Examining rehabilitation access disparities: an integrated analysis of electronic health record data and population characteristics through bivariate choropleth mapping

The study cohort included San Francisco County residents over 18 years of age who had received outpatient physical therapy services between the dates of January 1, 2016 and January 1, 2020 at a large urban academic medical center. Patients with all conditions with a referral to an outpatient setting were included in the study. Any residents outside of San Francisco County were excluded from the study. This study protocol was approved by the University of California San Francisco Institutional Review Board (19–28,255).

An EHR dataset was extracted from the institution’s clinical data warehouse (CDW) consisting of patient-level data: age, sex, race (White, Asian, Black, Other, Unknown), ethnicity (Hispanic, non-Hispanic, Unknown), insurance payer type (Private, Medicaid, Medicare, Other), Area Deprivation Index (ADI), and distance (miles) from home to clinic. The dataset also includes geocoded data of latitude and longitude coordinates for patient addresses, census tract IDs, and block group IDs prepared by the UCSF Population Health Data Initiative (PHDI) team [22, 23]. By using census block group ID number assigned to individual patient record, we were able also to link it to neighborhood-level data. From the patient-level data, three access timelines were evaluated. They included: 1) referral-to-schedule time, defined as the time elapsed between when patients receive a referral to physical therapy and when patients initiate care, 2) schedule-to-appointment time, defined as the time that elapses between when a patient calls the clinic to schedule and when they arrive for their first visit, and 3) referral-to-appointment time defined as the total time from referral to first completed appointment. Other utilization metrics included the number of visits, cancellations (total and number of same-day cancellations), and diagnosis type (orthopedic, neurologic, pelvic, other).

The Area Deprivation Index (ADI) is a multi-faceted proxy measure derived from several different variables, including education level, employment, income, housing quality, and access to services [24]. A higher score indicates a greater level of socioeconomic disadvantage [24] and can be derived for both a National and State-specific level for California [25]. Our analysis used State-specific rather than National scores to be locally sensitive and accurately capture the region’s local comparability [26]. Using Spatial Join in ArcGIS Pro (2.8.3) [27], the census block group ID numbers from patient geocoded data were linked with the ADI scores (ADI, 2020 dataset) to understand and visualize the potential socioeconomic disadvantage of geographical regions of patients’ residences. We mapped the geocoded patients with ADI scores [24, 25].

Similar methodology as described by Kethireedy et al. [28] was used to calculate geospatial statistics for patients’ time-based access metrics and ADI state-level scores [24]. By using individual geocoded patient data prepared by the UCSF Population Health Data Initiative (PHDI) team [22, 23], we used individual patient records that contained three time-based metrics aggregated to census block group level, we aggregated to the census block group. Using Getis–Ord G_i^∗statistic feature from ArcGIS Pro (version 2.8.3) [29, 30], geospatial clustering analysis of three distinct time-based metrics (i.e., referral-to-scheduled time, scheduled-to-appointment time, and referral-to-appointment time) was assessed to identify patterns and groupings in spatial data. Several steps were taken to calculate the required input parameters for G_i^∗statistics, and they are described in the supplementary file (See Supplemental File). A statistically significant “hot spot” represents a higher-value feature surrounded by other higher-value neighboring features. Conversely, a statistically significant “coldspot” refers to a lower-value feature surrounded by other lower-value neighboring features [29]. G_i^∗ statistic contain a z-score [31] and clusters with a 95% significance level from a two-tailed normal distribution. A z-score close to zero and a p-value greater than 0.05 suggest complete spatial randomness within the study area. On the other hand, a positive z-score and a p-value less than 0.05 signify the clustering of high values. The hot and coldspot analysis results were calculated with a 95% CI (Fig. 1). Lastly, clusters were compared using the False Discovery Rate correction in the analysis to account for potential false positive hot spots [32].

To understand neighborhood-level characteristics of the regions in the hot and coldspot analysis, several data sources were used, including CDC PLACES data, the American Community Survey, California Healthy Places index from Public Health Alliance. Neighborhood-level characteristic data are summarized in Table 1.

