Development and validation of a cuff size prediction model for pediatric blood pressure measurement outside the office

Abstract

Background:

Mid-arm circumference (AC) is typically measured in-person to determine cuff size for ambulatory blood pressure monitoring (ABPM). Our objective was to develop a model to predict cuff size using electronic health record (EHR) data to enable virtual ABPM programs.

Methods:

We developed a prediction model for youth 3–21 years in the National Health and Nutrition Examination Survey. Using linear regression, we considered piecewise and polynomial effects of age, sex, height, and weight as predictors of AC. We selected the model with the lowest bootstrapped root mean square error (RMSE) and predicted residual error sum of squares with leave-one-out cross-validation. We validated the model in pediatric hypertension clinics at an academic medical center.

Results:

Based on 34,517 youth in the derivation cohort (median 12 years [25^th, 75^th percentiles: 7, 16], 49.0% female, 16.6% obesity), the final model included age, age², sex, height, weight, weight², weight³, and all-possible interactions between age and height and between age and weight (adjusted R2: 0.97, RMSE: 1.20). In the external validation cohort of 107 youth (median 14 years [25^th, 75^th percentiles: 10, 17], 35.5% female, 61.7% obesity), observed and predicted AC were highly concordant (ρc: 0.94) with mean bias of −0.5 cm (95% limits of agreement: −5.2, 4.1). Overall agreement for the correct cuff size was 87.5% in the derivation cohort (weighted κ: 0.92) and 83.2% in the validation cohort (weighted κ: 0.90).

Conclusions:

We found substantial agreement between observed and predicted AC using basic EHR data. This model may facilitate ABPM when in-person visits are not feasible.

Introduction

Due to the limited accuracy of clinic blood pressure (BP), the 2017 American Academy of Pediatrics (AAP) guideline on hypertension recommends that children with high BP in clinic undergo ambulatory blood pressure monitoring (ABPM) outside clinic to diagnose hypertension [1]. However, access to ABPM remains limited as only 8% of youth with incident hypertension diagnosed after the 2017 guideline have completed ABPM [2].

To improve access to ABPM, our team is developing a virtual pathway for patients to complete ABPM without a clinic visit. Patients receive the ABPM device via a courier service, attend a brief telehealth visit on how to use the device, initiate the 24-hour monitoring period at their convenience, and then ship the device back to our team. One barrier to this virtual approach is how to select the proper cuff size for ABPM. Using an appropriately sized cuff is critical to obtaining accurate BP readings and to the proper classification of a patient’s hypertension status [3, 4]. Guidelines recommend measuring the mid-arm circumference (AC) to determine the cuff size [1], which is not possible in a virtual pathway. Although prediction models have been developed to estimate an adult patient’s cuff size using basic information from the electronic health record (EHR) [5–8], pediatric models do not exist. Pediatric populations have smaller AC and cuff sizes, requiring prediction models derived in children.

The objective of this study was to develop and validate a model to predict cuff size for ABPM in children and adolescents. We hypothesized that cuff size can be predicted with good accuracy using demographic and anthropometric data readily available in the EHR. A pediatric cuff size prediction model has the potential to improve utilization of ABPM by allowing patients to complete ABPM without an in-person specialty visit.

Methods

Derivation cohort

To develop the prediction model, we used data from the 1999–2018 National Health and Nutrition Examination Survey (NHANES). NHANES is a nationally representative, annual survey of the U.S. civilian, non-institutionalized population. We included youth aged 3 to 21 years as pediatric guidelines recommend universal screening for hypertension at annual well-visits starting at 3 years old [1]. We included non-pregnant youth with complete data for age, sex, height (in centimeters), weight (in kilograms), and AC (in centimeters). Consistent with the 2017 AAP guideline’s recommendation [1], NHANES uses a standardized procedure to measure AC at the midpoint of the upper arm between the acromion process and the olecranon with the right arm flexed 90 degrees at the elbow [9].

