‘inactive restless’ (26%) classes. Amongst those that did report one or more diagnoses (Figure S17), a diagnosis of either depression or GAD (or a combination of both) were the most numerous and were mostly in members of either the ‘minimally affected’ or the ‘inactive restless’ classes. Individuals with a diagnosis of mania/bipolar disorder, either alone or in combination with one or more of the remaining disorders, were mostly members of the ‘extensively affected’ class. Diagnosis of any of the six disorders was associated with increased risk of being in the ‘extensively affected’ class (Figure 3(A)), with the highest increases in risk observed for mania/bipolar disorder and schizophrenia/psychosis. Diagnosis of depression and GAD was associated with increased risk of being in the ‘inactive restless’ class, with weaker evidence for increased risk of being in this class for mania/bipolar disorder and schizophrenia/psychosis. Diagnosis of all six disorders was associated with increased risk of being in the ‘focused creative’ and the ‘active restless’ classes, with the strongest associations for each class observed for mania/bipolar disorder (Figures S18, Tables S18-20).

Figure 3.

(A) Associations of self-reported diagnoses of six disorders with most likely class membership, weighted for the probability of inclusion of an individual in that class. Effect estimates are presented as natural log risk ratio (RR) of inclusion in each class (relative to the reference class) for cases of each disorder. (B) Associations of polygenic risk scores (PRS) of six disorders with most likely class membership in a subset of N=33,604 with genetic data, weighted for the probability of inclusion of an individual in that class. Effect estimates are presented as risk ratios (RR) of inclusion in each class (relative to the reference class) per standard deviation (SD) increase in standardised polygenic risk score for each disorder. The “minimally affected” class (MA) is used as the reference (comparison) class in all analyses. Classes are AR = active restless class, EA = extensively affected class, FC = focused creative class, IR = inactive restless class, MA = minimally affected class.

Observed differences between classes when examining ICD-10 diagnoses of depression, mania/bipolar disorder and schizophrenia/psychotic disorder extracted from hospital records (N=36,258) largely corroborated findings of the analysis of self-reported diagnoses. For dementia diagnosis, we found little evidence for differences between classes. However, the number of cases of hospital diagnoses for all four disorders was low (Figures S20-21, Tables S21-23).

Associations of latent classes with self-reported diagnosis of personality disorder (N=42,183) indicated an increased risk of being in all classes when compared to the ‘minimally affected’ class, with particularly large effects observed for the ‘extensively affected’ (Figure S22-23 and Table S24). For derived neuroticism score, a one SD increase in score was associated with an increased risk of being in the ‘extensively affected’ class and with smaller but significant risk increases for all other classes, when compared to the ‘minimally affected’ class (Figure S24 and Table S25). We also explored overlap of each class with cases of probable bipolar disorder type I and II, as defined by Davis et al. and Smith et al. (Supplementary Results Section A, Tables S24-25, Figures S25-26).

Associations with PRS of psychiatric disorders

Polygenic risk scores (PRS) of psychiatric traits discriminated between classes (N=33,604) (Figure 3(B) and Tables S26-S27). PRS of schizophrenia was associated with increased risk of being in the ‘extensively affected’, ‘focused creative’ and ‘active restless’ classes. For bipolar disorder PRS there was an increased risk of being in the ‘extensively affected’ and ‘focused creative’ classes. Depression PRS conferred an increased risk of being in the ‘extensively affected’, ‘focused creative’ and ‘active restless’ classes. Results for ADHD were weaker, with an increased risk of being in the ‘active restless’ and to a lesser degree, ‘inactive restless’ classes. Anxiety and ASD showed no significant increase in observed risk of being in any of the classes. These results contrast with the high proportion of GAD and ASD diagnoses reported by the ‘extensively affected’ classes, but might also reflect lower power of the PRS, for these disorders, compared to the PRS of other disorders.

Replication

Latent class analysis

In the PROTECT replication cohort, there were N=4,445 participants with positive responses to the questions on ever experiencing a manic and/or irritable episode, ~10% of the sample size of UK Biobank. We observed some differences in characteristics between the studies, with a notably higher proportion of females in PROTECT than in UK Biobank (74% vs 58% in the analytical subsets) (Tables S28-S29).

Comparing latent class models with increasing numbers of classes; indicated that a 5-class model was again the optimum model, with an almost identical patterns of condition probabilities for the symptom indicators (Figure 1B, Table S30). The size of some classes was notably different to the discovery cohort (31.6% vs 17% for ‘inactive restless’ and 43.9% vs 56.9% for ‘minimally affected’).

