Machine Learning-Based Prediction of Intervention-Free Survival After Surgery for Adenomyosis: Development and Internal Validation of a Prognostic Model

Observational OA: gold CC-BY-NC-4.0
⚙ AI-generated deep summary by qwen3.7-flash, 2026-11-18 · read from full text ⓘ

This retrospective cohort study developed and internally validated a machine learning model to predict intervention-free survival in patients who underwent surgical treatment for adenomyosis. The researchers analyzed clinical, hematologic, and imaging data from women with pathologically confirmed adenomyosis to identify factors influencing the duration of time without further medical or surgical intervention. The findings highlight the utility of advanced analytical frameworks in capturing complex, non-linear relationships that traditional linear models fail to address, thereby improving prognostic accuracy for postoperative outcomes. This paper is centrally about adenomyosis — specifically focusing on surgical management and the development of predictive tools for long-term patient prognosis following uterine-sparing procedures or hysterectomy.

Read from the paper's body, not the abstract. Not a substitute for reading the paper. No clinical advice. How this works

Abstract

OBJECTIVE: To develop and validate a machine learning-based model for predicting long-term intervention-free survival (IFS) in patients with adenomyosis (AM) after surgery. METHODS: This retrospective cohort study enrolled patients with AM who underwent surgical treatment at our institution between March 2016 and December 2020. Data were retrospectively collected from medical records, while follow-up outcomes were prospectively gathered. Seven machine learning algorithms, including decision trees (DT), random forest (RF), eXtreme Gradient Boosting (XGBoost), support vector machine (SVM), multilayer perceptron (MLP), light gradient boosting machine (LightGBM), and k-nearest neighbors (KNN), were evaluated and compared using regression prediction performance (assessed by root mean squared error, RMSE) and time-dependent ROC analysis. The algorithm demonstrating the highest accuracy was selected as the final model, and the variables it identified were subsequently used to construct the prognostic model. A nomogram was developed using multivariate Cox regression to predict 3-year, 4-year, and 5-year IFS. Its performance was validated through receiver operating characteristic (ROC) curve, calibration curves, and decision curve analysis (DCA). RESULTS: A total of 541 patients were included, with a median follow-up of 77 months. The SVM model demonstrated the optimal performance, yielding the lowest RMSE among the candidate models and achieving the highest time-dependent AUC (0.671). SHAP analysis identified seven stable predictors: age, CA125 level, number of pregnancies, number of abortions, VAS score, hysterectomy, and GnRH-a course. Subsequently, these significant variables were integrated to formulate a nomogram for predicting 3-year, 4-year, and 5-year IFS. The resulting nomogram demonstrated good discrimination, with AUCs of 0.72-0.80 for 3- to 5-year IFS in both cohorts. CONCLUSION: This study developed and internally validated a machine learning-guided nomogram for predicting IFS in AM patients after surgery. The model may support individualized risk stratification and postoperative management. However, its clinical utility requires prospective validation before implementation.