A bivariate choropleth map was created to visually describe relationships between two distinct variable classes (clusters of time-based access metrics resulting in hot and coldspots and ADI State scores) on a projected layer. Each variable (i.e., access metric and ADI score) was assigned a different graded color scheme to delineate relationships of high and low respective values within a 3 × 3 table (Fig. 2). While there are different classification approaches for bivariate choropleth maps, we applied equal intervals to classify groupings with the same hot and coldspots with 95% CI and ADI scores [36, 37], and to highlight changes in the extreme points with a relatively intuitive view for readers [36, 38]. For example, Fig. 2 (Map A) highlights the “Longer” and “Shorter Times” of the referral to schedule metric with an overlay of graded colors of ADI scores reflecting least to most disadvantaged block groups labeled as “High” and “Low” ADI in shades of color green. Similarly, Figs. 2 (Map B & C) show longer and shorter times of the schedule to appointment metric and referral to the first appointment metrics with ADI scores, respectively.

The data queried yielded 10,363 unique patient records after removing missing or erroneous data pertaining to our primary outcome measuresThe cohort had a median (IQR) age of 54 (40, 67) years; it was 63% female, and 53% non-Hispanic White (Table 2). The payer type was predominately Private (42%), and the referral type was predominately for orthopedic diagnoses (69%). Median access times for care were 3 (0, 10), 9 ( 5, 15), and 15 (9, 26) days for referral-to-scheduled time, scheduled-to-appointment time, and referral-to-appointment time, respectively.

For referral-to-scheduled time (Fig. 1, Map A), between-group statistics found that the hotspot (i.e., longer time) cohort was older (51 vs 44 years) lived farther from the clinic (2.03 vs 1.42miles), and had a higher proportion of Black patients (25% vs 5%) and patients with Medicaid (25% vs 11%) and Medicare (28% vs 17%) (Table 3). For scheduled-to-appointment time (Fig. 1, Map B), the hot spot cohort was older (56 vs 46 years), lived farther from the clinic (2.59 vs 1.61 miles), and had higher proportions of Females (65% vs 59%), Asians (41% vs 25%), non-Hispanic patients (92% vs 88%), Medicare (29% vs 16%) and Other payer types (18% vs 9%), and neurologic-related diagnoses (9% vs 5%) (Table 3). For referral-to-appointment time (Fig. 1, Map C), the hot spot cohort was older (55 vs 47 years), lived farther from the clinic (2.07 vs 1.71 miles), lived in locations of higher ADI scores (2 vs 1), and had higher proportions of Asian (28% vs 23%), Black (15% vs 4%), and non-Hispanic patients (90% vs 88%), and Medicaid (23% vs 13%) and Medicare (33% vs 17%) payer types (Table 3).

Neighborhood characteristics of the hot and coldspots for each access metric are summarized in Table 4. Consistently higher proportions of disability, arthritis, and self-reported poor physical health and physical inactivity were found for all access metrics for hot spot neighborhoods (p < .001). Lower Healthy Places Indices were also found in the referral-to-schedule and referral-to-appointment times (p < .001). Proportions of obesity and reported poor mental health had inconsistent results, being higher in the hot spot neighborhood of the referral-to-scheduled time (p < .001) but lower in the scheduled-to-appointment time (p < .001).

Three bivariate choropleth maps (Fig. 2 Maps A-C) visualize the relationships between each access measure’s hot and coldspots and ADI Scores at the census block group, where each polygon was shaded with a unique color combination representing a distribution of high and low values from each feature variable of interest. The top bivariate map (Fig. 2, Map A) shows residential areas at census block group in San Francisco where a longer referral-to-schedule time and higher ADI scores diverge in spatial relationships – with the longest access time and highest ADI in dark purple and other combinations in varying colors.

