Depression subtypes and their response to cognitive behavioral therapy: A latent transition analysis

Depression is a heterogeneous condition, with multiple possible symptom‐profiles leading to the same diagnosis. Descriptive depression subtypes based on observation and theory have so far proven to have limited clinical utility.

only moderately effective. However, an alternative argument is that depression is a highly heterogeneous condition (Goldberg, 2011)-potentially characterized by various subtypes-and clinical outcomes could be improved if treatment was based on more precise assessments of each individual's symptom profile (Fried, 2017).
It is widely acknowledged that current diagnostic systems often fail to capture the underlying heterogeneity within a particular diagnostic label (Cuthbert & Insel, 2013). This is a particular issue for depression, where-in theory-over 100 different combinations of symptoms can result in the same unitary diagnosis of depression (Zimmerman et al., 2015). Research investigating the effectiveness of different treatments for different depression subtypes has offered mixed results, limiting the use of classification systems for determining the most appropriate treatment for a particular patient (Arnow et al., 2015;Uher et al., 2011). In particular, subtyping concepts that are derived from clinical observation and theory often have little empirical support or prognostic utility (Haslam & Beck, 1994).
In recent years, data-driven approaches, such as latent class analysis (LCA) that aim to define depression subtypes based on itemized scores in standardized questionnaires, without imposing any theoretical constructs on the statistical model a priori (e.g., Putnam et al., 2015;Ulbricht et al., 2015Ulbricht et al., , 2018 have provided valuable insights into different symptom profiles. However, LCA has been limited to rigid clustering of patients into static classes, providing no indication of their temporal stability, or how different subtypes may respond to treatment. Latent transition analysis (LTA) is an extension of LCA which uses longitudinal data to explore transitions between classes over time (Ni et al., 2017;Ulbricht et al., 2016). This technique is better suited to examine how patients with different depression subtypes respond to treatment, which is potentially informative for personalized treatment planning.
In a recent demonstration, Catarino et al. (2020) applied LTA in a large (N = 9,891) sample of patients who accessed internet-enabled cognitive behavioral therapy (CBT) for depression. The results classed patients into seven distinctive depressive states, loading preferentially on cognitive/affective versus somatic symptoms. Transition probability analysis revealed that patients starting in cognitive/affective states typically do not transition to somatic states, and vice-versa.
Although the distinction between cognitive/affective and somatic symptoms in depression is well supported by the literature (Barton et al., 2017;Carragher et al., 2009;Lee et al., 2012Lee et al., , 2014, this was the first study exploring how patients in these two distinct states respond to a highly standardized evidence-based psychological treatment. Importantly, Catarino et al. (2020) showed that patients who are classified into somatic states were less likely to improve, and more likely to be female, suffering from a long-term physical illness, and taking antidepressants (Catarino et al., 2020). Notwithstanding its clinically informative results, this study was based on a sample of patients who opted-in to receive CBT via text messages. It is unclear if patients who find this treatment modality acceptable may be systematically different to typical clinical samples who receive more traditional in-person therapy in healthcare settings. Thus, the generalizability of these findings to typical clinical samples is unknown.

| Data sources and sample selection
All participating NHS Trusts offered psychological care as part of the Improving Access to Psychological Therapies (IAPT) program. IAPT services deliver evidence-based psychological interventions for common mental health disorders within a stepped-care model (Clark, 2011). In this system, most patients are initially referred for brief (≤8 sessions), low-intensity guided self-help interventions.
Those who remain symptomatic, or who initially present with more severe and complicated problems, are referred for high intensity psychotherapies (up to 20 sessions), including CBT, person-centered experiential counseling or interpersonal psychotherapy as re-   Catarino et al. (2020), was that the present sample received traditional in-person CBT rather than internet-enabled CBT. The selected study sample consisted of 8380 patients (see Table 1 for sample characteristics). Study flowchart with reasons for exclusion is reported in the Supplemental Material.