Full text 58,517 characters · extracted from pmc-nxml · 6 sections · click to expand

Intro

Adenomyosis (AM) is an estrogen-dependent gynecological disorder characterized by the diffuse invasion and proliferation of endometrial tissue into the myometrium. 1 Although the pathogenesis of AM remains elusive, current research has primarily centered on several key theories, including the tissue injury and repair mechanism, localized estrogen excess, and chronic inflammatory stimulation. 2 AM can lead to a globally sensitized and enlarged uterus within a chronic inflammatory milieu, predisposing patients to persistent pain and dyspareunia. 3 This pathological invasion results in thickening of the uterine wall and is associated with a spectrum of clinical manifestations, including dysmenorrhea, chronic pelvic pain, lumbago, dyspareunia, perianal pain, menorrhagia, and infertility. 4 The prevalence of AM among women of reproductive age is estimated to range from 20% to 40%, exerting a substantial impact on their physical and mental well-being. 5 Consequently, endometriosis imposes a substantial economic burden on both patients and society, with the annual national economic burden estimated to range from $15.58 billion to $78.05 billion. 6 The Society of Obstetricians and Gynecologists of Canada has released clinical practice guidelines regarding the diagnosis and management of AM, underscoring the importance of symptom-oriented medical therapy and individualized treatment selection. 7 In recent years, the non-invasive diagnosis of AM using transvaginal ultrasound and magnetic resonance imaging (MRI) has been widely integrated into clinical practice. Given its cost-effectiveness, wide availability, and high diagnostic accuracy, pelvic ultrasound is recommended as the first-line imaging modality for AM. 8 Meanwhile, the diagnostic rate of AM has substantially increased with the use of MRI. Imaging indicators such as junctional zone (JZ) thickness, lesion distribution, and vascularity have been suggested to correlate with disease severity and therapeutic responsiveness. 9 AM is recognized as a chronic condition necessitating individualized, long-term management. Therapeutic strategies aim to alleviate symptoms, minimize recurrence, avoid repeat surgery, and preserve fertility. This is primarily achieved through a comprehensive approach centered on pharmacological and surgical interventions. 10 Pharmacological treatments, including gonadotropin-releasing hormone agonists and progestin-based therapies, can provide temporary symptom relief. However, symptoms frequently recur upon drug withdrawal, and the long-term use is often limited by adverse effects. 11 In cases where pharmacological therapy proves insufficient, minimally invasive modalities such as high-intensity focused ultrasound (HIFU) or microwave endometrial ablation (MEA) may be considered for selected patients, particularly when focal lesions are clearly delineated on MRI. 12 Traditionally, hysterectomy was regarded as the gold standard for treating AM. In addition, the increasing desire for fertility preservation and the trend of delayed childbearing have necessitated a clinical migration toward fertility-sparing surgery. 13 Management is primarily symptom-driven, and uterine-sparing treatments are increasingly used for women who wish to avoid hysterectomy. Nevertheless, sufficient long-term follow-up data are still lacking to fully elucidate the prognosis and influencing factors associated with these surgical interventions. Notably, AM is characterized by a heterogeneous spectrum of anatomical and clinical manifestations, ranging from a normal-sized uterus to significant enlargement, and from severe dysmenorrhea and hypermenorrhea to asymptomatic presentations. 4 Besides, pathology-proven endometriosis (EMS) was identified in the majority of patients with suspected AM, even after excluding those with preoperative imaging findings suggestive of deep infiltrating endometriosis (DIE). 14 A strong link between AM and EMS is well-documented, as concurrent disease is observed in 13%–89% of patients diagnosed with EMS. 15 The heterogeneity of AM, coupled with its frequent comorbidity with EMS, poses significant challenges to the analysis of follow-up data. Diverse anatomical presentations and overlapping symptoms generate substantial variability in patient responses, which limits the efficacy of traditional linear prognostic models. Traditional logistic regression models function fundamentally as linear classifiers, utilizing the Logit function to constrain outputs to the interval. This inherent linearity renders them ill-equipped to capture complex non-linear relationships embedded within the data. 16 Therefore, advanced analytical frameworks are urgently required to decipher this multifaceted clinical landscape and deliver accurate long-term prognostic predictions for AM. AM is characterized by substantial clinical and pathological heterogeneity, which contributes to variable recurrence patterns and complicates postoperative risk assessment. The machine learning models leverage iterative training to uncover latent patterns within data and may capture nonlinear relationships and complex interactions among predictors. However, machine learning is not inherently superior to traditional statistical models. 17 Existing research has pioneered the application of machine learning in the precision diagnosis and management of AM. Notably, a prospective study involving 1654 patients developed a deep learning model utilizing transvaginal ultrasound images to facilitate automated, non-invasive diagnosis, which demonstrated an accuracy of 92.33%. 18 Another retrospective study involving 170 AM patients constructed a radiomics-based machine learning model using multiparametric MRI, which successfully predicted the therapeutic efficacy of HIFU. 19 Indeed, the high prevalence, insidious symptoms, and significant risk of recurrence associated with AM present substantial clinical challenges. By integrating complex imaging signatures with clinical heterogeneity, machine learning techniques hold great promise for unraveling the multidimensional determinants of AM prognosis and guiding personalized therapeutic strategies. Symptom or imaging recurrence alone may not fully reflect patient-perceived benefit after surgery for adenomyosis. Prior systematic reviews have noted considerable heterogeneity in outcome reporting and a scarcity of patient-centered endpoints, particularly quality of life and satisfaction. 20 A composite endpoint combining recurrent or persistent symptoms with the need for further medical therapy or reintervention may therefore better indicate durable clinical benefit. This study involved a cohort of postoperative AM patients with up to 5 years of intervention-free survival (IFS) data. The primary research question of this study was whether machine learning models could accurately predict postoperative IFS. By analyzing these longitudinal records, we sought to develop a predictive prognostic model and systematically explore the underlying factors influencing disease recurrence and survival.

