부산대학교 도서관

BackgroundPolycystic ovary syndrome (PCOS) is an endocrine disrupting disorder affecting about 10% of reproductive-aged women. PCOS diagnosis may be delayed several years and may require multiple physicians, resulting in lost time for risk-reducing interventions. Menstrual tracking apps are a potential tool to alert women of their risk while also prompting evaluation from a medical professional. ObjectiveThe primary objective of this study was to develop and pilot test the irregular cycle feature, a predictive model that generated a PCOS risk score, in the menstrual tracking app, Clue. The secondary objectives were to run the model using virtual test subjects, create a quantitative risk score, compare the feature’s risk score with that of a physician, and determine the sensitivity and specificity of the model before empirical testing on human subjects. MethodsA literature review was conducted to generate a list of signs and symptoms of PCOS, termed variables. Variables were then assigned a probability and built into a Bayesian network. Questions were created based on these variables. A total of 9 virtual test subjects were identified using self-reported menstrual cycles and answers to the feature’s questions. Upon completion of the questionnaire, a Result Screen and Doctor’s Report summarizing the probability of having PCOS was displayed. This provided information about PCOS and data to facilitate diagnosis by a medical professional. To assess the accuracy of the feature, the same set of 9 virtual test subjects was assigned probabilities by the feature and the physician, who served as the gold standard. The feature recommended individuals with a score greater than or equal to 25% to follow-up with a physician. Differences between the feature and physician scores were evaluated using a t test and a Pearson correlation coefficient in 8 of the 9 virtual test subjects. A second iteration was conducted to assess the feature’s probability capabilities. ResultsThe irregular cycle feature’s first iteration produced 1 false-positive compared with the physician score and had an absolute mean difference of 15.5% (SD 15.1%) among the virtual test subjects. The second iteration had 2 false positives compared with the physician score and had an absolute mean difference of 18.8% (SD 13.6%). The feature overpredicted the virtual test subjects’ risk of PCOS compared with the physician. However, a significant positive correlation existed between the feature and physician score (Pearson correlation coefficient=0.82; P=.01). The second iteration performed worse, with a Pearson correlation coefficient of 0.73 (P=.03). ConclusionsThe first iteration of the feature outperformed the second and better predicted the probability of PCOS. Although further research is needed with a more robust sample size, this pilot study indicates the potential value for developing a screening tool to prompt high-risk subjects to seek evaluation by a medical professional.