Identifying subgroups of urge suppression in Obsessive-Compulsive Disorder using machine learning

Abstract

Obsessive-compulsive disorder (OCD) is phenomenologically heterogeneous. While predominant models suggest fear and harm prevention drive compulsions, many patients also experience uncomfortable sensory-based urges (“sensory phenomena”) that may be associated with heightened interoceptive sensitivity. Using an urge-to-blink eyeblink suppression paradigm to model sensory-based urges, we previously found that OCD patients as a group had more eyeblink suppression failures and greater activation of sensorimotor-interoceptive regions than controls. However, conventional approaches assuming OCD homogeneity may obscure important within-group variability, impeding precision treatment development. This study investigated the heterogeneity of urge suppression failure in OCD and examined relationships with clinical characteristics and neural activation.

Eighty-two patients with OCD and 38 controls underwent an fMRI task presenting 60-second blocks of eyeblink suppression alternating with free-blinking blocks. Latent profile analysis identified OCD subgroups based on number of erroneous blinks during suppression. Subgroups were compared on behavior, clinical characteristics, and brain activation during task.

Three patient subgroups were identified. Despite similar overall OCD severity, the subgroup with the most erroneous eyeblinks had the highest sensory phenomena severity, interoceptive sensitivity, and subjective urge intensity. Compared to other subgroups, this subgroup exhibited more neural activity in somatosensory and interoceptive regions during the early phase (first 30 seconds) of blink suppression and reduced activity in the middle frontal gyrus during the late phase (second 30 seconds) as the suppression period elapsed.

Heterogeneity of urge suppression in OCD was associated with clinical characteristics and brain function. Our results reveal potential treatment targets that could inform personalized medicine.

Introduction

OCD is characterized by recurrent thoughts, impulses, or images that cause distress and anxiety (obsessions) and repetitive behaviors or mental acts performed to relieve this distress (compulsions) (American Psychiatric Association, 2013). The symptomatology of OCD is highly heterogeneous. Although traditional cognitive-behavioral models for OCD generally posit that harm and fear prevention drive compulsions (Abramowitz et al., 2009; Frost and Steketee, 2002), up to 65% of individuals with OCD experience sensory-based symptoms (i.e., ‘sensory phenomena’) preceding or driving compulsions, regardless of a concrete fear (Ferrão et al., 2012; Shavitt et al., 2014). Sensory phenomena in OCD include uncomfortable physical sensations (for example, feeling one’s hands are contaminated), perception of symmetry, incompleteness or “not-just-right” experiences (for example, the need to repeat a behavior until achieving a ‘just right’ feeling), energy buildup, and sensory urges to perform compulsions. Out of the individuals experiencing sensory phenomena, approximately 35% experience sensory-based physical urges (Ferrão et al., 2012; Lee et al., 2009). Although first-line psychotherapeutic and pharmacological treatments for OCD are effective for many, up to 50% fail to respond adequately (Davidson and Bjorgvinsson, 2003; Simpson et al., 2006; Springer et al., 2018), possibly due to the heterogeneity of OCD and the resistance of sensory phenomena to standard treatments (Summerfeldt, 2004; Summerfeldt et al., 2014). Indeed, some research suggests that sensory phenomena in OCD do not respond as well to cognitive behavioral therapy approaches as other types of OCD symptoms, such as those related to fear of harm or negative outcomes, particularly if the treatment is not tailored to target sensory phenomena (Schwartz, 2018).

Previous research have identified similarities between sensory-based urges in OCD and premonitory urges in Tourette’s disorder (TD) (Brandt et al., 2018; Cavanna et al., 2017; Miguel et al., 2000). Premonitory urges preceding tics are phenomenologically and neurobiologically similar to everyday urges-for-action (UFA) (Brandt et al., 2018; Cavanna et al., 2017; Miguel et al., 2000), such as the urge to scratch, cough, and blink (Berman et al., 2012; Botteron et al., 2019; Jackson et al., 2011a; Mazzone et al., 2010; Mazzone et al., 2013). UFA, like premonitory urges, intensify the longer they are suppressed and temporarily subside when acted upon (Berman et al., 2012; Botteron et al., 2019; Leech et al., 2013). In healthy samples, UFA activate a network of brain regions involved in interoceptive and sensorimotor-related processes, including the insula, mid-cingulate cortex, and supramarginal gyri (Jackson et al., 2011a). These brain regions are also related to premonitory urges in TD (Bohlhalter et al., 2006; Neuner et al., 2014) and sensory phenomena in OCD (Brown et al., 2019; Subirà et al., 2015), suggesting similar mechanisms between pathological urges in OCD and more normative urges-for-action as indexed by the urge-to-blink task. The use of the urge-to-blink task allows for the experimental quantification of urge-related behavior and brain function in OCD patients without relying on self-report or evoking idiosyncratic OCD symptoms.