Validation cohort

For the validation cohort, we retrospectively identified youth aged 3 to 21 years referred to the pediatric hypertension clinics or pediatric nephrology clinics in the Yale–New Haven Health System from 2023 to 2025. We included youth if they had an AC recorded in the EHR during their referral clinic visit or at the time of placement of an ABPM device. AC was measured by either the specialty nurse or the attending pediatric nephrologist as part of routine clinical practice. We extracted additional demographic and anthropometric information from the EHR at the time of AC measurement.

Statistical analysis

Based on the existing literature in adults [8], we considered age, sex, height, and weight as candidate predictors of AC. We visualized all-possible pairwise scatterplots with a locally estimated scatterplot smoothing (loess) for AC (y-axis) with age, sex, height, and weight to explore polynomial and piecewise effects. We also examined plots of AC versus each predictor variable stratified by a third variable to investigate potential interactions. We considered 11 ordinary least squares regression models of increasing complexity (Supplemental Table 1). Through internal cross-validation, we assessed model fit by evaluating the prediction accuracy of each model using out-of-sample observations that were not used in model derivation. We selected the best-fitting model based on (1) the lowest root mean square error (RMSE) from the 0.632+ bootstrap method with 5000 bootstrapped samples [10], and (2) the lowest predicted residual error sum of squares (PRESS) statistic from leave-one-out cross-validation [11]. We also reported the adjusted R² for each candidate model. Following selection of the best-fitting model, we refit the model in the derivation cohort after excluding outlier observations (n=54 [0.2%]) with standardized residuals >5 standard deviations.

We applied the best-fitting model to the validation cohort and obtained predicted values of AC. We plotted observed versus predicted values of AC and calculated Lin’s concordance correlation coefficient (ρ_c) to evaluate agreement on the continuous scale [12]. We used Bland–Altman plots to estimate mean bias and 95% limits of agreement (LoA) for the prediction model [13]. To evaluate categorical agreement, we classified AC measurements into cuff sizes based on the SpaceLabs ABPM cuff sizing cutoffs: child (12 to <17 cm), small adult (17 to <24 cm), adult (24 to <32 cm), large adult (32 to <40 cm), and extra large adult (40 to <50 cm). We determined the percent overall agreement between observed and predicted cuff size category, with 95% confidence intervals calculated using the Wilson score interval approach. We also assessed agreement using Cohen’s weighted kappa statistic [14], which accounts for ordinal data and gives preference to closer values, with 95% confidence intervals calculated by 5000 bootstrapped repetitions. In exploratory subgroup analyses, we evaluated agreement in the validation cohort by subgroups of age, sex, race, and body mass index. Consistent with prior literature, we did not account for survey weights in NHANES because we were not making national estimates [7, 8]. We conducted all statistical analyses in Stata SE, version 17.0. We used the two-sided alpha of 0.05 for statistical inference. Based on the final model, we created a publicly available online calculator using the R Shiny App [15]. The study was deemed exempt by the Yale University Institutional Review Board.

Results

Based on 34,517 youth in the derivation cohort (median 12 years [25^th, 75^th percentiles: 7, 16], 49.0% female, 16.6% obesity) (Table 1), the final model (Model #10) included age, age², sex, height, weight, weight², weight³, and all-possible interactions between age and height and between age and weight (adjusted R²: 0.97, RMSE: 1.20, PRESS: 49,417) (Supplemental Table 1). Models #10 and #11 had nearly identical adjusted R² and RMSE, with marginally lower PRESS in Model #11, but Model #10 was selected for parsimony and to facilitate clinical use. All predictors were significantly associated with AC (Table 2). The final AC prediction model is shown below and can be accessed at: https://yale-ctra.shinyapps.io/BPCuffCalculator/.

Table 1.

Characteristics of Derivation and Validation Cohorts.

