In this article, we applied, for the first time, app-based behavioral training (experiment 1) to a sample of patients with OCD. We compared 32 patients and 33 healthy participants, matched for age, gender, IQ, and years of education in measures of motivation and app engagement (see Materials and methods for participants’ demographics and clinical characteristics). We also assessed to what extent performing such repetitive actions in 1 month impacted OCD symptomatology. In an initial phase (30 days), two action sequences were trained daily to produce habits/automatic actions (experiment 1). We collected data online continuously to monitor engagement and performance in real time. This approach ensured we acquired sufficient data for subsequent analysis of procedural learning and automaticity development.

In a second phase, we administered two follow-up behavioral tasks (experiments 2 and 3) addressing two important questions relevant to the habit theory of OCD. The first research question investigated whether repeated performance of motor sequences could develop implicit rewarding properties, hence gaining value, potentially leading to compulsive-like behaviors (experiment 2: explicit preference task, conducted without feedback). The hypothesis postulates that the repeated performance, initially driven by the goal of proficiency, may eventually become motivated by its own implicit reward, tied to proprioceptive and kinesthetic feedback (e.g. offering anxiety relief alongside skillful execution). The second question explored whether manipulations of extrinsic feedback, based on monetary reward or on the physical effort required (by varying the length of the sequence) affected choice for the familiar trained action sequence (experiment 3: re-evaluation task, conducted with feedback).

Finally, we administered a comprehensive set of self-reported clinical questionnaires, including a recently developed questionnaire (Ersche et al., 2017) on habit-related aspects. This aimed to investigate: (1) if OCD patients report more habits; (2) whether stronger subjective habitual tendencies predict enhanced procedural learning, automaticity development, and an (in)ability to adjust to changing circumstances; and (3) if app-based habit reversal therapy yields therapeutic benefits or has any subjective sequelae in OCD.

Anticipating implicit learning issues in OCD (Deckersbach et al., 2002; Kathmann et al., 2005; Rauch et al., 1997) and fine-motor difficulties (Bloch et al., 2011), we expected poorer procedural learning in patients compared to HV. However, once learned, we predicted OCD patients would reach automaticity faster, possibly due to a stronger tendency to form habitual/automatic actions (Gillan et al., 2016; Gillan et al., 2014). We also hypothesized differences in the learning rate and automaticity development between the two action sequences based on their associated (1) reward schedule (continuous versus variable), with faster automaticity in the continuous reward sequence, as suggested by past research (Bouton, 2021); and (2) sign of changes in reward scores, expecting enhanced performance improvements following a decrease in scores, particularly pronounced in OCD patients due to heightened sensitivity to negative feedback (Apergis-Schoute et al., 2024; Becker et al., 2014; Kanen et al., 2019). Additionally, we predicted that OCD patients would generally display stronger habits and assign greater intrinsic value to the familiar app sequences, evidenced by a marked preference for executing them even when presented with a simpler alternative sequences. Finally, we expected patients to show a greater tendency to perform the familiar/trained sequences, even though its extrinsic relative value was reduced and new, more valuable, options became available.

Participants completed self-reported questionnaires measuring various psychological constructs (see Materials and methods). Highly relevant for the current topic is the Creature of Habit Scale (COHS) (Ersche et al., 2017), recently developed to measure individual differences in routine behavior and automatic responses in everyday life. As compared to healthy controls, OCD patients reported significantly higher habitual tendencies in both the routine (t = –2.79, p = 0.01; HV: ${\displaystyle \overline{COHSroutine}}$ = 48.4, ${\displaystyle \sigma }$ = 9; OCD: ${\displaystyle \overline{COHSroutine}}$ = 55.7, ${\displaystyle \sigma }$ = 11) and the automaticity (t = –3.15, p < 0.001; HV: ${\displaystyle \overline{COHSautomaticity}}$ = 26.3, ${\displaystyle \sigma }$ = 8; OCD: ${\displaystyle \overline{COHSautomaticity}}$ = 32.9, ${\displaystyle \sigma }$ = 9) subscales.

All participants were presented with a calendar schedule and were asked to practice both sequences daily. They were instructed to practice as many times as they wished, whenever they wanted during the day and with the sequence order they would prefer. However, a minimum of two practices (P) per sequence was required every day; each practice comprised 20 successful sequence trials. Participants had to make up for missed training by completing both the current day’s session and the previous day’s if they skipped a day. If they missed training for over 2 days, the researcher gauged their motivation and incentivized their commitment. Participants were excluded if they missed training for more than 5 consecutive days.

At least 30 days of training was required, and all data were anonymously collected in real time, through an online server. On the 21st day of practice, the rewards were removed (extinction) to ensure that the action sequences were more dependent on proprioceptive and kinesthetic, rather than on external, feedback. Analysis of the reward removal (extinction) is presented in Appendix 1 and Appendix 1—figure 3. Other additional task components and analysis are also included in Appendix 1 and Appendix 1—figures 1 and 2.

Participants reliably committed to their regular training schedule, practicing consistently both sequences every day. Unexpectedly, OCD patients completed significantly more practices as compared with HV (p = 0.005) (Figure 2a). Descriptive statistics are as follows (values provided as median number of practices and interquartile range): HV: ${\displaystyle \tilde{P}}$ = 122, IQR = 7; OCD: ${\displaystyle \tilde{P}}$ = 130, IQR = 14. When visually inspecting the daily training pattern, we observed that HV tended to practice earlier than OCD. Circular statistics within each group demonstrated that HV practiced preferentially at a peak time of ~15:00 (mean resultant length 0.47, p = 0.000497, Rayleigh test for the uniformity of a circular distribution of points; Figure 2b). For OCD participants, the preferred practice time had a mean direction at ~18:00 (mean resultant length 0.58, p = 8.03 × 10^–6, Rayleigh test; Figure 2c). There were, however, no significant differences between both samples (p = 0.19, Watson’s U² test).

Figure 2

Download asset Open asset

Learning was evaluated by the decrement in sequence duration throughout training. To follow the nomenclature of the motor control literature, we refer to sequence duration as movement time (MT, in s), which is defined as

where $t_6$ and $t_1$ are the time of the last (6th) and first key presses, respectively.

For each participant and sequence reward type (continuous and variable), we measured MT of a successful trial, as a function of the sequence trial number, n, across the whole training. Across trials, MT decreased exponentially (Figure 3a). The decrease in MT has been widely used to quantify learning in previous research (Crossman, 1959). A single exponential is viewed as the most statistically robust function to model such decrease (Heathcote et al., 2000). Accordingly, each participant’s learning profile was modeled as follows:

where ${\displaystyle n_r}$ is the learning rate (measured in number of trials), which governs the rate of exponential decay. Parameter ${MT}_0$ is the movement time at asymptote (at the end of the training). Last, ${MT}_L$ is the speed-up achieved over the course of the training (referred to as amount of learning) (Figure 3a). The larger the value of ${MT}_L$ , the bigger the decline in the movement time and thus the larger the improvement in motor learning.

Figure 3

Download asset Open asset

The individual fitting approach we used has the advantage of handling the different number of trials executed by each participant by modeling their behavior to a consolidated maximum value of n, n_max = 1200. We used a moving average of 20 trials to mitigate any effect of outlier trials. This analysis was conducted separately for continuous and variable reward schedules.

To statistically assess between-group differences in learning behavior, we pooled the individual model parameters (${\displaystyle M{T}_L}$ , ${\displaystyle n_r}$ and ${MT}_0$), and conducted a Kruskal-Wallis H test to assess the effect of group (HV and OCD), reward type (continuous and variable), and their interaction on each parameter (Figure 3b).

There was a significant effect of group on the amount of learning (${MT}_L$) parameter, H = 16.5, p < 0.001, but no reward (p = 0.06) or interaction effects (p = 0.34) (Figure 3c). Descriptive statistics are as follows (values provided as median and interquartile range): HV: ${\displaystyle \widetilde{M{T}_L}}$ = 3.1 s, IQR = 1.2 s and OCD: ${\displaystyle \widetilde{M{T}_L}}$ = 3.9 s, IQR = 2.3 s for the continuous reward sequence; HV: ${\displaystyle \widetilde{M{T}_L}}$ = 2.3 s, IQR = 1.2 s and OCD: ${\displaystyle \widetilde{M{T}_L}}$ = 3.6 s, IQR = 2.5 s for the variable reward sequence.

