Previous experimental research on musical improvisation has focused on individual performers (e.g., Keller et al., 2011; Norgaard, 2011, 2014). More recently, researchers have emphasized the interaction in joint improvisation as an ecologically valid domain to investigate interpersonal processes, and spontaneous coordinated behavior such as interpersonal synchronization (IPS) in particular (Keller et al., 2014; Walton et al., 2015). In a musical joint improvisation, the playing behavior emerges within a context of social collaboration, and without musical scores. Joint musical improvisation is a complex interaction to study, but Jeung and Herpertz (2014) stress the importance of socially complex stimuli to study interpersonal processes in patients with BPD.

Fundamental to the interactions involved in joint musical improvisation are affective and temporal IPS (Iyer, 2004; De Backer and Foubert, 2011; Hennig, 2014). Affective IPS in musical improvisations consists of shared moments that are important in changing the relationship and moving it to a deeper level of intersubjectivity within a therapeutic process. There have been a number of studies concerning affective IPS, addressing synchronicity (De Backer, 2008), meaningful moments (Amir, 1996), significant moments (Trondalen, 2006), affect attunement (Trondalen and Skårderud, 2007), and inter-affective synchronization (Schumacher and Calvet, 2007).

In this study, we will focus on temporal IPS. Temporal IPS entails the capacity to plan and execute specific actions at precise times, in relation to other performers. People can synchronize spontaneously, such as when people start to walk unintentionally in the same gait cadence. Other forms of temporal IPS can be intentional, for instance when dancers attune their movements to those of a partner. Temporal synchronization in a joint action is generally assessed based on measurements of “asynchronies” or timing deviations between people (Mills et al., 2015).

Experimental research in the normal healthy population demonstrates a strong relationship between the quality of temporal IPS in (musical) joint action and experiences related to social cohesion (Marsh et al., 2009), cooperation (Anshel and Kipper, 1988; Wiltermuth and Heath, 2009), bonding and attachment (Hove and Risen, 2009; Wheatley et al., 2012). As for the BPD population, individuals appear to cooperate less in an experimentally manipulated interpersonal context than do controls (Lazarus et al., 2014). Further, BPD individuals are likely to have more difficulties in repair of relationship ruptures than controls (King-Casas et al., 2008). Ruptures in cooperation seem to be associated with diminished trust in the interacting partner (Seres et al., 2009; Unoka et al., 2009). Finally, oxytocin, a neuropeptide known to enhance cooperation and prosocial behavior for instance in musical joint action (e.g., Grape et al., 2002), may have paradoxical effects for BPD individuals. For example, a study of Bartz et al. (2010) showed that intranasal administration of oxytocin did not have its normal trust facilitating effects in response to a hypothetical partner cooperation in BPD individuals.

From a theoretical viewpoint, BPD is typically characterized by disturbed attachment (Agrawal et al., 2004; Gunderson and Lyons-Ruth, 2008; Beeney et al., 2016). According to attachment theory (Bowlby, 1988), the quality of relationships, such as measured by child-caregiver IPS, results in the development of mental representations, including beliefs about the self, expectations about interpersonal relationships and their quality, all of which act as prototypes or attachment patterns (e.g., secure/insecure) in later adult social interactions (Fraley, 2002; Shaver and Mikulincer, 2005; Scott et al., 2009; Lindsey and Caldera, 2014). This attachment theory is supported by research suggesting that the quality of child-caregiver IPS is critical to the emergence of other socio-cognitive and socio-affective abilities (Crandell et al., 2003; Feldman, 2007a; Newman et al., 2007; Feldman, 2007b, 2012; Gratier, 2009; Hobson et al., 2009; Guedeney et al., 2011; Kiel et al., 2011; Kleinspehn-Ammerlahn et al., 2011; Dumas et al., 2014).

Empirical research shows that BPD patients have difficulties in maintaining close relationships, and attachment relationships in particular (e.g., romantic partner, Melges and Swartz, 1989; Levy, 2005; Gunderson and Lyons-Ruth, 2008; Choi-Kain et al., 2009; Fonagy and Luyten, 2009; Beckes and Coan, 2011; Levy et al., 2015; Beeney et al., 2016). Difficulties in attachment relationships are characterized by oscillations between opposing fears of abandonment and dependency, between neediness and angry withdrawal (Melges and Swartz, 1989). This leads to unstable relationships and difficulties in maintaining strong attachments with others (Bodner et al., 2011). For example, a recent study by Lazarus and Cheavens (2016) found that women with BPD reported more relationship ruptures within the previous month compared to healthy control women.