Result

The workflow about patient selection and prognostic analysis was summarized in Figure 1 . We retrospectively collected clinicopathologic data of 647 patients diagnosed with AM receiving surgical treatment between March 2016 and December 2020. The 37 patients with severe organ damage, 23 patients with malignant lesions, and 46 patients who lost to follow-up were excluded. Thus, a total of 541 AM patients were included in this study, comprising 159 with comorbid ovarian EMS. Among the participants, 203 underwent hysterectomy, while 338 received fertility-sparing surgery. Patients who underwent hysterectomy were significantly older, had a higher number of pregnancies and a higher number of abortions than those who underwent fertility-sparing surgery ( Table S1 ). The 198 IFS events included 93 cases (47.0%) of severe clinical manifestations, 84 cases (42.4%) of disease progression via imaging, and 21 cases (10.6%) with significant serum CA125 elevation. Notably, the incidence of IFS events was substantially lower in the hysterectomy group (n=27) compared to the fertility-sparing surgery group (n=171, P < 0.001). Their median follow-up time was 77 months. Although the minimum follow-up of 12 months was insufficient to ensure that all patients could reach the 3-, 4-, or 5-year endpoint, the majority of patients had long-term follow-up. Other baseline characteristics are summarized in Table 1 . Figure 1 Flow diagram of patient selection, risk factors screening and nomogram development. Flowchart sequence: Patients diagnosed with adenomyosis receiving surgical treatment between March 2016 and December 2020 were retrospectively enrolled, N equals 647, Severe organ damage, N equals 37, Malignant lesions, N equals 23, Loss to follow up, N equals 46, Patients with complete medical data, N equals 541, Training set, 70 percent, N equals 378, Screening of candidate risk factors for predicting intervention free survival, Machine learning algorithms: DT, KNN, LightGBM, LSVM, MLP, RF, XGBoost, Construction of intervention free survival nomogram, Test set, 30 percent, N equals 163, Evaluation of model effectiveness by ROC, calibration curve and DCA. A flowchart of patient selection, training and test split and model evaluation steps. Table 1 Comparison of Baseline Characteristics Between Patients with and without Intervention-Free Survival (IFS) Event Variables Total (n = 541) IFS event P -value No (n = 343) Yes (n = 198) Age (years) 42 (37, 46) 43 (39, 47) 38 (34, 44) < 0.001 Height (cm) 160 (157, 163) 160 (157, 163) 160 (158, 163) 0.985 Weight (Kg) 57 (52, 63) 57 (53, 63) 56 (51, 63) 0.031 Duration (years) 3.0 (0.5, 6.0) 2.0 (0.5, 5.0) 3.0 (1.0, 7.0) 0.008 Age at menarche (years) 13 (12, 14) 13 (12, 14) 13 (12, 14) 0.675 Regular menstruation 0.509  No 105 (19) 70 (20) 35 (18)  Yes 436 (81) 273 (80) 163 (82) Menorrhagia 0.702  No 266 (49) 166 (48) 100 (51)  Yes 275 (51) 177 (52) 98 (49) Marital history 0.058  No 27 (5) 12 (3) 15 (8)  Yes 514 (95) 331 (97) 183 (92) Pregnancy history 0.002  No 53 (10) 23 (7) 30 (15)  Yes 488 (90) 320 (93) 168 (85) Number of pregnancy < 