While several studies have used the urge-to-blink elicited with an eyeblink suppression task as a model for sensory-based urges in healthy samples and TD (Berman et al., 2012; Botteron et al., 2019; Mazzone et al., 2010; Stern et al., 2020), research in OCD has been limited. Using an eyeblink suppression paradigm, we recently showed that OCD patients as a group were less successful than controls in suppressing their eyeblinks for 60 seconds, i.e., patients made more erroneous blinks when instructed to withhold blinking for an extended duration (Bragdon et al., 2023; Stern et al., 2020). Additionally, patients were more likely to experience their first blink (i.e., initial suppression failure) earlier, and experience more subsequent erroneous blinks, than controls (Bragdon et al., 2023). We further identified an “urge network” comprising the insula, cingulate, prefrontal cortices, occipital, primary somatosensory cortex (postcentral gyrus), and secondary somatosensory areas (parietal and central operculum and inferior parietal lobule including supramarginal gyrus) in which activity increased as the blink suppression period elapsed, potentially reflecting the neural instantiation of an intensifying urge sensation over time along with the increasing effort required to continually suppress the urge (Stern et al., 2020). Patients with OCD showed greater activation in several regions of this putative urge network specifically during the early phase of suppression (first 30 seconds) than controls, which may reflect the neurocircuit mechanisms underlying a heightened urge-for-action in OCD (Stern et al., 2020).

Mechanistically, the experience of an urge-for-action may involve interoception, which is the process of sensing, interpreting, and integrating signals from within the body (Khalsa et al., 2018; Tsakiris and Critchley, 2016). Patients with OCD reported heightened subjective sensitivity to body sensation (interoceptive sensitivity; IS) than controls, and greater IS was associated with more severe sensory phenomena (Eng et al., 2020). Beyond OCD, IS and the ability to accurately perceive body sensations were related to premonitory sensations and urges in patients with TD (Ganos et al., 2015; Rae et al., 2019). These findings, along with the observation that interoception, sensory phenomena, and urges-for-action show overlapping neurocircuitry in the insula and sensorimotor regions (Brown et al., 2019; Stern et al., 2020), suggest that IS may be related to sensory-based urges in OCD (Bragdon et al., 2021).

Despite recent work providing an initial model for understanding behavioral and neural mechanisms of pathological urges in OCD (Bragdon et al., 2023; Stern et al., 2020), little is known about the heterogeneity of urge suppression amongst patients with the disorder, as prior work analyzed the OCD sample as a homogeneous sample. Machine learning approaches, such as latent variable mixture modeling, are increasingly used to identify clinically distinct profiles among clinical samples (De Nadai et al., 2023; Hutchins et al., 2022). These data-driven approaches identify patterns in the data and elucidate distinct subgroups, thereby providing the opportunity to develop targeted treatments tailored to specific groups. Given the heterogeneity of symptoms in OCD, it is possible that only a subgroup of patients with OCD shows alterations of urge suppression and associated neural mechanisms. This is supported by the fact that approximately 20% of individuals with OCD experience sensory urges (Ferrão et al., 2012; Lee et al., 2009), indicating heterogeneity in the phenomenology. The current study is novel and aims to utilize latent profile analysis, a form of unsupervised machine learning, to uncover subgroups of patients based on their ability to suppress urges, operationalized as performance in an eyeblink suppression task. We also evaluated associations between urge behavior, and phenotypic characteristics and brain activation. We hypothesized that there would be different subgroups of patients based on urge behavior. The acquisition of task-related functional magnetic resonance imaging (fMRI) data as well as behavioral and clinical assessments enabled a detailed comparison of phenotypic characteristics and neural activation between patient subgroups. Findings from this investigation will refine our understanding of behavioral and neural heterogeneity in OCD, potentially contributing to the novel development of personalized medicine targeting individual patient phenotypes.

Materials and methods

Participants and Procedure

Participants were recruited and scanned at Icahn School of Medicine at Mount Sinai (ISMMS), Nathan Kline Institute for Psychiatric Research (NKI), and New York University Grossman School of Medicine (NYUSoM). Final data were analyzed on 82 patients with OCD and 38 controls, all of whom had valid neuroimaging data. The study protocol was approved by the Institutional Review Boards at each institution and all subjects provided written informed consent (Supplement 1.1–1.2).

Clinical interviews by a trained rater assessed current DSM-5 diagnoses (Mini International Neuropsychiatric Interview (Sheehan et al., 1998)), overall OCD severity (Yale-Brown Obsessive Compulsive Scale; Y-BOCS (Goodman et al., 1989)), and sensory phenomena (Sensory Phenomena Scale; SPS (Rosario et al., 2009; Sampaio et al., 2014)). Self-report measures assessed IS (Noticing subscale of the Multidimensional Scale of Interoceptive Awareness; MAIA (Mehling et al., 2012)), OCD symptom dimensions (Dimensional Obsessive-Compulsive Scale; DOCS (Abramowitz et al., 2010) and Obsessive-Compulsive Inventory-Revised; OCI-R (Foa et al., 2002)), state anxiety (Beck Anxiety Inventory; BAI (Beck et al., 1988)) and depressive symptoms (Quick Inventory of Depressive Symptomatology; QIDS) (Rush et al., 2003)) (Supplement 1.3). Information regarding comorbid conditions and the use of psychotropic medication in the sample is available in Supplement 1.4–1.5.