Regarding the learning rate (${\displaystyle n_r}$) parameter, we found no significant main effects of group (p = 0.79), reward (p = 0.47), or interaction effects (p = 0.46). Descriptive statistics: sequence trials needed to asymptote HV: ${\displaystyle \widetilde{n_r}}$ = 176, IQR = 99 and OCD: ${\displaystyle \widetilde{n_r}}$ = 200, IQR = 114 for the continuous reward sequence; HV: ${\displaystyle \widetilde{n_r}}$ = 182, IQR = 123 and OCD: ${\displaystyle \widetilde{n_r}}$ = 162, IQR = 141 for the variable reward sequence. These non-significant effects on the learning rate were further assessed with Bayes factors (BF) for factorial designs (see Materials and methods). This approach estimates the ratio between the full model, including main and interaction effects, and a restricted model that excludes a specific effect. The evidence for the lack of main effect of group was associated with a BF of 0.38, which is anecdotal evidence. We additionally obtained moderate evidence supporting the absence of a main effect of reward or a reward × group interaction (BF = 0.16 and 0.17, respectively).

In analyzing the asymptote (${MT}_0$) parameter, we found no significant main or interaction effects (group effect: p = 0.17; reward effect: p = 0.65 and interaction effect: p = 0.64). Descriptive statistics are as follows: HV: ${\displaystyle \widetilde{M{T}_0}}$ = 1.7 s, IQR = 0.4 s and OCD: ${\displaystyle \widetilde{M{T}_0}}$ = 1.8 s, IQR = 0.5 s for the continuous reward sequence; HV: ${\displaystyle \widetilde{M{T}_0}}$ = 1.8 s, IQR = 0.5 s and OCD: ${\displaystyle \widetilde{M{T}_0}}$ = 1.8 s, IQR = 0.5 s for the variable reward sequence. BF analysis indicated anecdotal evidence against a main group effect (BF = 0.53). Meanwhile, there was moderate evidence suggesting neither reward nor reward × interaction factors significantly influenced performance time (BF = 0.12 and 0.17, respectively).

The results indicate that OCD patients do not exhibit learning deficits. While they initially performed action sequences slower than the HV group, their learning rates ultimately matched those of HV. Both groups showed comparable movement durations at the asymptote. This suggests that, though OCD patients began at a lower baseline level of performance, they enhanced their motor learning to a degree that reached the same asymptotic performance as the controls.

To assess automaticity, the ability to perform actions with low-level cognitive engagement, we examined the decline over time in the consistency of inter-keystroke interval (IKI) patterns trial to trial. We mathematically defined IKI consistency as the sum of the absolute value of the time lapses between finger presses from one sequence to the previous one.

where $n$ is the sequence trial number and $k$ is the inter-keystroke response interval (Figure 4a). In other words, C quantifies how consistent/reproducible the press pattern is from trial to trial. The assumption here is that the more reproducible the sequences are over time, the more automatic the person’s motor performance becomes.

Figure 4

Download asset Open asset

For each participant and sequence reward type (continuous and variable), automaticity was assessed based on the decrement in C, as a function of n, across the entire training period. Since C decreased in an exponential fashion, we fitted the C data with an exponential decay function (following the same reasoning and procedure as MT) to model the automaticity profile of each participant,

where $n_C$ is the automaticity rate (measured in number of trials), $C_0$ is the sequence consistency at asymptote (by the end of the training), and $C_L$ is the change in automaticity over the course of the training (which we refer to amount of automation gain). The model fitting procedure was conducted separately for continuous and variable reward schedules.

A Kruskal-Wallis H test was then conducted to assess the effect of group (OCD and HV) and reward type (continuous and variable) on each parameter resulting from the individual exponential fits (${\displaystyle C_L}$, ${\displaystyle n_C}$ and ${\displaystyle C_0}$ ).

There was a significant effect of group on the amount of automation gain (${\displaystyle C_L}$) parameter: H = 11.1, p < 0.001 but no reward (p = 0.12) or interaction effects (p = 0.5) (Figure 4b). Descriptive statistics are as follows: HV: ${\displaystyle \widetilde{C_L}}$ = 1.4 s, IQR = 0.7 s and OCD: ${\displaystyle \widetilde{C_L}}$ = 1.9 s, IQR = 1.0 s for the continuous reward sequence; HV: ${\displaystyle \widetilde{C_L}}$ = 1.1 s, IQR = 0.8 s and OCD: ${\displaystyle \widetilde{C_L}}$ = 1.5 s, IQR = 1.1 s for the variable reward sequence.

There was also a significant group effect on the automaticity rate ($n_C$) parameter: H = 4.61, p < 0.03 but no reward (p = 0.42) or interaction (p = 0.12) effects. Descriptive statistics: sequence trials needed to asymptote HV: ${\displaystyle \widetilde{n_C}}$ = 142, IQR = 122 and OCD: ${\displaystyle \widetilde{n_C}}$ = 198, IQR = 162 for the continuous reward sequence; HV: ${\displaystyle \widetilde{n_C}}$ = 161, IQR = 104 and OCD: ${\displaystyle \widetilde{n_C}}$ = 191, IQR = 138 for the variable reward sequence.

At asymptote (${\displaystyle C_0}$), no group (p = 0.1), reward (p = 0.9), or interaction (p = 0.45) effects were found. We found anecdotal evidence against a main group effect (BF = 0.65). In addition, there was moderate evidence in favor of no main effects of reward or interaction (BF = 0.12 and 0.18, respectively).

Of note is the median consistency in consecutive sequences achieved at asymptote: HV: ${\displaystyle \widetilde{C_0}}$ = 287 ms, IQR = 127 ms, OCD: ${\displaystyle \widetilde{C_0}}$ 301 ms, IQR = 186 ms for the continuous reward sequence and HV: ${\displaystyle \widetilde{C_0}}$ = 288 ms, IQR = 110 ms, OCD: ${\displaystyle \widetilde{C_0}}$ = 300 ms, IQR = 114 ms for the variable reward sequence. These values of the C at asymptote are generally shorter than the normal reaction time for motor performance (Kosinski, 2008), reinforcing the idea that automaticity was reached by the end of the training.

In conclusion, compared to HV, patients took significantly longer to achieve a similar level of automaticity in both reward schedules. They began at a slower pace, exhibited more variability, and progressed to automaticity at a slower rate.

Our next goal was to investigate the sensitivity of performance improvements over time in our participant groups to changes in scores, whether they increased or decreased. To do this, we quantified the trial-by-trial behavioral changes in response to a decrement or increase in reward from the previous trial using the sequence duration (in ms), labeled as MT (movement time). Note that in our experimental design, MT was negatively correlated with the scores received. Following Pekny et al., 2015, we represented the change from trial n to n+1 in MT simply as:

Reward (R) change at trial n was computed as:

We next aimed to analyze separately ∆MT values that followed an increase in reward from trial n − 1 to n, ∆R+, denoting a positive sign in ∆R; and those that followed a drop in reward, ∆R−, indicating a negative sign in ∆R. An issue arises with poor performance trials (those with a slower duration, or a large MT⁽ⁿ⁾). These could inherently result in a systematic link between ∆R− and smaller (negative) ∆MT⁽ⁿ⁺¹⁾ values due to the statistical effect known as ‘regression to the mean’. Essentially, a trial that is poorly performed, marked by a large MT⁽ⁿ⁾, is likely to be followed by a smaller MT⁽ⁿ⁺¹⁾ just because extreme values tend to be followed by values closer to the mean. As training progresses and MT reduces overall, the potential for significant changes relative to reward increments or decrements may diminish. To account for and counteract this statistical artifact, we normalized the ∆MT⁽ⁿ⁺¹⁾ index using the baseline MT⁽ⁿ⁾:

We used this normalized measure of ∆MT⁽ⁿ⁺¹⁾ (adimensional) for further analyses. It reflects the behavioral change from trial n to n+1 relative to the baseline performance on trial n. Following Pekny et al., 2015, we estimated for each participant the conditional probability distributions p(∆T|∆R+) and p(∆T|∆R−) (where T denotes a behavioral measure, MT in this section or IKI consistency in the next section) by fitting a Gaussian distribution to the histogram of each data sample (Appendix 1—figure 5). The standard deviation (σ) and the center μ of the resulting distributions were used for statistical analyses (Appendix 1—figure 5). Similar analyses were carried out on a normalized version of index C (Equation 3), which already reflected changes between consecutive trials. See next section.