0.001  0 55 (10) 26 (8) 29 (15)  1 72 (13) 36 (10) 36 (18)  2 105 (19) 62 (18) 43 (22)  3 121 (22) 89 (26) 32 (16)  4 95 (18) 71 (21) 24 (12)  ≥5 93 (17) 59 (17) 34 (17) Number of Delivery 0.004  0 79 (15) 38 (11) 41 (21)  1 323 (60) 207 (60) 116 (59)  ≥2 139 (26) 98 (29) 41 (21) Number of abortions 0.005  0 158 (29) 83 (24) 75 (38)  1 122 (23) 81 (24) 41 (21)  2 138 (26) 99 (29) 39 (20)  ≥3 123 (23) 80 (23) 43 (22) Cesarean section 0.208  No 377 (70) 246 (72) 131 (66)  Yes 164 (30) 97 (28) 67 (34) Dysmenorrhea 0.159  No 103 (19) 72 (21) 31 (16)  Yes 438 (81) 271 (79) 167 (84) VAS score 0.046  0 343 (63) 213 (62) 130 (66)  1–2 127 (23) 91 (27) 36 (18)  ≥3 71 (13) 39 (11) 32 (16) History of painkillers 0.083  No 435 (80) 284 (83) 151 (76)  Yes 106 (20) 59 (17) 47 (24) Hormone therapy 1.000  No 502 (93) 318 (93) 184 (93)  Yes 39 (7) 25 (7) 14 (7) Intrauterine device 0.702  No 420 (78) 264 (77) 156 (79)  Yes 121 (22) 79 (23) 42 (21) Preoperative GnRH-a 0.457  No 517 (96) 330 (96) 187 (94)  Yes 24 (4) 13 (4) 11 (6) Surgical history of endometriosis 0.075  No 490 (91) 317 (92) 173 (87)  Yes 51 (9) 26 (8) 25 (13) Leukocyte count (10 9 /L) 5.42 (4.50, 6.66) 5.42 (4.50, 6.69) 5.42 (4.50, 6.54) 0.718 Neutrophil count (10 9 /L) 3.11 (2.38, 4.05) 3.13 (2.39, 4.11) 3.08 (2.37, 3.97) 0.775 Lymphocyte count (10 9 /L) 1.66 (1.34, 2.02) 1.66 (1.35, 1.99) 1.67 (1.32, 2.06) 0.448 Monocyte count (10 9 /L) 0.39 (0.32, 0.49) 0.39 (0.32, 0.50) 0.39 (0.32, 0.48) 0.865 Hemoglobin (g/L) 113 (94, 126) 113 (94, 126) 113 (96, 125) 0.948 Platelet count (10 9 /L) 256 (214, 312) 255 (212, 312) 256 (215, 314) 0.654 Fibrinogen (g/L) 2.81 (2.48, 3.18) 2.82 (2.45, 3.20) 2.81 (2.52, 3.13) 0.704 Albumin (g/L) 42.2 (39.6, 44.7) 42.1 (39.4, 44.7) 42.3 (40.0, 44.7) 0.413 Globulin (g/L) 28.7 (26.2, 31.2) 28.8 (26.3, 31.3) 28.7 (26.13, 31.1) 0.755 Glutamyltransferase (U/L) 14 (11, 19) 14 (11, 19) 14 (11, 18) 0.719 Creatinine (μmol/L) 54 (49, 60) 54 (49, 60) 54 (49, 60) 0.806 CA125 level (U/mL) 0.010  ≤35 97 (18) 73 (21) 24 (12)  >35 444 (82) 270 (79) 174 (88) CA199 level (U/mL) 0.205  ≤37 458 (85) 296 (86) 162 (82)  >37 83 (15) 47 (14) 36 (18) Diameter of uterus (cm) 7.2 (6.3, 8.1) 7.4 (6.3, 8.4) 7.0 (6.2, 7.9) 0.022 Location of lesions 0.153 Anterior 103 (19) 64 (19) 39 (20) Fundus 53 (10) 40 (12) 13 (7) Posterior 246 (45) 147 (43) 99 (50) Two or more lesions 139 (26) 92 (27) 47 (24) Uterine adenomyoma 0.457  No 24 (4) 13 (4) 11 (6)  Yes 517 (96) 330 (96) 187 (94) Uterine fibroids 0.479  No 385 (71) 240 (70) 145 (73)  Yes 156 (29) 103 (30) 53 (27) Endometrial polyps 0.204  No 487 (90) 304 (89) 183 (92)  Yes 54 (10) 39 (11) 15 (8) Ovarian endometriosis 0.009  No 382 (71) 256 (75) 126 (64)  Yes 159 (29) 87 (25) 72 (36) Pelvic peritoneal lesion 0.131  No 468 (87) 303 (88) 165 (83)  Yes 73 (13) 40 (12) 33 (17) Painful nodule 0.007  No 499 (92) 325 (95) 174 (88)  Yes 42 (8) 18 (5) 24 (12) Pelvic adhesions 0.29  No 247 (46) 163 (48) 84 (42)  Yes 294 (54) 180 (52) 114 (58) Hysterectomy < 0.001  No 338 (62) 167 (49) 171 (86)  Yes 203 (38) 176 (51) 27 (14) GnRH-a course < 0.001  0 278 (51) 214 (62) 64 (32)  1–3 60 (11) 29 (8) 31 (16)  4–6 203 (38) 100 (29) 103 (52) Follow-up time (months) 77 (36, 93) 87 (76, 101) 24 (12, 48) < 0.001 Note : Data are presented as median (Q1, Q3) or number (percentage). Bold text indicates statistically significant differences (P < 0.05). Flow diagram of patient selection, risk factors screening and nomogram development. Comparison of Baseline Characteristics Between Patients with and without Intervention-Free Survival (IFS) Event Note : Data are presented as median (Q1, Q3) or number (percentage). Bold text indicates statistically significant differences (P < 0.05). For internal validation, the patients with complete medical data were randomly allocated into a training set (N=378, 70%) and an internal split-sample test set (N=163, 30%). The detailed demographic data of the total cohort are presented in Table 2 , which showed no statistically significant differences between the two datasets. According to the iterative analysis of baseline characteristics, we employed correlation matrix analysis to optimize predictive performance ( Figure S1A ). While most variables exhibited relative independence, significant positive correlations existed within the pregnancy-related cluster, such as number of pregnancy, number of delivery, and number of abortions. In order to avoid overfitting, LASSO analysis was employed only in the training set to screen parameters, and the coefficient of the 44 variables were shown in Figure S1B . The 20-fold cross-validation method determined that the optimal model for predicting IFS should incorporate 11 variables: age, duration, CA125 level, number of pregnancy, number of abortions, VAS score, diameter of uterus, ovarian EMS, painful nodule, hysterectomy, GnRH-a course ( Table 1 and Figure S1C ). Table 2 Demographics of Patients with Adenomyosis Variables Total (n = 541) Train (n = 378) Test (n = 163) p Age (years) 42 (37, 46) 42 (36, 46) 43 (38, 47) 0.138 Height (cm) 160 (157, 163) 160 (157, 163) 160 (158, 163) 0.850 Weight (Kg) 57 (52, 63) 57 (52, 63) 57 (53, 63) 0.907 Duration (years) 3 (0.5, 6) 3 (0.5, 6) 2 (0.5, 5) 0.719 Age at menarche 13 (12, 14) 13 (13, 14) 13 (12, 14) 0.472 Regular menstruation 0.249  No 105 (19) 68 (18) 37 (23)  Yes 436 (81) 310 (82) 126 (77) Menorrhagia 0.234  No 266 (49) 179 (47) 87 (53)  Yes 275 (51) 199 (53) 76 (47) Marital history 1.000  No 27 (5) 19 (5) 8 (5)  Yes 514 (95) 359 (95) 155 (95) Pregnancy history 0.883  No 53 (10) 38 (10) 15 (9)  Yes 488 (90) 340 (90) 148 (91) Number of pregnancy 0.763  0 55 (10) 39 (10) 16 (10)  1 72 (13) 51 (13) 21 (13)  2 105 (19) 68 (18) 37 (23)  3 121 (22) 90 (24) 31 (19)  4 95 (18) 66 (17) 29 (18)  ≥5 93 (17) 64 (17) 29 (18) Number of Delivery 0.541  0 79 (15) 58 (15) 21 (13)  1 323 (60) 220 (58) 103 (63)  ≥2 139 (26) 100 (26) 39 (24) Number of abortions 0.287  0 158 (29) 115 (30) 43 (26)  1 122 (23) 81 (21) 41 (25)  2 138 (26) 102 (27) 36 (22)  ≥ 3 123 (23) 80 (21) 43 (26) Cesarean section 0.553  No 377 (70) 260 (69) 117 (72)  Yes 164 (30) 118 (31) 46 (28) Dysmenorrhea 0.911  No 103 (19) 71 (19) 32 (20)  Yes 438 (81) 307 (81) 131 (80) VAS score 0.737  