Health Disparities in Orthobiologics: Factors Influencing Access to Platelet-Rich Plasma for Knee Osteoarthritis

Abstract

Background:

Platelet-rich plasma (PRP) is an emerging therapy that demonstrates potential benefits for knee osteoarthritis (OA); however, the accessibility of PRP remains limited, in part, because its use is not covered by insurance.

Purpose:

To identify factors contributing to health disparities in PRP treatment for knee OA using machine learning models.

Study Design:

Cross-sectional study; Level of evidence, 3.

Methods:

The authors collected data from electronic medical records at 2 upper Midwest medical centers (Mayo Clinic and Detroit Medical Center) from 2022 to 2023 encompassing demographic, socioeconomic, and clinical factors. Data collection encompassed age, sex, race, ethnicity, area deprivation index, marital status, income, rurality, insurance, employment, language, religion, treatment site, and region. Logistic regression and random forest models were applied to 3667 samples (80% training, 20% testing), comparing PRP recipients against a standard care group. Models were evaluated using 10-fold cross-validation to ensure robust performance assessment.

Results:

Logistic regression revealed that increased age, widowed status, lower income, rural residence, unemployment/disability, and Medicaid/Medicare insurance were associated with lower odds of receiving PRP. Upper-middle income, contracted insurance, and treatment at the Mayo-Arizona and Mayo-Florida sites were associated with higher odds. The model achieved areas under the curve of 0.83 (training set) and 0.82 (testing set). Random forest analysis identified distance to care, age, area deprivation index, income, and treatment site as top predictors, with areas under the curve of 0.92 (training) and 0.87 (testing).

Conclusion:

The current study demonstrated that both models discerned key barriers (income, distance to care, retirement status) and facilitators (regional differences) affecting PRP accessibility. These findings provide potential policy initiatives aimed at improving PRP access for underserved demographics, offering a quantitative blueprint to foster equitable access to regenerative therapies for knee OA.

Knee osteoarthritis (OA) represents a pervasive health challenge affecting 32.5 million US adults and 528 million adults globally, representing an increase of 113% since 1990. ²³ The burden of knee OA is projected to rise owing to aging populations and increasing obesity rates. ²³ Research indicates that knee OA disproportionately affects African Americans, who are 37% less likely to receive total knee replacement and subsequently experience greater hospitalization and postprocedural pain than White patients. ⁸ This disparity is reinforced by other studies, such as an 18-year analysis in which African Americans had 36%-40% lower usage rates of primary total knee arthroplasty as compared with White patients. ³² This gap appears to persist even when controlling for other variables: a large prospective cohort study found that Black women were 25% to 30% less likely to receive total knee arthroplasty than White women after adjustment for clinical need and socioeconomic factors. ⁷ Additionally, African Americans facing knee OA often harbor a desire to delay surgical interventions, given the invasive nature and the associated recovery. ¹⁶ In an attempt to translate nonsurgical interventions, orthobiologics have been heralded as a potential novel avenue to aid tendon and joint healing for musculoskeletal disease such as OA. In particular, platelet-rich plasma (PRP) may be an alternative for patients who are less responsive to standard medical management and injectable treatments and who wish to delay surgery.

PRP involves the collection, isolation, and injection of a high concentration of autologous platelets into the knee to foster tissue healing and reduce pain. While the mechanism of action is not fully understood, PRP is believed to elicit immunomodulation and reduce inflammation inside the joint, temporarily healing the affected area. ¹² Early clinical studies highlight PRP's potential, especially in the initial stages of knee OA, suggesting it as a promising therapeutic option.^3,13,15,32 Despite the potential future promise, a significant gap exists in understanding health disparities in accessing orthobiologics, including clinical, sociodemographic, and attitudinal factors. While PRP treatments are available in medical centers across the United States, they are less common in community settings, range in cost from $175 to $4,973, and are not typically covered by insurance.^4,18,25,27 Finally, clinical guidelines for PRP eligibility and use differ widely: the European Society of Sports Traumatology, Knee Surgery and Arthroscopy, and American Academy of Orthopaedic Surgeons recommend PRP for mild to moderate knee OA, yet vary in their preferred injection volumes, frequencies, and type of PRP advised. Standardization is still lacking across organizations.^6,21,41

By proactively examining determinants of PRP accessibility, we can identify potential barriers and inform interventions to ensure equitable access for future implementation of PRP and other novel orthobiologics. The purpose of this study is to examine sociodemographic and clinical predictors to identify facilitators and barriers in accessing PRP versus standard care for knee OA at 3 Mayo Clinic academic medical centers (Minnesota, Arizona, Florida campuses) and the Mayo Clinic Health System, as well as the Detroit Medical Center (DMC) in Michigan. We hypothesized that sociodemographic and clinical factors such as income, distance to care, and insurance status would significantly predict access to PRP treatment for knee OA.