UFA Task

The rationale for selecting this task is based on prior work from our group and others using eyeblink suppression as a model for sensory-based urges in OCD and TD (Botteron et al., 2019; Jackson et al., 2011a; Stern et al., 2020). A key advantage to this task lies in the normative behavior it elicits, which allows control participants to perform the task even when they do not have pathological urges unlike some individuals with OCD. This task elicits urges by asking participants to suppress eyeblinks for prolonged periods of time (also see (Stern et al., 2020)) (Figure 1). Eight blocks of alternating 60-second blocks of blink suppression and 30-second blocks of free-blinking were presented over two runs while fMRI data was acquired. Following each suppression block, participants rated the subjective intensity of the urge experienced during the prior suppression block on a 5-point scale (Figure 1). Eyeblinks were measured via pupil occlusion using Eyelink 1000-Plus (SR Research, 2016). Participants were instructed to blink normally during free-blinking, withhold blinking during suppression, and resume withholding eyeblinks should they accidentally blinked during suppression (Figure 1). Inter-stimulus and inter-trial intervals were jittered between 2 to 5 seconds. Notably, the alternating design of the task, combined with jittering, increases experimental rigor for detecting brain activations associated with the task conditions.

Figure 1: Stimulus flow of the blink suppression task.

This study utilized an urge-to-blink paradigm as a model for examining sensory-based urges in patients with OCD. Participants were presented with alternating 60-second blocks of blink suppression (“HOLD”), and 30-second blocks of free-blinking (“NORMAL”). Following suppression blocks, participants were permitted to blink (“OK TO BLINK”; blink-recovery, 4 seconds) and then rated the subjective strength or intensity of the urge experienced during the prior suppression block on a 5-point scale (1 = “Not strong at all / no urge” to 5 = “Extremely strong”; ~4 seconds). Inter-stimulus and inter-trial intervals are jittered from 2 to 5 seconds, with any leftover time from the rating scale (if the rating is made before the full 4 seconds have elapsed) added to the inter-trial interval. Participants were instructed to blink normally during free-blinking, withhold blinking during suppression, and resume withholding eyeblinks should they accidentally blinked during suppression.

Following completion of the task, participants completed a post-task questionnaire rating the overall strength of the urge experienced during suppression blocks and the difficulty experienced in refraining from blinking after making an initial erroneous blink during suppression.

Neuroimaging data acquisition

Participants were scanned on a Siemens 3T MAGNETOM Skyra (ISSMS) or 3T MAGNETOM TrioTim (NKI/NYU) (Supplement 1.6).

Data analysis

To reflect the time-varying nature of an urge (Berman et al., 2012; Botteron et al., 2019; Mazzone et al., 2010), each 60-second suppression block was analyzed in separate early (first 30 seconds) and late (second 30 seconds) phases. This division was performed to investigate whether subgroup effects differed between early and late phases of suppression, following previous findings of behavioral (Bragdon et al., 2023) and neural (Stern et al., 2020) differences across the phases. Erroneous blinks committed during each early and late suppression phase were summed across all eight suppression blocks as the primary measures of urge suppression failure. In-scanner ratings of urge intensity in each suppression block were averaged across all blocks and compared between subgroups.

Identifying OCD subgroups

Latent profile analysis was performed using Mplus (version 8) (Muthén and Muthén, 2017) to identify latent groups (termed “subgroups”) of patients with OCD based on count data of erroneous blinks committed during early and late phases of blink suppression. Latent profile analysis is a form of unsupervised machine learning that utilizes finite mixture modeling to reveal multiple underlying distributions to explain the observed pattern of response. Due to the presence of meaningful zero-value data points (zero erroneous blink count during suppression indicates participants were able to successfully withhold blinking), count data of erroneous blinks during early and late suppression were modeled using a negative binomial model distribution. Model parameters were estimated using maximum likelihood estimation with robust standard errors. Because the primary goal was to identify distinct profiles within the patient group, controls were assigned into a single subgroup separated from patients, and patients were freely assigned to model-estimated groups based on estimated probabilities of latent group membership. Critically, this approach allowed data from control participants to contribute to overall variance estimation and for direct comparison between patients and controls without contaminating group formation results for the patients. Next, we adopted an iterative process to fit the data. We ran two-profile (one patient group, one control group), three-profile (two patient groups, one control group) and four-profile (three patient groups, one control group) models. To determine the optimal model, we utilized the Bayesian Information Criterion (BIC) and qualitative interpretation. BIC is a model fit statistic that compares models, taking into account of the number of parameters and goodness of fit. A change in BIC (ΔBIC) greater than 10 indicates strong evidence in favor of the model with lower BIC value (Kass and Raftery, 1995; Raftery, 1995). This criterion is consistent with previous studies (Curci et al., 2022; Schneider et al., 2018). However, as the BIC value alone may not have strong clinical relevance, we also used qualitative interpretability of the identified subgroups as an additional model selection criterion, in line with contemporary recommendations (Weller et al., 2020).

