Borderline personality disorder (BPD) is a mental health condition characterized by 9 core symptoms usually involving intense and volatile emotional reactivity, identity disturbance, and interpersonal difficulties; however, symptom presentation is highly heterogeneous [1]. People seeking treatment for BPD represent a significant proportion of those referred to psychiatric clinical services: about 22% of inpatients and 12% of outpatients globally [2]. The demand for BPD treatment vastly outstrips availability, with nearly 6000 people seeking treatment for the condition per evidence-based care provider in the United States. Furthermore, the cost per patient is estimated to be more than double that of patients with depression [3,4].

This treatment gap could be reduced by digital interventions, which inform users and support behavior changes to reduce symptom severity and functional impairment [5,6]. Although few digital interventions are specifically designed for BPD, many address transdiagnostic symptoms, making them potentially effective for BPD treatment due to the disorder’s diverse symptom presentation. For example, many digital interventions have a treatment target of emotion regulation, a core symptom of BPD that is also common in mood, substance use, and eating disorders [7]. People with BPD tend to use less effective emotion regulation strategies than healthy controls [8]. Poor emotion regulation is associated with behavioral control and drives maladaptive behaviors such as impulsive anger and nonsuicidal self-injury (NSSI), which are also symptoms of BPD [9-11]. Paranoia is another transdiagnostic symptom that can be treated effectively, although few interventions focus specifically on paranoia in BPD [12]. Symptom-based care offers treatment without focusing on a specific diagnosis, which could help reduce pathologization—or the unnecessary labeling of illness—in individuals diagnosed with BPD, many of whom are opposed to the label of personality disorder [13]. Digital interventions offer advantages such as access without diagnosis, lower cost, greater anonymity, and continuous availability. Despite concerns about limited clinician contact, self-guided digital interventions can be as effective as traditional face-to-face alternatives [14]. However, several challenges remain, such as nonadherence, limited assessment of efficacy in controlled trials, and poor user experience [15,16].

The importance of user experience in digital health has been demonstrated by Kelders et al [17], who showed that persuasive system design (PSD) was associated with increased adherence to digital interventions. The concept of PSD proposes that user experience and interface elements, detailed in Tables 1 and 2, support behavior change in users [18]. For example, reminders to engage with the intervention and opportunities to rehearse new skills may help users reach their treatment goals by maintaining their adherence to behavioral and cognitive interventions. Research on adherence typically compares characteristics between adherers and nonadherers, thereby putting the burden of adherence onto the user [19]. However, given the evidence for PSD, adherence may be a product of the interaction between the user and the interface rather than a reflection of the users’ characteristics.

The proposed mechanism of action is that PSD elements increase user adherence, which in turn supports greater treatment effect. Different elements have different effects: some may reduce goal behaviors into smaller, manageable changes, while others promote frequent use or lend increased credibility to the intervention [17]. Accordingly, this review aims to (1) identify and describe the therapeutic approach, duration, intended frequency of use, and user interface of digital interventions treating BPD symptoms in community and secondary care settings; (2) conduct a meta-analysis to determine the efficacy of digital interventions compared to active or passive control groups; and (3) determine using subgroup analysis whether there were associations between treatment efficacy and characteristics of the intervention, such as the therapeutic modality, the degree of human support, and the use of PSD elements. Identifying characteristics associated with treatment efficacy will help optimize the design of future digital psychological interventions.

Existing reviews of digital interventions in this area have not thoroughly considered the implementation and user experience of digital interventions, nor have they determined treatment effects for individual symptoms of BPD. Frias et al [31] conducted a scoping review of 15 studies of digital interventions in participants with BPD, which were primarily adjunctive to dialectical behavior therapy (DBT) but did not calculate their treatment effect. A meta-analysis of smartphone apps for BPD-related symptoms did not find a significant treatment effect. However, this meta-analysis pooled outcome measures across symptoms, potentially obscuring item-level changes for individual BPD symptoms [32]. Other reviews have focused on the user experience of these interventions, with a scoping review of 8 interventions, primarily DBT tools, reporting positive user feedback, while another review found serious issues with functionality and interface design, such as the exclusion of DBT diary cards, technical issues with enrollment, and difficulties with navigation [33,34]. This review provides a comprehensive evaluation of digital interventions that target both specific and transdiagnostic symptoms of BPD. Treatment effects for individual symptom measures and an analysis of the effect of user experience features and PSD elements on treatment efficacy are also examined.

