Assessing the structure of mental health service networks in Belgium: implications for recovery-oriented systems of care

Background

Given the chronic and complex nature of mental disorders (MD), which often require long-term coordinated interventions involving multiple care providers, ensuring continuity of care is a key challenge for mental healthcare systems [1–3]. Care continuity is a multi-dimensional concept defined as the “orderly, uninterrupted movement of patients among the diverse elements of the service delivery system” [4]. Whereas relational continuity refers to a durable relationship between an individual and his/her care providers, other dimensions of care continuity involve comprehensive and coordinated care provision by different care providers, either during a single healthcare episode or multiple episodes over time [1, 2, 5, 6]. Continuity of care therefore requires collaboration between care providers, both within and between services [2, 7, 8].

In recent decades, evidence has increasingly supported a personal recovery approach to mental health and substance use [9], so greater capacity for collaboration between care providers is needed [10]. Although the benefits of recovery-oriented interventions are widely documented at the individual level [11–13], the conditions these interventions require at the system level, i.e., a shift towards a Recovery-Oriented System of Care (ROSC), remain unclear [14]. A ROSC is a coordinated set of services designed to address the multiple complex needs of individuals with MD, improve quality of life, and foster social integration [9, 10, 14–16]. Although the concrete implications of a ROSC have not been fully identified, some key principles can guide how care provision is considered. First, recovery is a unique and non-linear process, so individuals’ needs may evolve as circumstances and life experiences change [15]. Collaboration between different service types is therefore necessary to ensure flexible, comprehensive care pathways, tailored to each individual [10, 14]. Second, care should be organised around individuals’ needs and preferences, not diagnostic categories or providers’ perspectives, to avoid limiting care to standardised treatment pathways [3, 10, 14]. Third, the focus should shift from clinical symptom management to broader objectives such as improving quality of life and social integration, acknowledging social and economic determinants of mental health [3, 10, 14]. This requires further deinstitutionalisation and the development of community-based services, as longer hospital stays are associated with reduced social integration [14, 17]. With regard to these ROSC principles, we explored the existing characteristics of care provision for individuals with MD. We focused on the divide between mental health services and specialised services for substance use disorders (SUD), which have long been organised separately, creating well-documented barriers to collaboration, despite evidence that integrating them improves access to care, treatment adherence, continuity, and person-centredness [9, 10, 12, 18, 19].

Mental health service networks

Despite recommendations in the literature and reforms aimed at better integrated community-based care that have been implemented in recent decades, most mental healthcare systems remain highly fragmented and struggle to offer comprehensive care pathways tailored to individual needs. To address this issue, several countries have established service networks as a tool for organising mental healthcare delivery and fostering integration between autonomous community-based services [20–23]. Service networks balance market dynamics, whereby autonomous actors compete for resources, regulated by supply and demand, and hierarchy, with centralised control and formal regulation [24–27]. This hybrid governance model offers flexibility and a degree of predictability [27]. Its primary goal is to improve healthcare system performance by addressing the complex needs of individuals with MD, ensuring access to comprehensive services and continuous care, while ensuring a stable cost [24].

Mental health care provision in Belgium

In Belgium, since 2010, adult mental health care has been organised into service networks under the “Title 107” reform, aimed at supporting the development of community-based services, reducing hospital admissions and the length of hospital stays, enhancing collaboration between services, and better addressing users’ needs and social integration [28, 29]. Twenty regionally embedded mental health service networks (MHSN) were established across the country. MHSN were allocated five care functions: [1] prevention and early detection of MD, [2] outreach, [3] social rehabilitation, [4] intensive residential treatment, and [5] specific long-term housing facilities [28, 29]. Although care functions were meant to be implemented through collaboration between different services, in practice, they were devolved to specific service types, with primary care and community mental health teams implementing function 1, all types of outreach teams classified as function 2, social services and rehabilitation services as function 3, hospital units as function 4, and other residential services as function 5. Mobile teams were introduced to improve outreach, while existing services were reorganised to improve integration [28, 29]. The establishment of MHSN followed a bottom-up approach in a system that preserved the considerable autonomy of the mainly private care providers, and traditional financial channels [29, 30]. Psychiatric hospitals were tasked with developing their own projects regarding network collaboration with community-based services [28]. Therefore, not all services within a given catchment area were involved in the network. Specialised SUD services, for example, were traditionally separate from mental health services, so not always formally included in these networks. As a result, MHSN vary substantially in size and composition, and in their practices and interventions [28, 29].

The structure of service collaboration within these networks is particularly relevant from a ROSC perspective. Previous studies suggest that collaboration is facilitated in smaller, more homogeneous networks, as high diversity can hinder consensus [31]. Continuity of care is greater in larger networks with connections between similar services, whereas social integration benefits from smaller networks linking complementary services [32]. This study aims to examine how service collaboration is structured within the Belgian mental health service networks and assess how well these patterns align with ROSC principles. To our knowledge, no studies have explored the structure of collaboration in relation to ROSC principles. Furthermore, although many healthcare systems deliver mental health services separately from specialised services for SUD, no research has specifically examined how collaboration between these sectors is structured.