Based on attachment theory and associated empirical research, we predict that in our study involving an accompanied musical improvisation, the (insecure) attachment system will be activated in BPD patients, and this will affect temporal IPS between therapist and BPD patients. More specifically, we predict:

poorer temporal IPS, represented by higher overall timing deviations, for BPD patients compared to normal controls;

Additionally, from the perspective of BPD pathology, we assume that impulsivity, a core feature of BPD, will influence temporal IPS in a joint musical improvisation. Impulsivity is one of the 9 diagnostic criteria in the Diagnostic and Statistical Manual of Mental Disorders, 4th edition (DSM-IV; American Psychiatric Association, 1994). In the literature, BPD is often described and conceptualized as a disorder characterized by high levels of impulsivity (Silk, 2000; Depue and Lenzenweger, 2001; Widiger and Costa, 2002; Scott et al., 2009). These findings motivate a further prediction for our own study, that (4) BPD patients will play in a more impulsive manner than normal controls. In other words, we predict that BPD individuals will be less inhibited in their playing than normal controls and adapting to the therapist's playing more readily.

A sample of 16 carefully screened BPD patients and 12 matched normal controls participated in the study. Participants in the BPD group were patients consecutively admitted in the psychiatric hospital UPC KULeuven, Kortenberg (Belgium), who met the following inclusion criteria: (a) a primary diagnosis of BPD according to the structured clinical interview for Diagnostic and Statistical Manual of Mental Disorders, fourth edition (DSM–IV) Axis II disorders (SCID-II), (b) age between 21 and 60 years, and (c) not participated in music therapy sessions previously. Twenty eight patients were screened for BPD, 20 patients who fulfilled the inclusion criteria were asked to participate in the study, and, subsequently, 16 patients confirmed willingness to do so.

Participants in the matching normal control sample were recruited from the community, based on characteristics from the BPD group. Control participants were pairwise-matched with the BPD sample on gender, age, level of education, level of musical education, and musical principal instrument. Sixteen potential participants were asked and agreed to participate. Four of these participants were excluded because they fulfilled criteria for at least one personality disorder based on the Assessment of DSM-IV personality Disorders (ADP-IV), a self-report questionnaire for personality pathology (see below).

In summary, we collected and carried forward to the analysis data from 16 BPD patients and12 matched controls. The absence of matches for four BPD patients was not of particular concern, because from a methodological standpoint, even if matching has been performed, this does not necessitate a matched analysis (Pearce, 2016).

The study was carried out in accordance with the recommendations of the local ethics committee, UPC KULeuven, and the central ethics committee, UZ KULeuven. All subjects gave written informed consent in accordance with the Declaration of Helsinki. After being provided with the necessary information, all participants (BPD group and matched normal control group) signed informed consent forms and were given an appointment to participate in the improvisation within 4 days. After completing the musical improvisation, participants were asked to fill out the questionnaires as detailed below.

Most of the BPD patients were female (12 female; 3 male; 1 transgender). The mean age was 31 years (sd = 9.41; range 21–51). Three patients completed primary high school, four secondary high school, six higher education (professional bachelor), and three higher education (academic master). Seven patients indicated experience of playing a musical instrument, among which five patients indicated that they had received musical education and two patients described themselves as autodidacts. Two patients had 1 year of musical education; two patients had 2 years of musical education; one patient had 7 years of musical education; one patient had one and a half years autodidactic experience; and one patient had 7 years autodidactic experience. Three patients indicated voice as principal musical instrument, two patients played guitar, one patient piano, and one patient flute.

At the time of the study, 50% of the BPD patients were receiving psychotropic medication. Most patients were using more than one type of psychotropic medication (n = 5); only three patients (19%) were using one type of psychotropic medication.

Both BPD patients and normal controls completed (a) a listening test of beat perception, (b) a questionnaire assessing age, gender, educational level, musical principal instrument, musical educational level, music therapy history, motoric restrictions, hearing problems, and sensitivity for sound, (c) self report measures of musical sophistication to analyse the confounding influence of musical experiences (d) self report measures of attachment (see below). Data on depression and current psychotropic medication were only gathered for the BPD group. Our reasoning for testing the influence of depression was because depression has been hypothesized as a possible confounder in interpersonal functioning in BPD (Fonagy and Bateman, 2008; Lowyck et al., 2016). Current psychotropic medication was gathered based on the medical records of the BPD patients.