Our methods and study results may inform administrators and policymakers of disparities contributing to variability in time-based access metrics in an urban outpatient physical therapy setting. By merging individual patient data with neighborhood characteristics linked to social determinants of health at the census block group [24], we gain insights into associations between disadvantaged communities and patients experiencing extended access times. Our analysis in San Francisco revealed a cluster of patients with prolonged access times in regions exhibiting neighborhood attributes with higher disability rates [33], lower mental and physical well-being [34], and lower socioeconomic status [24, 39] (Table 4). These findings align with a systematic review by Dawkins et al., who found that in high-income countries like the US, patients with more severe physical and mental comorbidities had more limited access to healthcare [40]. Further, our analysis from hot and coldspots juxtaposes the two extreme ends of access patterns consistent with prior research by Gao et al., who found limited access to rehabilitation services in areas of higher disability prevalence where potential demands are most needed [20].

Considering that as much as 50 percent of health outcomes are affected by social, economic, and environmental factors [41], it’s unsurprising that such profound discrepancies in a delay of care are shown in our study. Various data sources from health equity tools like the Healthy Places Index arose from the need to establish objective index score rankings to target interventions towards the most pressing needs of policymakers [35, 39], supporting integration of neighborhood-level and patient datasets to continue the efforts to combat health disparities.

Delineating the timelines of three access metrics captures important distinctions of provisioning care between the patient and the scheduler. For example, the referral-to-schedule measure depends on patients initiating care after providers create the referral. Conversely, the schedule-to-appointment measure is influenced by both patients and clinics because scheduling teams need to find a time that accommodates both availability of patients and providers. Based on the hot spot patterns from the three distinct measures, we were able to quantify the extreme differences in the timing of access patterns starting with the time to schedule by patients. Operationally, the referral-to-schedule measure showed a median value of four business days for the hot spot group and only one day for the coldspot group (P < .001). Variance of four business days may seem minor, but there is significant difference between hot and coldspot groups with respect to the proportion of time devoted to scheduling relative to the overall time taken: 31% to 12%, respectively, as indicated in Table 3 (P< .001). Moreover, any additional delay in timely care could result in adverse effects on pain, quality of life and psychological symptoms for patients waiting for physical therapy services [42].

While the urgency to schedule may also depend on patients’ personal values [43], recent studies highlight several barriers to access, particularly those from marginalized groups: navigating complex health systems to specialty care, financial burdens associated with gaps in insurance coverage, loss of time from work, and travel costs [40, 44‐46]. Further research is needed to fully identify the need of our local neighborhood to deliver interventions that acknowledge their values for rehabilitation services. Nevertheless, our findings raise awareness of health disparities in care access, which warrants discussions with healthcare leaders and policymakers to bridge these gaps.

Combining two distinct variables from patient and neighborhood-level data – access metrics and ADI scores, respectively – can be helpful to visually provide spatial patterns of different access times with neighborhood characteristics. The gradient colors in the 3 × 3 table (Fig. 2, Maps A-C) reflect the range values from each and the combination of two variables. For example, our bivariate map identifies areas with hot spot regions (i.e., longest access time) that also spatially overlap with areas with higher ADI scores at the census group block. In Fig. 2 (Maps A-C), the purple colors represent the patient resident location with the longest access time and the highest ADI scores. While the median ADI values between hot and cold spot regions may not seem significantly different, the interquartile range values are more positively skewed for hot spot regions (Table 3). Further, other social determinants of health variables such as income, employment, and housing quality may be contributing factors. Numerous studies suggest that socioeconomic neighborhood disadvantage, as expressed through ADI scores, can predict health outcomes such as hospital readmission rates [47], observation stays [48], and mortality rates [49]. ADI scores could serve as a valuable screening tool to inform clinicians and health systems to proactively engage patients returning from their challenging environments [26]. Moreover, we can investigate the degree of neighborhood disadvantages potentially contributing to differences in health utilization using visuals of these extreme ends of access patterns. Such additional neighborhood characteristic data provides further context to the factors that may influence the care delay [50].