OVID Embase, OVID MEDLINE, OVID PsycINFO, and the Cochrane Central Register for Controlled Trials were searched for literature. The search strategy [35], further discussed in Multimedia Appendix 1, was developed in consultation with coauthor EH and consisted of three themes: (1) the core symptoms of BPD, (2) digital health, and (3) intervention studies. The digital health theme was based on a National Institute for Health and Care Excellence (NICE)–validated filter for health care applications, with additional terms added for web-based interventions [36]. Automatic filters were used to remove non-English language studies and those published before the year 2000, as the technology of interest did not exist before that time. The search was first executed on July 19, 2022, and was then re-executed on February 28, 2023, to check for new literature. Furthermore, we performed forward reference checking on the included publications to enhance the literature search.

All records were screened by coauthors JABL and TH independently, and disagreements were settled in consultation with NMM. Rayyan (Rayyan Systems, Inc) was used to manage the screening process [37]. Studies were eligible if they were peer-reviewed, published in English after the year 2000, and assessed an automated digital intervention with a treatment target of BPD or one of its symptoms. Here, automated indicates an intervention for which content is primarily delivered without human support. Given the possibility of preliminary studies in this area, both randomized controlled trials (RCTs) and nonrandomized studies were eligible for inclusion in the descriptive analysis. All participants were aged ≥18 years. They were either healthy volunteers or individuals with a diagnosis of BPD or its common comorbidities of depressive disorders, anxiety disorders, substance abuse disorders, posttraumatic stress disorder, or complex posttraumatic stress disorder. Studies drawing on community samples were included because they tended to recruit participants with mild to moderate symptom severity, which corresponded with the degree of BPD symptom severity that would be reasonable to treat with an automated digital intervention. This aligns with the NICE guidelines for digital interventions, suggesting professional oversight should increase proportionally with clinical risk [38]. Exclusion criteria were as follows: (1) interventions without a primary digital component; (2) nonautomated interventions (eg, therapist-centered treatment via telehealth); (3) interventions requiring equipment inaccessible to the public (eg, functional magnetic resonance imaging or professional-caliber virtual or augmented reality); (4) participants recruited based on a mental health or neurodevelopmental disorder not mentioned in the inclusion criteria; or (5) studies focusing on participants belonging to special groups such as combat veterans, people who are displaced, or law enforcement officers. These groups were excluded because the life experiences of these participants were unlikely to be representative of a community sample of people with BPD.

The risk of bias was assessed with Cochrane’s Risk of Bias 2 for RCTs including cluster-randomized trials and the National Institutes of Health Quality Assessment Tool for pre-post studies for all single-arm studies [39,40]. All studies were separately evaluated by coauthors JABL and TH, and discrepancies were settled by consensus.

The following features were extracted for each intervention: method of delivery, therapeutic approach, overall intended treatment duration and frequency of use, degree of human support, and any PSD elements present. Data were extracted from the published literature and, where possible, supplemented by direct access to publicly available interventions and interviews with authors. A reliability check of the coding of PSD elements was carried out in a random sample of 5 papers, resulting in 91% interrater reliability. Therapeutic approaches were coded individually to describe each study. To facilitate the subgroup analysis, therapeutic approaches were then classified into evidence-based treatments (EBTs) and non-EBTs. EBTs comprised DBT, cognitive behavioral therapy (CBT), mentalization-based treatment, acceptance and commitment therapy, schema therapy, and transference-focused therapy. The degree of human support of the intervention was binary coded for either full automation or some degree of support based on the descriptions provided in the publications. Facilitated interventions ranged from technical support to therapeutic support to adjunct to in-person care.