0 343 (63) 238 (63) 105 (64)  1–2 127 (23) 92 (24) 35 (21)  ≥3 71 (13) 48 (13) 23 (14) History of painkillers 0.918  No 435 (80) 303 (80) 132 (81)  Yes 106 (20) 75 (20) 31 (19) Hormone therapy 0.928  No 502 (93) 350 (93) 152 (93)  Yes 39 (7) 28 (7) 11 (7) Intrauterine device 0.070  No 420 (78) 302 (80) 118 (72)  Yes 121 (22) 76 (20) 45 (28) Preoperative GnRH-a 1.000  No 517 (96) 361 (96) 156 (96)  Yes 24 (4) 17 (4) 7 (4) Surgical history of endometriosis 0.716 No 490 (91) 344 (91) 146 (90) Yes 51 (9) 34 (9) 17 (10) Leukocyte count (10 9 /L) 5.42 (4.50, 6.66) 5.38 (4.48, 6.68) 5.49 (4.53, 6.56) 0.989 Neutrophil count (10 9 /L) 3.11 (2.38, 4.05) 3.10 (2.4, 4.09) 3.14 (2.32, 3.93) 0.644 Lymphocyte count (10 9 /L) 1.66 (1.34, 2.02) 1.65 (1.34, 1.99) 1.69 (1.34, 2.03) 0.733 Monocyte count (10 9 /L) 0.39 (0.32, 0.49) 0.4 (0.32, 0.49) 0.38 (0.31, 0.50) 0.693 Hemoglobin (g/L) 113 (94, 126) 113 (96, 125) 113 (93, 127) 0.988 Platelet count (10 9 /L) 256 (214, 312) 259 (216, 312) 248 (207, 323) 0.229 Fibrinogen (g/L) 2.81 (2.48, 3.18) 2.77 (2.48, 3.13) 2.88 (2.49, 3.27) 0.153 Albumin (g/L) 42.2 ± 3.7 42.3 ± 3.7 41.8 ± 3.7 0.108 Globulin (g/L) 28.7 (26.2, 31.2) 28.8 (26. 3, 31.2) 28.6 (26, 31.3) 0.752 Glutamyltransferase (U/L) 14 (11, 19) 14 (11, 19) 14 (11, 19) 0.858 Creatinine (μmol/L) 54 (49, 60) 55 (49, 60) 54 (49, 59) 0.499 CA125 level (U/mL) 0.198  ≤35 97 (18) 62 (16) 35 (21)  >35 444 (82) 316 (84) 128 (79) CA199 level (U/mL) 0.364  ≤37 458 (85) 324 (86) 134 (82)  >37 83 (15) 54 (14) 29 (18) Diameter of uterus (cm) 7.2 (6.3, 8.1) 7.2 (6.3, 8.1) 7.1 (6.1, 8.2) 0.762 Location of lesions 0.492  Anterior 103 (19) 74 (20) 29 (18)  Fundus 53 (10) 35 (9) 18 (11)  Posterior 246 (45) 166 (44) 80 (49)  Two or more lesions 139 (26) 103 (27) 36 (22) Uterine adenomyoma 1.000  No 24 (4) 17 (4) 7 (4)  Yes 517 (96) 361 (96) 156 (96) Uterine fibroids 0.757  No 385 (71) 271 (72) 114 (70)  Yes 156 (29) 107 (28) 49 (30) Endometrial polyps 1.000  No 487 (90) 340 (90) 147 (90)  Yes 54 (10) 38 (10) 16 (10) Ovarian endometriosis 0.118  No 382 (71) 275 (73) 107 (66)  Yes 159 (29) 103 (27) 56 (34) Pelvic peritoneal lesion 0.892  No 468 (87) 326 (86) 142 (87)  Yes 73 (13) 52 (14) 21 (13) Painful nodule 1.000  No 499 (92) 349 (92) 150 (92)  Yes 42 (8) 29 (8) 13 (8) Pelvic adhesions 0.125  No 247 (46) 185 (49) 62 (38)  Yes 294 (54) 193 (51) 101 (62) Hysterectomy 0.518  No 338 (62) 240 (63) 98 (60)  Yes 203 (38) 138 (37) 65 (40) GnRH-a course 0.907  0 278 (51) 192 (51) 86 (53)  1–3 60 (11) 42 (11) 18 (11)  4–6 203 (38) 144 (38) 59 (36) Demographics of Patients with Adenomyosis Seven machine learning algorithms, including DT, RF, XGBoost, SVM, MLP, LightGBM, and KNN were performed to construct model. As illustrated in Figure 2A , the regression plots revealed that both SVM and MLP models demonstrated the tightest alignment between predicted and actual values along the diagonal. Regarding model stability, the 5-fold cross-validation results ( Figure 2B ) showed that the KNN model exhibited significant instability, characterized by substantial fluctuations in RMSE, whereas the other models maintained