Methods

Data Collection

In this study, we examined sociodemographic and clinical predictors to identify facilitators and barriers in accessing PRP versus standard care for knee OA at Mayo Clinic and DMC. In this cross-sectional study, patient data were obtained (2022-2023) from electronic medical records from all Mayo Clinic locations (n = 3603) and DMC (n = 64), culminating in a data set of 3667 patients. For standard of care, we obtained a random sample of 1832 patients as control. Standard of care was inclusive of oral medications, topical medications, corticosteroid and hyaluronic acid injections, arthroscopy, and arthroplasty, and the control group included patients with posttraumatic OA and those with previous knee surgery. The inclusion criteria were defined as having been diagnosed with knee OA and being ≥18 years of age. Patients were excluded if bone marrow aspirate concentrate treatment had been received or if >50% of their data variables were missing from their records. Missing data ranged from 0.1% to 3% for study variables.

Imputation was performed with the missRanger algorithm in R, which applies the MICE method ³⁵ (multivariate imputation by chained equations), a random forest–based approach, to predict and impute missing values for each variable in turn. Specifically, we used 200 trees (num.trees = 200) and a set seed for reproducibility. This method iteratively imputes missing features by leveraging all observed data, preserving the underlying multivariate relationships in the data set, and reducing the risk of bias attributed to missingness. Diagnostic checks confirmed that imputed data distributions were comparable to the original observed values. Best practice predictive modeling requires splitting data into a training set (80% for model development/internal validation) and a test set (20% held out for final accuracy). Within the 80% training set, 10-fold cross-validation was used: the training set was divided into 10 parts, and iteratively, 9 of 10 were used to build the model (logistic regression or random forest) and 1 of 10 to validate performance. This rotation produced an unbiased average accuracy and area under the curve (AUC) for model selection and hyperparameter tuning. Crucially, cross-validation is done only on the training set to prevent information leakage; using the test set risks inflated performance. After model selection, the final models were fit on the entire training data and evaluated once on the untouched 20% test set to simulate generalization to new data. This was applied identically for logistic regression and random forest.

Demographic, clinical, and socioeconomic variables were collected, encompassing age, sex, race, ethnicity, marital status, insurance, employment, language, treatment site, and region. Zip codes were used to extract scores based on the area deprivation index (ADI), which is a composite percentile-based measure of neighborhood socioeconomic disadvantage, calculated at the zip code level using census variables on income, education, housing quality, and employment. In this study, each patient's ADI was assigned according to one's residential zip code, with higher scores indicating greater area-level deprivation.^20,31 Also, we collected rurality classifications using RUCA codes (Rural-Urban Commuting Area), which were established by the US Department of Agriculture and Federal Office of Rural Health Policy ³⁴ to classify geographies based on population density, urbanization, and commuting patterns. We grouped RUCA into 3 analytic categories for this analysis: urban (codes 1-3), large rural city/town (codes 4-6), and finally small and isolated rural town (codes 7-10). Finally, distance to care was measured from each patient's zip code city to clinical site in miles, and neighborhood median income data had 4 category divisions: low income, <$35,000; middle income, $35,000-$75,000; upper-middle income, $75,000-$150,000; high income, >$150,000. Additionally, we referenced our collected ADI scores against the national percentile rankings calculated by Broadstreet ⁵ to contextualize where our ADI scores fell within the broader national distribution of deprivation levels.

No generative artificial intelligence or large language models were used in the conceptualization, data analysis, or drafting of this manuscript. Grammarly was utilized solely for routine grammatical checking. The natural language processing (NLP) algorithms used for data classification were programmed in R. This study was approved by Mayo Clinic and DMC (IRB-21-011328, IRB-21-11-4196).