| Measures
Patients accessing IAPT services complete standardized questionnaires at the start of each session, as part of routine outcome monitoring. The Patient Health Questionnaire-9 (PHQ-9) is a nineitem questionnaire assessing symptoms of major depression, including affective (items 1, 2), cognitive (items 6, 7, 9) and somatic domains (items 3, 4, 5, 8). Each item is rated on a 0-3 scale (0 = "not at all," 3 = "nearly every day"), with a total score between 0 and 27 (Kroenke et al., 2001). A cut-off point of ≥10 has been recommended to detect clinically significant depression symptoms (Kroenke et al., 2001). The GAD-7 is a seven-item questionnaire assessing symptoms of generalized anxiety. Each item is also rated from 0 to 3, producing a total score between 0 and 21 (Spitzer et al., 2006). A cut-off point of ≥8 has been recommended to identify clinically significant anxiety symptoms (Kroenke et al., 2007).
Additional anonymized demographic and clinical data included age, gender, ethnicity, employment status, medication status, presence of a long-term physical condition, and baseline impaired functioning severity measured using the Work and Social Adjustment Scale (WSAS; Mundt et al., 2002).

| Outcomes of interest
To allow comparison with the study by Catarino et al. (2020), treatment outcomes were assessed according to the same criteria for treatment engagement and reliable change. An additional dropout outcome was also assessed in the current study. Patients were deemed treatment engagers if they attended two or more treatment sessions. Consistent with outcome definitions used routinely by IAPT services to examine reliable change (National Collaborating Center for Mental Health, 2018), reliable improvement was present when at least one of the two primary outcome measures showed a statistically reliable reduction in scores (≥6 points on the PHQ-9 and/or ≥4 points on the GAD-7), in the absence of a reliable increase in the other measure (Gyani et al., 2013). Reliable deterioration was recorded when at least one of the measures showed a reliable increase in scores (≥6 points on the PHQ-9 and/or ≥4 points on the GAD-7).
Finally, patients were classed by therapists as having dropped out if they unilaterally discontinued treatment and had an unplanned ending to their episode of care.

| Latent Markov modeling procedure
Analyses were conducted in R (version 3.6.1) using the package LMest (Bartolucci et al., 2017). Following the procedure reported in Catarino et al. (2020), a Hidden Markov Model (HMM) was applied to the longitudinal item-level PHQ-9 data from the entire sample (N = 8380) to estimate latent depressive subtypes and the corresponding state-to-state transition probabilities. Unlike Catarino et al. (2020), we decided to interpret the model that had best empirical support (smallest Bayesian Information Criterion value), rather than to identify a reduced state model that aimed to balance fit and interpretability. This was to ensure complete objectivity in model selection and to explore differences and/or similarities between states and transitions with greater granularity. After determining the optimal model, global decoding was performed to approximate a depressive state for each patient at every time point. Transition probabilities were extracted and plotted to explore between-state transitions in response to treatment.

| Post-hoc analyses of state groupings
Qualitatively similar states were grouped into a smaller set of overarching depression subtypes. Patients' starting states were compared using post-hoc analyses in relation to treatment engagement, duration and posttreatment clinical outcomes. These analyses were based on chi-square (for binary outcomes) and analysis of variance (for continuous outcomes). In addition, we applied logistic regressions to investigate which clinical and demographic features were associated with patients' starting states. In the interest of parsimony and statistical power, post hoc comparisons were performed between the three subtypes that encompassed most of the sample (>85%). Abbreviations: GAD-7, Generalized Anxiety Disorder-7; LTC, long-term health condition or illness; PHQ-9, Patient Health Questionnaire-9; WSAS, Work and Social Adjustment Scale.

