Data Analysis and Model Development
The study was conducted conform both the STROBE 22 and the TRIPOD statements 23 (see supplementary files C and D for the appropriate checklists). All statistical analyses were conducted with IBM SPSS Statistics, version 28.0 (IBM Corp., Armonk, NY, USA). Flowcharts were created using Miro (Miro, Amsterdam, the Netherlands). Except for univariate logistic regression analysis, a p-value of <.05 was considered statistically significant for all variables.
Between-group differences were compared between patients with and without a histopathological adenomyosis diagnosis after hysterectomy. For clinical characteristics and primary MRI parameters, counts and frequencies were reported. For normally distributed continuous variables, means and standard deviations were calculated. For continuous variables that were not normally distributed, medians and inter-quartile ranges were given. To assess between-group differences for continuous variables, Student’s t-Test and Mann-Whitney U test were used. For categorical variables, the Chi Squared test was used.
For all possible predictive factors, sensitivity, specificity, PPV, NPV, positive likelihood ratio (PLR), negative likelihood ratio (NLR), and accuracy were calculated. Potential threshold values of continuous variables were investigated using Receiver Operator Characteristics (ROC) curves and Area Under the Curve (AUC) to identify appropriate cut-off values, and to test the prognostic diagnostic potential for histopathological adenomyosis diagnosis.
For the development of the prediction model, the methodology as described by Grant et al. 24 and the TRIPOD guidelines were followed 23. For all individual potential predictors for a histopathological adenomyosis diagnosis, a univariate logistic regression analysis was first performed. The odds ratios (ORs) with their corresponding 95% confidence intervals (CIs) were reported. Missing values were dealt with by multiple imputation. Furthermore, interaction terms were used to test possible interaction between individual predictive factors. Tests for multicollinearity were performed as well to assess potential correlation between predictors. Individual variables were used for inclusion into the multivariate logistic regression model if they had a p- value <.10 in the univariate logistic regression analysis, or if they were considered clinically relevant, and if they had a high diagnostic performance (sensitivity/specificity>70% or AUC >0.70). Overfitting of the model was avoided by reducing the number of variables included in the model and by using shrinkage factors. Model fit was further improved by including additional predictive power of continuous variables based on locally weighted smoothing (LOESS).
The final model was evaluated for discrimination and calibration performance. The AUC was obtained to discriminate between women with and without a histopathological adenomyosis diagnosis after hysterectomy. To assess the calibration of the predicted probabilities, and to show the relation between predicted and observed probabilities for the histopathological adenomyosis diagnosis, an observed to expected ratio was calculated and a Hosmer and Lemeshow Test was performed.