Clinical and behavioral characteristics

Planned primary behavioral analyses focused on differences between subgroups of patients with OCD. Key clinical and phenotype characteristics were probed, including overall OCD symptom severity (Y-BOCS), sensory phenomena severity (SPS and Symmetry/NJRE subscale of the DOCS) and IS (Noticing subscale of the MAIA). Behavioral outcomes associated with the UFA task, including ratings of urge intensity (in-scanner) and difficulty refraining from repeated blinks (post-task) were compared. Shapiro-Wilk tests revealed that only the SPS and DOCS symmetry/NJRE subscale had normal distribution in all subgroups; independent-samples t-tests were used to compare these measures between subgroups. The distribution of all other measurements significantly differed from normal in at least one subgroup, P<0.05; Dunn’s tests were used to compare subgroups for these variables.

Effect sizes using Cohen’s d as a parametric measure of effect size and delta median absolute deviation (Δ_MAD) (Ricca and Blaine, 2022) as a non-parametric measure of effect size were presented for subgroup comparisons significant at P<0.05 (Sullivan and Feinn, 2012). We assumed discrete subgroup assignment for effect size computations, given the high entropy in the latent profile analysis (Table S1), reflecting strong confidence in subgroup assignment.

Secondary analyses correlated in-scanner ratings of urge intensity with clinical characteristics within the full OCD sample. Additionally, mediation analysis evaluated IS as a statistical mediator between sensory phenomena and OCD subgroup membership (Figure S1), as motivated by prior work reporting associations of IS with both sensory phenomena and premonitory urges (Eng et al., 2020; Ganos et al., 2015; Rae et al., 2019). Mediation analysis used maximum likelihood estimation in MPlus and 95% confidence intervals for indirect effects were generated using 10,000 bootstrap samples (Hayes, 2009; Muthén et al., 2017; Zhao et al., 2010). Notably, the mediation analysis was performed on cross-sectional data, preventing our ability to ascertain causal relationships (Winer et al., 2016).

Neuroimaging data preprocessing and analyses

Preprocessing was performed using a combination of Statistical Parametric Mapping (SPM) v.12, AFNI v.10.6, and FSL v.5.0.10 (Supplement 2.1).

Following preprocessing, a fixed-effects general linear model was created at the individual subject level. Early (first 30 seconds of 60-second suppression period) and late (second 30 seconds of suppression period) phases of suppression were modeled separately with block regressors to differentiate brain activation based on the build-up of the urge over time (Stern et al., 2020). Group level analyses used random-effects models to compare brain activation between subgroups for each phase of suppression separately using two-sample t-tests with scan site as a covariate-of-no-interest (Forsyth et al., 2014; Glover et al., 2012; McNeish and Kelley, 2019). fMRI results were corrected for multiple whole-brain comparisons using AFNI’s 3dFWHMx and 3dClustSim (with -acf option) with NN1 bi-sided thresholding through 10,000 Monte Carlo simulations to achieve a familywise error rate (FWER) of P<0.05 with cluster-defining threshold of P<0.005 (Cox et al., 2017). This cluster estimation procedure and threshold are consistent with approaches used in prior studies (Hodgdon et al., 2021; Yang et al., 2021) (Supplement 2.1).

Brain clusters resulting from subgroup comparisons were labelled using the Harvard-Oxford (H-O) and Automated Anatomical Labelling (AAL) atlases (respectively provided through the CONN Toolbox v.17f (Whitfield-Gabrieli and Nieto-Castanon, 2012) and xjview v.10.0 (Cui et al.)) to aid in interpretation.

Supplemental analyses

Additional analyses that support the present approach and assist with interpretability are detailed in the Supplement. These include an examination of latent model selection and alternatives (Supplement 2.2, Tables S1–S2), subgroup comparisons for clinical characteristics not planned a priori (Supplement 2.3, Table S3), post-task urge ratings collected from outside the scanner (Supplement 2.4), comorbid disorders and psychotropic medication use within subgroups (Supplement 2.5), brain activity in control subjects (Supplement 2.6.1–2.6.2, Figure S2), comparison of brain activity between early (first 30 seconds) and late (second 30 seconds) phases as the suppression period elapsed (Supplement 2.6.3, Figure S3), and whether subgroup differences in brain activity were related to a higher rate of blinking (Supplement 2.7, Figure S4). A summary of the analyses and findings described in the Supplement are presented in Table 1.

Table 1.