We used the coding scheme by Kelders et al [17] for PSD elements, based on a framework developed by Oinas-Kukkonen and Harjumaa [18] (Tables 1 and 2). All interventions were coded dichotomously, indicating the presence or absence of primary task, dialogue, and social support elements. In cases where the presence of PSD elements was unclear in the publication, we contacted the authors and accessed the intervention directly where possible. Failing these means, the element was coded as absent.

Treatment effects for the RCTs were calculated using between-group standardized mean differences (SMDs) at the study end point, with Hedges small sample correction applied [41]. Hedges g values of 0.0-0.2 were interpreted as small, 0.21-0.8 as moderate, and 0.81-1.00 as large. Outcome data were extracted directly from the publications or requested from the study lead author. In cases (2/32, 6%) where the authors did not reply or were unable to provide the requested data, the study was excluded from the meta-analysis [42,43].

Six separate meta-analyses were used to assess the effect of treatment for which there were at least 2 studies reporting these as outcomes. These involved the following outcomes: BPD symptom severity assessed via the Zanarini Rating Scale for Borderline Personality Disorder [44] and Borderline Personality Disorder Severity Index [45]; suicidal ideation via the Beck Scale for Suicidal Ideation (BSSI) [46], Suicidal Ideation Attributes Scale [47], Self-Injurious Thoughts and Behaviours Interview–Suicidal Ideation [48], Columbia-Suicide Severity Rating Scale [49], Suicidal Behaviours Questionnaire [50], and Suicide Status Form [51]; NSSI via the Self-Injurious Thoughts and Behaviours Interview–Nonsuicidal Self-Injury Episodes [48]; paranoia via the Paranoia Scale [52] and Adapted Paranoia Checklist [53]; and anger via the Modified Overt Aggression Scale [54] and Trait Anger Scale [55]. As emotion regulation is considered a central mechanism in BPD, a meta-analysis was also conducted with the following outcome measures: the Difficulties with Emotion Regulation Scale in both long and short forms [56] and the German version of the Emotion Regulation Skills Questionnaire [57]. Following the Cochrane Collaboration recommended procedures [58], in cases where multiple different outcome measures of the same symptom were used in a single study, the effect sizes of each measure were calculated and then aggregated before meta-analysis using the approach by Borenstein et al [59].

Random effects models were chosen over fixed-effects as considerable between-study heterogeneity was expected due to differences in control arms [60]. I² was used to estimate between-study heterogeneity, and restricted maximum likelihood was used to estimate heterogeneity variance τ² [61]. The Knapp-Hartung adjustment was applied, given the small number of studies in some of the meta-analyses [62]. Funnel plots were visually inspected for risk of publication bias. Egger tests were conducted in meta-analyses with sufficient studies [63].

Following recommendations from Schwarzer et al [64] and Fu et al [65], in meta-analyses with ≥10 studies, subgroup analyses were conducted to determine the effects of specific intervention elements, provided there were a minimum of 4 studies per subgroup. Intervention features assessed using subgroup analysis included PSD elements, degree of human support, and use of EBT versus non-EBT interventions. τ² was pooled across subgroups because of the small number of studies in some subgroups [59]. The Cochran Q test was used to evaluate the significance in subgroup differences with P<.10 considered significant for the Q test owing to low event rates [66-68]. The Benjamini-Hochberg false discovery rate was used to correct for multiple comparisons [69].

All meta-analytic procedures were done using the R (R Foundation for Statistical Computing) packages: meta (version 6.2.1), dmetar (version 0.1.0), and Mad (version 0.8-3) packages [64,66,70].