Distributions of responses to the stem question of ever experiencing a period of manic and/or irritable mood were also similar to the discovery results. The ‘inactive restless’ and ‘minimally affected’ classes comprised mostly of individuals reporting an irritable episode, whereas the ‘extensively affected’ class was comprised mostly of both manic and irritable episodes. The ‘focused creative’ and ‘active restless’ classes were more mixed (Figure S27, Table S31).

Associations of latent classes in PROTECT

Similar associations to the discovery analyses were found between episode duration (N=3,706) and episode disruptiveness (N=3,290) with the five latent classes in PROTECT (Figures S28-S31, Table S32-35). Associations with sociodemographic characteristics (N=4,411) suggested similar distinctions between classes to the discovery analyses, although associations were often weaker and of smaller magnitude (Figure S32-39, Table S36-39). For self-reported diagnoses of disorders (N=4421), there were an adequate number of cases (n>20) to analyse four disorders; depression, schizophrenia/psychosis, mania/bipolar disorder and GAD (Figures S40-43). There was increased risk of being in all classes with a diagnosis of depression or GAD that mirrored the associations found in the discovery analysis. A diagnosis of schizophrenia/psychosis or mania/bipolar disorder led to an increased risk of being in the ‘extensively affected’ class in particular (Figures S44-47, Table S40-41). For PRS of six disorders (N=1494), directions of effect were mostly consistent with the discovery cohort but with confidence intervals that overlapped the null, except for an increased risk of being in the ‘extensively affected’ class for bipolar disorder PRS (Figure S48, Table S42).

Discussion

Using a self-reported questionnaire based on diagnostic criteria for bipolar disorder, we have identified latent structure in participants reporting symptoms experienced during periods of manic and/or irritable mood. In both the main discovery cohort and the replication cohort, the participants were assigned into five latent classes. Class membership was associated with episode duration, episode disruptiveness, sociodemographics, diagnoses of psychiatric disorders, and genetic risk of those disorders. These classes likely capture a broad range of disorders, but also sub-diagnostic threshold conditions and non-pathological experiences.

The ‘extensively affected’ class comprises individuals who are the most markedly clinically affected, with particularly high prevalence of diagnoses of bipolar disorder and schizophrenia but also including cases of depression, anxiety, ADHD and ASD. The ‘inactive restless’ class comprises individuals with diagnoses of depression and anxiety, but fewer individuals with diagnoses of schizophrenia, bipolar disorder, ADHD or ASD. The ‘active restless’ class comprises individuals with diagnoses of all disorders, to a lesser extent than the ‘extensively affected’ class. The ‘focused creative’ class comprises individuals with diagnoses of mostly bipolar disorder and schizophrenia, and to a lesser extent than the ‘inactive restless’ class, anxiety, depression. Genetic analyses using PRS corroborate these findings, suggesting the ‘focused creative’ class has higher genetic liability for bipolar disorder and schizophrenia and the ‘inactive restless’ class has a higher genetic liability for depression and ADHD. The ‘minimally affected’ class may comprise individuals reporting normal variations in mood, with episodes of brief duration and low disruptiveness, with no increase in risk of disorder diagnosis or genetic liability to any of the disorders. This ‘minimally affected’ class may comprise individuals that experience symptoms that are not captured using the pre-defined questionnaire). responses. As this was the largest class, our findings underline low specificity of the stem question in capturing clinically relevant periods of manic and/or irritable mood. Likewise, most participants who reported a manic and/or irritable episode but not a mental health disorder were in the ‘minimally affected’ or ‘inactive restless’ classes. The remainder of these individuals were members of the other three classes, indicating either (1) under-diagnosis of mental health disorders, (2) the presence of sub-diagnostic threshold symptoms, or (3) participant misreport of symptoms. Although we found little evidence of differences in dementia diagnosis between classes, mild cognitive impairment as a precursor to dementia diagnosis, may lead to periods of irritable and /or manic mood. Longitudinal collections of cognitive measures in the UK Biobank will enable future investigations of cognitive decline with class membership.