Summary of Supplementary Analyses and Main Findings

Section	Analysis	Main Findings	Relevant tables / figures
Supplement Section 2.2 Examination of latent model selection and alternatives	2.2.1 Comparison of two competing latent group models - a three-latent group model and a four-latent group model, to determine the model that best fit the data. 2.2.2 Conducted two separate latent profile analyses - one for blink counts during early suppression and one for blink counts during late suppression and examined classification diagnostics.	2.2.1 The four-latent group model (3 patient groups, 1 control group) was the best final model based on qualitative interpretability of the latent class solution and classification diagnostics (entropy and average latent class posterior probabilities). 2.2.2 Classification diagnostics for the two-model approach were not satisfactory.	•Table S1 •Table S2
Supplement Section 2.3 Subgroup comparisons for clinical characteristics not planned a priori	OCD Subgroup comparisons for clinical characteristics that were not planned a priori: DOCS Germs and Contamination, DOCS Responsibility for Harm, DOCS Unacceptable Thoughts, OCI-R subscales, BAI anxiety, QIDS depressive symptoms.	No statistically significant differences in clinical characteristics between the OCD subgroups.	•Table S3
Supplement Section 2.4 Post-task urge ratings collected from outside the scanner	Subgroup comparisons for post-task urge ratings collected from outside the scanner.	Subgroup differences in post-task urge ratings were similar as in-scanner urge ratings reported in the main manuscript.	•Table S3
Supplement Section 2.5 Comorbid disorders and psychotropic medication use within subgroups	Subgroup comparisons for the proportion of patients who had current comorbidity based on DSM-5 and were taking psychotropic medications.	No statistically significant differences between subgroups in the proportions of current morbidity and psychotropic medication use.	-
Supplement Section 2.6 Visualization of brain activity in control subjects and subgroups during suppression	2.6.1 Conjunction analysis to determine whole-brain maps of subgroup differences overlapping with Harvard-Oxford (H-O) atlas parcels. Areas of overlap created ROIs representing the fMRI findings were segregated according to H-O atlas. Parameter estimates were extracted from the ROIs. 2.6.2 Parameter estimates for the control group and each subgroup were visualized. 2.6.3 Computed difference in parameter estimates from early to late suppression.	2.6.1 35 ROI clusters were identified. 2.6.2 Refer to Figure S2 for visualization of brain activation for OCD subgroups and control group. 2.6.3 Refer to Figure S3 for visualization of changes in brain activation as the suppression phase elapsed.	•Figure S2 •Figure S3
Supplement Section 2.7 Examined whether OCD subgroup differences in brain activity were related to higher rate of blinking	Conjunction analysis to determine whole-brain maps of subgroup differences overlapping with activations related to blinking behavior (free-blinking>suppression) within the OCD sample.	We found an overlapping cluster of 1056 voxels in the occipital cortex, suggesting that while some occipital activation may be related to blinking behavior, the majority of activations identified through subgroup differences may be specific to suppression failures.	•Figure S4 •Table S6

Results

Number of subgroups

A four-latent group model comprising three patient subgroups and one control group best fit the data (Group 1_OCD n=19; Group 2_OCD n=49, Group 3_OCD n=14, Group 4_Controls n=38) (Supplement 2.2, Table S1). Table 2 presents demographic, behavioral, and clinical characteristics of the OCD subgroups and controls. There were no statistically significant subgroup differences in age, male-female ratio, or years of education (P>0.05).

Table 2.

Demographics, clinical, and behavioural information

	OCD_Lo (n=19)		OCD_Mod (n=49)		OCD_Hi (n=14)		Controls (n = 38)		OCD Subgroup Comparisons
	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Demographics
Age (Years)a	28.32	9.70	30.61	11.01	32.14	11.17	31.03	11.35	n.s.
Education (Years)a	15.63	2.22	15.86	2.00	15.43	1.60	16.03	1.67	n.s.
Sex assigned at birth (M:F [% F])	7:12 [63.2%]		18:31 [63.3%]		4:10 [71.4%]		9:29 [76.3%]		n.s.
Clinical
Y-BOCS Totala	23.47	4.79	22.92	5.12	25.57	6.21	-	-	n.s.
Sensory Phenomena Scale Total	7.29	4.31	7.30	3.33	9.57	2.72	-	-	OCD_Hi>OCD_Mod, d = 0.71*
MAIA-Noticinga	2.91	1.11	3.47	0.91	4.11	0.89	2.16	1.40	OCD_Hi>OCD_Mod, ΔMAD = 1.50* OCD_Hi>OCD_Lo, ΔMAD = 1.94**
DOCS Symmetry/NJRE	1.39	1.07	1.54	0.88	2.17	1.19	-	-	OCD_Hi>OCD_Mod, d = 0.66* OCD_Hi>OCD_Lo, d = 0.70*
UFA Task b
Blinks Counts: Early-Suppressionc	1.68	1.67	24.51	12.72	78.57	32.09	19.05	17.01	-
Blinks Counts: Late-Suppressionc	2.53	2.61	31.76	15.52	96.07	31.07	23.39	22.03	-
Urge Intensity Rating (In-scanner)a	3.15	0.93	3.91	0.94	4.35	0.62	3.72	1.05	OCD_Hi>OCD_Lo, ΔMAD = 1.62*** OCD_Mod > OCD_Lo, ΔMAD = 1.25***
Difficulty Refrain From Blinking Following First Blinka	2.84	1.38	3.71	1.10	4.29	0.99	2.84	1.17	OCD_Hi>OCD_Lo, ΔMAD = 3.44*** OCD_Mod>OCD_Lo, ΔMAD = 1.00*

Effect sizes of subgroup differences using Cohen’s d as a parametric measure of effect size and delta median absolute deviations (Δ_MAD) as a non-parametric measure of effect size were presented for significant comparisons. Note that distinct subgroups of OCD patients were identified by latent profile analysis based on erroneous blinks during blink suppression: OCD_Lo (lowest number of erroneous blinks; n = 19), OCD_Mod (moderate; n = 49), and OCD_Hi (highest; n = 14).