This systematic review was preregistered with PROSPERO (CRD42022358270). The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklist can be found in Multimedia Appendix 2. Some deviations were made from the registered protocol: first, gray literature was not included. This decision was made after reviewing the gray literature, as we believed including it could lower the average quality of the studies and increase their overall risk of bias. Furthermore, most of the gray literature was not peer-reviewed and thus did not meet the inclusion criteria for this review. Adherence was not used as an outcome measure of the review because only 10% (4/42) of the included studies described adherence as the number of participants who completed the intervention. Other adherence metrics were reported, but these were inconsistent and could not be organized into a meaningful measure. Finally, some studies listed multiple outcome measures without identifying a primary measure. For this review, all studies reporting at least 1 appropriate outcome measure were included. Studies of interventions with multiple treatment targets and outcome measures were eligible to be included in multiple syntheses. For example, some interventions addressed both suicidal ideation and NSSI and thus are included in both syntheses.

Figure 1 illustrates the PRISMA flowchart of the initial and repeated searches. The first search, executed on July 19, 2022, returned 7531 records, of which 1645 (21.84%) were duplicates. A further 428 (5.68%) were automatically removed due to their year of publication (before 2000) or initial publication in a non-English language. Of the remaining 5458 records, 69 (1.26%) were retrieved for full-text screening and 36 (0.66%) of these were included. A repeated search before analysis with publication dates restricted from the initial search date to February 28, 2023, returned 989 records, of which 665 (67.24%) were duplicates. The remaining 324 records were screened, 22 (6.79%) were selected for full-text screening, and 4 (1.23%) of those were included. The details of all included studies are provided in Tables S1-S6 in Multimedia Appendix 1.

The most common reason for exclusion was an inappropriate outcome measure, including unvalidated measures and subscales. Other reasons for exclusion were lack of peer review, enrollment of participants aged <18 years, interventions that were delivered in person or via telehealth meeting software, and research about combat veterans. A total of 40 reports met the eligibility requirements of this review, 1 (3%) of which detailed 3 studies in the same report, such that 42 unique studies were included. Across these studies, 38 unique digital interventions were assessed and discussed in this review. Some studies assessed the same intervention, for example, translations of Living with Deadly Thoughts into multiple languages [71-75]. One study had 3 arms, each assigned a different intervention [76]. In total, 38 unique interventions are discussed in this review. Of the 42 studies included in the review, 32 (76%) were RCTs, 8 (19%) were single-arm studies (pre-post), 1 (2%) was an open-label trial, and 1 (2%) was a cluster RCT. Four studies reported BPD symptoms or key mechanisms as outcome measures, 20 reported suicidal ideation and suicidal thinking, 5 studies each reported paranoia and NSSI, 4 reported emotion regulation, and 4 reported anger and hostility measures. Some interventions listed multiple outcome measures without indicating a primary. No interventions were identified targeting identity disturbance, impulsivity, feelings of emptiness, fears of abandonment, or relationship instability. Participants were recruited based on a variety of conditions: of the 6611 participants, 3372 (51%) were recruited based on suicidal ideation, 661 (10%) based on NSSI, 397 (6%) on a diagnosis of BPD, 331 (5%) on trait anger or hostility, and 198 (3%) on paranoia. In total, 19% (1256/6611) of the participants were recruited without specific baseline symptoms and 6% (396/6611) were recruited to studies with multiple possible symptom thresholds (eg, they met symptom thresholds for either suicidal ideation or NSSI).

The results of the Cochrane’s Risk of Bias 2.0 and the National Institutes of Health Quality Assessment Tool for each study can be found in Figures S1-S5 in Multimedia Appendix 1. In brief, there was a widespread risk of bias: only 9% (3/32) of the RCTs and 13% (1/8) of the single-arm studies were deemed low risk.

The included studies were primarily conducted in the United States (16/42, 38%), the United Kingdom (6/42, 14%), or Australia (5/42, 12%). The methods of intervention delivery included websites (27/42, 64%), mobile phone apps (10/42, 24%), both (3/42, 7%), and email (2/42, 5%). The most common therapeutic approaches used were based on DBT (8/42, 19%), CBT (6/42, 14%), a combination of CBT and DBT (5/42, 12%), mindfulness (3/42, 7%), and acceptance and commitment therapy (2/42, 5%). Out of the 42 studies, 18 (43%) evaluated other or unspecified therapeutic approaches.