Contrasting dimensions of mood disorder symptomatology were evident between classes. The ‘active restless’ class and the ‘inactive restless’ class included disorganized, unproductive and unfocused characteristics, whereas the ‘focused creative’ class included more creative characteristics, with higher education levels (similarly to the ‘extensively affected’ class) and lower levels of episode disruptiveness. Some psychiatric disorders have been suggested to share genetics with traits such as educational attainment (31) and creativity (32,33). Participants responses to the questions in the MHQ are subjective and some participants may perceive the symptoms they experience during episodes of manic/irritable mood less negatively than an external observer would (34,35)(36). However, this would not explain the more objective characteristic of higher educational attainment observed in the ‘extensively affected’ and ‘focused creative’ classes. Reported creative episodes and higher educational attainment in these two classes may precede onset and diagnosis of bipolar disorder, where the average age of onset for mood disorders is 29-43[IQR: 35-40] years of age (2). Episodes of elevated mood experienced earlier in life may precede later-life bipolar disorder diagnosis and explain the observation. Further investigations into age of disorder onset and age at which episodes were experienced may aid in resolving these questions, with future follow-up questionnaires in UK Biobank extending the range of questions asked. Whereas results support a distinction between a less disruptive subtype of manic and/or irritable mood (the ‘focused creative’ class) and more disruptive subtype(s), e.g. the ‘active restless’ and ‘extensively affected’ classes, these classes cannot be mapped directly to bipolar disorder type I/II definitions. Instead, they suggest that the underlying symptoms can be used to group individuals into more homogenous classes, independently from a diagnosis of bipolar disorder. Future work should aim to further explore whether these homogenous groups can inform the debate on the distinction between type I/II bipolar disorder(37), or feed into new classification systems.

Symptom groupings in the LCA suggested some redundancy between possible responses in the questionnaire. Symptoms did not all contribute equally to class separation; for example, increased confidence and creativity appeared to differentiate the ‘focused creative’ and ‘extensively affected’ classes from the other classes but did not separate out across classes. The five classes suggest that just four responses would suffice to distinguish the classes from each other, with symptoms forming the following groups: (1) increased active/talkative, (2) increased confident/creative, (3) increased restless/thoughts-racing/distracted and (4) less sleep. These results may also inform research for future updates of the diagnostic classification systems. Rather than the current simple summation of number of symptoms present, a weighted approach to diagnostic criteria may be appropriate, constituting a step towards a more dimensional classification of bipolar spectrum disorders. Although the five classes are categorical constructs, the underlying probabilities of individuals belonging to each class are on a continuous scale. The derived classes, as well as the more general latent structure reported amongst symptoms in our results, inform the ongoing development of novel classification systems, aiming to systematically evaluate the hierarchical taxonomy of disorders within psychopathology and to collate and integrate evidence generated across studies to date, such as HiTOP (16)and RDoC (38). Future work could assess the merits of the current LCA approach against the use of continuous measurement instruments for symptom domains beyond manic/irritable episodes in bipolar spectrum disorders.

Our investigations have revealed differences in kind rather than just in degree between classes. Although a spectrum of increasing severity overlays the five classes, with the ‘minimally affected’ having the least severe presentation, we found higher numbers of cases of depression/anxiety in the ‘inactive restless’ class and lower disruptiveness in the focused creative’ class for example. Future work may further explore the effect of increasing psychopathology on class membership, particularly in relation to latent constructs such as the p-factor (general psychopathology factor) (39).

Strengths and limitations

There are a number of strengths to the present study. Firstly, the use of a large, well-characterised cohort, the UK Biobank, ensures that results of this study will inform future mental health research in a well-powered, extensively studied and continually updated research resource. Secondly, the use of a model-based method enabled an agnostic bottom-up approach to defining latent subtypes that mitigates investigator bias of pre-defined criteria and uses the data to inform selection of the number of optimum classes. Finally, the replication of the identified latent classes in an independent dataset, PROTECT, demonstrates robustness and replicability of the findings

There are also several limitations that should be noted. Firstly, the relative entropy of the optimum model in UK Biobank and PROTECT was below 0.7 indicating that classes may not be particularly homogenous, with some “fuzziness” between classes. To account for this, we have weighted associations with the probability of belonging to each class in multinomial regressions. Entropy is usually not considered a model selection criterion and varies depending on the data under study (29). _Secondly, the study is limited by the scope of the questions that UK Biobank participants were asked on manic and/or irritable episodes experienced. Responses were dependent on the selection of multiple-choice answers presented, and it is possible that other questions better characterise participants experiences, ultimately defining classes differently. However, since DSM-5 uses similar symptom reports, the value of additional questions would be of limited clinical relevance at present. Thirdly, given the use of two UK-based volunteer cohorts in restricted age-groups (generally >50 years of age), generalisability beyond these populations is unknown. However, we would not expect age to substantially influence classes, since episode and symptom reports were lifetime retrospective. Finally, conclusions about associations with psychiatric diagnoses are limited by small numbers of individuals with hospital diagnoses, and in the replication dataset low statistical power to fully replicate associations found in the discovery study.

We have used a data-driven approach, with replication in an external sample to derive latent classes differentiated by self-reported symptoms experienced during periods of manic and/or irritable mood that approximate the diagnostic criteria for bipolar disorder. Our findings will inform future studies of mood disorders by guiding self-reported symptom data collection and interpretation, and research aimed at an improved characterisation of bipolar disorder in future classification systems of psychopathology.