A total of 158 lower extremities from 79 children were included in the study. Of those, 28 (35.4%) had bilateral Blount’s disease, 28 (35.4%) had unilateral involvement (9 (11.4%) right side, and 19 (24.1%) left side), and 23 (29.1%) had bilateral physiologic bowlegs (Table 1). Demographic and clinical information on lower extremities categorized by the study endpoint (Blount’s disease (n = 84) and physiologic bowlegs (n = 74)) were summarized and compared. Patients diagnosed with Blount’s disease were significantly older (27 ± 5.2 vs. 24.9 ± 6.9 months, p = 0.030), and had greater FTA (13.5 ± 6.2° vs. 9.2 ± 7.3°, p < 0.001), greater MDA (14.5 ± 4.0° vs. 10.0 ± 4.4°, p < 0.001), and higher MMB (127.4 ± 6.1° vs. 118.3 ± 6.2, p < 0.001) (Table 2). The distribution of variables after categorization with a pre-specified cut-off point is presented. Of all observations, only patient BMI information was missing for 62 (39.2%) patients. Therefore, multiple imputation analysis was performed using all other predictors (age, gender, FTA, MDA, and MMB) as independent predictors by the PMM method. The interobserver reliability of radiographic parameter measurement showed a substantial agreement with an ICC greater than 0.9 for all radiographic measurements.
Univariable logistic regression analysis revealed age, FTA, MDA, and MMB to be statistically significant predictors of Blount’s disease (Table 3
). Nevertheless, all candidate predictors were included in the full model multivariable logistic regression analysis using the multiple imputed datasets. Of the six predictors, three were identified as independent predictors including age ≥ 24 months (mOR 2.75, 95% CI 1.09 to 6.95, p
= 0.03), MDA > 16° (mOR 11.65, 95% CI 2.44 to 55.63, p
= 0.002), and MMB ≥ 122° (mOR 4.47, 95% CI 1.59 to 11.52, p
= 0.005). However, previous studies identified BMI as a strong predictor for Blount’s disease. Therefore, after discussion with all investigators, we decided to include patient BMI along with other independent predictors in the final predictive model. The discriminative ability of the final model was found to be excellent, with an AuROC of 0.85 (95% CI 0.79 to 0.91) (Figure 1
). The regression coefficient for each predictor from the final model was then transformed into a weighted score (Table 4
). The scoring scheme with a total score from 0 to 8 was then classified into three risk groups for clinical implementation. The groups were defined as low-risk, moderate-risk, and high-risk based on a total score > 2.5, within 2.5 to 5.5, or >5.5, respectively (Table 5
). The mean total score was significantly different between the Blount’s disease group and the physiologic bowlegs group (5.2 ± 0.2 vs. 2.5 ± 0.2, p
< 0.001). Model calibration is presented via calibration plots, as recommended by the TRIPOD statement in Figure 2 
. Internal validation using the bootstrap resampling method revealed an optimism of 0.018 (range 0.018 to 0.028).
Figure 1. The area under the receiver operating characteristic (ROC) of the final proposed diagnostic model, including age, body mass index, metaphyseal-diaphyseal angle, and medial metaphyseal beak angle.