The mean course of treatment, unweighted by the number of participants, was 55 (SD 75) days, with a mean unweighted recommended frequency of use of 5 (SD 2.7) times per week. This frequency calculation excluded 26% (11/42) of the studies with open dosage or intended frequency of use not stated, as well as 10% (4/42) of single-session interventions.

Explanations for each PSD element and examples of relevant interventions are given in Tables 1 and 2, and the frequency of PSD elements used in each of the digital interventions is illustrated in Figure S6 in Multimedia Appendix 1. The presence of PSD elements varied from suggestion, used in 86% (36/42) of the studies to praise, for which we only found evidence in 10% (4/42) of the interventions. Elements of social support were not analyzed, as “e-motion” was the only intervention that allowed users to interact with one another [77].

We first compared interventions targeting suicidal ideation that used EBTs (12/18, 67%) to those that did not (6/18, 33%). EBT interventions were associated with significantly greater treatment effect than non-EBT interventions (SMD –0.21 vs 0; Cochran Q=4.87; P=.03). We then compared suicidal ideation interventions that were fully automated to those that involved at least some degree of human support. There was not a significant difference in treatment effect between these 2 subgroups (automated SMD –0.09 vs facilitated SMD –0.16; Cochran Q=0.37; P=.54). Table 3 reports these results in detail, as well as subgroup analysis comparing treatment effect between interventions that did or did not use each PSD feature. For statistical power reasons, this analysis was only conducted for PSD elements that were present in at least 4 interventions and absent in at least 4 others. Interventions using reminders, opportunities for self-monitoring, and opportunities for rehearsal were associated with significantly greater reductions in suicidal ideation than those without. No significant effects were identified for other PSD elements.

A random effects model fitted to the data found a moderate, significant effect of digital interventions for paranoia (N=4; Hedges g=–0.52, 95% CI –0.86 to –0.18; P=.01). The between-study heterogeneity variance was estimated at τ2=0.04 (95% CI 0-0.55), and between-study heterogeneity was estimated at I2=51.3% (95% CI 0%-82.1%; Figure 6 [87-91]). The funnel plot in Figure 7 [87-91] shows possible evidence of publication bias, but there were too few studies to conduct the test.

Meta-analysis of interventions targeting other symptoms (NSSI, emotion regulation, and anger) did not suggest a significant treatment effect of digital interventions. Full meta-analytic results for these interventions can be found in Figures S7-S12 in Multimedia Appendix 1. Identified interventions are tabulated in Multimedia Appendix 1: NSSI interventions are in Table S4 [27,43,76,86], emotion regulation interventions are in Table S5 [23,29,77,92-94], and anger interventions are in Table S6 [26,95,96].

This review suggests that digital interventions may be efficacious in reducing BPD symptoms of paranoia and suicidal ideation. By contrast, digital interventions addressing emotional regulation, anger, and NSSI did not demonstrate a significant treatment effect. Among interventions targeting suicidal ideation, those based on evidence-based therapeutic modalities for BPD were associated with greater improvement, but effectiveness was not found to be dependent on the degree of human support. Importantly, these findings highlight which symptoms may be better suited for digital interventions and could be candidates for further investigation and implementation. This is particularly relevant in clinically confirmed BPD, where limited access to in-person therapy and high service costs may be barriers to receiving timely care.

A secondary objective of this review was to assess the effect of PSD on treatment efficacy. There is an overlap between the aims of psychological interventions and the PSD framework: both support change in behavior and beliefs. At least some PSD element involvement was nearly ubiquitous among the studies reviewed; changes to thought and behavior (suggestion) and reduction of these changes into component processes (reduction) were used in 79% (33/42) and 86% (36/42) of the reviewed interventions, respectively. Importantly, our meta-analytic review of interventions targeting suicidal ideation found that the specific PSD elements rehearsal, self-monitoring, and reminders were associated with significantly greater treatment effect. Each element may confer different benefits to the design of digital interventions for suicidal ideation and per se parallel the therapeutic components of EBTs for suicidal ideation.