^aAssumption of normality was not assumed. Statistical tests were performed using non-parametric Dunn’s tests.

^bRefer to Supplemental Section 2.7.

^cEarly Suppression refers to early phase of suppression (first 30 seconds of the 60-second suppression period); Late Suppression refers to late phase of suppression (second 30 seconds of the 60-second suppression period).

^*P < 0.05;

^**P < 0.01;

^***P < 0.001

Abbreviations: Y-BOCS, Yale-Brown Obsessive Compulsive Scale; MAIA, Multidimensional Scale of Interoceptive Awareness; UFA, Urges-for-action; n.s., not significant; d, Cohen’s d; Δ_MAD, Delta median absolute deviation.

Erroneous blink counts and task ratings during suppression between subgroups

The first subgroup (“OCD_Lo”, n=19) had the lowest number of erroneous blinks (i.e. better blink suppression performance than all other subgroups), the second (“OCD_Mod”, n=49) had a moderate number of blinks (i.e., higher than OCD_Lo and controls), and the third (“OCD_Hi”, n=14) had the highest number of blinks during both early and late suppression (Table 2, Figure S5).

OCD_Hi exhibited the highest mean rating of both urge intensity experienced during suppression (in-scanner rating) and difficulty refraining from blinking following the first blink (post-task rating) (Table 2, Figure 2A). The OCD_Hi and OCD_Mod subgroups did not significantly differ for either rating, but both subgroups scored higher on both ratings compared to the OCD_Lo subgroup (all P<0.03).

Figure 2: Clinical characteristics of OCD subgroups.

A) Bar graphs showing sensory phenomena severity, interoceptive sensitivity, and ratings for urge intensity (rated in-scanner) and difficulty in refraining from blinking following first blink during suppression (post-task rating) in OCD subgroups. B) Scatterplot showing associations between interoceptive sensitivity, sensory phenomena, and urge intensity (rated in-scanner). Note that distinct subgroups of OCD patients were identified by latent profile analysis based on erroneous blinks during blink suppression: OCD_Lo (lowest number of erroneous blinks; n = 19), OCD_Mod (moderate; n = 49), and OCD_Hi (highest; n = 14).

^a Assumption of normality was not assumed. Statistical tests were performed using non-parametric Dunn’s tests.

^b Question was administered at post-scan, rated on a 5-point Likert scale (ranging from 1 [least] to 5 [most]): “After blinking, how difficult was it to refrain from blinking again?”

*P < 0.05; **P < 0.01; ***P < 0.001

Clinical characteristics between subgroups and mediation analysis

While the OCD subgroups did not significantly differ in overall OCD severity (Table 2), the OCD_Hi subgroup had higher IS than the other two subgroups (Table 2, Figure 2A), indicating a greater tendency to notice bodily sensations. The OCD_Hi subgroup also reported more OCD symptoms related to concerns about symmetry, completeness, and NJREs than the other subgroups (Figure 2A). Consistently, the OCD_Hi subgroup had more severe sensory phenomena than the OCD_Mod subgroup (P=0.035) (Figure 2A).

Within the full sample of 82 patients with OCD, greater IS was associated with more severe sensory phenomena (r=0.248, P=0.025), stronger urge intensity ratings (in-scanner: r=0.268, P=0.015) (Figure 2B), and higher overall difficulty in refraining from blinking following the first erroneous blink during suppression (r=0.243, P=0.022). Contrary to expectation, sensory phenomena severity was not significantly associated with urge intensity (P>0.05).

Mediation analysis showed that IS significantly mediated the relationship between sensory phenomena and subgroup membership (Table 3).

Table 3.

Results of mediation analyses

Model	Path	b	p	Bootstrap 95% CI
X: Sensory Phenomena	a	0.07	0.027
M: IS	b	0.78	0.007
Y: OCD Subgroups
	c’	0.06	0.42
	a*b	0.06		0.002, 0.140
	c	0.11	0.14

95% confidence interval for indirect effects were generated using 10,000 bootstrap samples. Also refer to a mediation model diagram illustrated in Figure S1 (Supplement). Path a is a direct effect of X to M; Path b is a direct effect of M to Y; Path c’ is a direct effect of X to Y controlled for M; Path a*b is an indirect effect of X to Y through the mediator M; Path c is the total effect of X on Y. Abbreviations: IS, Interoceptive Sensitivity.

Neuroimaging results

During the early phase (first 30 seconds) of the suppression period, both the OCD_Mod and OCD_Hi subgroups showed significantly greater activation than the OCD_Lo subgroup in regions previously associated with the build-up of an urge (putative “urge network”) (Stern et al., 2020) including occipital (cuneus, calcarine), insula (anterior and posterior regions), somatosensory (supramarginal gyrus), and parahippocampal gyrus in a whole-brain analysis (Figure 3, Table S4). The OCD_Mod subgroup additionally showed greater activation than the OCD_Lo subgroup in several frontal regions including middle frontal gyrus and medial superior frontal gyrus during early suppression. There were no areas where the OCD_Lo subgroup had significantly greater activity than the other two subgroups (Figure 3, Table S4).

Figure 3: Differences in brain activation between OCD subgroups during early suppression (first 30 seconds).