relatively stable performance. A comprehensive comparison of the mean RMSE values ( Figure 2C ) further indicated the SVM showed the lowest RMSE among the candidate models. Finally, time-dependent ROC analysis at 5-year IFS ( Figure 2D ) revealed that SVM model achieved the highest discriminative ability with an AUC of 0.671. In contrast, the KNN model exhibited the poorest predictive performance (AUC = 0.426), falling below the random chance level. Figure 2 Performance comparison of seven machine learning models. ( A ) Scatter plots of predictive vs actual IFS risk scores; ( B ) Line plot depicting the RMSE of each model across five cross-validation folds; ( C ) Bar plot showing the RMSE for each model; ( D ) Comparison of time-dependent ROC curves for different machine learning models at 5-year IFS. A) Seven scatter plots comparing Predictive value with Actual value for model: dt, model: knn, model: lightgbm, model: lsvm, model: mlp, model: rf and model: xgboost. The x-axis label is Actual value with tick labels negative 2, negative 1, 0, 1. The y-axis label is Predictive value with tick labels negative 2, negative 1, 0, 1. Each subplot includes a diagonal reference line from bottom left to top right. Model: dt shows many points scattered around the diagonal across the full range. Model: knn shows a wider spread of points around the diagonal, with many points deviating above and below the line. Model: lightgbm shows points closely aligned to the diagonal with small deviations. Model: lsvm shows points almost exactly on the diagonal line. Model: mlp shows points almost exactly on the diagonal line. Model: rf shows points clustered near the diagonal with a narrow spread. Model: xgboost shows points near the diagonal with moderate spread. A legend titled model lists dt, knn, lightgbm, lsvm, mlp, rf, xgboost. B) A line graph of rmse across five folds for multiple models. The x-axis label is Fold with tick labels Fold1, Fold2, Fold3, Fold4, Fold5. The y-axis label is rmse with tick labels 0.0, 0.1, 0.2, 0.3. A legend titled model lists dt, knn, lightgbm, lsvm, mlp, rf, xgboost. The knn line is highest, around 0.27 at Fold1, about 0.32 at Fold2, about 0.33 at Fold3, about 0.23 at Fold4 and about 0.19 at Fold5. The dt line is around 0.15 at Fold1, about 0.16 at Fold2, about 0.16 at Fold3, about 0.17 at Fold4 and about 0.11 at Fold5. The lightgbm line is around 0.07 at Fold1, about 0.08 at Fold2, about 0.09 at Fold3, about 0.09 at Fold4 and about 0.07 at Fold5. The rf line is around 0.09 at Fold1, about 0.09 at Fold2, about 0.13 at Fold3, about 0.11 at Fold4 and about 0.08 at Fold5. The xgboost line is around 0.08 at Fold1, about 0.08 at Fold2, about 0.07 at Fold3, about 0.06 at Fold4 and about 0.07 at Fold5. The lsvm line lies near 0.00 across all folds. The mlp line lies near 0.00 across all folds. C) A bar chart of rmse by model. The x-axis label is model with categories dt, knn, lightgbm, lsvm, mlp, rf, xgboost. The y-axis label is rmse with tick labels 0.0, 0.1, 0.2, 0.3. Numeric labels appear above bars: dt is 0.16, knn is 0.3, lightgbm is 0.08, lsvm is 0, mlp is 