Geospatial Investigations in Colombia Reveal Variations in the Distribution of Mood and Psychotic Disorders

The study addresses how mood and psychotic disorders are distributed geographically in a middle-income country context and whether access to mental healthcare contributes to observed patterns. Prior work in upper-income countries has demonstrated distance-decay effects—healthcare utilization declines with increasing distance or travel-time to facilities—and suggested that geographic variation in disorder frequency can reflect localized social, environmental, and genetic factors. In Colombia, national mental health surveys have provided valuable prevalence and perception-of-access data but lack detailed geospatial resolution and information on severe presentations and actual service utilization. Leveraging a comprehensive EHR from CSJDM, the authors aim to quantify geographic accessibility to specialty mental healthcare across Caldas, examine the relationship between travel-time and incidence of treated mood and psychotic disorders (bipolar disorder, schizophrenia, major depressive disorder) by severity (inpatient vs outpatient), and identify spatial hotspots of severe cases. The purpose is to inform strategies to reduce inequities in access to mental healthcare and to lay groundwork for research into localized risk factors.

The paper situates its work within a long history of research documenting distance-decay in mental healthcare utilization, dating back to the 19th century, and multiple studies in upper-income countries showing geographic variation in mental health outcomes linked to social, environmental, and genetic factors. It highlights the lack of population registries in many LMICs and the consequent scarcity of geospatial mental health studies, noting that Colombia’s National Mental Health Survey (most recently 2015) provides broad epidemiological insights but cannot support fine-grained geospatial analyses or capture severe/acute presentations and actual service utilization. Emerging EHR resources in LMICs, such as CSJDM’s database, now enable population-level geospatial analyses of treated mental illness.

Design and data sources: Retrospective cross-sectional study using CSJDM EHR data (2005–2018) encompassing demographics, residential address, and ICD-10 diagnostic codes. IRBs at CSJDM and UCLA approved the study; anonymized data were hosted on a HIPAA-compliant server. Adult patients aged 18–90 residing in Caldas were included. Case definition: Individuals with an initial (first-visit) diagnosis of bipolar disorder (BPD; F31), schizophrenia (SCZ; F20), or major depressive disorder (MDD; F32/F33). Analyses were stratified by hospitalization status (inpatient vs outpatient). Diagnostic code accuracy in this EHR has been previously validated. Cohort and completeness: Initially identified N=16,376; addresses available for N=16,308; after geocoding and quality thresholding, N=16,295 patients were included. Geocoding: Residential addresses were standardized and geocoded using OpenCage (Spanish setting) via the R package opencage. Grid-pattern addresses (N=11,745) were formatted as “Calle [X], Carrera [Y], [Municipality], Caldas”; rural/vereda addresses (N=4563) appended vereda name, municipality, and department. OpenCage returned candidate coordinates with confidence scores; the best match per OpenStreetMap relevance was selected. Addresses with confidence <6 were excluded. Resulting coordinates localized all included addresses within a 7.5×7.5 km bounding box (71% to 0.5×0.5 km). Incidence calculation: Annual incidence per 100,000 for each of Caldas’ 27 municipalities was computed as newly diagnosed cases per year divided by estimated municipality population for that year (WorldPop gridded estimates). Due to low annual counts, analyses used 14-year aggregated incidence (reported as rate per 100,000). For finer resolution, a 5×5 km grid was imposed across Caldas; cell-level incidence was defined as newly diagnosed cases (2005–2018) divided by cell population in 2018, per 100,000. Visualization used ggplot2 and DANE shapefiles. Geographic accessibility (friction surface) and travel-time: A friction surface (100 m resolution) was built with AccessMod 5.6.0, integrating topography (ASTER), land cover (GLC2000), permanent water bodies (Digital Chart of the World), and road networks (DANE). Minimum travel-time per pixel assumed driving on roads and walking where roads absent, with elevation-adjusted (anisotropic) walking speeds; transport mode speeds per Supplementary Table 1. A geographic accessibility map was generated by computing travel-time from each 1×1 km cell centroid to CSJDM. Modeling distance-decay: Zero-inflated negative binomial regression (R pscl v1.5.5) modeled expected case counts by travel-time (hours), offsetting by population in each grid cell. Separate models were fitted for inpatients and outpatients for all diagnoses combined and for each diagnosis (total of eight tests). Inference used 1200 bootstrap resamples (R boot v1.3-20) for 95% CIs and two-sided p-values; Bonferroni threshold 0.00625 (0.05/8). Sensitivity analyses included gender main effects and gender×travel-time interactions. Hotspot detection: Kulldorff’s Spatial Scan Statistic (SaTScan v9.6) scanned the 5×5 km grid to identify clusters unconstrained by municipal boundaries. Circular windows varied up to a maximum of 25% of Caldas’ population. Significance used one-sided Monte Carlo testing; Bonferroni threshold 0.0125 (0.05/4 tests). Uncertainty in hotspot borders was assessed using Oliveira’s F function (cell-level cluster membership intensity), visualized in QGIS with Jenks natural breaks. Sensitivity analyses varied the maximum cluster size and stratified by gender to evaluate robustness. Dataset summary: Of the 16,295 patients, 5218 were inpatients and 11,077 outpatients; women comprised 65% overall (inpatients: 57% women; outpatients: 68% women). By diagnosis across all patients: BPD N=4,985 (36% men, 64% women), MDD N=10,257 (31% men, 69% women), SCZ N=1,053 (76% men, 24% women).