Scan site was included as a covariate-of-no-interest. Distinct subgroups of OCD patients were identified by latent profile analysis based on erroneous blinks during blink suppression: OCD_Lo (lowest number of erroneous blinks; n = 19), OCD_Mod (moderate; n = 49), and OCD_Hi (highest; n = 14).

During the late phase (second 30 seconds) of the suppression period, both the OCD_Mod and OCD_Hi subgroups had greater activation than the OCD_Lo subgroup in occipital regions (lingual gyrus, cuneus, calcarine), parahippocampal gyrus, and posterior cingulate (Figure 4, Table S5). Compared with the OCD_Mod subgroup, the OCD_Hi subgroup showed reduced activation in the middle frontal gyrus (Figure 4, Table S5).

Figure 4: Differences in brain activation between OCD subgroups during late suppression (second 30 seconds).

Scan site was included as a covariate-of-no-interest. Distinct subgroups of OCD patients were identified by latent profile analysis based on erroneous blinks during blink suppression: OCD_Lo (lowest number of erroneous blinks; n = 19), OCD_Mod (moderate; n = 49), and OCD_Hi (highest; n = 14).

Discussion

Subtyping OCD by urge suppression failures

Unlike our prior work which compared patients with OCD as a homogeneous group to control participants (Stern et al., 2020), the current study applied unsupervised machine learning techniques to elucidate heterogeneity in urge suppression in patients with OCD. Latent profile analysis identified three distinct subgroups of OCD patients based on the number of erroneous blinks committed during the early (first 30 seconds) and late (second 30 seconds) phases of a 60-second period of eyeblink suppression. Interestingly, the largest subgroup (‘OCD_Mod’) had similar numbers of erroneous blinks as control participants, while smaller subgroups had the highest (‘OCD_Hi’) and lowest (‘OCD_Lo’) number of failures compared to all other patient subgroups and the control sample. Patient subgroups did not significantly differ in overall OCD severity, current psychotropic medication use, comorbidity, and demographic characteristics. These findings highlight the heterogeneity of urge suppression behavior in OCD, which may be independent of the severity of OCD symptoms, and suggest that treatments focusing on improving urge suppression may be efficacious for a subset of patients.

Relationships between urge suppression, interoceptive sensitivity, and sensory phenomena

OCD subgroups did not differ in overall OCD severity, but the subgroup with the most failures of urge suppression (OCD_Hi) had the most severe sensory phenomena (clinician-rated) and highest IS (self-reported). OCD subgroups with greater suppression failures (OCD_Hi and OCD_Mod) also reported greater intensity of urge sensation during suppression and difficulty refraining from blinking after their first erroneous eyeblink. Within the OCD sample, higher IS was associated with greater urge sensation and higher sensory phenomena, consistent with prior studies reporting associations between IS and premonitory urges in TD (Ganos et al., 2015; Rae et al., 2019) and between IS and sensory phenomena in OCD (Eng et al., 2022). IS statistically mediated the relationship between sensory phenomena and OCD subgroup membership, revealing IS as a potential psychological mechanism underlying the relationship between sensory phenomena and failures of urge suppression. However, the mediation analysis was conducted on cross-sectional data, and causal interpretations cannot be ascertained (Winer et al., 2016). Nevertheless, findings from the current study indicate that IS may be an important treatment target for a subset of patients with pathological sensory urges. One may speculate that these patients may benefit from therapeutic interventions targeting IS such as mindfulness therapy (Gibson, 2019; Kerr et al., 2013) and interoceptive exposure (Blakey and Abramowitz, 2018; Craske et al., 2008).

Overview of neural differences between subgroups during the early phase of suppression

Differences in neural activation among OCD subgroups during the early phase of the suppression period (first 30 seconds) were widespread throughout the brain, including the insula, posterior cingulate, secondary somatosensory regions (central and parietal operculum, supramarginal gyrus), occipital regions (calcarine, cuneus and lingual gyrus) and posterior precuneus. These regions are also part of the putative urge network reported in our prior work (Stern et al., 2020) and are consistent with studies investigating urges-for-action (Berman et al., 2012; Bohlhalter et al., 2006; Jackson et al., 2011b; Neuner et al., 2014).

Somatosensory and interoceptive activity during the early phase is linked to suppression failure

Compared to the subgroup with the lowest number of urge suppression failures (OCD_Lo), subgroups with relatively higher numbers (OCD_Hi and OCD_Mod) showed greater activation during the early phase of suppression in the insula (anterior and posterior) and secondary somatosensory regions including supramarginal gyrus and central and parietal operculum. These regions are involved in processing and integrating sensory information from within the body (Blefari et al., 2017; Bragdon et al., 2021; Craig, 2002, 2003, 2009; Eickhoff et al., 2007; Mălîia et al., 2018). The mid-posterior insula and secondary somatosensory areas were previously reported in studies investigating urges-for-action (Berman et al., 2012; Jackson et al., 2011a; Stern et al., 2020) and are activated immediately before and during tics in patients with TD (Bohlhalter et al., 2006; Neuner et al., 2014). Hyperactivity in these regions during the early phase of urge build-up may reflect the heightened physical experience of the urge sensation associated with more suppression failures.

