Before the Burn: Predicting Endometrial Ablation Failure.

There were 681 patients coded as having an endometrial ablation. Thirty‐five were excluded (coded incorrectly for EA). Table  2 summaries all included variables. The average age at the time of EA was 45 years between 28 and 64 years. The average BMI was 30 kg/m 2 . Summary data for all variables. Row percentages shown for categorical and binary variables. NovaSure ( n  = 606, 94%) Other ( n  = 22, 3%) Yes ( n  = 95, 15%) No ( n  = 429, 66%) Previous ( n  = 118, 18%) Axial ( n  = 23, 4%) Yes ( n  = 306, 47%) No ( n  = 271, 42%) Missing ( N  = 69, 11%) Yes ( n  = 232, 36%) No ( n  = 370, 57%) Missing ( N  = 44, 7%) Yes ( n  = 341, 53%) No ( n  = 170, 26%) Yes ( n  = 154, 24%) No ( n  = 489, 76%) Most EA procedures were completed with NovaSure (93.8%). Most subjects had not had a tubal ligation ( n  = 429, 66%). Irregular bleeding and dysmenorrhoea were common (47% and 53%, respectively) but had the highest proportion of missing values (11% and 21%, respectively). Insertion of Mirena at the time of EA occurred in 24% of cases. A total of 136 patients had a failed EA (21%). Of these, 102 underwent hysterectomy, 4 were on the waiting list for a hysterectomy, 3 had a repeat endometrial ablation and 53 patients required ongoing medical management of AUB. Some used a combination of these. The median time to hysterectomy was 16 months. Time from EA to data extraction ranged from 2 to 8 years (mean 5 years). Patients were not followed up for a defined timeframe due to variation in clinical practice. The mean number of follow‐up appointments was 2.9 (SD 2.8). After imputing for missing data, a univariate logistic regression was conducted for each predictor variable against failure of EA. The output is shown in Table  3 . The variables age, Mirena, BMI and fibroids were all individually associated with failure ( p  ≤ 0.1). Failure was less likely with increasing age, and the odds of failure with a Mirena at the time of EA were lower. Increasing BMI and fibroids, increased the odds of failure. Univariate and multivariate analysis. Note: The bolded values demonstrate those variables included in subsequent analyses. Indicates variables with p  ≤ 0.1 which were then included in the multivariate logistic regression analysis. Variables with p  ≤ 0.1 on the above univariate logistic regression were then included in a multivariate logistic regression analysis, see Table  3 (model 1). At this stage, no variables were considered just on clinical grounds. Model 1 demonstrates that increasing BMI was associated with failure of EA. Increasing age and insertion of the Mirena at the time of EA made failure less likely. A sensitivity analysis was conducted. Given the association on univariate logistic analysis and the robust documentation in the literature that Mirena insertion at the time of EA reduces the likelihood of failure, the decision was made to keep it in the model on clinical grounds despite p  > 0.05. Each variable that was not included in the multivariate logistic regression analysis (model 1) was then added back in to assess if they were statistically significant predictors of failure of EA. The presence of fibroids was associated with failure ( p  = 0.04) and was added to the model (model 2, Table  3 ). Another sensitivity analysis was conducted. There was no evidence that the other variables were associated with failure or were confounders. Next, interactions were evaluated to determine if they were required in the prediction model [ 20 ]. Interactions between all variables in model 1 were assessed (age, Mirena, BMI and fibroids). No significant interactions were found (all p  > 0.1). The Brier score, assessing the overall model performance, was 0.17. Discrimination was assessed using the C‐index and was 0.62. This C‐index shows that the model is better than chance at predicting failure after an EA. The discrimination plot shows that the predicted probabilities for failure are slightly higher for those that did fail, although there is a significant overlap between the two outcome groups. Using the Hosmer‐Lemeshow goodness‐of‐fit test for calibration, there was no evidence of lack of fit (all p  > 0.05). The optimal cut point for high versus low risk of failure is a probability of 0.26. At this cut point the sensitivity and specificity are 64% and 57%, respectively, with a C‐index of 0.60. Internal validation was completed with 500 bootstrapped samples using the 0.632 and 0.632+ methods. The mean area under the ROC curve was 0.64 (95% interval 0.63–0.65) for both the 0.632 and 0.632+ methods. This is a minimal difference of 0.019 between the C‐index of the original model and the pessimistic estimates of the bootstrapped sample. A nomogram is a visual representation of multiple variables, each on a different scale, and their relationship with each other. To make the final predictive model (model 2 shown in Table  3 ) user friendly, a nomogram was generated using Stata (Figure  1 ). For an example of using the nomogram: a patient who was 44 years old (score~3) with a BMI of 40 (score~3.5), has fibroids (score~1.75) and did not get a Mirena at the time of the EA (score~1.75) has a total score of 10. This corresponds to an approximate probability of 35% chance of failure of the EA. Nomogram for the model for predicting failure of endometrial ablation. For each variable a score is given, these are added up and the total relates to a probability of failure of endometrial ablation.