	Derivation Cohort (N=34,517)	Validation Cohort (N=107)
Age, median (25th, 75th percentiles), years	12 (7–16)	14 (10–17)
Sex, No. (%)
Male	17,595 (51.0%)	69 (64.5%)
Female	16,922 (49.0%)	38 (35.5%)
Race/ethnicity,a No. (%)
Hispanic	12,179 (35.3%)	46 (43.0%)
Non-Hispanic Black	9595 (27.8%)	32 (29.9%)
Non-Hispanic White	9532 (27.6%)	18 (16.8%)
Other or Multiracial	3,211 (9.3%)	11 (10.3%)
BMI, median (25th, 75th percentiles), kg/m2	19.8 (16.6–24)	29.2 (22.7–34.6)
BMI Category, No. (%)
Underweight	2023 (5.9%)	1 (0.9%)
Normal	19,832 (57.5%)	22 (20.6%)
Overweight	6945 (20.1%)	18 (16.8%)
Obesity	5717 (16.6%)	66 (61.7%)
Arm circumference, median (25th, 75th percentiles), cm	24.5 (19.4–29.1)	32 (27–36)
Cuff Size, No. (%)
Arm circumference <12 cm	0 (0%)	0 (0%)
Child (12–17 cm)	3011 (8.7%)	1 (0.9%)
Small adult (17–24 cm)	13,240 (38.4%)	14 (13.1%)
Adult (24–32 cm)	13,282 (38.5%)	38 (35.5%)
Large adult (32–40 cm)	4273 (12.4%)	43 (40.2%)
Extra large adult (40–50 cm)	686 (2.0%)	10 (9.4%)
Arm circumference >50 cm	25 (0.1%)	1 (0.9%)

^aRace and ethnicity information were extracted from the electronic health record. BMI, body mass index; kg, kilograms; cm, centimeters

Table 2.

Final Regression Model for Prediction of AC in Derivation Cohort

Variable	Coefficient	Standard Error	P-value
Height	−0.151	0.006	<0.001
Female sex	−0.308	0.013	<0.001
Age	1.238	0.063	<0.001
Age2	−0.049	0.003	<0.001
Weight	1.221	0.024	<0.001
Weight2	−0.013	4.8 × 10−4	<0.001
Weight3	5.6 × 10−5	3.2 × 10−6	<0.001
Age × Height	−0.002	9.5 × 10−4	0.008
Age2 × Height	2.4 × 10−4	3.6 × 10−5	<0.001
Age × Weight	−0.059	0.004	<0.001
Age × Weight2	0.001	5.8 × 10−5	<0.001
Age × Weight3	−5.1 × 10−6	3.7 × 10−7	<0.001
Age2 × Weight	0.001	1.5 × 10−4	<0.001
Age2 × Weight2	−2.5 × 10−5	2.0 × 10−6	<0.001
Age2 × Weight3	1.3 × 10−7	1.1 × 10−8	<0.001
Constant	14.623	0.418	<0.001

AC, mid-arm circumference

$$\widehat{\mathit{\text{AC}}}=14.623-0.151(\mathit{\text{Height}})-0.308(\mathit{\text{Female}}sex)+1.238(Age)-0.049\left(Ag{e}^2\right)+1.221(\mathit{\text{Weight}})-0.013\left(\mathit{\text{Weigh}}{t}^2\right)+\left(5.6\times {10}^{-5}\right)\left(\mathit{\text{Weigh}}{t}^3\right)-0.002(Age\times \mathit{\text{Height}})+\left(2.4\times {10}^{-4}\right)\left(Ag{e}^2\times \mathit{\text{Height}}\right)-0.059(Age\times \mathit{\text{Weight}})+0.001\left(Age\times \mathit{\text{Weigh}}{t}^2\right)-\left(5.1\times {10}^{-6}\right)\left(Age\times \mathit{\text{Weigh}}{t}^3\right)+0.001\left(Ag{e}^2\times \mathit{\text{Weight}}\right)-\left(2.5\times {10}^{-5}\right)\left(Ag{e}^2\times \mathit{\text{Weigh}}{t}^2\right)+\left(1.3\times {10}^{-7}\right)\left(Ag{e}^2\times \mathit{\text{Weigh}}{t}^3\right)$$