| Optimal depressive states model
The HMM analysis produced an optimal model with 14 separate depressive states, each exhibiting different symptom-profiles and levels of severity. Figure 1 displays the overall mean score and a plot of the corresponding intensity of each PHQ-9 item for all 14 states.
State 1 displayed a profile of minimal symptoms with very low severity. States 2, 4, 6, and 10 showed similar symptom-profiles at differing overall severities, with peak intensity on PHQ-9 items 3, 4, and 5, representing somatic symptoms. We therefore considered these to belong to an overarching somatic depression subtype. States 3, 5, and 11 also displayed similar symptom-profiles with different levels of severity, but with patterns of peak intensity on items 1, 2, and 6, representing cognitive-affective symptoms. We considered these to belong to an overarching cognitive-affective depression subtype. States 7, 8, 9, 12 and 13 all showed relatively even intensity across most items (1-8) for different levels of severity; we considered these to belong to an overarching typical depression subtype.
However, there was a further distinction between the symptomprofiles of states 8 and 13 and states 9 and 12, with the latter two states showing greater intensity on the suicidal risk item (item 9). We therefore draw a further distinction between low-risk typical (8,13) and high-risk typical (9, 12) subtypes (7 being    The most severely depressed patients appeared to be most likely to transition into the two typical states, either low-risk typical (states 8 and 13) or high-risk typical (states 9 and 12). However, there were rarely any crossover transitions between these two states, with patients merging into state 7 when symptoms were at a subthreshold severity. Similar trends were seen within the small number of cases who experienced deterioration in their symptoms, with transitions to higher severity states within the same symptom cluster (e.g., state 3-5; state 8-13; state 9-12). Interestingly there were no prominent deterioration transitions within the somatic states.
3.2 | Relationship between clinical outcomes, demographics, and depressive states and cognitive-affective and somatic states (p = .056) were not significant. Of those who engaged in treatment (n = 6364), patients in somatic and typical states were found to have significantly lower rates of reliable improvement than those in cognitive-affective states, but did not differ from each other (χ 2 = 13.764, p = .001).
Rates of deterioration did not differ between the three depression subtypes (χ 2 = 1.121, p = .571). Significantly more patients in the typical states dropped out of treatment compared to the cognitiveaffective and somatic states (χ 2 = 15.047, p = .001).
Six demographic variables were significantly associated with starting treatment in a somatic state (relative to a cognitive-affective state) (χ 2 = 239.57, p < .001, Nagelkerke R 2 = 0.127). Patients with somatic subtypes were more likely to be females from a white British background, with a long-term physical illness, additionally having pharmacotherapy, and who had lower baseline depression and anxiety severity (Table 2).
Analyses comparing patients entering treatment in typical states with differing risk profiles (low vs. high-risk) and for states with similar elevated risk scores, but differing state profiles (high-risk typical vs. high-risk cognitive-affective) are reported in the Supplemental Material and indicate that high-risk typical subtypes had lower dropout rates compared to low-risk typical and high-risk cognitive-affective subtypes. overarching subtype. We also found that state transitions denoting a deterioration of symptom-severity occurred predominantly within the cognitive-affective and typical subtypes. Some cases transitioned to other subtypes, but the probability tended to be low (<20%).
Post-hoc comparisons of indicators of treatment engagement and outcomes revealed that these depression subtypes do not merely have descriptive value, but also have prognostic utility. As shown in Table 3, symptom-states characterized by moderate-tosevere levels of depression (8,9,10,11) showed the highest rates of reliable improvement (>60%). States 1 and 2 had the lowest improvement rates (<31%), though these had the lowest baseline severity, so this is likely to be a floor effect (i.e., little room for improvement). Deterioration rates were generally low (<10%) across all states, with the highest probability (9.6%) observed for state 7 and the lowest (0.5%) for state 14-although a ceiling effect is likely for the latter. The four states with highest depression severity (11-14) also had the highest dropout rates (>35%). Comparisons between overarching depression subtypes revealed that patients with a cognitive-affective subtype were more likely to engage, attended more sessions, and attain reliable improvement compared to the typical and somatic subtypes. Patients with a typical subtype were more likely to drop out of treatment compared to those with cognitive-affective and somatic subtypes. Furthermore, cases in the high-risk typical subtype had lower dropout rates compared to the low-risk typical subtype and the cognitive-affective subtype cases with comparably acute suicidal risk.
F I G U R E 3 Within and between-state transition probabilities in response to CBT; the range of probabilities over the course of treatment are reported for each transition; p < .05 at more than half the time points are hidden; thickness of the arrow depicts the magnitude of the probability, with thicker arrows representing more likely transitions; states are color coded accordingly-green = somatic, pale orange = cognitive-affective, blue = typical, pale blue = typical with low-risk and dark blue = typical with high-risk. CBT, cognitive behavioral treatment

