83 Prediction of histopathological adenomyosis Table 3.4 presents the results of both univariate and multivariate logistic regression analysis. Univariate logistic regression analysis showed p-values <.10 for: age at MRI, history of curettage, mean JZ thickness, JZ Max, JZ Diff, JZ/MYO, mean uterine volume, JZ Max ³12 mm, JZ Diff ³5 mm, and the presence of HSI foci. The potential predictors showed no two-way interaction; however, mean JZ thickness, JZ Max, and JZ Diff did show multicollinearity. These variables were not included in the multivariate regression model to avoid overoptimism. Nevertheless, high diagnostic performance was found for dysmenorrhoea and AUB (sensitivity/specificity >70%). Additionally, due to clinical relevance, BMI was manually forced into the multivariate model. The final model included age at MRI, BMI, history of curettage, dysmenorrhoea, AUB, mean JZ thickness, JZ Diff ³5 mm, JZ/MYO >.40, and the presence of HSI foci. In this model, mean JZ thickness, JZ/MYO >.40 and the presence of HSI foci reached statistical significance. Preference was given to variables with the most statistical significance in univariate analysis, and the number of included variables in the model was kept to a minimum. To further correct the model for overfitting, a shrinkage factor of .747 was applied. Since LOESS already showed a good model fit for the continuous variables of interest, no modifications were necessary. The formula for the final prediction model therefore is as follows: != ! "( A$; 7% B& (%&@ (9 .:*:+5,9$% (--(.;%%∗7. 0?("*@) $9 ?( 203) ∗4. ∗0. C0 (* ),$) $( D( E5 627 8(9;:%;7 ?9 <" =@>9:?% 808 )-∗.. 0% *( F; )%$7 ?( B" %@-9@? G0H)∗∗. .",((F())$$ (GH I6<< ³J BB (;%7?" @9?0)∗.(*0)&K# ! $ " %L.0+ (;%7?" @9?0)∗".**,M$ (NO4 P9=6 (;%7?" @9?0)∗"."+F) Discrimination performance evaluation of this prediction model showed an AUC of .776. A sub-analysis was conducted to assess whether the clinical query presented to the pathologist affected model diagnostic performance. The pathologist was directly asked to assess for the presence of adenomyosis in 142/296 patients. Model diagnostic performance did not improve significantly when adenomyosis was specifically evaluated for (data not shown). Calibration performance evaluation showed an observed to expected ratio of 1.2, and the Hosmer and Lemeshow test that did not reach statistical significance (p=.688).
RkJQdWJsaXNoZXIy MTk4NDMw