Predicting menstrual bleeding patterns with levonorgestrel-releasing intrauterine systems

To develop a bleeding-pattern prediction model to inform counselling on amount and regularity of bleeding after levonorgestrel-releasing intrauterine system (LNG-IUS) placement.

Fixed-cluster and regression-tree models were developed using bleeding data pooled from two clinical trials of LNG-IUSs. Models were trained and cross-validated on LNG-IUS 12 data, then applied to LNG-IUS 20 and LNG-IUS 8 data. Three clusters were generated for the fixed-cluster model: predominantly amenorrhoea; predominantly spotting; and predominantly bleeding. A random-forest model predicted the future-bleeding cluster, then the probability of cycle regularity was calculated. In the regression-tree model, women were assigned by the model to less- or more-bleeding groups.

With LNG-IUS 12 (n = 1351) in the fixed-cluster model, 70.4% of women were correctly classified. The correct classification rates for LNG-IUS 20 (n = 216) and LNG-IUS 8 (n = 1300) were 72.2% and 69.0%. The probability distribution for cycle regularity showed regular and irregular bleeding were best separated with LNG-IUS 12 data, and less well with LNG-IUS 20 and LNG-IUS 8 data. In the regression-tree model there was high variability in the more- and less-bleeding group distributions with LNG-IUS 12 data.

A fixed-cluster model predicted bleeding patterns better than a regression-tree model in women using LNG-IUS, yielding understandable, informative output.