Codebook Development and Validation for Knee OA Severity

In the clinical notes section of the electronic health records, providers described the severity of knee OA using inconsistent terminology. To ensure systematic characterization of severity, we developed a codebook categorizing knee OA severity as mild, moderate, and severe (Appendix). The codebook was developed by randomly analyzing 60 patient charts (50 from Mayo Clinic and 10 from DMC) with an orthopaedic nurse practitioner (J.A.) and was reviewed by 2 senior orthopaedists (S.S. and S.H.). After the codebook was established, 3 coders (M.A., L.B., and J.A.) independently and blindly classified knee severity from a random sample of 50 charts based on the newly established codebook. Cohen $\mathrm{\kappa }$ intercoder reliability testing revealed significant agreement between coders (${\rm{\kappa }}$ = 0.97, J.A. and M.A., indicating almost perfect agreement; $\mathrm{\kappa }$ = 0.87, M.A. and L.B., indicating almost perfect agreement). M.A. served as the anchor coder, tested independently against J.A. and L.B., demonstrating consistently high agreement across multiple coder pairs.

We developed an NLP algorithm to automatically categorize knee OA severity from unstructured physicians’ notes, converting them into structured data. The NLP algorithm was modified on the basis of small random samples several times where manual coding was compared with NLP to ensure a high match of severity diagnosis. A random sample of 50 notes comparing NLP with manual coding by M.A. resulted in excellent intercoder reliability ( $\mathrm{\kappa }$ = 0.839, indicating almost perfect agreement). The algorithm, equipped with the capacity to parse through unstructured text data, analyzed each note and classified severity accordingly. The NLP algorithm was also programmed to match severity with the date that PRP was administered to ensure proper linkage of knee OA severity at the time when the biologic was administered.

Statistical Analysis

After normality tests, descriptive statistics were initially used to understand fundamental attributes of the data set, and inferential tests (parametric and nonparametric) were subsequently used to glean insights about the wider population from the data set. Machine learning models, specifically multiple logistic regression and random forest, were deployed in R software. Model training and cross-validation were performed using the caret package, ²² random forest models were fit using ranger, ⁴² and variable importance and predictor interactions were displayed using vivid. ¹⁷ Data management and figures were produced using dplyr, ³⁹ tidyr, ⁴⁰ and ggplot2. ³⁸ Logistic regression ⁹ and random forest ⁹ were trained on the training data set. Model assumptions of independence of data, normality, linearity, heteroscedasticity, outliers, and collinearity were examined. Furthermore, the rigor and predictive acumen of both models were tested through 10-fold cross-validation using the testing data set, ¹⁹ with the AUC from the receiver operating characteristic curve serving as the yardstick for model performance, illuminating the delicate balance between sensitivity and specificity.

Results

Sociodemographic and descriptive statistics are outlined in Table 1. Inferential analysis revealed significant contrasts between the standard care group (n = 1466) and the PRP group (n = 1468) across numerous variables. The median age for the PRP group (70 years) was lower than the standard group (75 years), indicating that younger individuals were more likely to receive PRP treatment (P < .001). Similarly, individuals in the PRP group manifested a lower median deprivation percentile of 30, as opposed to 42 in the standard group, indicating higher socioeconomic standing among those receiving PRP (P < .001). Related to deprivation percentile, statistically significant differences were observed in income. A higher median income of $96,208 and a substantial proportion of individuals classified as upper-middle class (46%) were observed in the PRP group, as compared with a median income of $82,386 and 20% upper-middle class in the standard group. Employment status and insurance type also diverged significantly between the groups. The PRP group had a higher proportion of employed individuals (44%) and those with contracted insurance (68%), which is associated with better financial capability to access emerging treatments (P < .001). Additionally, the PRP group had a higher concentration of urban dwellers (83%) and a varied distribution across treatment sites and regions, underscoring potential geographical barriers in accessing PRP treatment. Distance to care was substantially higher for the PRP group with a median distance of 62 miles, as opposed to 31 miles in the standard group, indicating the distance that individuals are willing to travel to obtain PRP treatment (P < .001). Racial distribution exhibited slight variations, with a higher percentage of Asians in the PRP group (2.6%) as compared with the standard group (1.2%; P = .039). Additionally, marital status was significant for PRP versus standard of care (P < .01). The distribution of sex, ethnicity, and language between the groups did not exhibit pronounced differences, which contrasts with prior literature on health disparities in knee OA. For example, the proportion of females was 54% in the PRP group and 57% in the standard group, with a P value of .11. Interestingly, the severity of knee OA also did not exhibit a significant difference between the groups (P = .9).