- Geographic accessibility: Approximately 50% of Caldas’ population can reach CSJDM within <1 hour driving; ~90% within ~4 hours; some eastern areas require >5 hours (up to >10 hours), though sparsely populated. - Distance-decay: Significant for outpatients but not inpatients. For outpatients (all diagnoses combined), each additional hour of travel-time is associated with a 20% decrease in expected cases (RR=0.80; 95% CI: 0.71–0.89; p=5.67×10^-5). No significant distance-decay for inpatients overall or by diagnosis. - Municipal incidence variation (inpatients): Overall inpatient incidence highest in Aranzazu (~55 km from CSJDM), driven by mood disorders: BPD inpatient incidence 862/100,000 vs 219/100,000 department-wide; MDD inpatient incidence 364/100,000 vs 235/100,000; MDD inpatient incidence there is nearly as high as in Manizales (387/100,000). For SCZ, incidence is highest in three neighboring municipalities ~111 km from CSJDM: ~93/100,000 vs 45/100,000 department-wide. - Hotspots: Nine statistically significant inpatient hotspots identified. Cluster 1 (around Manizales/CSJDM) shows overrepresentation across diagnoses, with particularly high MDD overrepresentation (Observed 1476 vs Expected 531; RR=5.47; p<1×10^-17). Cluster 2 (spanning Aranzazu and Filadelfia, ~1.5 h from CSJDM) exhibits the most extreme BPD overrepresentation (Observed 111 vs Expected 20; RR=5.83; p<1×10^-17) and also overrepresented MDD (Observed 57 vs Expected 21; RR=2.72; p=9.19×10^-7), suggesting a mood-disorder-specific hotspot. Cluster 3 (~5 h from CSJDM) shows overrepresentation across all three diagnoses; SCZ overrepresentation is striking (Observed 13 vs Expected 1; RR=8.98; p=2.11×10^-5). Additional clusters (4–9) show varying overrepresentation patterns; Cluster 9 is specific to BPD (Observed 31 vs Expected 12; RR=2.69; p=4.03×10^-3). - Robustness: Hotspots for MDD and SCZ were largely insensitive to the maximum scanning window size; BPD hotspots were stable except that Cluster 2 is subsumed by Cluster 1 when allowing >40% maximum cluster size due to SaTScan’s selection procedure. Gender-stratified analyses yielded the same clusters without new hotspots. - Utilization pattern: Outpatients predominantly reside nearer to CSJDM; inpatients are more widely distributed and cluster in discrete hotspots, indicating factors beyond geographic accessibility (e.g., local sociodemographic, environmental, or genetic influences).

The findings demonstrate inequitable geographic accessibility to specialty mental healthcare in Caldas and a strong distance-decay effect for outpatient utilization, indicating that travel-time is a key barrier for individuals requiring less intensive care. This aligns with evidence from upper-income countries and underscores the potential value of Colombia’s strategy to embed mental healthcare in primary care networks; future geospatial analyses of primary care access and EHRs could evaluate its impact. In contrast, inpatient utilization does not exhibit distance-decay, instead clustering into nine hotspots. This pattern suggests that local contextual factors—sociodemographic, environmental, and possibly genetic—drive geographic variation in severe presentations or in access and help-seeking for such cases. The marked overrepresentation of BPD relative to SCZ among inpatients, especially within certain hotspots (e.g., Cluster 2), is unusual compared with international inpatient distributions and warrants targeted investigation. Potential contributors include high regional socioeconomic inequality, past exposure to armed conflict and displacement, and the demographic/genetic history of the Paisa population, a regional isolate where founder effects and rapid expansion may elevate local frequencies of risk variants. Additionally, stigma and fear of discrimination may influence treatment-seeking, particularly for outpatient care, though current data are limited. Addressing these factors through integrated datasets combining EHRs with detailed sociodemographic, conflict exposure, and genetic data could clarify mechanisms underlying observed spatial patterns and inform equitable service planning.

This study leverages a comprehensive EHR and geospatial modeling to quantify mental healthcare accessibility in Caldas, Colombia, reveal a pronounced distance-decay effect for outpatient care, and identify nine inpatient hotspots with diagnosis-specific patterns. These findings highlight inequities in access to outpatient services and suggest localized determinants of severe mental illness requiring inpatient care. The work provides actionable insights for resource allocation (e.g., enhancing outpatient services in remote areas and targeted outreach in hotspots) and establishes a framework for future research integrating sociodemographic, environmental, and genetic data. Future directions include longitudinal analyses of patient trajectories and diagnostic transitions, evaluation of primary-care–based mental health integration, incorporation of dynamic travel-time data, and expanded multi-department studies to power genetic investigations within the broader Paisa region.

- Cross-sectional analysis focused only on initial (first-visit) diagnoses, not leveraging longitudinal EHR trajectories; geographic patterns might differ when considering clinical course and diagnostic changes. - Smaller inpatient sample compared with outpatients may limit power to detect modest distance-decay among inpatients. - Travel-time estimates use average speeds by surface type and do not account for temporal variability (seasonal, diurnal, traffic). - Spatial scan statistics (SaTScan) sequentially select most likely clusters rather than optimizing a joint global cluster pattern, potentially affecting hotspot delineation under certain parameter settings. - Additional unmeasured factors (e.g., stigma, discrimination) that influence treatment-seeking were not available at sufficient scale for inclusion.