We use a novel, structured piano improvisation paradigm distinguishing between two different accompaniment frameworks—a predictable repetitive interaction, and a more dynamic, socially complex interaction. The therapist's accompaniment was designed to be in a three-part ABA′ structure (see Figure 1): In part A, the accompanist played a single low note that sounded for one beat before a two-note chord was played and both were sustained for three beats to make up a four-beat pattern (see the staff notation below A in Figure 1). The musical term for this type of accompaniment figure is “bourdon,” and it was repeated throughout section A at a steady tempo; as implied by the label A', this bourdon pattern returned in the third section of the therapist's accompaniment; the content of section B was somewhat freer, but it generally contained an increase in tempo and dynamic level, as well as a change from Phrygian to Aeolian modes.

The rationale for this accompaniment design is that in the more dynamic B part of the improvisation, the interaction comes to the fore. Our premise is that in part B, the attachment system will be more activated in BPD patients than in either parts A or A'. As such, differences in temporal IPS between patients and controls may well be revealed in the B part of the accompanied improvisation. The most convenient way to determine whether changes in IPS have occurred within part B is to split the music data for this section in two, B1 and B2, and calculate variables for these subparts separately. In experiments on visual working memory (e.g., Brady et al., 2009), it is quite common to establish regularities in stimuli, upon which participants may come to rely in order to improve task performance, before subverting those regularities and measuring participants' sensitivities. Our ABA′ accompaniment structure, where A establishes the regular bourdon and B subverts it, can be seen as a less common and therefore relatively novel musical analog of experimental paradigms that establish and then subvert regularities in order to measure participants' sensitivities.

As mentioned above, we made use of an ABA′ structured piano improvisation, distinguishing between two different accompaniment frameworks. In the next section, we will give more music-theoretic details related to our improvisation design.

In the previous section, we stated the potential utility of MIR variables for investigating how aspects of a participant's piano playing vary across an accompanied improvisation. In the interests of clarity, we defer details of the music data processing and mathematical definitions of all MIR variables to Supplementary Materials. In brief, the music data was beat-tracked by a professional musician/music therapist, and each improvisation was then quantized automatically using the Lisp package MCStylistic and Matlab package PattDisc (Collins, 2011). The purpose of these steps (beat-tracking and quantization) is to map and/or compare each performed note to a start time (called ontime) commensurate with how it would be written in staff notation, as a basis for measuring participants' IPS. The variables we considered are shown in Table 1 (see Supplementary Materials for full definitions). Below we mention only those that became most relevant in our analyses. To avoid our analyses becoming too exploratory, we employed a common, principled variable selection technique called stepwise selection. As can be seen from Table 1, the focus was on variables associated with IPS (seven out of 15), but for the sake of thoroughness we included several from other categories (tempo, rhythmic patterns, and interpersonal imitation) that were either straightforward (Occam's razor) or could be obliquely related to IPS.

(1) MD_m stands for mean metrical deviation. This calculates the average deviation between each note performed by a participant and the underlying eighth-note beat of the therapist to which it is closest. It is indicated by the blue horizontal lines in Figure 3. The larger the value of MD_m, the more the participant deviates from the beat over an improvisation section, and the more “out of time” their playing will sound. We use this as one operational definition for the participant's overall asynchronies.

(2) The above variable says nothing about whether the participant tends to play ahead of or behind the beat, or leader-follower behavior. The variable LP, standing for lag proportion, is the proportion of times that a participant's notes are behind the beat over the course of an improvisation section. If a participant is always ahead of the beat, then LP = 0, and thus the participant shows more leader behavior; if a participant is always behind the beat, then LP = 1, and thus the participant shows more follower behavior.

(7) The variable MDB_sd, standing for standard deviation of metrical deviations behind the beat, measures the consistency of timing deviations of those notes played late by the participant. If a participant tends to play late (behind the beat) in a consistent manner, then this variable will take a relatively small value; if a participant tends to play late in an erratic manner, then this variable will take a relatively large value.

(10) CR_dur stands for compression ratio applied to (ontime, duration)-pairs. Existing work posits that the more it is possible to compress data, the more structure or patterning the original data contains (Collins et al., 2010, 2016; Collins and Meredith, 2013). The more rhythmic motifs or patterns in a participant's playing, the more their corresponding (ontime, duration)-point set tends to be compressible, and the higher the compression ratio will be.