Table 1.

Sociodemographic Variables and Descriptive Statistics ^a

Characteristic	Standard	PRP	P Value
Age, y	75 (68-84)	70 (62-77)	<.001 b
Area deprivation index	42 (31-54)	30 (16-46)	<.001 b
Sex			.11 c
Female	840 (57)	798 (54)
Male	626 (43)	670 (46)
Ethnicity			.3 c
Not Hispanic	1444 (98)	1439 (98)
Hispanic	22 (1.5)	29 (2.0)
Race			.039 c
White	1399 (95)	1373 (94)
Black	21 (1.4)	18 (1.2)
American Indian	7 (0.5)	11 (0.7)
Asian	17 (1.2)	38 (2.6)
Other	22 (1.5)	28 (1.9)
Language			.6 c
English	1451 (99)	1456 (99.2)
Non-English	15 (1.0)	12 (0.8)
Marital Status			<.001 c
Single	136 (9.3)	154 (10)
Married	904 (62)	1126 (77)
Divorced	138 (9.4)	75 (5.1)
Widowed	280 (19)	109 (7.4)
Separated	8 (0.5)	4 (0.3)
Income, continuous	82,386 (73,000-95,000)	96,208 (76,000-121,000)	<.001 b
Income, categorical			<.001 d
Middle class	1152 (79)	763 (52)
Poor class	1 (<0.1)	2 (0.1)
Lower-middle class	16 (1.1)	27 (1.8)
Upper-middle class	297 (20)	676 (46)
Rurality (Rural-Urban Commuting Area)			<.001 c
Urban	883 (60)	1224 (83)
Large rural city and town	306 (21)	147 (10)
Small and isolated small rural	278 (19)	97 (6.6)
Sites			<.001 c
Mayo-Minnesota	916 (62)	878 (60)
Mayo-Arizona	86 (5.9)	347 (24)
Mayo-Florida	88 (6.0)	179 (12)
Mayo–Health System	357 (24)	35 (2.4)
Detroit Medical Center	19 (1.3)	29 (2.0)
Region			<.001 c
Midwest	1261 (86)	842 (57)
West	93 (6.3)	364 (25)
South	112 (7.6)	262 (18)
Employment			<.001 c
Employed	383 (26)	640 (44)
Self-employed	71 (4.8)	111 (7.6)
Not employed	105 (7.2)	113 (7.7)
Retired	845 (58)	587 (40)
Disabled	62 (4.2)	17 (1.2)
Insurance			<.001 c
Contracted insurance	811 (55)	1004 (68)
Medicaid	44 (3)	22 (1.5)
Medicare	582 (40)	415 (28)
Tricare and Champus	9 (0.6)	16 (1.1)
Health maintenance organization	20 (1.4)	11 (0.7)
Severity			.9 c
Mild	590 (40)	603 (41)
Moderate	285 (19)	285 (19)
Severe	591 (40)	580 (40)
Distance to care	31 (6-63)	62 (11-201)

^aData presented as Median (IQR) or No. (%). Mayo Clinic is a large multisite academic medical center system in the upper Midwest; Detroit Medical Center is a smaller community medical center in the same region. Bold indicates P < .05. PRP, platelet-rich plasma.

^bWilcoxon rank sum test.

^cPearson chi-square test.

^dFisher exact test.

