Prospective evaluation of IOTA logistic regression models LR1 and LR2 in comparison with subjective pattern recognition for diagnosis of ovarian cancer in an outpatient setting

ABSTRACT Objective To determine whether International Ovarian Tumor Analysis (IOTA) logistic regression models LR1 and LR2 developed for the preoperative diagnosis of ovarian cancer could also be used to differentiate between benign and malignant adnexal tumors in the population of women attending gynecology outpatient clinics. Methods This was a single‐center prospective observational study of consecutive women attending our gynecological diagnostic outpatient unit, recruited between May 2009 and January 2012. All the women were first examined by a Level‐II ultrasound operator. In those diagnosed with adnexal tumors, the IOTA‐LR1/2 protocol was used to evaluate the masses. The LR1 and LR2 models were then used to assess the risk of malignancy. Subsequently, the women were also examined by a Level‐III examiner, who used pattern recognition to differentiate between benign and malignant tumors. Women with an ultrasound diagnosis of malignancy were offered surgery, while asymptomatic women with presumed benign lesions were offered conservative management with a minimum follow‐up of 12 months. The initial diagnosis was compared with two reference standards: histological findings and/or a comparative assessment of tumor morphology on follow‐up ultrasound scans. All women for whom the tumor classification on follow‐up changed from benign to malignant were offered surgery. Results In the final analysis, 489 women who had either or both of the reference standards were included. Their mean age was 50 years (range, 16–91 years) and 45% were postmenopausal. Of the included women, 342/489 (69.9%) had surgery and 147/489 (30.1%) were managed conservatively. The malignancy rate was 137/489 (28.0%). Overall, sensitivities of LR1 and LR2 for the diagnosis of malignancy were 97.1% (95% CI, 92.7–99.2%) and 94.9% (95% CI, 89.8–97.9%) and specificities were 77.3% (95% CI, 72.5–81.5%) and 76.7% (95% CI, 71.9–81.0%), respectively ( P > 0.05). In comparison with pattern recognition (sensitivity 94.2% (95% CI, 88.8–97.4%), specificity 96.3% (95% CI, 93.8–98.0%)), the specificities of the IOTA models were significantly lower ( P < 0.0001). A significantly higher number of women would have been offered surgery for suspected cancer if the women had been assessed using the IOTA models instead of pattern recognition (213/489 (43.6%) vs 142/489 (29.0%); P < 0.001). Conclusions The IOTA models maintained their high sensitivity when used in an outpatient setting. Specificity was relatively low, which indicates that a significant proportion of the women would have been offered unnecessary surgery for suspected ovarian cancer. These findings show that the IOTA models could be used as a first‐stage test to diagnose ovarian cancer in an outpatient setting, but a different second‐stage test is required to minimize the number of false‐positive findings. Copyright © 2017 ISUOG. Published by John Wiley & Sons Ltd.