In the external validation cohort of 107 youth (median 14 years [25^th, 75^th percentiles: 10, 17], 35.5% female, 61.7% obesity), observed and predicted AC were highly concordant (ρ_c: 0.94 [95% CI: 0.91, 0.96], P<0.001) with a mean bias of −0.5 cm (95% LoA: −5.2, 4.1) (Figure 1). Overall agreement for selecting the appropriate cuff size was 87.5% (95% CI: 87.1, 87.8) in the derivation cohort with weighted kappa of 0.92 (95% CI: 0.92, 0.92) and 83.2% (95% CI: 75.0, 89.1) in the validation cohort with weighted kappa of 0.90 (95% CI: 0.85, 0.95) (Table 3). All the incorrect predicted cuff sizes in the validation cohort were within 1-size of the observed cuff size, with 3.7% of patients undercuffed and 13.1% of patients overcuffed. The characteristics of patients with disagreement between their predicted and observed cuff sizes are shown in Supplemental Table 2. For most patients with an incorrect predicted cuff size, the predicted AC was <2 cm from the AC cutoff for the next cuff size (Supplemental Figure). The exploratory subgroup analyses are presented in Table 4. For the subgroup analysis by body mass index, Lin’s concordance correlation coefficient ranged from 0.90 (95% CI: 0.86, 0.94) in youth with overweight/obesity to 0.95 (95% CI: 0.91, 0.99) in youth without overweight/obesity; corresponding weighted kappa values ranged from 0.83 (95% CI: 0.74, 0.92) to 0.94 (95% CI: 0.79, 1.00).

Figure 1.

(A) Predicted vs. Observed Arm Circumference in Validation Cohort. (B) Bland–Altman Plot in Validation Cohort with Mean Bias (dashed grey) and 95% Limits of Agreement (solid grey).

Table 3.

Agreement Between Observed and Predicted Cuff Size in Validation Cohort

		Observed Cuff Size
		Child	Small Adult	Adult	Large Adult	XL Adult	>50 cm
Predicted Cuff Size	Child	1 (100%)	0	0	0	0	0
	Small Adult	0	12 (86%)	1 (3%)	0	0	0
	Adult	0	2 (14%)	29 (76%)	1 (2%)	0	0
	Large Adult	0	0	8 (21%)	38 (88%)	2 (20%)	0
	XL Adult	0	0	0	4 (9%)	8 (80%)	0
	>50 cm	0	0	0	0	0	1 (100%)

Overall agreement: 83.2% (95% CI: 75.0, 89.1)

Weighted kappa: 0.90 (95% CI: 0.85, 0.95)

For reference, overall agreement was 87.5% (95% CI: 87.1, 87.8) in the derivation cohort with weighted kappa of 0.92 (95% CI: 0.92, 0.92).

Table 4.

Subgroup Analyses of Agreement Between Observed and Predicted Arm Circumference and Cuff Size in Validation Cohort