To mitigate the effects of confounding factors and better isolate the impact of individual variables, a multiple logistic regression model was applied highlighting several significant predictors of access to PRP treatment for knee OA (Table 2). Age was a notable factor, with each additional year reducing the odds of accessing PRP by 4% (odds ratio [OR], 0.96; P < .001). Income also played a critical role: individuals in the upper-middle income group had nearly double the odds of receiving PRP as compared with those in the middle class (OR, 1.94; P < .001). Employment status significantly influenced access, as retired and disabled individuals had a reduction of 49% (OR, 0.51; P < .001) and 84% (OR, 0.16; P < .001) in odds of receiving PRP, respectively. Geographic factors were prominent, with patients at the Mayo-Arizona and Mayo-Florida sites being 4.71 times (OR, 4.71; P < .001) and 2.19 times (OR, 2.19; P < .001) more likely to receive PRP than those at the Mayo-Minnesota site. Conversely, those at Mayo Clinic–Health System sites had an 85% reduction in odds (OR, 0.15; P < .001). Other significant factors included marital status, with widowed individuals being 44% less likely to access PRP (OR, 0.56; P = .005), and insurance type, where Medicaid and Medicare recipients had reduced odds of 59% (OR, 0.41; P = .005) and 31% (OR, 0.69; P < .001), respectively, when compared with those with contracted insurance. The model's performance was rigorously evaluated using a 10-fold cross-validation method, resulting in a training set AUC of 0.83 and a testing set AUC of 0.82 (Figure 1).

Table 2.

Multiple Logistic Regression Adjusted ^a

Predictor	Odds Ratio	95% CI	P Value
Age	0.96	0.95-0.97	<.001
Area deprivation index	1.00	0.99-1.00	.2
Sex
Female	1 [Reference]
Male	1.03	0.86-1.23	.7
Ethnicity
Not Hispanic	1 [Reference]
Hispanic	0.84	0.41-1.74	.6
Race
White	1 [Reference]
Black	0.55	0.26-1.17	.12
American Indian	0.65	0.20-2.07	.5
Asian	1.51	0.76-3.16	.3
Other	0.94	0.46-1.94	.9
Language
English	1 [Reference]
Non-English	1.09	0.41-2.92	.9
Marital Status
Single	1 [Reference]
Married	1.17	0.87-1.57	.3
Divorced	0.66	0.43-1.00	.052
Widowed	0.56	0.37-0.84	.005
Separated	0.31	0.07-1.11	.080
Income categorical
Middle class	1 [Reference]
Poor class	2.23	0.16-60.1	.6
Lower-middle class	1.74	0.81-3.84	.2
Upper-middle class	1.94	1.46-2.59	<.001
Rurality (Rural-Urban Commuting Area)
Urban	1 [Reference]
Large rural city and town	1.03	0.78-1.36	.8
Small and isolated small rural	0.59	0.43-0.80	<.001
Sites
Mayo-Minnesota	1 [Reference]
Mayo-Arizona	4.71	3.56-6.30	<.001
Mayo-Florida	2.19	1.62-2.97	<.001
Mayo–Health System	0.15	0.10-0.21	<.001
Detroit Medical Center	0.91	0.43-1.96	.8
Employment
Employed	1 [Reference]
Self-employed	1.09	0.74-1.60	.7
Not employed	0.67	0.47-0.94	.021
Retired	0.51	0.42-0.63	<.001
Disabled	0.16	0.08-0.29	<.001
Insurance
Contracted insurance	1 [Reference]
Medicaid	0.41	0.22-0.75	.005
Medicare	0.69	0.57-0.84	<.001
Tricare and Champus	1.99	0.74-5.70	.2
Health maintenance organization	0.45	0.19-1.05	.069
Severity
Mild	1 [Reference]
Moderate	0.94	0.74-1.19	.6
Severe	0.99	0.82-1.20	>.9
Akaike information criterion			3160
Total (training sample size)			2934

^aBold indicates P < .05.

Figure 1.

Logistic regression area under the curve (AUC). A receiver operating characteristic curve shows the predictive performance of the logistic regression model in classifying platelet-rich plasma versus standard care, with a training set AUC of 0.83 and a testing set AUC of 0.82.

To assess interactions among predictors, we adopted a random forest machine learning model. The model's variable importance was assessed by the mean decrease in Gini, identifying the top 5 influential variables as follows (Figure 2): distance to care (238 decrease in Gini), age (213), deprivation percentile (148), income (160), and site (126). The complex interactions among predictors were further explored using the Friedman H statistic, ¹⁴ as depicted through a network plot (Figure 3). Employment variables heavily interacted with age, deprivation percentile, distance to care, income, and region. This network plot also illuminated the complex interactions among all predictors, providing a more comprehensive understanding of variables influencing the accessibility to PRP treatment. The model's performance was gauged using a 10-fold cross-validation reflecting a training set AUC of 0.92 and a testing set AUC of 0.87 (Figure 4).