(12) The variable RS, standing for rhythmic simplicity, measures the prevalence of a participant's most prevalent rhythm. We tally their inter-onset times (the time differences between the notes played), determine their modal (most prevalent) time difference, and define RS as the proportion of all time differences that belong to this mode. If a participant plays only isochronous (evenly spaced) notes (possibly of differing pitches), then their RS = 1; if a participant plays n notes such that the time difference between two consecutive notes is never the same, then RS = 1/n, i.e., is close to 0.

Variables were calculated from the separate parts of the accompanist's ABA′ structure, with the additional bisection of section B into B1 and B2 (on the basis of the overall duration of part B), to enable investigation of participant sensitivities to the changes in musical content at the beginning of section B.

The piano improvisation was recorded using a Yamaha Disklavier MPX70 piano. Each key was connected with a specially designed optical sensor, and these were connected to a USB MIDI interface (Motu Midi Express 128). Improvisations were recorded with Logic Pro X (Mac system) and exported as MIDI files for subsequent analyses. The MIR variables were calculated in Matlab, and R was used for conducting statistical analyses.

Participants were asked to play intuitively and freely on the piano's white keys, without playing well-known songs, but with the aim of exploring joint interaction with the accompanist. They were informed about the ABA′ structure of the improvisation.

The accompanist was a senior registered music therapist, and undertook all the improvisations. The accompanist was blind in the sense that no knowledge about the background of the participants (control or BPD group) was known. The accompanist had 35 years of clinical experience, was experienced in the use of clinical improvisation, and had expertise in the therapeutic musical interventions described by De Backer et al. (2014).

The piano was chosen as an instrument based on a previous study about choices of musical instruments used in individual music therapy sessions with BPD patients (De Backer et al., 2016). In that study, seven Belgian music therapists were asked to fill in questionnaires about the musical instruments chosen by BPD patients in individual sessions over a period of 1 year. Piano was the most frequently used instrument in this population.

The accompanist was sitting on the left side of the piano and played the lower registers of the keyboard. The participant was sitting on the right side, and was playing the upper registers (as shown in Figure 4). This setting was based on the concept of the left and right hand position within music therapy (De Backer et al., 2014)—that the therapist can sustain and support (harmonically) the play of the participant. The keyboard had a split point on G4, which enabled (mostly) convenient splitting of the therapist and participant's playing in Logic.

In terms of statistical analyses, we conducted logistic regressions on a dependent variable of BPD (patient = 1, control = 0), using independent variables as described in the section “MIR variables.” In other words, we investigated whether it was possible to predict the category of a given participant (BPD or control), based solely on quantifications of their playing habits. Initially, we employed a stepwise selection procedure. Incorporated in the first stage of stepwise selection is a comparative assessment of the discriminative power of each variable in isolation. Supplementary Figure 2 shows the distribution for each variable, split into patients and controls, and that there was only one variable (lag proportion in section B1, or LP_B1) with significant differences between BPD_patients and controls, [t_(25.04) = −2.32, p = 0.029]. As such, LP_B1 is the first variable to enter the model.

While rigorous, stepwise selection and the inclusion of further variables tended to result in overfitting of the data and coefficient blowup.² En route to overfitting, however, we identified a parsimonious model that provided strong predictions of BPD or control, as summarized in the following equation and Table 2A. The model consists of mean metrical deviation in section B1 (MD_m_B1) and mean metrical deviation in B2 (MD_m_B2): y=3.35-257.44 MD_m_B1+210.28 MD_m_B2+ε where y is the log odds of having BPD (patient ≈ 1, control ≈ 0) and ε is an error term. Nagelkerke's R² = 0.57 for this model, and the Hosmer-Lemeshow test indicates that the actual diagnoses (patient or control) are not significantly different from those predicted by the model, χ²(8) = 5.31, p = 0.72. The signs of the coefficients, −257.44 and +210.28, are opposite, suggesting it is the difference between metrical deviation in sections B1 and B2 that drives prediction of borderline personality disorder. On further inspection, patients' metrical deviations either tended to become bigger in section B2 than in B1, or remain the same, meaning their log odds of having BPD was driven toward 1 in the above formula by the constant term being not much reduced by −257.44 × MD_m_B1 + 210.28 × MD_m_B2. Controls, on the other hand, had smaller metrical deviation in B2 than in B1, meaning their log odds of having BPD was driven toward 0 by −257.44 × MD_m_B1 being negative and of greater magnitude than 210.28 × MD_m_B2.

