OC147: Mathematical decision trees vs. clinician‐based algorithms in the diagnosis of endometrial disease
Thierry Van den Bosch, Anneleen Daemen, Olivier Gevaert, Dirk Timmerman · Ultrasound in Obstetrics and Gynecology · 2007
Most diagnostic algorithms have been built by clinicians using accuracy figures for the different tests and also based on their ‘clinical intuition’, but few decision trees have been challenged by mathematicians. The data from 402 consecutive patients presenting at the department's ‘one-stop bleeding clinic’ were used in this study. The patients subsequently underwent a transvaginal gray-scale ultrasound examination with color Doppler imaging, hydrosonography, office hysteroscopy and endometrial sampling. The data were entered into an algorithm designed by clinicians (Algorithm 1) and into a mathematical decision tree (Algorithm 2). The data from the first 281 cases (70%) were used to train the mathematical algorithm; the remaining cases were used to validate this algorithm. Both algorithms were evaluated in terms of diagnostic accuracy, cost, patient discomfort and clinical feasibility. The clinical algorithm proposed ultrasound as the primary investigation whereas the mathematical model proposed office hysteroscopy as the initial diagnostic test. The sensitivity and specificity for diagnosis of endometrial disease was 75.5% and 82.5% respectively for Algorithm 1 vs. 86.7% and 93.3% respectively for Algorithm 2. The mean number of tests needed to reach the diagnosis was 2.1 for Algorithm 1 and 1.8 for Algorithm 2. The average cost per patient was € 91.5 for Algorithm 1 vs. € 70.6 for Algorithm 2. The mean pain score (using the visual analogue scale) was 5.9 for the clinical algorithm vs. 6.5 for the mathematical decision tree. The mathematical decision tree scored better in terms of diagnostic accuracy and cost, but caused more patient discomfort. This decision tree does not address clinical concerns such as the disturbance of the ultrasound image caused by previous intrauterine manipulation and might not be feasible as a one-stop diagnostic pathway.