As a general result, we expected that healthy participants would introduce larger behavioral changes (more pronounced reduction in MT, more negative ∆MT) following a decrease in scores, as shown previously (Chen et al., 2017; van Mastrigt et al., 2020). Accordingly, we predicted that the p(∆T|∆R−) distribution would be centered at more negative values than p(∆T|∆R+), corresponding to greater speeding following negative reward changes. Given previous suggestions of enhanced sensitivity to negative feedback in patients with OCD (Apergis-Schoute et al., 2024, Becker et al., 2014; Kanen et al., 2019), we predicted that the OCD group, as compared to the control group, would demonstrate greater trial-to-trial changes in movement time and a more negative center of the p(∆T|∆R−) distribution. Additionally, we examined whether OCD participants would exhibit more irregular changes to ∆R− and ∆R+ values, as reflected in a larger spread of the p(∆T|∆R+) and p(∆T|∆R−) distributions, compared to the control group.

The conditional probability distributions were separately fitted to subsamples of the data across continuous reward practices, splitting the total number of correct sequences into four bins. This analysis allowed us to assess changes in reward sensitivity and behavioral changes across bins of sequences (bins 1–4 by partitioning the total number of sequences, from the whole training, into four). We focused the analysis on the continuous reward schedule for two reasons: (1) changes in scores on this schedule are more obvious to the participants and (2) a larger number of trials in each subsample were available to fit the Gaussian distributions, due to performance-related reward feedback being provided on all trials.

We observed that participants speeded up their sequence duration more (negative changes in trial-wise MT) following a drop in scores, as expected (Figure 5a). Conducting a three-way analysis of variance (ANOVA) with reward change (increase, decrease) and bin (1:4, each bin denoting ~110 sequences) as within-subject factors, and group as between-subject factor, we found a significant main effect of reward (F (504,1) = 319.383, p = 2.0 × 10^–16). This outcome indicated that, in both groups, participants reduced MT differently as a function of the change in reward. There was also a significant main effect of bin (F (504,3) = 19.583, p = 5.06 × 10^–12), such that participants sped up their sequence performance over practices. The main effects are illustrated in Figure 5b. There was no significant main group effect (F (504,1) = 1.099, p = 0.2951) and omitting the group factor from the full model using a BF ANOVA improved the model moderately (BF = 6.08, moderate evidence in support of the model with the main group effect removed relative to the full model). Thus, both OCD and HV individuals introduced comparable changes in MT overall during training.

Figure 5

Download asset Open asset

In addition, there was a significant interaction between reward and bin in predicting the trial-to-trial changes in movement time (F (504,3) = 3.652, p = 0.0126). This outcome suggested that the relative improvement in MT over sequences depended on whether the reward increased or decreased from the previous trial. To explore this interaction effect further, we conducted a dependent-sample pairwise t-test on MT, after collapsing the data across groups. In each sequence bin, participants speeded up MT more following a drop in scores than following an increment, as expected (corrected p_FDR = 2 × 10^–16).

On the other hand, assessing the effect of bins separately for each level of reward, we observed that the large sensitivity of normalized MT changes to reward decrements was attenuated from the first to the second bin of practice (corrected p_FDR = 0.00034, significant attenuation for pairs 1–2; dependent-sample t-tests between consecutive pairs of bins). No further changes over practice bins were observed (p > 0.53, no change for pairs 2–3, 3–4). Similarly, the sensitivity of MT changes to reward increments – consistently smaller – did only change from bin 1 to bin 2 (p_FDR = 0.01092; no significant changes for pairs 2–3 and 3–4, p > 0.31670).

Overall, these findings indicate that both OCD and HV participants exhibited an acceleration in sequence performance following a decrease in scores (main effect). Furthermore, the sensitivity to score decrements or increments was reduced as participants approached automaticity through repeated practice. Crucially, however, the increased sensitivity to reward decrements relative to increments persisted throughout the practice sessions in both groups.

Assessment of the std (σ) of the Gaussian distributions p(∆T|∆R−) and p(∆T|∆R+) in the continuous reward condition (Figure 5c) with a similar three-way ANOVA revealed a significant main effect of group (F (504,1) = 19.928, p = 9.93 × 10^–6). As shown in Figure 5d, the std (σ) of the distribution of trial-to-trial MT changes was smaller in HV than in OCD. In addition, we observed a significant change over bins of sequences in σ, and independently of the group or reward factors (main effect of bin, F (504,3) = 39.078, p = 2 × 10^–16). This outcome reflected that over the course of training, both groups exhibited less variable changes in MT in response to both reward increments and decrements, in line with improvements in skill learning (Wolpert et al., 2011). Reward also modulated σ, with ∆R− being associated with a more variable distribution of behavioral changes than ∆R+(main effect of reward, F (504,1) = 17.110, p = 4.13 × 10^–05). No interaction effects were found (there was moderate to strong evidence that removing any of the possible interaction effects improved the model: BF ranged from 5.67 to 41.3). Control analyses demonstrated that the group, reward, or bin effects were not confounded by differences in the size of the subsamples used for the Gaussian distribution fits (not shown; Appendix 1 -Sample size for the reward sensitivity analysis).

To further explore the potential impact of reward changes on the previously reported group effects on automaticity, we quantified the trial-by-trial behavioral changes in IKI consistency (represented by C) in response to changes in reward scores relative to the previous trial. Note that a smaller C indicates a more reproducible IKI pattern trial to trial. As for ∆MT, we normalized the index C (termed normC to avoid confusion with the main analysis on C) with the baseline IKI values on the previous trial n

where $k$ is the inter-keystroke response interval and $n$ is the sequence trial number. During continuous reward practices, both patients and healthy controls exhibited an increased consistency of IKI patterns trial to trial across bins of correct sequences (decreased normC, Equation 8, Figure 6a; main effect of bin on the center of the Gaussian distribution, F (497, 3) = 4.188, p = 0.00607; three-way ANOVA). Performance in OCD and HV, however, differed with regard to how reproducible their timing patterns were (main effect of group, F (497, 1) = 8.130, p = 0.00454). The timing patterns were less consistent trial to trial in OCD, relative to HV (Figure 6b, left panel). Moreover, the IKI consistency improved more (smaller normC) following reward increments than after decrements, as shown in Figure 6b (right panel; main reward effect, F (497, 1) = 23.283, p = 1.86 × 10^–6). No significant interaction effects between factors were found (BF analysis demonstrated that when any of the interaction effects among factors was removed from the ANOVA design, there was moderate to strong evidence that the model improved: BF in range from 6.83 to 14.1). Accordingly, although OCD participants exhibited an attenuated IKI consistency in their performance relative to HV, the main effects of reward and bins of sequences were independent of the group.

Figure 6

Download asset Open asset

Regarding the spread of the p(∆T|∆R) distributions, we found a significant main effect of bin factor (F (504,3) = 23.350, p = 3.63 × 10^–14; Figure 6c and d). These outcomes suggest that the σ of the Gaussian distribution for normC values was reduced across bins of practiced sequences. There was only a trend for a significant main effect of group (F (504,1) = 3.412, p = 0.0653) and no main effect of reward (F (504,1) = 2.327, p = 0.1278). These non-significant effects were explored further using BF. This analysis provided anecdotal and moderate evidence that omitting either the group or reward effects was beneficial to the model (BF = 1.98 for removing group, BF = 3.20 for removing reward). We did not observe any interaction effect either (BF values increased moderately to strongly when any of the interaction effects among factors was removed from the ANOVA design: BF ranged from 4.89 to 43.3). The results highlight that over the course of training participants’ normalized IKI consistency values stabilized, and this effect was not observed to be modulated by group or reward factors. Similarly to the MT analyses, the sensitivity analyses of normC were not influenced by differences in the size of the subsamples used for the ∆R+ and ∆R– Gaussian distribution fits (Appendix 1 - Sample size for the reward sensitivity analysis).