Subgroup	Lin’s Concordance Correlation Coefficient (95% CI)	Overall Agreement (95% CI)	Weighted Kappa (95% CI)
Overall (n=107)	0.94 (0.91, 0.96)	83.2% (75.0, 89.1)	0.90 (0.85, 0.95)
Age
<13 years old (n=38)	0.93 (0.88, 0.97)	81.6% (66.6, 90.8)	0.86 (0.74, 0.98)
≥13 years old (n=69)	0.92 (0.89, 0.96)	84.1% (73.7, 90.9)	0.89 (0.82, 0.96)
Sex
Female (n=38)	0.94 (0.90, 0.98)	84.2% (69.6, 92.6)	0.91 (0.82, 0.99)
Male (n=69)	0.93 (0.90, 0.96)	82.6% (72.0, 89.8)	0.89 (0.82, 0.96)
Race
Hispanic (n=46)	0.93 (0.89, 0.97)	78.3% (64.4, 87.7)	0.87 (0.77, 0.96)
Non-Hispanic Black (n=32)	0.93 (0.89, 0.98)	81.3% (64.7, 91.1)	0.91 (0.82, 1.00)
Non-Hispanic White (n=18)	0.95 (0.89, 1.00)	88.9% (67.2, 96.9)	0.91 (0.77, 1.00)
Other or Multiracial (n=11)	0.89 (0.78, 1.00)	100.0% (74.1, 100.0)	1.00
Body mass index
No overweight/obesity (n=23)	0.95 (0.91, 0.99)	95.7% (79.0, 99.2)	0.94 (0.79, 1.00)
Overweight/obesity (n=84)	0.90 (0.86, 0.94)	79.8% (70.0, 87.0)	0.83 (0.74, 0.92)

Discussion

We have developed a model using age, sex, height, and weight from the EHR to predict AC in youth. In external validation using a real-world cohort of youth referred to pediatric hypertension clinics, we found substantial agreement between the observed and predicted AC. To the best of our knowledge, this is the first prediction model for pediatric BP cuff size based on information that is readily available in the EHR.

Prediction equations for AC have been developed for adults using data from NHANES III (1988–1994) [8]. These adult prediction models were stratified by sex and included weight, height, age, and age² for males (adjusted R² 0.80) and weight, height, age, age², and the interactions between weight and age and between height and age for females (adjusted R² 0.86). In model validation using data from NHANES 2001–2012 [6], these prediction equations assigned 83.0% of males and 80.0% of females to the correct cuff size, although performed poorly in the few individuals that needed small adult cuffs. Our prediction model achieved overall agreement of 83.2% in the validation cohort, with consistent results across cuff sizes.

One of the most common errors during office BP measurement is use of an incorrect cuff size (“miscuffing”) [16], which has been reported in up to 40% of BP measurements in pediatric primary care [17]. Selecting the correct cuff size is critical to the accurate measurement of BP [4], with the degree of measurement error increasing as the selected cuff size gets further away from the appropriate cuff size. A recent randomized crossover trial in adults demonstrated that using a cuff that was 2 sizes too-small overestimated automated systolic BP by 19.5 mmHg in individuals requiring extra large cuffs while using a cuff that was 1 size too-small overestimated systolic BP by 9.6 mmHg [3]. This “undercuffing” tends to overestimate BP to a greater degree than “overcuffing” underestimates BP, as this trial showed that using a cuff that was 2 sizes too-large underestimated systolic BP by 7.5 mmHg in individuals requiring a small adult cuff and 1 size too-large underestimated systolic BP by 3.6 mmHg [3]. Given that undercuffing accounts for 84% of miscuffings in adults [18], the effect of miscuffing can substantially impact the diagnosis of hypertension, particularly in patients with larger arm circumferences.

In the external validation of our prediction model, all the predicted cuff sizes were within 1-size of the actual cuff size. Most (78%) of the miscuffings in the validation cohort represented overcuffing, which would not have biased BP as much as undercuffing. To minimize the risk of miscuffing, a virtual ABPM pathway could send patients multiple BP cuffs within 1-size of the predicted cuff size along with measuring tape so that patients could select the appropriate cuff based on their self-measured AC. Nevertheless, our results show that this prediction model can correctly assign >80% of BP cuff sizes in a cohort of youth referred to pediatric hypertension clinics, with 13.1% of patients being overcuffed by 1 size and only 3.7% of patients being undercuffed by 1 size. Thus, nearly all patients can use the predicted cuff size without significantly biasing their BP measurements.