There is longstanding evidence supporting different factors associated with failure after an EA. Despite this, the rates of failure remain relatively unchanged. This study has developed a predictive model for failure of EA using Australian data and incorporates the use of the Mirena in the model. Age and Mirena use reduced the odds of failure of EA, whereas increasing BMI and presence of fibroids increased them. These variables informed the predictive model. This model was generated with many patients and used the available literature to inform predictors. It has fair discriminatory value with a C‐index of 0.64 using internal validation. Consistent with previous studies, younger age and increasing BMI predicted EA failure [ 3 , 5 , 6 , 7 , 8 ]. The presence of fibroids impacting failure of EA remains debated [ 7 , 8 ]. Tubal ligation which has frequently been associated with post‐operative pain and failure of EA was not significant in this cohort [ 3 , 5 , 6 , 14 ]. The reduction in failure of EA with the use of Mirena is consistent with current research [ 21 , 22 ]. Limitations include missing data, with dysmenorrhoea having the most (21%). The capture of failure as the outcome variable was limited to what was available in the electronic record and those who responded to the phone questionnaire, possibly underestimating failure if patients sought care with their general practitioner or privately. An mean lag of up to 2–3 years between a failed EA and a subsequent hysterectomy has been published [ 7 , 8 ]. Participants in this analysis included a follow up 2–8 years, which may have missed some later failures. Postmenopausal patients were not specifically excluded from this analysis; however, documentation regarding menopausal status was limited in some cases. The primary indications for presentation were abnormal uterine bleeding and heavy menstrual bleeding. COVID‐19 is another significant factor that has likely impacted the overall failure. There was a significant reduction in elective surgeries in 2020 and 2021 because of the COVID‐19 pandemic [ 23 ]. Alonso et al. reported a reduction of up to 44% of elective gynaecological procedures during the pandemic. It is also possible that more patients would have sought care privately because of the delay in the public system. Manually building a model is a subjective process, however, we used a systematic approach to include statistically and clinically relevant predictors [ 18 , 19 ]. It is important to include all clinically relevant predictors despite non‐significance on univariate logistic regression analysis or collinearity [ 16 ]. Automated methods of variable selection may exclude confounders or clinically relevant predictors if their p ‐values are non‐significant [ 2 , 3 , 5 , 6 , 7 , 15 ]. Imputing for missing values is important to reduce bias and improve model development [ 16 , 18 , 19 ]. The alternative to imputing for missing values would be to exclude variables with missing values from the analysis. Because of the number of missing variables and their random distribution this may exclude many variables from the analysis and reduce the power. It also then introduces bias. The Brier score (0.17) suggests moderate model performance [ 20 , 24 ]. The clinical relevance of this, however, is unclear and is a consistent limitation with other traditional methods of assessing calibration and discrimination including the C‐index and the Hosmer‐Lemeshow goodness of fit test [ 16 , 20 , 25 ]. Risks and benefits may differ between patients and clinical situations and thus the weights of false positive and false negative values should differ [ 16 ]. For example, a patient wishing to avoid a hysterectomy at all costs would be more likely to accept the risk of failure of EA at a higher predicted probability than a patient whose priority is to avoid any further treatment. These situations are not accounted for by traditional performance measures. This study developed a predictive model within the Australian context to assist patients and clinicians in decision making surrounding the use of EA for the treatment of AUB. Population characteristics may, however, differ between health services, states, and countries. Information for all predictive variables are available in gynaecological outpatient consultations to easily inform treatment counselling. Ongoing research is needed to improve model performance and then validate the model externally prior to using the nomogram in a clinical context. The nomogram is a demonstration of a possible application of a predictive model. Given the persistently static failure rates of EA over several decades and its associated complications, further refinement of this predictive model may enhance the quality of information available to patients, supporting more informed decision‐making. It is essential that clinicians individualise counselling for the treatment of AUB based on their risk factors for failure of EA.