Once the month-long app training was completed, participants attended a laboratory session to conduct additional behavioral tests aimed at assessing preference for familiar versus novel sequences (experiments 2 and 3) including a re-evaluation test to assess ability to adapt to environmental changes (experiment 3 only). Below we briefly describe these two experiments and report the results. See Materials and methods and Table 3 for a more detailed description of the tasks. Since these follow-up tests required observing additional stimuli while performing the action sequences, it was impractical to use participant’s individual iPhones to simultaneously present the task stimuli and be an interface to play the action sequences. We therefore used a ‘Makey-Makey’ device to connect the testing laptop (presenting the task stimuli) to four playdough keys arranged on a table (used as an interface for action sequence input). This device ensured precise key registration and timing. The playdough keys matched the size of those on the participants’ iPhones used for the 1-month training. Participants practiced the action sequences in this new setup until they were comfortable. Hence, the transition to a non-mobile/laboratory context was conducted with great care. These tasks were conducted in a new context, which has been shown to promote re-engagement of the goal system (Bouton, 2021).

Experiment 2: Preference for familiar versus novel action sequences

This experiment tests the hypothesis stated in the outline, that the trained action sequence gains intrinsic/rewarding properties or value. We used an explicit preference task, assessing participants’ preferences for familiar (hypothetically habitual) sequences over goal-seeking sequences. We assume that if the trained sequences have acquired rewarding properties (e.g. anxiety relief, or the inherent gratification of skilled performance or routine), participants would express a greater preference to ‘play’ them, even when alternative easier sequences are offered (i.e. goal-seeking sequences).

After reporting which app sequence was their preferred, participants started the explicit preference task. On each trial, they were required to select and play one of two sequences. The two possible sequences were presented and identified using a corresponding image. Participants had to choose which one to play. There were three conditions, each comprising a specific sequence pair: (1) app preferred sequence versus app non-preferred sequence (control condition); (2) app preferred sequence versus any 6-move sequence (experimental condition 1); (3) app preferred sequence versus any 3-move sequence (experimental condition 2). The app preferred sequence was their preferred putative habitual sequence while the ‘any 6’- or ‘any 3’-move sequences were the goal-seeking sequences. These were considered less effortful for two reasons: (1) they could comprise any key press of participant’s choice, even repeated presses of the same key (six or three times, respectively), and (2) they allowed for variations in key combinations each time the ‘any-sequence’ was input, rather than a fixed sequence on every trial. The conditions (15 trials each) were presented sequentially but counterbalanced among participants. See Materials and methods and Figure 7a for further details.

Figure 7

Download asset Open asset

A Kruskal-Wallis H test indicated a main effect of condition (H = 23.2, p < 0.001) but no group (p = 0.36) or interaction effects (p = 0.72) (Figure 7b). Dunn’s post hoc pairwise comparisons revealed that experimental condition 2 (app sequence versus any three sequence) was significantly different from control condition (p_FDR < 0.001) and from experimental condition 1 (app sequence versus any six sequence) (p_FDR = 0.006). No differences were found between the latter two conditions (p = 0.086). Bayesian analysis further provided moderate evidence in support of the absence of main effects of group (BF = 0.129) and interaction (BF = 0.054). These results denote that both groups evaluate the trained app sequences as being equally attractive as the alternative novel-but-easier sequence when of the same length (Figure 7b, middle plot). However, when given the option to play an easier-but-shorter sequence (in experimental condition 2), both groups significantly preferred it over the app familiar sequence (Figure 7b, right plot). A positive correlation between COHS and the app sequence choice (Pearson r = 0.36, p = 0.005) showed that those participants with greater habitual tendencies had a greater propensity to prefer the trained app sequence under this condition.

Given the high variance of participants’ choices on this preference task, particularly in the experimental conditions, and the findings reported below related to the mobile-app performance effect on symptomatology, we further conducted an exploratory Dunn’s post hoc test splitting the OCD group into two subgroups based on their Yale-Brown Obsessive-Compulsive Scale (YBOCS) score changes after the app training: 14 patients with improved symptomatology (reduction in YBOCS scores) and 18 patients who remained stable or felt worse (i.e. respectively, same or increase in YBOCS scores). Patients with lowered YBOCS scores after the app training had significantly greater preference for the app trained sequence in both experimental conditions as compared to patients with same or increased YBOCS scores after the app training: experimental condition 1 (p_FDR = 0.015, Figure 7c, left) and experimental condition 2 (p_FDR = 0.011, Figure 7c, right). In addition to this subgrouping analysis, we conducted a correlation analysis between changes in YBOCS scores and patient preferences for the app sequences. This helped us determine whether patients who experienced greater changes in YBOCS scores tended to prefer the learned sequences, and vice versa. We observed a positive correlation, meaning that the higher the symptom improvement after the month training, the greater the preference for the familiar/learned sequence. This is particularly the case for the experimental condition 2, when subjects are required to choose between the trained app sequence and any 3-move sequence (r_s = 0.35, p = 0.04). A trend was observed for the correlation between the YBOCS score change and the preference for the app sequences in experimental condition 1 (r_s = 0.30, p = 0.09). In conclusion, most participants preferred to play shorter and easier alternative sequences, thus not showing a bias toward the trained/familiar app sequences. Contradicting our hypothesis, OCD patients followed the same behavioral pattern. However, some participants still preferred the app sequence, specifically those with greater habitual tendencies, including patients who improved their symptoms during the month training and considered the app training beneficial (see also below exploratory analyses of ‘Mobile-app performance effect on symptomatology’). Such preference presumably arose because some intrinsic value may have been attributed to the trained action sequence.

Experiment 3: Re-evaluation of the learned action sequence

In experiment 3, we employed a two-choice appetitive learning task. We modified the conditions by manipulating extrinsic feedback to assess participants’ capacity to adopt a different response choice, after re-evaluating their options. By providing more value to alternative action sequences (as opposed to the previously automatized ones), participants were thus encouraged to reassess their choices and respond appropriately. Of note, we did not use a conventional goal devaluation procedure here, as this could possibly have disrupted the behavioral control of the sequences and thus invalidated the test.

On each trial, participants were required to choose between two ‘chests’ based on their associated reward value. Each chest depicted an image identifying the sequence that needed to be completed to be opened. After choosing which chest they wanted, participants had to play the specific correct sequence to open it. Their task was to learn by trial and error which chest would give them more rewards (gems), which by the end of the experiment would be converted into real monetary reward. There was no penalty for incorrectly keyed sequences because behavior was assessed based on participants’ choice regardless of the sequence accuracy.

Four chest-pairs (conditions, 40 trials each) were tested (see Figure 8a and Materials and methods for detailed description of each condition): three conditions pitted the trained/familiar app sequence against alternative sequences of higher monetary outcomes (given by variable amount of reward that did not overlap [deterministic]). The fourth condition kept the monetary value equivalent for the two options (maintaining a probabilistic rather than deterministic contingency) but offered a significantly easier/shorter alternative sequence. This set up a comparison between the intrinsic value of the familiar sequence and a motor-wise less effortful sequence. The conditions were presented sequentially but counterbalanced among participants.

Figure 8

Download asset Open asset

Both groups were highly sensitive to the re-evaluation procedure based on monetary feedback, choosing more often the non-app sequence, irrespective of the novelty of that sequence (Figure 8b, no group effects; p = 0.210 and BF = 0.742, anecdotal evidence supporting no main effect of group). However, when re-evaluation required motor effort (condition 4), participants were less inclined to choose the ‘any 3’ alternative, which is the sequence demanding less motor effort (Kruskal-Wallis main effect of condition: H = 151.1, p < 0.001). Moreover, OCD patients significantly favored the trained app sequence over HV (post hoc group × condition 4 comparison: p = 0.04). In conclusion, following the month of training, both groups exhibited the ability to update their behavior based on monetary re-evaluation. Yet, OCD patients more frequently selected the familiar sequence, even when a less effortful and shorter alternative was available.