| Wider empirical context
Previous empirical studies using the PHQ-9 have proposed that these symptoms load onto distinctive cognitive-affective and somatic factors, particularly in studies that included participants with comorbid physical illnesses/chronic health problems (e.g., Chilcot et al., 2013;Doi et al., 2018;Krause et al., 2010). Approximately 29% of our sample reported comorbid long-term health problems, and this was one of the significant features that characterized the somatic depression subtype. However, studies specifically seeking to identify depressive subtypes using data-driven LCA have thus far yielded mixed and inconclusive results (see review by Ulbricht et al., 2018).
Methodological advances, such as LTA have rarely been applied in the wider field of mental health (e.g., McElroy et al., 2017;Rodgers et al., 2014;Ulbricht et al., 2016), and less so in the specific field of CBT for depression. Like other studies (e.g., Rodgers et al., 2014) we found that gender played a role in determining class membership. In the present sample, females were more likely to have the somatic depression subtype whereas males were more likely to have the cognitive-affective subtype. Within the majority class of typical depression, males were more likely to have acute suicidal risk. Of those that exhibited high-levels of suicidal risk, females were more likely to have a cognitive-affective subtype whereas males were more likely to have a typical subtype.
To our knowledge, the study by Catarino et al. (2020) was the first application of LTA in a large clinical sample of depressed patients accessing internet-enabled CBT, albeit in an atypical format.
Our findings show remarkable similarities. Catarino et al. (2020) also found distinctive states that clustered around three overarching subtypes characterized by cognitive-affective, somatic and "hybrid"

| Strengths, limitations, and future directions
This study used a large and adequately powered sample, covering multiple regions of England, thus enhancing the external validity and generalizability of findings to a typical population of patients accessing CBT in routine care. Relative to other cluster methods, LTA is a state-of-the-art method that makes best use of time-series data simultaneously enabling the modeling of symptom-states, their temporal stability and transition probabilities to other states (Bartolucci et al., 2017). In addition, our sample selection, modeling strategy, outcome measures and outcome definitions were closely aligned to prior research (Catarino et al., 2020), thus enabling direct comparability. Note: A positive relationship signifies that the variable is more likely to occur in patients entering treatment in a cognitive-affective state. Continuous variables were mean centered. Reference categories for categorical variables: gender "female," ethnicity "minority," employment "employed," medication "not prescribed pharmacotherapy," self-report LTC "no LTC." Abbreviations: GAD-7, Generalized Anxiety Disorder-7; LTC, long-term health condition or illness; PHQ-9, Patient Health Questionnaire-9; WSAS, Work and Social Adjustment Scale.

| 913
Despite these strengths, these findings should be interpreted in light of some limitations. The depression subtypes identified in this study are inevitably constrained by the limited number and specific type of items contained in the PHQ-9. Other studies using LTA on different measures have reported different concepts and symptompatterns, for example an atypical depression subtype associated with eating disorder and psychotic symptoms (e.g., Rodgers et al., 2014).
Outcomes were defined using patient-reported symptom measures, and no formal diagnoses or observer-rated outcomes were available.
Furthermore, outcomes could only be defined based on end-oftreatment measures, so the longer-term prognosis of patients with different subtypes remains unclear.
Despite these caveats, the replicated evidence from this and a previous study (Catarino et al., 2020)

| CONCLUSION
Replicated evidence indicates that depression is a heterogeneous condition characterized by several subtypes which are stable over time, which are more likely to change in severity but less likely to transition into other subtypes, and which show differential treatment engagement and response patterns.
T A B L E 3 Treatment engagement and clinical outcomes by starting state (1-14) and overarching depression subtypes Abbreviation: PHQ-9, Patient Health Questionnaire for major depressive disorder. a Percentage within subset of treatment engagers.