Abnormal uterine bleeding (AUB) affects up to 35% of women globally [ 1 ]. While EA offers a minimally invasive alternative to a hysterectomy for women with abnormal uterine bleeding (AUB), clinicians lack reliable predictive tools to identify which patients will experience treatment failure, leaving providers and patients to make treatment decisions with incomplete prognostic information [ 2 ]. Failure of EA is usually defined as a subsequent hysterectomy. Commonly reported risk factors for failure include; younger age at time of EA [ 3 , 4 , 5 , 6 , 7 ], tubal sterilisation, dysmenorrhoea [ 3 , 5 , 6 ], intramural fibroids [ 6 ], previous caesarean [ 8 ], type of EA [ 3 , 4 ] and adenomyosis [ 3 ]. Endometrial ablation has been used for nearly 40 years [ 9 ]. Despite changes in technology, failure rates have remained static. One study reported a hysterectomy rate of 21% after first‐generation EA techniques (involves direct visualisation and use bipolar endometrial resection, roller ball ablation or laser ablation for endometrial destruction) [ 7 ]. Two other papers (second‐generation techniques) reported failure rates of 19% and 19.3% respectively [ 6 , 8 ]. Second‐generation techniques, including bipolar radiofrequency, thermal balloon, cryoablation and microwave energy, blindly destroy the endometrium [ 8 ]. A Cochrane review comparing EA with minimally invasive hysterectomy found that patients undergoing EA were less satisfied, experienced more bleeding and were more likely to require ongoing surgery. Quality of life was lower in the EA group than following minimally invasive hysterectomy [ 10 ]. Several complications are associated with EA. One study described a 4.4% complication rate, including infection, uterine perforation, cervical damage, fluid overload and pain [ 3 , 11 ]. More significant complications include thermal injuries and organ damage [ 11 ]. Complications from pregnancy after EA include increased: perinatal mortality, prematurity, morbidly adherent placenta and a higher caesarean section rate [ 12 ]. Endometrial sampling post‐EA is also challenging due to adhesions [ 3 , 13 ]. Chronic post‐ablation pain has been increasingly reported in the literature with rates up to 29.7%. One study reported a 10.5% hysterectomy rate secondary to post‐ablation pain [ 2 , 14 ]. Because of the possible complications associated with EA, it is important to determine which patients are least likely to require further treatment after EA. A Norwegian study developed predictive models for failure and surgical re‐intervention after EA using age, dysmenorrhoea, duration of menses > 7 days, parity ≥ 5, menorrhagia, a raised pre‐treatment endometrial thickness and previous caesarean as predictors. They excluded post‐menopausal patients and did not assess the insertion of a levonorgestrel intrauterine device (Mirena) at the time of EA [ 15 ]. The objective of this study is to identify factors associated with failure of EA, and to use these to develop and internally validate a model predicting failure. A predictive model for failure of EA would be a useful clinical tool for decision making regarding treatment options for AUB.

The authors declare no conflicts of interest.

Retrospective cohort study. A Tertiary Health Service, Australia. In Australia, a tertiary health service is typically a large referral centre that manages specialised medical care with a comprehensive range of services. Women who had undergone an EA between the years of 2015 and 2021. Ethics approval through the local Human Research and Ethics Committee (LNR/2021/QGC/84953). A description of the variables collected are listed in Table  1 . The outcome variable, failure, is a composite measure of hysterectomy, repeat EA, or ongoing medical treatment after an EA. Project 1, description of variables included in analysis. Details of patients who had an EA procedure coded in the intraoperative record were extracted. Patients were excluded if their procedure was coded incorrectly (i.e., did not have an EA). P. Electronic records were instituted at the health service in 2015. The study end date was set as 2021 to allow adequate follow‐up for the detection of EA failure. Prior research indicates a median time to hysterectomy following EA failure of approximately 1.6 years [ 8 ]; therefore, a two‐year period beyond 2021 was included to capture relevant outcome data. Participant data for variables as listed above were manually extracted from electronic records during 2022 and 2023. To improve the accuracy of the data, patients that consented to be contacted were telephoned in 2023 and undertook a phone survey to collect data for the variables listed above. To reduce bias, most predictors could be reliably measured with good interobserver reliability. A total of 681 records were coded as EA procedures conducted between 2015 and 2021. An EA failure rate of approximately 20% is reported in the literature. For the database of 681 participants, this would be approximately 136 participants. For model development, using at least 10 outcome events per predictor [ 16 ], a maximum of 13 predictors could be included in the analysis. Stata/IC v16.1 statistical software was used for statistical analysis [ 17 ]. Descriptive statistics were produced for baseline patient characteristics, predictor variables and outcome variables. Variables included in the model building process were selected based on clinical reasoning and relevant literature. To avoid selection bias, missing data was addressed by multiple imputation using chained equations with 10 iterations [ 16 ]. The predictive model building process described by Hosmer et al. was used to build the predictive model [ 18 ]. Univariate logistic regression was completed for each variable. Those with p  ≤ 0.1 were then included as variables in the multivariate logistic regression analysis. Multivariate logistic regression was then used to generate the predictive model. Variables were retained if their p  < 0.05 or if they were considered important on a clinical basis. Confounding and clinically relevant interactions were evaluated. Sensitivity analyses were conducted to consider if variables excluded after the univariate analysis were predictors or confounders. After the final model was developed, the model performance was assessed, and it was internally validated [ 18 , 19 ]. The overall model performance was assessed using the Brier score and determinants of calibration and discrimination. Calibration was assessed using the goodness‐of‐fit measure, the Hosmer‐Lemeshow test. Discrimination was assessed using the area under the curve (AUC) which is statistically equivalent to the C‐index, the receiver operating characteristic (ROC) curve and the discrimination slope [ 20 ]. The optimal cut point model performance, sensitivity and specificity was determined. The 0.632 and 0.632+ bootstrap methods with 500 bootstrapped samples were used to internally validate the model [ 20 ].