In a debriefing questionnaire, participants were asked to give feedback about their app training experience and how it interfered with their routine: (1) how stressful/relaxing the training was (rated on a scale from –100% highly stressful to 100% very relaxing); (2) how much it impacted their life quality (Q) (rated on a scale from –100% maximum decrease to 100% maximum increase in life quality). Appendix 1—table 1 and Appendix 1—figure 4 depict participants’ qualitative and quantitative feedback. Of the 33 HV, 30 reported the app was neutral and did not impact their lives, neither positively nor negatively. The remaining 3 reported it as being a positive experience, with an improvement in their life quality (rating their life quality increase as 10%, 15%, and 60%). Of the 32 patients assessed, 14 unexpectedly showed improvement (I) in their OCD symptoms during the month as measured by the YBOCS difference, in percentage terms, pre-post training ($\overline{I}$ = 20 ± 9%), 5 felt worse ($\overline{I}$ = –19 ± 9%) and 13 remained stable during the month (all errors are standard deviations). Of the 14 who felt better, 10 directly related their OCD improvement to the app training (life quality increase: $\overline{Q}$ = 43 ± 24%). Nobody rated the app negatively. Of note, the symptom improvement was positively correlated with patients’ habitual tendencies reported in the Creature of Habit questionnaire, particularly with the routine subscale (Pearson r = 0.45, p = 0.01) (Figure 9a, left). A three-way ANOVA test showed that patients who reported less obsessions and compulsions after the month training were the ones with more pronounced habit routines (group effect: F = 13.7, p < 0.001, Figure 9a, right). A strong positive correlation was also found between the OCD improvement reported subjectively as direct consequence of the app training and the OCI scores and reported habit tendencies (Pearson r = 0.8, p = 0.008; Pearson r = 0.77, p < 0.01, respectively) (Figure 9b): i.e., patients who considered the app somewhat beneficial were the ones with higher compulsivity scores and higher habitual tendencies. In HV, participants who also had greater tendency for automatic behaviors, regarded the app as more relaxing (Pearson r = 0.44, p < 0.01). However, such correlation between the self-reported relaxation measure attributed to the app and the COHS automaticity subscale was not observed in OCD (p = 0.1). Finally, patients’ symptom improvement did not correlate with how relaxing they considered the app training (p = 0.1) nor with the number of total practices performed during the month training period (p = 0.2).

Figure 9

Download asset Open asset

We also checked whether the preferred app sequence, chosen by participants at the beginning of Phase B, was consistently the one that had yielded more reward during the app training (i.e. the continuously rewarded sequence). We found no evidence for this case: 54.5% of HV and 29% of the OCD sample considered the continuous sequence to be their preferred one, a non-statistically significant difference. This result suggests that participants’ preference may not solely be linked to programmed reward. Other factors, such as the aesthetic appeal of, or ease of performing specific combinations of finger movements, may also influence overall preference.

In addition to the Creature of Habit findings, of the remaining self-reported questionnaires assessed (see Materials and methods), OCD patients also reported enhanced intolerance of uncertainty, elevated motivation to avoid aversive outcomes and higher perfectionism, worries and perceived stress, as compared to healthy controls (see Table 1 for statistical results and Figure 10 for overall summary).

Figure 10

Download asset Open asset

Rapid execution, invariant response topography, action chunking, and low cognitive load have all been considered essential criteria for the definition of habits (Balleine and Dezfouli, 2019; Haith and Krakauer, 2018). We have successfully achieved all these elements with our app using the criteria of extensive training and context stability, both previously shown to be essential to enhance formation and strengthening of habits (Haith and Krakauer, 2018; Verplanken and Wood, 2006). Context stability was provided by the tactile, visual, and auditory stimulation associated with the phone itself, which establishes a strong and similar context for all participants, regardless of their concurrent circumstances. Overtraining has been one of the most important criteria for habit development, and used by many as an operational definition on how to form a habit (Dickinson et al., 1995; Haith and Krakauer, 2018; Tricomi et al., 2009) (for a review, see Balleine and O’Doherty, 2010), despite current controversies raised by de Wit et al., 2018, on its use as an objective test of habits. A recent study has demonstrated though that even short overtraining (1 day) is effective at producing habitual behavior in participants high in affective stress (Pool et al., 2022), confirming previous suggestions for the key role of anxiety and stress on the behavioral expression of habits (Dias-Ferreira et al., 2009; Hartogsveld et al., 2020; Schwabe and Wolf, 2009). Here, we have trained a clinical population with moderately high baseline levels of stress and anxiety, with training sessions of a higher order of magnitude than in previous studies (de Wit et al., 2018; Gera et al., 2022). By all accounts our overtraining is valid: to our knowledge the longest overtraining in human studies achieved so far. All participants attained automaticity, exhibiting similar and stable asymptotic performance, both in terms of speed and the invariance in the kinematics of the motor movement.

We succeeded in achieving automaticity – which at a neural level is known to reliably engage the brain’s habitual circuitry (Ashby et al., 2010; Bassett et al., 2015; Graybiel and Grafton, 2015; Lehéricy et al., 2005) – and fulfilled three of the four criteria for the definition of habits according to Balleine and Dezfouli, 2019 (rapid execution, invariant topography, and chunked action sequences). However, we were not able to test the fourth criterion of resistance to devaluation. Therefore, we are unable to firmly conclude that the action sequences are habits rather than, for example, goal-directed skills. According to a very recent study, also employing an app to study habitual behavior, the criterion of devaluation resistance was shown to apply to a three-element sequence with less training (Gera et al., 2022). Thus, overtraining of our six-element sequence might also have achieved behavioral autonomy from the goal in addition to behavioral automaticity. While we did not employ the conventional goal devaluation test, it is possible that some experts may interpret our follow-up experiment 3 (the re-evaluation test) as a measure of Balleine and Dezfouli, 2019, fourth criterion, which defines habits as ‘insensitive to changes in their relationship to their individual consequences and the value of those consequences’. Consequently, they may conclude that the app-trained sequences exhibited some features of goal-directed behavior. While this interpretation holds merit, the logical conclusion is that the app-trained sequences encompass both habitual and goal-directed qualities. This aligns with contemporary perspectives on skilled/habitual sequences (Du and Haith, 2023).

Regardless of whether the trained action sequences are labeled as procedural habits or goal-directed motor skills, one must question why OCD patients preferred familiar sequences in specific situations, even when it seemed counterproductive (e.g. in the effort condition). This observation leads to the hypothesis that motivation for action sequences might include other factors besides explicit goals, such as monetary rewards. The apparent (intrinsic) therapeutic value of performing these sequences further blurs the attribution of a singular goal such as monetary reward to human action sequences. One implication of this analysis may be to consider that behavior in general is ‘goal-directed’ but may vary in the balance of control by external and internal goals. This perspective aligns with motor control theories that classify the successful completion of a motor action, in the spatio-temporal sense, as ‘goal-related’. Hence, underlying any action sequence is possibly a hierarchy of objectives, ranging from overt rewards like money to intrinsic relief from an endogenous state (e.g. anxiety or boredom). In light of these insights, the dual associative process framework of behavioral control might be better understood in terms of the relative importance of extrinsic versus intrinsic outcomes. Another possible formulation is that habits, which depend initially on cached or historically acquired rewarding action values, may not necessarily lose current value, but instead acquire alternative sources of value (Hommel and Wiers, 2017; Kruglanski and Szumowska, 2020; O’Doherty, 2014).

We observed a slower and more irregular performance in patients with OCD as compared to healthy participants at the beginning of training. This was expected given previous reports of visuospatial and fine-motor skill difficulties in patients with OCD (Bloch et al., 2011). However, despite this initial slowness, no procedural learning deficits were found in our patient sample. This finding is inconsistent with other implicit learning deficits previously reported in OCD using the serial reaction time (SRT) paradigm (Deckersbach et al., 2002; Joel et al., 2005; Kathmann et al., 2005; Rauch et al., 2001; Rauch et al., 1997). Nevertheless, this result aligns with recent studies demonstrating successful learning both in patients with OCD (Soref et al., 2018) and in healthy individuals with subclinical OCD symptoms (Barzilay et al., 2022) when instructions are given explicitly, and participants intentionally search for the underlying sequence structure. In fact, our task does not tap into memory processes as strongly as SRT tasks because we explicitly demonstrate the sequence to participants before they begin their 30-day training, which likely decreases demands on procedural learning.

Quantifying trial-to-trial behavioral changes in response to a decrease or increase in reward suggested that the slower progression toward automaticity observed in OCD patients might be related to their more inconsistent response to changes in feedback scores compared to healthy participants. The adjustments that OCD participants made to sequence duration after a score change were more variable (with a larger σ) than those made by healthy individuals. Additionally, the normalized consistency index was higher for the OCD group than for HV, indicating more fluctuating changes from trial to trial in IKI. Despite these group differences, we observed that in both samples the consistency of IKI patterns improved after reward increments. This observation contrasts with the more pronounced MT acceleration in both groups when faced with negative reward changes.