Rehearsal refers to practicing skills in a comfortable environment with the objective of building the habit of using them as needed. Rehearsing a response to periods of intense suicidal ideation may reduce the duration and risk of these events; this is a component in the Collaborative Assessment and Management of Suicidality approach [97]. Indeed, several evidence-based psychological interventions that address suicidal ideation such as DBT and CBT for suicide prevention involve rehearsal of applying the skills learned in therapy in advance of a challenging situation [98,99].

Self-monitoring in the context of PSD refers to keeping users motivated by encouraging reflection on progress toward goals. Both processes may address feelings of hopelessness and difficulty maintaining a course of action, which are associated with suicidal ideation and BPD [100,101]. Furthermore, self-monitoring, often with an emphasis on mood monitoring, is also a core aspect of interventions targeting emotional regulation and suicidality, such as DBT and CBT for suicide prevention.

Reminders are a design element that help reduce forgetful nonadherence [102]. Further work is necessary to understand which types of reminders are most effective for digital interventions. This PSD element is most likely to be digital-intervention specific, with no immediate comparative components in in-person EBTs.

In the context of existing therapeutic frameworks, these PSD features could be leveraged in the development of future digital interventions for suicidal ideation for BPD. Due care should be taken in the design of digital mental health interventions; the inclusion of PSD elements is a good starting point, but user experience should be evaluated and iterated upon, a process with established merits [103].

As in face-to-face settings, digital EBT interventions for suicidal ideation, such as CBT and DBT, were more effective than alternative therapeutic approaches [104]. EBTs appear to be effective for reasons that transcend delivery method, perhaps by introducing active components that directly facilitate treatment goals, such as behavioral activation in CBT, rather than general nondirective support [105]. Despite the demonstrated efficacy of EBTs, some included interventions used promising alternative therapeutic approaches, such as stimulus pairing or journaling [27,76].

Efficacy of digital intervention may also vary depending on the treatment target. Included visualization interventions were effective, at least temporarily, in reducing symptoms of paranoia. Paranoia in BPD is transient, meaning immediate symptom reduction is relevant even if temporary. Recent work with virtual reality in community and psychosis samples has found that compassionate imagery, cognitive therapy, and mental relaxation exercises all appear to be effective in the reduction of paranoia symptoms [106,107]. The commonality between these approaches is reduced anxious cognitions, which likely underpin paranoia and form a suitable target for digital interventions [108,109]. All included paranoia interventions were brief (ranging from 1 session to 2 weeks). The NICE guidance for BPD advises against brief interventions for BPD partially because few brief EBTs were available at the time of the original 2009 publication, and partially because of the possibility of negative reaction upon withdrawal of therapist support [110]. The latter issue is less applicable to automated digital interventions, particularly if they remain accessible after treatment. Given the efficacy of paranoia interventions in this review, and as brief interventions minimize the risk of user attrition over time, we suggest that further research should evaluate brief digital interventions for symptom relief and teaching specific skills. Such interventions could be deployed in adjunct or anticipation of in-person treatment.

Suicidal ideation also appears feasible as a digital treatment target. For example, Living with Deadly Thoughts, an intervention based on the principles of CBT and DBT that has been translated into multiple languages, has been found effective in several large RCTs. Collectively, the studies of Living with Deadly Thoughts contributed 35% (1604/4583) of the meta-analytical suicidal ideation results. The intervention focuses on reducing negative automatic thoughts, which are associated with suicidal ideation through several mediation pathways [111]. Reduction in negative automatic thoughts after CBT is associated with reduced suicidal ideation [112], a treatment pathway that appears to have been successfully digitally adapted in this series of interventions. For interventions targeting paranoia and suicidal ideation, cognitive therapy components may be best suited to digital formats. These components can help reduce anxious thoughts in paranoia and negative automatic thoughts in suicidal ideation.