Our study has several limitations. First, the derivation cohort had AC measured according to the NHANES protocol, which is consistent with the 2017 AAP hypertension guideline’s recommendation [1]. In contrast, patients in the validation cohort had their AC measured by specialty nurses or physicians but these measurements were not standardized. Second, the derivation cohort had weight and height measured while wearing a paper examination gown without any shoes according to the NHANES protocol [19]. This standardized process for weight and height measurement was not necessarily followed in the pediatric hypertension clinics. Despite these differences in measurement technique between the derivation and validation cohorts, our model performed well in the clinical environment in which it is meant to be used. Third, we included children starting at 3 years old even though reference norms for pediatric ABPM start at 5 years old [20]. ABPM may not be used for these younger children but the predicted AC can be used when selecting a home BP self-monitor; although, the availability of validated pediatric home BP monitors is limited [21] and most home BP monitors come with only a single adult-sized cuff [22]. Fourth, the validation cohort had a higher proportion of male patients and patients with obesity than the derivation cohort. However, the validation cohort reflects the referral population of youth with hypertension at an academic medical center. The exploratory subgroup analyses demonstrated strong agreement based on weighted kappa values >0.80 in all subgroups of age, sex, race, and body mass index [23]. Fifth, due to the number of terms and complexity of the final model, we created an online calculator that can be easily accessed by pediatric clinicians and nurses.

Lastly, we did not compare BP measurements using the predicted versus measured cuff size to determine the impact of an incorrect predicted cuff size on the measured BP. The Health Measures at Home Study [24] has addressed this question, to some extent, for the adult regression equations by Ostchega et al. [8]. In this pilot study of 128 adults recruited from NHANES in 2011–2012, home BP was measured using the predicted cuff size based on the regression equations [8] and compared with the BP measured at the NHANES mobile exam center using the actual cuff size based on the patient’s AC [24]. Although these BP measurements may not be directly comparable across the home and mobile exam settings, the mean difference in systolic/diastolic BP was <1 mmHg and overall agreement in hypertension status was 91% between the home BP and mobile exam BP measurements, with 86% agreement between the predicted and measured cuff size [19, 24]. While direct measurement of AC remains the gold standard for selecting the appropriate cuff size, using the predicted cuff size can still obtain BP measurements with a high degree of accuracy.

Our prediction model can be implemented to promote the timely completion of ABPM. In the current workflow, the primary care clinician identifies a child with high BP in the office, coordinates repeat visits to confirm sustained hypertension, and refers the patient to the pediatric hypertension clinic where they typically have an initial appointment before scheduling another date for ABPM. Delays in referral are common [25], which contributes to the high prevalence of left ventricular hypertrophy at the time of subspecialty evaluation [26]. This clinical inertia may also reflect clinicians’ limited confidence in the accuracy of office BP, which can be mitigated by increasing access to ABPM [27, 28]. In addition, children may undergo unnecessary workups for white coat hypertension while they are waiting for their subspecialty appointment and ABPM [29]. In a virtual pathway, the ABPM coordinator can use this online calculator to select an appropriately sized cuff, ship the device to the patient, conduct a brief training session via telehealth, and then complete ABPM without requiring multiple in-person specialty clinic visits. Although there are other barriers to a virtual ABPM program like the potential for equipment loss, prolonged shipping times, and billing for telehealth sessions, the ability to accurately predict a patient’s cuff size can inform the design of innovative pathways for more efficient completion of ABPM. Reducing the time to ABPM has the potential to prevent the delays in care and loss to follow-up associated with the current referral-based approach. Future studies should prospectively evaluate whether predicting cuff size as part of a virtual ABPM program maintains diagnostic accuracy and decreases the time to diagnosis, treatment, and control of hypertension.

Conclusion

We have developed and validated a regression equation to predict the BP cuff size for ABPM. Although direct measurement of AC is the optimal approach to select an appropriate BP cuff size, we found substantial agreement between the observed and predicted AC from a model using basic information available in the EHR. This model may be used to facilitate BP monitoring outside the clinic for telehealth visits or when in-person visits are not feasible.

Data Availability Statement:

The datasets analyzed during the current study are available from the corresponding author on reasonable request.