Prefrontal cortex is more active for OCD subgroup with a moderate number of suppression failures

During the early phase of suppression, the OCD_Mod subgroup showed greater activation than the OCD_Lo subgroup in prefrontal cortical regions typically associated with cognitive control, including middle frontal gyrus and medial superior frontal gyrus (Figure 3, Table S4, Figure S2) (Menon and Uddin, 2010; Seeley et al., 2007; Sridharan et al., 2008). It is not clear what these activations might reflect. It is possible that the increased prefrontal activity in OCD_Mod along with higher urge suppression failure and urge intensity compared to OCD_Lo reflects cognitive effort related to unsuccessful attempts at increased behavioral control. However, the lack of greater prefrontal activations in the OCD_Hi subgroup, who experienced the highest suppression failure and urge intensity among all subgroups, contradicts this interpretation. Further work investigating the role of prefrontal cortex in urge suppression is needed to better understand this finding.

Occipital cortex is consistently more active in OCD subgroups with more urge suppression failures for both phases of suppression

OCD subgroups with higher numbers of erroneous blinks (OCD_Hi and OCD_Mod) consistently showed greater activation in occipital regions during both early (first 30 seconds) and late (second 30 seconds) phases of the suppression period (Figure 3–4, Table S4–S5, Figure S2). This observation aligns with prior findings of greater occipital activations associated with the build-up of the urge-to-blink (Berman et al., 2012; Stern et al., 2020). Hyperactivity in the occipital regions could be related to increased attention to visual stimuli and greater arousal during blink suppression (Demiral et al., 2023). Consistent with this interpretation, the OCD subgroups with greater occipital activations (OCD_Hi and OCD_Mod) also reported heightened urge intensity.

Middle frontal gyrus showed decreased activity in high-failure subgroup as suppression phase elapsed

During the late phase of suppression (second 30s), the subgroup of patients highest in erroneous blinks (OCD_Hi) showed reduced activation in middle frontal gyrus compared to the subgroup with a moderate number of blinks (OCD_Mod). In an exploratory comparison between early and late phases (Supplement 2.6.3; Figure S3), the two subgroups with relatively low and moderate blink suppression failures (OCD_Mod, OCD_Lo) had increased activation in this middle frontal gyrus area as the suppression period elapsed, potentially reflecting an increase in inhibitory control as the urge builds up over time. By contrast, OCD_Hi actually showed a reduction of activation in this area as the suppression period elapsed, suggesting impairment in the continued engagement of inhibitory control amongst OCD patients with higher urge suppression failures (Mazzone et al., 2010). This finding is consistent with prior work identifying impairments in inhibitory control in OCD (Chamberlain et al., 2007; Kang et al., 2013; McLaughlin et al., 2016; Menzies et al., 2007). Notably, our previous analysis (which included an overlap of ~30% of the current sample) did not find evidence of reduced prefrontal activity that would suggest an altered inhibitory control system when comparing OCD patients as a homogeneous group to control participants (Stern et al., 2020). Although sample differences between the two studies make it hard to directly compare findings, these data highlight the importance of considering heterogeneity of urge suppression in future work on this topic.

Limitations

To our knowledge, this is the first study to investigate heterogeneity of urge suppression in OCD, with results identifying behavioral and neural targets for future treatments. However, the study has several limitations. First, because in-scanner urge intensity ratings were only collected at the end of each 60-second suppression block, we could not characterize the change in urge intensity as the suppression period elapsed. While continuous or more frequent recording of urge intensity self-reports may inform how urge sensations change over time, such reporting paradigms could distract individuals and interfere with the task. Second, it is unclear whether mechanisms associated with the build-up of an urge-to-blink are the same as those related to the urge to perform a compulsion in OCD. We use the blink suppression paradigm specifically because it elicits an urge-for-action in all participants regardless of idiosyncratic urge symptoms. Incorporating participant-specific OCD-related urges would enhance construct validity of the approach, but it would also introduce unwanted variability to our study. This is especially problematic as the study already focuses on heterogeneity of urge suppression-related behavior and brain function. The blink suppression task has been used numerous times in studies of urge-related behavior and the associated neural mechanisms in healthy and patient samples (Berman et al., 2012; Bragdon et al., 2023; Stern et al., 2020), enabling comparisons with prior published work. Nevertheless, future work should investigate whether the neural mechanisms underlying build-up of the urge-to-blink overlap with other forms of urges more related to OCD symptoms.

Despite these limitations, our current work highlights the importance of considering heterogeneity when evaluating clinical phenomenology in OCD. Subtyping approaches are particularly beneficial in identifying shared features in heterogeneous conditions that would otherwise be less likely found. Using latent profile analysis, we identified subgroups of patients with OCD based on behavior in an urge suppression task that differed in clinical characteristics and brain activation during suppression. Our results suggest that therapeutic interventions (for example, neuromodulation) aimed at decreasing activation of an overactive sensorimotor/interoceptive circuit or increasing activity of an underactive inhibitory control circuit may be helpful in a subset of patients who experience difficulty with suppressing sensory-based urges.