A heightened sensitivity to negative feedback within the motor domain has been documented in the general population, influencing initial motor improvements, while an increase in reward primarily boosts motor retention (Abe et al., 2011; Galea et al., 2015; Pekny et al., 2015; van Mastrigt et al., 2020). OCD individuals have also been shown to have an amplified sensitivity to negative feedback (Becker et al., 2014). Our findings indicate that decreased feedback scores affect sequence duration and IKI consistency in distinct ways. Specifically, reduced score feedback hampered automatization (reducing the IKI consistency, increasing normC), even though it generally had a positive effect on movement speed. This heightened responsiveness of MT (the rewarded variable) to decreased feedback scores is consistent with recent studies (Abe et al., 2011; Galea et al., 2015; Pekny et al., 2015; van Mastrigt et al., 2020). Our results, however, do not support differences in OCD and healthy individuals with regard to sensitivity to negative score changes, unlike previous work highlighting increased response switching after negative feedback, hyperactive monitoring systems, and amplified prediction errors in OCD (Hauser et al., 2017; Marzuki et al., 2021). One possible interpretation of these divergent results relates to the type of feedback. Previous work in OCD employed explicit negative feedback. In contrast, our participants received positive reward feedback, which increased or decreased trial by trial in the continuous reward schedule. The implicit nature of a reduced positive score, which is fundamentally different from overt negative feedback, might not elicit the same heightened sensitivity in OCD patients. They may primarily respond to explicit indications of failure or error, as opposed to subtle reductions in positive feedback. Another possibility is that the salient responses to negative feedback in OCD could be specific to the early stages of learning and may not persist after training for more than 1 hr or on subsequent days. Follow-up work will address these questions explicitly.

Considering the hypothetically greater tendency in OCD to form habitual/automatic actions described earlier (Gillan et al., 2014; Voon et al., 2015), we predicted that OCD patients would attain automaticity faster than healthy controls. This was not the case. In fact, the opposite was found. Since this was the first study to our knowledge assessing action sequence automatization in OCD, our contrary findings may confirm recent suggestions that previous studies were tapping into goal-directed behavior rather than habitual control per se (Gillan et al., 2015b; Vaghi et al., 2019; Zwosta et al., 2018) and may therefore have inferred enhanced habit formation in OCD as a defaulting consequence of impaired goal-directed responding. On the other hand, we are describing here two potential sources of evidence in favor of enhanced habit formation in OCD. First, OCD patients show a bias toward the previously trained, apparently disadvantageous, action sequences. In terms of the discussion above, this could possibly be reinterpreted as a narrowing of goals in OCD (Robbins et al., 2019) underlying compulsive behavior, in favor of its intrinsic outcomes. Second, OCD patients self-reported greater habitual tendencies in both the ‘routine’ and ‘automaticity’ subscales. Previous studies have reported that subjective habitual tendencies are associated with compulsive traits (Ersche et al., 2019; Wuensch et al., 2022) and act, in addition to cognitive inflexibility, as a predictor of subclinical OCD symptomatology in healthy populations (Ramakrishnan et al., 2022). There is an apparent discrepancy between self-reported ‘automaticity’ and the objective measure of automaticity we provided. This may result from a possible mis-labeling of this factor in the Creature of Habit questionnaire, where many of the relevant items indicate automatic S-R elicitation by situational triggering stimuli rather than motor topographic features of the behavior (e.g. ‘when walking past a plate of sweets or biscuits, I can’t resist taking one’).

Finally, we also expected that OCD patients would show a greater resistance than controls in adjusting their behavior when the extrinsic relative value of the trained familiar sequences is diminished, in the re-evaluation procedure. Our findings show that this is partially the case, depending on the type of reward considered. Although we showed that all participants, including OCD patients, were apparently goal-directed in terms of gaining money this was not so clear when goal re-evaluation involved the physical effort expended. In this latter manipulation, participants were less goal-oriented and OCD patients preferred to perform the longer, familiar, to the shorter, novel sequence, thus exhibiting significantly greater habitual tendencies, as compared to controls. Such group differences may be driven hypothetically by the intrinsic value associated with the automatic sequences.

OCD patients engaged significantly more with the Motor Sequencing App and enjoyed it more than HV. Additionally, the patients more prone to routine habits (COHS), with higher OCI scores, and who additionally showed a preference for familiar sequences (possibly by attributing to them intrinsic value), found the use of the app beneficial, exhibiting symptomatic improvement based on the YBOCS. One hypothesis for the therapeutic potential of this motor sequencing training is that the trained action sequences may disrupt OCD compulsions, either via ‘distraction’ or habit ‘replacement’, by engaging the same neural ‘habit circuitry’. This habit ‘replacement’ hypothesis is in line with successful interventions in Tourette syndrome (Hwang et al., 2012), tic disorders (Bate et al., 2011), and trichotillomania (Morris et al., 2013).

As mentioned above, we were unable to employ the often-mooted ‘gold standard’ criterion of resistance to devaluation because it would have invalidated the subsequent tests. This meant that we were unable to conclusively define the trained action sequences as habitual according to the full set of Balleine and Dezfouli, 2019 criteria, although they satisfied other important criteria such as automatic execution, invariant response topography, action chunking and low cognitive load. Nevertheless, the utility of the devaluation criterion has been questioned especially when applied to human studies of habit learning. This is because achieving devaluation can be difficult given that human behavior has multiple goals, some of which may be implicit, and thus difficult to control experimentally, as well as being subject to great individual variation. In fact recent analyses of habitual behavior have not employed devaluation or revaluation as a criterion (Du and Haith, 2023). That study ascertains habits using different criteria and provides supporting evidence for trained action sequences being understood as skills, with both goal-directed and habitual components.

Although we found a significant preference for the trained action sequence in OCD patients in the condition where it was pitted against a simpler and shorter motor sequence, as compared to the monetary discounting condition, the reason for this difference is not immediately obvious. However, it may have arisen because of the nature of the contingencies inherent in these choice tests. Specifically, the ‘monetary discounting’ condition involved a simple deterministic choice between the two alternatives, which should readily be resolved in favor of the option associated with the greater, non-overlapping, range of rewards provided (e.g. 1–7 versus 8–15 gems). In contrast, in the ‘effort discounting’ condition, the reward ranges for the two options were equivalent (e.g. 1–7 gems), which raised uncertainty concerning which of the chosen sequences was optimal. The probabilistic constraint over this choice may therefore account for the greater sensitivity of the task in highlighting preference in OCD, given the greater susceptibility of such patients to uncertainty (Pushkarskaya et al., 2015).

Finally, some of the conclusions relating to the effects of OCD severity on sequence preference without feedback were based only on a post hoc exploratory analysis. Specifically, only those patients with higher compulsivity (OCI) and COHS scores exhibited this preference, therefore consistent with the hypothesis described above of the importance of intrinsic value of the habitual sequence to the development of compulsions. Evidence of this intrinsic value was provided by the greater engagement with, and therapeutic findings for, the app training in these patients. However, the latter effect needs to be confirmed in a registered clinical trial in a controlled manner, which is ongoing.

We employed a battery of behavioral tasks designed to investigate two key hypotheses of the goal/habit imbalance theory of compulsion, specifically pertaining to enhanced habit formation and automaticity and impaired goal re-evaluation in individuals with OCD. Our findings did not support greater habit formation nor heightened automaticity in patients with OCD. Moreover, evidence for patients’ ability to adapt behavior when facing environmental changes was mixed. In certain contexts, OCD patients were able to behaviorally re-adjust (e.g. when reward is monetary) but in others (e.g. when involving motor effort) patients demonstrated a distinct augmented inclination to perform their trained/familiar action sequences, attributing higher intrinsic value to them. Interestingly, this preference was more pronounced in patients with higher compulsivity and habitual tendencies, who engaged significantly more with the motor habit-training app, reporting symptom relief after the experiment. This suggests a promising avenue for investigating the therapeutic potential of this application as a tool for habit reversal in the context of OCD.

We recruited 33 OCD patients and 34 healthy individuals, matched for age, gender, IQ, and years of education. Two participants (one HV and one OCD) were excluded because they did not perform the minimum required training (i.e. two daily practices for a period of 30 days). Therefore, a total of 32 OCD patients (19 females) and 33 healthy participants (19 females) were included in the analysis. Most participants were right-handed (left-handed: four OCD and six HV). Participants’ demographics and clinical characteristics are presented in Table 2 and Figure 10.