Figure 2.

Random forest variable importance among factors influencing access to platelet-rich plasma. The model's variable importance is assessed by the mean decrease in Gini, identifying the top 5 influential variables in predicting access to platelet-rich plasma treatment for knee osteoarthritis: distance to care (238 decrease in Gini), age (213), area deprivation index (ADI; 148), income (160), and site (126).

Figure 3.

Random forest interaction plot among factors influencing access to platelet-rich plasma. Complex interactions among predictors explored by the Friedman H statistic, as depicted through a network plot showing employment variables heavily interacting with age, deprivation percentile, distance to care, income, and region.

Figure 4.

Random forest area under the curve (AUC). A receiver operating characteristic curve shows the predictive performance of the random forest model in classifying platelet-rich plasma versus standard care, with a training set AUC of 0.92 and a testing set AUC of 0.87.

Discussion

The major findings in our study showed that sociodemographic and clinical factors significantly predicted access to PRP for knee OA. Logistic regression identified age (OR, 0.96; P < .001), upper-middle income (OR, 1.94; P < .001), retirement (OR, 0.51; P < .001), disability (OR, 0.16; P < .001), treatment site (OR, 4.71 for Arizona; P < .001), and widowed status (OR, 0.56; P = .005) as significant predictors, achieving AUCs of 0.83 (training) and 0.82 (testing). Random forest analysis confirmed distance to care, age, ADI, income, and treatment site as the top 5 predictors, with AUCs of 0.92 (training) and 0.87 (testing).

The investigation into PRP treatment for knee OA offers insight into the sociodemographic complexities influencing access to orthobiologics. Our findings resonated with the broader literature that identifies health care disparities, particularly reduced access to care based on key variables including distance, income, area deprivation, rurality, and the setting in which the care is offered. ³⁶ One of the most significant predictors to accessing PRP was distance to the care and, relatedly, geographic location where patients from small and isolated rural communities had a nearly 40% decrease in access. This urban-rural bias mirrors challenges faced by rural populations in the United States who often have limited access to advanced treatments and specialists, highlighting the need for expanded health care infrastructure in underserved areas.^1,11

Age emerged as a significant factor influencing access to PRP treatment for knee OA. Clinically, PRP has been shown to have temporary effectiveness at reducing inflammation and pain among patients with mild and moderate knee OA.^10,33 In addition, many patients prefer nonsurgical options, which is particularly pronounced among younger individuals.^2,33 The analysis revealed that younger patients were more likely to receive PRP treatment. This could be attributed to several factors, including increased clinical effectiveness, greater likelihood of seeking to maintain an active lifestyle, and stronger preference for less invasive treatment options. While PRP is generally recommended for patients with mild to moderate OA, our data indicated that it was being offered equally to those with moderate and severe knee OA. Given patient preferences for nonsurgical options, this trend suggests that patients, irrespective of the severity of their condition, were willing to explore PRP as a viable option before considering surgical interventions. ²⁶ This behavior may be driven by a desire to avoid the risks and lengthy recovery times associated with surgery or, potentially, by fear of surgery. Many patients prefer orthobiological treatments such as PRP in an effort to delay or completely avoid surgical procedures. ¹¹

Another major determinant to accessing PRP for knee OA surrounded financial barriers as wealthier, employed, and insured individuals were more likely to receive PRP treatment. Currently, PRP for knee OA is neither covered under Centers for Medicare and Medicaid Services programs nor accepted by most private insurers. ³⁷ PRP treatment generally costs $175 to $4973, ²⁵ and such out-of-pocket expenses may be substantive for many patients on low or fixed incomes, especially if patients are required to travel distances and possibly take time off work. The finding of economic disparities in our study aligns with broader trends that link economic stability to health care access. ²⁴ Unemployment and lower income levels are associated with poorer health outcomes, ³⁰ a fact underscored by the National Healthcare Quality and Disparities Report, which points to economics and insurance as crucial determinants of health care access. ²⁸ Our regression model highlighted that insurance was a major factor and that patients with Medicare or Medicaid were significantly less likely to have PRP, but this variable showed less importance perhaps because it is overshadowed by other variables surrounding financial means. Financial barriers seen as income, employment, and insurance were major factors that significantly predicted access to PRP.