A plot of metrical deviation in section B1 is shown in blue in Figure 6, metrical deviation in section B2 is shown in red, the difference MD_m_B2−MD_m_B1 is shown in green, and a dashed black line indicates how the difference acts as an effective discriminator between BPD patient and control. All but three patients have a difference above the cut off and all but two controls have a difference below. While the difference may seem small (e.g., cut off is ~4 ms), the pattern in results in Figure 6 is clear: The blue line is always above the red for controls, but not always so for BPD patients.

Applying a leave-one-out cross-validation procedure with this model, we found a prediction error of 0.18. In other words, this model successfully predicts whether a given participant has or does not have BPD in 23 (= 0.18 × 28) out of 28 cases (82% success). The chance that a baseline (guessing) model succeeds in predicting more than 18 cases is less than 0.05 [P(B_{n = 28, p = 0.5} > 18) = 0.043]. Even if we “assist” the baseline model further, by including the knowledge that the proportion p of BPD patients is 16/28, success in predicting more than 20 cases is less than 0.05 [P(B_{n = 28, p = 16/28} > 20) = 0.040]. That is, our predictive model for BPD performs significantly better than chance.

As suggested by the first stage of stepwise selection, another model that might provide strong predictions of BPD is based on lag proportion in section B1, LP_B1 (Table 2B). Nagelkerke's R² = 0.22 for this model, and the Hosmer-Lemeshow test indicates that the actual diagnoses are not significantly different from those predicted by the model, χ²(8) = 4.77, p = 0.78. The negative coefficient on LP_B1 suggests that the more a participant lags behind the beat in section B1, the more likely that participant is to be a control. AIC was not as good for the lag variable (= 37.16) as for the model in Table 2A (AIC = 28.80), however, and also prediction error on cross-validation was worse (0.24). To investigate whether we might improve the lag proportion model further, we built a third model based on it and variation in playing behind the beat in section B1 (MDB_sd_B1, see Table 2C). Nagelkerke's R² = 0.34 for this model, and the Hosmer-Lemeshow test indicates that the actual diagnoses are not significantly different from those predicted by the model, χ²(8) = 4.45, p = 0.81. This was motivated by seeding a stepwise selection procedure with LP_B1 and including the strongest predicting variable in the next stage, which happened to be MDB_sd_B1. Despite the inclusion of an extra variable, AIC (= 36.77) was not as low as for the metrical deviation model in Table 2A (AIC = 28.80). The MDB_sd_B1 variable was not significant in its own right (p = 0.116 in Table 2C) and prediction error on cross-validation was worse (0.23).

Overall, therefore, we recommend the metrical deviation model as a parsimonious and, according to cross-validation, robust predictor for BPD. As described above, we explored various possibilities in an attempt to find a better model. Now we use the t-test results mentioned briefly at the beginning of this subsection to address questions of significant differences between patients and controls for the metrical deviation (MD_m_B1) and lag (LP_B1) variables: (1) is there a significant difference in MD_m_B1 between BPD patients and controls? According to Welch's two-sample t-test, there is no significant difference [t_(19.28) = −1.49, p = 0.153]. If we restrict the data to matched participants so that we can conduct a (generally more powerful) paired t-test, still there is no significant difference [t₍₁₁₎ = −1.09, p = 0.297]; (2) is there a significant difference in LP_B1 between BPD_patients and controls? As stated previously, there is a significant difference [t_(25.04) = −2.32, p = 0.029], with controls lagging significantly more in section B1 than do BPD patients. In summary, when the music accompaniment changes markedly in section B1, BPD patients do not play significantly less or more in time than do controls, but controls do tend to lag behind the beat more often than do BPD patients. As a final remark in this results section, we point out that in the second stage of a stepwise selection procedure seeded with LP_B1, there are other interesting variables that could make significant improvements to the model (e.g., rhythmic simplicity in section B or RS_B, compression ratio of ontime-duration pairs in sections A and A' or CR_dur_A, CR_dur_A'). These variables did not contribute as significantly as MDB_sd_B1, however, so we did not explore them further, but they could be investigated by the interested reader via the URL given in the caption of Figure 5.