Method

Statistical analysis

Service category and care function were available for both participating and non-participating services. However, information on current substance use as an exclusion criterion was only available for the participating services. To address this issue, we hypothesised that the likelihood of a service applying this exclusion criterion could be predicted based on services of the same category, care function, and network membership. To test this hypothesis, we ran a logistic regression on the participating services, which yielded a McFadden R² of 0.28, suggesting acceptable model fit, with care function and network emerging as significant predictors. Estimated results of the model are provided in the supplementary material (see Additional file 1). Based on these results, we then used this model to impute missing data for non-participating services, using multiple imputation by chained equations, implemented using the mice package in R [45].

As part of the data processing, we removed some obvious errors (such as loop ties), and symmetrised the network using the union method: any tie in at least one direction was considered reciprocal. We assumed that if a service reported regular contact with another, this relationship would be mutual. Since our network comprised non-participating services that did not report collaboration ties with other services, it remains possible that two non-participating services were connected, although that information was unavailable. Therefore, whereas ties between two participating services, or between a participating service and a non-participating service, were coded as either present or absent, ties between two non-participating services were treated as missing data.

Descriptive network statistics were computed for each MHSN, including the number of nodes, number of non-missing ties, average degree per attribute (i.e. the average number of ties services had based on their characteristics), density (i.e. the proportion of actual ties among all possible ties), degree centralisation (i.e. the tendency for ties to accrue among a small number of popular services), transitivity (i.e. the propensity for services to create cohesive, interconnected subgroups within the network), homophily (i.e. the tendency of similar services to form ties), and modularity (i.e. the extent to which a network is structured into distinct subgroups or communities) [46]. Subsequently, we used Exponential Random Graph Model (ERGM) to model the structure of ties in the five MHSN and assess whether the observed structures differ significantly from those in a random network [46]. Whereas descriptive network measures summarise patterns observed in existing ties, ERGM estimates the probability of tie formation between two services based on both service-level and network structural characteristics [46]. Importantly, this approach accounts for the interdependence of ties, a key feature of network analysis, and for potential missing ties [46]. Differences may arise between observed patterns and model-based estimates. Table 1 presents the network configurations mentioned in this study along with their associated ERGM terms, visual representations, and interpretations. In ERGM, the probability of a tie between two nodes is explained by dyad-independent (focused on attributes) or dyad-dependent (focused on ties) network configurations [46]. We used a simple random graph model as the baseline to compare more complex models. In a simple graph model, ties between nodes are formed randomly and are independent of each other. More complex ERGM models were created, adding node attributes and network configurations, and testing different combinations. Based on theoretical considerations and in an effort to avoid model degeneracy, the final model was chosen using the Bayesian Information Criterion (BIC), with a lower BIC indicating a better fit. The final model included covariates accounting for network density, transitivity, degree distribution, the main effect of service attributes, and the interaction effect of homophily. Unfortunately, we were unable to fit a model that included covariates for both transitivity and degree centralisation simultaneously, as this resulted in convergence issues. Therefore, we tested two different models and selected the most relevant based on the criteria described above. Networks were constructed and descriptive network statistics were computed using the igraph package in R. To estimate the ERGM, we used the statnet and ergm.multi packages, applying Markov chain Monte Carlo maximum likelihood estimation across five partially observed, independent networks [47]. The use of ergm.multi allows the estimation of a single ERGM model that captures common structural processes across networks, while accounting for differences in network characteristics.

Table 1.

Theoretical network configurations

Network features	Ergm term	Visualisation	Interpretation
Dyad dependent
Density	edges		The number of actual ties among all possible ties (usually not interpreted)
Transitivity	gwesp		The tendency to form multiple triangles
Degree centralisation	gwdegree		The tendency of the network as a whole to concentrate ties around a few popular nodes
Degree distribution	kstar(2)		The tendency of nodes to form 2-stars, i.e. a node connected to two other nodes
Dyad independent
Main effect of categorical node attribute	nodefactor		The degree effect of categorical node attributes
Interaction effect of homophily for categorical attributes	nodematch		The tendency for similar nodes to connect

Open in a new tab

Results

Discussion

We found that services tended to collaborate primarily with services in the same category or with the same care function. Services tended to cluster in highly connected groups, and hospital units were more central than community-based services.

