Predictive tools are lacking for pain-related outcomes after endometriosis surgery. The objective of this study was to develop and validate a machine learning-based clinical model to predict poor pain-related quality of life after endometriosis surgery. Registry data from a prospective longitudinal cohort at a tertiary referral center (2013-2020) was used for model development and validation. Participants underwent an index endometriosis surgery, and completed the pain subscale of the Endometriosis Health Profile-30 (EHP-30) at baseline and 1-2-year follow-up. The outcome was poor pain-related quality of life defined as EHP-30 pain subscale above the 75th percentile for North America, at 1 to 2 years postsurgery. Thirty-two preoperative factors were evaluated, with final models retaining the top 10 most important predictors. Elastic net logistic regression, random forest (RF) and multilayer perceptron neural network models were developed. Internal validation was performed using 500 bootstrap samples, and a held-out test cohort. The study included 650 participants: 488 in the training cohort and 162 in a held-out test cohort. The RF model exhibited the most consistent discrimination, measured by the area under the receiver operating characteristic curve, between the training cohort (0.768; 95% CI: 0.690-0.837) and test cohort (0.766; 95% CI: 0.676-0.863, Δ = -0.002). The RF model also demonstrated the best integrated calibration index (0.029) and highest net benefit. Final preoperative predictors for the RF model included baseline EHP-30 score, surgery type (conservative fertility-sparing vs hysterectomy), anxiety scores, depression scores, pain catastrophizing scale scores, abdominal wall pain, pelvic floor myalgia, smoking status, back pain, and race/ethnicity. We present the RF model as the Endometriosis Pain Index to aid preoperative counselling for endometriosis surgery.
Tucker et al. (Wed,) studied this question.