Healthy individuals were recruited from the community, were all in good health, unmedicated, and had no history of neurological or psychiatric conditions. Patients with OCD were recruited through an approved advertisement on the OCD action website (https://ocdaction.org.uk/) and local support groups and via clinicians in East Anglia. All patients were screened by a qualified psychiatrist of our team, using the Mini International Neuropsychiatric Inventory (MINI) to confirm the OCD diagnosis and the absence of any comorbid psychiatric conditions. Patients with hoarding symptoms were excluded. Our patient sample comprised 6 unmedicated patients, 20 taking selective serotonin reuptake inhibitors (SSRIs), and 6 on a combined therapy (SSRIs+antipsychotic). OCD symptom severity and characteristics were measured using the YBOCS scale (Goodman et al., 1989), mood status was assessed using the Montgomery-Asberg Depression Rating Scale (MADRS) (Montgomery and Asberg, 1979) and Beck Depression Inventory (BDI) (Beck et al., 1961), anxiety levels were evaluated using the State-Trait Anxiety Inventory (STAI) (Spielberger et al., 1983), and verbal IQ was quantified using the National Adult Reading Test (NART) (Nelson and Willison, 1982). All patients included suffered from OCD and scored >16 on the YBOCS, indicating at least moderate severity. They were also free from any additional axis I disorders. General exclusion criteria for both groups were substance dependence, current depression indexed by scores exceeding 16 on the MADRS, serious neurological or medical illnesses, or head injury. All participants completed additional self-report questionnaires measuring:

1. Impulsiveness: Barratt Impulsiveness Scale (Barratt, 1994)

2. Compulsiveness: Obsessive Compulsive Inventory (Foa et al., 1998) and Compulsive Personality Assessment Scale (Fineberg et al., 2007)

3. Habitual tendencies: Creature of Habit Scale (Ersche et al., 2017)

4. Self-control: Habitual Self-Control Questionnaire (Schroder et al., 2013) and Self-Control Scale (Tangney et al., 2004)

5. Behavioral inhibition and activation: BIS/BAS Scale (Carver and White, 1994)

6. Intolerance of uncertainty (Buhr and Dugas, 2002)

7. Perfectionism: Frost Multidimensional Perfectionism Scale (Frost and Marten, 1990)

8. Stress: Perceived Stress Scale (Cohen et al., 1983)

9. Trait of worry: Penn State Worry Questionnaire (Meyer et al., 1990).

All participants gave written informed consent prior to participation, in accordance with the Declaration of Helsinki, and were financially compensated for their participation. This study was approved by the East of England – Cambridge South Research Ethics Committee (16/EE/0465).

Participant’s characteristics and self-reported questionnaires were analyzed with χ² and independent t-tests, respectively. The Motor Sequencing App automatically uploaded the data to a cloud-based database. This task enabled us to compare patients with OCD and HV in the following measures: training engagement (which included as primary output measures of the total number of practices completed and app engagement as defined as the number of sequences attempted, including both correct and incorrect sequences); procedural learning, automaticity development, sensitivity to reward (see definitions and description of data analyses in Results section); and training effects on symptomatology as measured by the YBOCS difference pre-post training. The Phase B experiments enabled further investigation of preference and re-evaluation strategies. The primary outcome was the number of choices.

Between-group analyses were conducted using Kruskal-Wallis H tests when the normality assumption was violated. Parametric factorial analyses were carried out with ANOVA. Our alpha level of significance was 0.05. On the descriptive statistics, main values are represented as median, and errors are reported as interquartile range unless otherwise stated, due to the non-Gaussian distribution of the datasets. When conducting several tests related to the same hypothesis, or when running several post-hoc tests following factorial effects, we controlled the FDR at level p = 0.05. Significant values after FDR control are denoted by p_FDR. Analyses were performed using Python version 3.7.6 and JASP version 0.14.1.0.

In the case of non-significant effects in the factorial analyses, we assessed the evidence in favor or against the full factorial model relative to the reduced model with BF (ratio BFfull/BFrestricted) using the bayesFactor toolbox (https://github.com/klabhub/bayesFactor, copy archived at Klabhu, 2020) in MATLAB. This toolbox implements tests that are based on multivariate generalizations of Cauchy priors on standardized effects (Rouder et al., 2012). As recommended by Rouder et al., 2012, we defined the restricted models as the full factorial model without one specific main or interaction effect. The ratio BFfull/BFrestricted represents the ratio between the probability of the data being observed under the full model and the probability of the same data under the restricted model. BF values were interpreted following Andraszewicz et al., 2015. The relationship between primary outcomes and clinical measures was calculated using a Pearson correlation.

The diurnal patterns of app use (Figure 2b and c) were assessed in each group using circular statistics (Mardia, 1975), with the ‘circular’ package in R (R version 4.3.1; 2023-06-16). This provided the group-level mean vector length and direction. To assess on the group level whether the daily practice data were uniformly distributed or, alternatively, oriented toward a specific time, we used a Rayleigh test (Landler et al., 2021; Mardia, 1975). We adapted code from Galvez-Pol et al., 2022. To test differences between two circular distributions (OCD, HV), we followed the recommendations of Landler et al., 2021, and employed the high-powered Watson’s U² test, a non-parametric rank-based test (function watson.two.test in R).

The code for the main analyses is provided with this paper. It is available in the Open Science Framework, in the following link: https://osf.io/9xrdz/.

The Motor Sequencing App comprised additional components and analysis, which we describe and report below respectively. Once the minimum daily practice sessions were completed, we additionally asked participants to further conduct a short retention speed test, to assess that day’s performance, and a switching practice session. In the five-trial retention speed session, participants were instructed to repeatedly tap a sequence as rapidly as possible while making as few errors as possible. The 10-trial switch session required switching between the two sequences in a pseudo-random order. The sequence to be played was cued by the respective associated picture. Speed and switch tests never received reward feedback (only the practice sessions). Finally, participants were asked to rate daily, on a percentage scale, the following two questions: (1) How much did you enjoy playing this sequence? and (2) How confident are you that you know this sequence by heart? This sequence of events (practice, speed, switch sessions, and ratings) happened every day.

We conducted a mixed-design repeated measures ANOVA to investigate potential group differences (OCD versus HV) on ratings of confidence (C) and enjoyment (E) over time (4 weeks of app practice). We observed a main effect of time on confidence (F (3, 177) = 92.45, p < 0.001, η_p² = 0.19) but no Group (p = 0.61) or interaction (p = 0.95) effects (Appendix 1—figure 1a). This means that both groups significantly increased their confidence on their sequence knowledge over the course of the training. A Greenhouse-Geisser (ε) correction was applied given that sphericity was violated. Descriptive statistics are as follows (values provided as mean and standard deviation): HV: ${\displaystyle C_{week1}}$ = 64.03 ± 16.09%, ${\displaystyle C_{week2}}$ = 74.90 ± 17.52%, ${\displaystyle C_{week3}}$ = 80.14 ± 14.58%, ${\displaystyle C_{week4}}$ = 85.60 ± 12.90% and OCD: ${\displaystyle C_{week1}}$ = 61.66 ± 20.62%, ${\displaystyle C_{week2}}$ = 72.89 ± 18.40%, ${\displaystyle C_{week3}}$ = 78.55 ± 15.48%, ${\displaystyle C_{week4}}$ = 83.68 ± 14.23%. Regarding the enjoyment ratings, we found no significant main effects of group (p = 0.16), reward (p = 0.45), nor interaction effects (p = 0.25) (Appendix 1—figure 1b). Descriptive statistics are as follows (values provided as mean and standard deviation): HV: ${\displaystyle E_{week1}}$ = 56.98 ± 18.06%, ${\displaystyle E_{week2}}$ = 53.15 ± 24.72%, ${\displaystyle E_{week3}}$ = 51.61 ± 26.24%, ${\displaystyle E_{week4}}$ = 53.70 ± 30.28% and OCD : ${\displaystyle E_{week1}}$ = 59.23 ± 15.77, ${\displaystyle E_{week2}}$ = 61.19 ± 17.78%, ${\displaystyle E_{week3}}$ = 60.49 ± 22.93%, ${\displaystyle E_{week4}}$ = 64.41 ± 26.37%.