Limited collaboration between service types is inconsistent with ROSC principles, which advocate for service diversity to support tailored care pathways [12, 14]. Rather than complementing one another, services appear to follow a logic of substitution. This could be due to service overload, lengthy waiting lists, or limited awareness of other providers’ roles, leading professionals to refer systematically to similar services rather than considering all available options [30]. Limited collaboration between mental health and SUD services is consistent with previous studies showing that the source of referral is likely to influence service orientation: hospital units, GPs, and family members are more likely to refer users to generic mental health services, whereas specialised SUD services and services related to justice are more likely to refer users to SUD services [48]. This is particularly concerning, given evidence that fragmentation between the two sectors is detrimental for individuals experiencing both SUD and other MD [9]. That divide is not specific to Belgium; it has been documented in different care systems, though empirical assessment of how much mental health and SUD services collaborate remains scarce. Limited collaboration across care functions is also consistent with existing evidence documenting difficult transitions between residential and outpatient care, often driven by poor information sharing or gaps in support between service types [49, 50]. Several mechanisms have been suggested to address such issues. Collaboration tends to be more effective when services in an area share common funding schemes and regulation [2], as in many National Health Systems. However, regulated-market systems such as the Belgian one have less organisational capacity to regulate collaboration. Structured and defined care pathways, financial incentives for collaboration, detached liaison staff in generic services, and structural case management are potential levers that could strengthen collaboration [11, 16, 51].

The tendency for services to collaborate within highly connected groups suggests that collaboration is driven more by interpersonal trust and relationships between individual care providers than by formalised inter-organisational arrangements [8, 29]. Such ties can sustain collaboration in the absence of formal structure but the resulting care pathways may be determined more by the provider and their network of relationships than by the individuals’ needs.

The central role played by hospital units in the MHSN is also inconsistent with ROSC principles, which promote care delivery in community settings. This finding may reflect the structural and financial importance of psychiatric hospitals to the reform [28] and the limited availability of recovery-oriented alternatives to traditional treatment [52]. According to Day and colleagues, a ROSC for people with SUD should encompass peer-based recovery support services, employment support, recovery housing, continuing care and recovery check-ups, community centres, and educational services [53]. In Belgium, most of these services remain generic and are often reluctant to admit individuals with SUD [52]. Our findings suggest two ways to strengthen deinstitutionalisation in Belgium. First, community-based mental health services could be further developed and consolidated. For example, mobile teams were intended to bridge hospital and community settings, but they tend to remain structurally dependent on hospitals with limited collaboration with community-based services [28, 54]. Funding mechanisms could be adapted to incentivise community-based mental health services to engage with individuals who traditionally receive hospital-based care, and governance arrangements could be strengthened, as hospitals and community-based services often fall under different administrative authorities. Second, mental health care could extend beyond specialised care to other welfare sectors through stronger intersectoral collaboration [3, 55], moving toward a stepped care model prioritising the least specialised intervention tailored to individuals’ needs [56].

Finally, the tendency for services which apply current substance use as an exclusion criterion to have more collaboration ties raises important concerns, particularly given the high proportion of services that apply this criterion. From a ROSC perspective, the predominance of abstinence-based criteria reflects clinical and professional objectives, not users’ needs and preferences [52]. Such exclusionary practices might increase stigma. More broadly, they reflect a persistent tendency to reduce complex conditions like MD to diagnostic categories, disregarding the diversity of recovery experiences [3, 10]. Nevertheless, abstinence-oriented services appeared more likely to collaborate with services applying more flexible criteria, suggesting a degree of complementarity, potentially driven by referral practices, whereby they redirect individuals who do not meet their criteria toward more flexible alternatives. Whether this translates into person-centred pathways, however, requires further investigation.

Our findings suggest that the structure of collaboration in these MHSN is driven more by organisational logics and providers’ interests than users’ needs. This is consistent with previous research documenting, among individuals with MD in Belgium, limited access to diverse service types, particularly housing support [2], fragmented care pathways [51], and poorer social integration when similar services collaborate [32]. Further research comparing the structure of collaboration in different care systems is needed to better understand how network organisation shapes individuals’ recovery pathways.

Conclusion

To enhance care continuity and promote recovery-oriented care, mental healthcare systems should adopt collaborative, community-based, and person-centred models. Examining the structure of collaboration between services provides valuable insight into how well care provision aligns with these organisational principles. In the service networks we studied, care provision remained highly fragmented and predominantly hospital-centred, as if prioritising the interests of care providers over those of individuals with MD. Changes in care provision must be sustained by appropriate funding and governance mechanisms; otherwise, old habits and practices persist. This study suggests that supporting individuals’ recovery is partly related to system-level factors, particularly the structure of collaboration between services. Finally, this study highlights the potential for Social Network Analysis to assess collaboration patterns and underscores the need to evaluate the healthcare system’s capacity to support personal recovery.

Supplementary Information

Author contributions

Conceptualization: MC, VL, PN; Data curation: MC, PN; Formal analysis: MC; Funding acquisition: PN; Investigation: MC, JM, IG, JDM, WV, VL, PN; Methodology: MC, VL, PN; Project administration: JDM, WV, PN; Resources: JDM, WV, PN; Software: MC, VL; Supervision: VL, PN; Validation: MC, JM, IG, JDM, WV, VL, PN; Visualization: MC; Writing – original draft: MC, VL, PN: Writing – review and editing: MC, JM, IG, JDM, WV, VL, PN.

Funding

This work was supported by the Belgian Science Policy Office (BELSPO) [grant number DR/89/SUMHIT].

Data availability

Data are not available in a publicly available repository due to networking data that could lead to service identification. However, interested parties may contact the corresponding author for further information.

Declarations