Ultrasound Based Decision Tree Analysis Modelling to Predict AAGL Endometriosis Surgical Complexity Levels

Abstract

Study Objective Develop ultrasound-based endometriosis severity staging models to predict AAGL surgical complexity levels. Design Univariate analysis used to identify ultrasound features of endometriosis that best predict surgical complexity and build decision tree models of staging. Setting Multicentre. Patients or Participants Pre-existing dataset (n= 640) with comprehensive ultrasound data, collected prospectively. Interventions Observational. Measurements and Main Results 4 stage (model 1) and 3 stage (model 2) ultrasound-based decision tree analysis models were developed. The data were divided into training data (n=448 (70%)) and test data (n=192 (30%)). Univariate analysis via chi-squared test was performed on 34 sonographic finding categories. The C4.5 algorithm was used to identify optimal features for inclusion in decision trees. Models built on training data, applied to test data. Pruning and tuning parameters applied. Concordance of endometriosis staging models and AAGL surgical complexity level assessed using kappa and weighted kappa coefficients as well as sens, spec, PPV and NPV. Model 1 correlated with AAGL surgical complexity levels (A, B, C and D) and model 2 correlated with AAGL surgical complexity levels (A, B+C and D). Model 1 identified four US features (any bowel DE, endometrioma, POD partial obliteration & bladder DE); Model 2 identified five surgical features (any endometrioma, any bowel DE, POD partial obliteration, uterosacral DE and torus uterinus DE). Model 1 (4 stage) accuracy predicting AAGL surgical complexity level (95%CI's): Kappa 0.52 (0.43-0.60), Weighted Kappa 0.59 (0.50-0.68). Sens/spec/PPV/NPV for A (94.74/72.27/62.07/96.63), B (60.47/91.73/70.27/87.77), C (4.76/100/100/77.01), D (94.12/87.32/64.00/98.41). Model 2 (3 stage) accuracy predicting AAGL surgical complexity level: Kappa 0.45 (0.33-0.56), Weighted Kappa 0.53 (0.43-0.63). Sens/spec/PPV/NPV for A (84.62/79.84/63.77/92.52), B+C (64.04/68.97/67.86/65.22), D (42.86/94.33/65.22/86.93). Conclusion Decision tree models for ultrasound-based endometriosis severity staging built on univariate analysis has a high level of agreement with AAGL surgical complexity levels. The 4-stage system outperformed the 3-stage system.