Based on the significant findings of the metrical deviation model, we calculated a new variable “MD_m_B1−MD_m_B2,” which describes the behavior of improvement in IPS during the B part of the improvisation. The interpretation of this measure is that positive scores on this measure imply a trend in improving IPS behavior. A negative score implies a trend in worsening IPS behavior.

We conducted a series of separate analyses to test possible influences of medication, severity of depression, and musical capacities on IPS behavior. To measure musical capacities, we assessed both perceptual and experiential capacities. Finally, we included a psychological measurement of general attachment, to explore if the variance in “MD_m_B1−MD_m_B2” could be at least partially explained by the underlying fundamental two dimensions of attachment: Avoidance and anxiety.

With a lifetime prevalence of 5.9% and serious consequences for emotion regulation, impulse control, and interpersonal relationships, BPD is a condition that has been and remains an important subject of research in psychology and neuroscience. Whereas, existing research on BPD—such as self-report questionnaires and unidirectional studies—has not focused on measuring the exchange of social signals between individuals, we attempted to do so via the use of an ABA′ accompanied improvisation paradigm. Impairments in IPS are a known characteristic of BPD, and this paradigm made it possible to measure timing habits in IPS (e.g., the exchange of social signals) over the course of the musical interaction. In the B part of the improvisation (freer improvisation), the intervention from the therapist (accompaniment) invited a greater degree of (social) interaction from the participant. We quantified 15 aspects of each participant's playing across the improvisation sections (A, B, split also into B1and B2, and finally A′), focusing on temporal characteristics that may act as behavioral measures of IPS, as well as some aspects intended to measure impulsivity.

Our main predictions were that there would be: (1) poorer temporal IPS, represented by higher overall timing deviations, for BPD patients compared to normal controls; (2) more oscillations (e.g., more variability) in timing deviations between therapist and BPD patients compared to normal controls; (3) problems in maintaining and improving IPS between therapist and BPD individuals in the course of the joint improvisation compared to normal controls; (4) more impulsivity (less inhibition) in the playing of BPD patients than normal controls. Among our main findings were that: (i) the control group showed significant improvements in IPS over the course of section B2 (variable name MD_m_B2) compared with section B1 (MD_m_B1) of the improvisation, contrary to the BPD group who showed less improvement in IPS over the course of part B (freer improvisation). This finding substantiates prediction 3; (ii) normal controls were significantly more likely to play behind the beat in section B1 (variable name LP_B1) than were BPD individuals, which substantiates prediction 4; (iii) a logistic regression model built on the difference in mean metrical deviation between sections B1 and B2 performed significantly better than chance at categorizing given participants as either having BPD or being a control (82% success rate). So while there was not clear evidence to support predictions 1 and 2 in our findings, we did find evidence to substantiate predictions 3 and 4, as well as a model whose discriminatory power suggests that our behavioral measures of IPS are relevant to the diagnosis of BPD.

This study has some limitations. First our sample is too small to claim any generalizability of our findings. Second, the individuals with BPD in our sample group participated in the context of an inpatient treatment facility, so our results may not be generalizable to other BPD patients. Further research in bigger and other samples is needed. Third, the beat tracking in this study was done manually, which may cause subjective interferences. Apart from using a beat-tracking algorithm (which may be error prone and so require manual, subjective corrections anyway), we suggest synchronization tasks with computer-generated pacing signals as a possible means of reducing subjectivity. While such tasks miss the human deviations (i.e., variations in timbre and intensity) in the joint improvisation, studies on action simulation have demonstrated that even very reduced stimuli can be experienced as a human product with social meaning (Steinbeis and Koelsch, 2009; Sevdalis and Keller, 2011). Finally, we included only the subscale of the general attachment style in this study. In future research, the assessment of attachment style across several distinct relationships is necessary to gain better insights into attachment-related correlations.

In terms of assessment and diagnosis of BPD, the methods developed in our study could become part of a comparatively lightweight tool to detect possible cases of BPD, in order to reduce the need to administer more onerous questionnaires. With regards treatment of BPD, longitudinal research (such as randomized controlled trials) in music therapy is needed to investigate the extent to which improvement in implicit interpersonal processes of IPS is correlated with the improvement of affectively oriented interpersonal functioning in BPD. In terms of both assessment and treatment, recent improvements in machine learning techniques and musical improvisation invite the possibility that accompaniments for participants could be computer-generated in real time, and thus musical parameters of the accompanist could be controlled more exhaustively, which might lead to more objective measures of (the development of) a participant or patient's musical behavior.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.