Appendix 1—figure 1

Download asset Open asset

Similarly to the learning analysis of the practice sessions, we conducted an individual exponential fitting approach to the switch sessions and assessed between-group differences on performance, both on sequence duration (or movement time [MT]) and reaction time (RT, i.e. time taken to initiate the sequence, latency). No significant group differences were found in any of the three estimated fitting parameters for MT: amount of learning (${\displaystyle M{T}_L}$): U = 395, p = 0.082; learning constant (${\displaystyle n_r}$): U = 575, p = 0.544 and asymptote (${\displaystyle M{T}_0}$): U = 450, p = 0.311 (Appendix 1—figure 2, a). Descriptive statistics are the following (values provided as median and interquartile range): HV: ${\displaystyle \widetilde{M{T}_L}}$ = 1.72 s, IQR = 0.66 s; ${\displaystyle \widetilde{n_r}}$ = 57, IQR = 50; ${\displaystyle \widetilde{M{T}_0}}$ = 1.56 s, IQR = 0.41 s and OCD: ${\displaystyle \widetilde{M{T}_L}}$ = 2.16 s, IQR = 1.24 s; ${\displaystyle \widetilde{n_r}}$ = 49, IQR = 53; ${\displaystyle \widetilde{M{T}_0}}$ = 1.67 s, IQR = 0.45 s. Similarly, no significant group differences were found in the three estimated fitting parameters for RT (Appendix 1—figure 2b): reaction time speed-up achieved over the course of the switch sessions (${\displaystyle R_L}$): U = 470, p = 0.45; reaction rate (${\displaystyle n_R}$): U = 450, p = 0.31 and reaction time at asymptote (${\displaystyle R_0}$): U = 552, p = 0.75 (Appendix 1—figure 2b). Descriptive statistics are the following (values provided as median and interquartile range): HV: ${\displaystyle \widetilde{R_L}}$ = 1.12 s, IQR = 0.72 s; ${\displaystyle \widetilde{n_R}}$ = 100, IQR = 168; ${\displaystyle \widetilde{R_0}}$ = 0.57 s, IQR = 0.50 s and OCD: ${\displaystyle \widetilde{R_L}}$ = 1.08 s, IQR = 0.70 s; ${\displaystyle \widetilde{n_R}}$ = 125, IQR = 129; ${\displaystyle \widetilde{R_0}}$ = 0.56 s, IQR = 0.32 s. Since the switching between the two sequences was done in a pseudo-randomised order, we have also assessed RT restricting our analysis to the specific trials where the switch occurred (i.e. when sequence 1 was followed by sequence 2 and vice versa). No group differences in the switching performance were found, meaning that patients with OCD could switch between the two sequences without any difficulties or slowness.

Appendix 1—figure 2

Download asset Open asset

We also investigated potential group differences in accuracy during the switch sessions. After a similar individual exponential fitting approach, statistical analysis of the three estimated fitting parameters for accuracy indicated no group differences (Appendix 1—figure 2c): amount of learning as measured by accuracy achieved over the course of the switch sessions (Acc_L): U = 538, p = 0.56; learning rate: U = 517, p = 0.77 and asymptote (Acc₀): U = 419, p = 0.29.

We did not analyze the short retention speed sessions because no new components were introduced here, as compared to the practice sessions. Therefore, we did not expect any differences between the practice and the speed sessions.

With the goal of promoting habitual actions, the app was designed to remove the explicit reward feedback (points) on the 21st day of practice (extinction procedure). We analyzed the effects of extinction on accuracy, sequence duration (MT), and reaction time (RT) by comparing the two blocks of practice pre- and post-removal of the rewarding feedback. We analyzed both the practice and switch conditions. While the latter had not previously received reward feedback, it could potentially have been influenced by extinction as a factor. A 2×2 ANOVA with extinction (pre- and post-extinction) as a within-subject factor and group as a between-subject factor indicated no group, extinction, or interaction effects on accuracy, RT, or MT (Appendix 1—figure 3). Statistical results are as follows: (1) practice sessions; variable reward: no effect of group (p = 0.31), extinction (p = 0.28) or interaction (p = 0.57) on accuracy; no effect of group (p = 0.07), extinction (p = 0.99) or interaction (p = 0.61) on MT; no effect of group (p = 0.06), extinction (p = 0.53) or interaction (p = 0.44) on reaction time; continuous reward: no effect of group (p = 0.74), extinction (p = 0.16) or interaction (p = 0.82) on accuracy; no effect of group (p = 0.56), extinction (p = 0.78) or interaction (p = 0.31) on MT; no effect of group (p = 0.55), extinction (p = 0.67) or interaction (p = 0.47) on reaction time; (2) Switch sessions; no effect of group (p = 0.19), extinction (p = 0.17), or interaction (p = 0.74) on accuracy; no effect of group (p = 0.47), extinction (p = 0.89), or interaction (p = 0.27) on MT; no effect of group (p = 0.46), extinction (p = 0.78), or interaction (p = 0.42) on reaction time.

Appendix 1—figure 3

Download asset Open asset

Appendix 1—figure 4

Download asset Open asset

The conditional probability distributions p(∆T|∆R+) and p(∆T |∆R−) were separately fitted to subsamples of the data across continuous reward practices (Appendix 1—figure 5). One practice corresponded to 20 correctly performed sequences. The p(∆T|∆R) distributions for MT (Equation 7) and the normalized consistency index normC were fitted with a non-significantly different number of correct sequences in OCD and in the HV sample (mean 127 [SEM 12] sequences in OCD; 109 [9] sequences in HV; BF < 1/3, moderate evidence against group differences in the number of correct sequences available for this analysis). However, more sequences were available to fit the p(∆T|∆R+) distribution (mean 107 [SEM 10]) than the p(∆T|∆R–) distribution (98 [8]). This outcome reflected that participants overall observed more increases than decrements in feedback scores, as expected. Importantly, matching both reward and group samples in the number of sequences yielded the same ANOVA results as reported in the main manuscript for the center and spread of the normalized ∆MT distribution and normC distribution.

Appendix 1—figure 5

Download asset Open asset

Our analysis protocol was designed to ensure that incorrect trials do not contaminate or confound the results. To estimate the trial-to-trial difference in the normalized ∆MT (or normC) and ∆R, we exclusively included pairs of contiguous trials where participants achieved correct performance and received feedback scores for both trials. For example, if a participant made a performance error on trial 23, we did not include ∆R or ∆MT estimates for the pairs of trials 23-22 and 24-23. Instead of excluding incorrect trials from our analyses, we retained them in our time series but assigned them an NaN (not a number) value in MATLAB. As a result, ∆R and ∆MT were not defined for those two pairs of trials. We followed the same protocol for normC. This approach ensured that our analyses are not confounded by incremental or decremental feedback scores between non-contiguous trials. In our previous work, when assessing the timing of correct actions during skilled sequence performance, we also considered events that were preceded and followed by correct actions. This excluded effects such as post-error slowing from contaminating our results (Ruiz et al., 2009; Bury et al., 2019).

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. For the purpose of Open Access, the authors have applied a CC BY public copyright license to any Author Accepted Manuscript version arising from this submission.

This research was funded by the Wellcome Trust: a Sir Henry Postdoctoral Research Fellowship (Grant 204727/Z/16/Z) to PB and a Wellcome Trust Senior Investigator Award (Grant 104631/Z/14/Z) to TWR. For open access, the author has applied a CC BY public copyright licence to any Author Accepted paper version arising from this submission. MB was supported by MHRUK and Angharad-Dodds Bursaries. AAM was supported as a research assistant funded by the aforementioned Wellcome Trust grant to TWR. We thank all participants for their contributions to this study. We would also like to acknowledge Dr. Sharon Morein-Zamir for fruitful brainstorm discussions on the tasks design.

All participants gave written informed consent prior to participation, in accordance with the Declaration of Helsinki, and were financially compensated for their participation. This study was approved by the East of England - Cambridge South Research Ethics Committee (16/EE/0465).

1. Preprint posted: February 24, 2023 (view preprint)
2. Sent for peer review: March 21, 2023
3. Preprint posted: June 26, 2023 (view preprint)
4. Preprint posted: December 15, 2023 (view preprint)
5. Preprint posted: March 20, 2024 (view preprint)
6. Version of Record published: May 9, 2024 (version 1)

You can cite all versions using the DOI https://doi.org/10.7554/eLife.87346. This DOI represents all versions, and will always resolve to the latest one.

© 2023, Banca et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.