Abstract
Polycystic ovary syndrome (PCOS) is a heterogenous condition that is estimated to effect up to 21% of reproductive aged people with ovaries. In previous work, a dataset of PCOS features was derived from approximately 100,000 PCOS subreddit users via machine learning. In this study, an exploration of treatment response within the PCOS subreddit was undertaken with the derived dataset. The treatment or symptom features in the dataset had sentiment labels indicating when a treatment was perceived to improve or worsen a condition or symptom. When different features were mentioned within two sentences of each other without conflicting sentiment, it could be assumed that they were related. This assumption allowed for a broad analysis of the perceived effect of popular treatments on the most frequently mentioned symptoms. In general, lifestyle changes and supplements were the most positively regarded, while contraceptives were frequently associated with considerable negative sentiment. For PCOS weight loss, unspecified dieting (RR 5.19, 95% CI 3.28-8.19, n = 99) and intermittent fasting (RR 33.50, 95% CI 8.54-131.34, n = 69) were the most successful interventions. Inositol was associated with a large range of favourable outcomes and was one of the few treatments associated with improved mental health [depression (RR 4.25, 95% CI 1.72-10.51, n = 21), anxiety (RR 5.83, 95% CI 2.76-12.35, n = 41) and mood issues (RR 25.00, 95% CI 3.65-171.10, n = 26)]. Combined oral contraceptive pills as a whole were strongly associated with adverse effects such as worsening depression (RR 0.06, 95% CI 0.02-0.25, n = 33), anxiety (RR 0.10, 95% CI 0.03-0.36, n = 23), fatigue (RR 0, n = 45) and low libido (RR 0.03, 95% CI 0.01-0.24, n = 30). However, combined contraceptives with anti-androgenic progestins were associated with more favourable experiences. This study demonstrates the utility of machine learning to derive measurable patient experience data from an internet forum. While patient experience data derived using machine learning is not a substitute for traditional clinical trials, it is useful for mass validation and hypothesis generation. This paper may serve as the first exploration into this category of clinical internet forum research.