Unlike paranoia and suicidal ideation, meta-analytic evidence did not show improvement in emotion regulation, anger, or NSSI after digital intervention. DBT, the gold standard treatment for BPD, has been shown to improve these symptoms likely by improving behavioral control [9,10]. Behavioral control supports emotion regulation, including impulsive anger, which is common in BPD, and, in turn, predicts NSSI behaviors [9,11,113]. DBT typically involves protracted membership in a therapeutic group, through which patients learn to maintain relationships with peers. This process of supervised exposure and interaction with others is not replicable in automated digital interventions but may be essential to developing behavioral control. As a result, treatment targets of emotion regulation, anger, and NSSI may be less suited to the affordances of digital interventions than targets that can be effectively addressed through cognitive therapies, such as paranoia and suicidal ideation [12,114].

Previous research has found that human-supported digital mental health interventions are more effective than fully automated ones [14,115]. This was not replicated in our subgroup analysis of suicidal ideation interventions; however, human support could deter engagement with suicidal ideation treatments due to stigma and shame, thus reducing treatment efficacy for human-supported interventions [116]. This may not be true for other treatment targets; nonetheless, the effects of anonymity and human support should be weighed carefully in the development of digital interventions.

Our results suggest that effective digital interventions focus on building specific skills through transdiagnostic evidence-based approaches or provide symptom management tools [47]. This review did not find evidence that digital interventions should attempt comprehensive BPD treatment, as these interventions likely require support from a trained clinician and repeated opportunities to build emotion regulation skills among peers. The affordances of digital interventions are likely better suited to targeted, single-symptom treatments that could be deployed adjunctively to in-person care. For people with BPD, this type of intervention may be less pathologizing as the goal is to treat the symptoms of the individual rather than the disorder. The single-symptom approach is increasingly relevant as the diagnosis of personality disorder moves toward a dimensional model [117]. For users with and without personality disorders, the brevity and content relevance of transdiagnostic single-symptom interventions may improve adherence to digital interventions relative to extended courses of comprehensive treatment. Change in single-symptom severity may also be easier to measure because single-symptom measures provide more granularity than composite diagnostic measures [118].

There are several limitations to this review, the first being the limited applicability of findings across symptoms. For example, the association between PSD elements and treatment effect was only analyzed in interventions for suicidal ideation and may not hold for all treatment targets. In addition, while the scope of this review includes interventions for BPD symptoms, many of the included studies draw on non-BPD samples. As a result, our findings address the optimization of digital interventions for symptom reduction, rather than BPD psychopathology specifically. These results remain relevant to BPD given that treatment mechanisms in the disorder are poorly understood, and there is little evidence to suggest that efficacy in existing treatments is specific to BPD rather than a general reduction in psychiatric distress [119]. This work may also have been restricted by a lack of statistical power in some analyses, particularly the smaller meta-analyses of treatment effect and the subgroup analyses [66]. In addition, the binary coding of interventions into human-supported and fully automated subgroups may have obscured some of the effect of human support. For example, interventions that only offered as-needed support were grouped with interventions that included regular support, although the former might not impact the treatment effect as much as the latter. All negative subgroup results should be interpreted as an absence of evidence for efficacy rather than evidence of inefficacy. The meta-analyses and subgroup analyses also depend on the quality, completeness, and accuracy of the data in the included studies. As discussed, many studies were at risk of bias and had considerable loss to follow-up. Finally, when intervention descriptions were ambiguous or absent and authors did not reply to our queries, PSD elements were coded conservatively. This may have lowered the frequency estimates of PSD elements and the accuracy of the subgroup analyses.

This review has found meta-analytic evidence that digital interventions, particularly cognitive interventions, are effective at reducing the BPD symptoms of paranoia and suicidal ideation. The most effective suicidal ideation interventions were those that used evidence-based therapies and certain features of PSD. Automated digital interventions are unlikely to replace face-to-face care as primary interventions for BPD because they do not facilitate sufficient opportunities to develop emotion regulation skills. However, digital interventions still represent an opportunity for highly accessible symptom-specific BPD treatment.