An interesting disparity in PRP treatment for knee OA relates to treatment location. Patients at Mayo-Arizona and Mayo-Florida were more likely to receive PRP than those in Mayo-Minnesota, while patients in Mayo Clinic–Health System or DMC were less likely to receive PRP than those in Mayo-Minnesota. Despite the low sample size of DMC, all Mayo Clinic locations (Minnesota, Arizona, Florida, Health System) reflect variations in health care infrastructure, specialist availability, or regional practices. The concentration of PRP access in specific locations raises equity concerns, particularly for underserved and remote areas. Not all musculoskeletal specialists offer PRP for knee OA treatment, thus indicating practice variation and the potential to standardize practice among physicians. ²⁹ Furthermore, cost-effectiveness analyses show that while PRP provides greater improvement in quality-adjusted life-years than hyaluronic acid injection, its higher cost results in a less favorable cost per quality-adjusted life-year. ⁷ All these variations suggest that health care access is shaped by individual economic conditions, geographic location, and local health care practices, rather than being uniform across the country.

This study has notable limitations. First of all, DMC had a considerably smaller sample size as compared with other locations. Second, while including DMC site increased sample diversity, data from only 2 medical centers (2022-2023) do not represent the entire US population. The cross-sectional design limits causal inferences. Logistic regression encountered overfitting when interaction terms were introduced, necessitating the use of random forest. Additional limitations include the potential for residual confounding and the absence of patient-reported outcome measures in our data set. While the NLP intercoder reliability test was high, that score could be increased more with increased sample size and diverse sites. Additionally, intracoder reliability was not formally assessed; future studies should include repeated coding by the same coder at different time points to evaluate consistency within individual coders. These limitations highlight the need for further methodological refinements to fully understand the complex nature of health disparities in PRP treatment accessibility for knee OA.

Conclusion

Our study demonstrated that both models discerned key barriers (income, distance to care, retirement status) and facilitators (regional differences) affecting PRP accessibility. These findings provide potential policy avenues aimed at improving PRP access for underserved demographics, offering a quantitative blueprint to foster equitable access to orthobiologic therapies for knee OA.

Appendix

Code Book Rating Severity of Knee OA

Clinicians use various languages to denote the severity of each knee among patients with knee OA. The severity is based on diagnostic imaging and may be interpreted by a radiologist or the treating clinician.

We iteratively observed the number of randomized electronic health records of patients with knee OA to identify the nomenclature and language used within. The following terms form a language that is routinely used to describe the severity of knee OA.

• Only minor

• Only mild

• Only moderate

• Only moderate degenerative

• Only severe

• Only advanced

• Only narrowing

• Mild to moderate

• Moderate to severe

• Moderate to advanced

• Mild narrowing

• Moderate narrowing

• Advanced narrowing

• Complete loss

We modified labeling for unilateral and bilateral knee OA to a single severity diagnosis as mild, moderate, or severe.

Codebook

	Within Electronic Health Record
Codebook Label	Unilateral Knee	Bilateral Knee
Mild: 1	Only minor Only mild Mild narrowing	Mild + mild
Moderate: 2	Only moderate Only degenerative Only moderate degenerative Mild to moderate Mild-moderate Moderate narrowing	Mild + moderate Mild to moderate + mild Mild-moderate
Severe: 3	Only severe Only advanced Moderate to severe Moderate-severe Moderate to advanced Moderate-advanced Advanced narrowing Complete loss Only severe degenerative Only advanced degenerative Moderate to severe degenerative Moderate-severe degenerative Moderate to advanced degenerative Moderate-advanced degenerative Advanced narrowing degenerative Complete loss	Mild + severe Moderate + severe Mild + advanced Moderate + advanced Severe + advanced Mild + narrowing Moderate + narrowing Severe + narrowing Moderate to severe + moderate Moderate-severe + moderate Moderate to advanced + severe Moderate-advanced + severe Mild to moderate + advanced Mild-moderate + advanced Moderate to severe + advanced Moderate-severe + advanced Moderate to advanced + advanced Moderate-advanced + advanced Mild to moderate + narrowing Mild-moderate + narrowing Moderate to severe + narrowing Moderate-severe + narrowing Moderate to advanced + narrowing Moderate-advanced + narrowing Complete loss + anything