0.01, rf is 0.11, xgboost is 0.07. D) A roc curve plot comparing multiple models. The x-axis label is 1-Specificity with tick labels 0.0 to 1.0. The y-axis label is Sensitivity with tick labels 0.0 to 1.0. A diagonal reference line runs from left bottom to right top. Multiple step-like curves are shown. The legend lists dt (AUC equals 0.610), knn (AUC equals 0.426), lightgbm (AUC equals 0.627), lsvm (AUC equals 0.671), mlp (AUC equals 0.618), rf (AUC equals 0.627) and xgboost (AUC equals 0.613). Different plots comparing seven machine learning models using scatter, line, bar and roc curve charts. Performance comparison of seven machine learning models. ( A ) Scatter plots of predictive vs actual IFS risk scores; ( B ) Line plot depicting the RMSE of each model across five cross-validation folds; ( C ) Bar plot showing the RMSE for each model; ( D ) Comparison of time-dependent ROC curves for different machine learning models at 5-year IFS. To interpret the optimal SVM model, feature importance and SHAP analyses were conducted. As shown in Figure 3A , the feature importance ranking based on RMSE loss identified hysterectomy as the predominant predictor, exhibiting a significantly higher importance score compared to other variables, followed by number of pregnancy and VAS score. The SHAP summary plot ( Figure 3B ) further revealed the direction and magnitude of each variable’s contribution to the predicted IFS risk score. For instance, the presence of a hysterectomy was strongly associated with negative SHAP values, indicating a substantial protective effect and a reduced risk of IFS. These SHAP-based rankings were used to prioritize variables for subsequent multivariable Cox regression and nomogram construction. Figure 3 Interpretability analysis of the SVM model. ( A ) Global feature importance ranking based on the mean absolute shap value; ( B ) SHAP summary plot illustrating the impact of each feature on the model output. A) Dot plot of feature importance ranking. The y axis lists feature names: Hysterectomy, VAS_score, Number_of_pregnancy, Number_of_abortions, CA125_level, GnRH_a_course, Painful_nodule, Duration, Age. The x axis label is Feature Importance (loss: rmse). The x axis tick labels shown are 1.0, 1.2, 1.4, 1.6. Approximate dot positions by feature: Hysterectomy about 1.7; VAS_score about 1.1; Number_of_pregnancy about 1.08; Number_of_abortions about 1.02; CA125_level about 1.01; GnRH_a_course about 1.01; Painful_nodule about 1.01; Duration about 1.00; Age about 0.99. B) SHAP summary plot with jittered points for each feature. The y axis label is feature_label. The y axis categories shown are: Hysterectomy, VAS_score, Number_of_pregnancy, Number_of_abortions, CA125_level, GnRH_a_course, Age, Painful_nodule, Duration. The x axis label is SHAP value. The x axis tick labels shown are negative 0.2, 0.0, 0.2. A vertical reference line is drawn at 0.0. A vertical scale bar at the right is labeled Feature value, with endpoints labeled High and Low. Approximate point ranges by feature row: Hysterectomy spans from about negative 0.3 to about 0.25, with dense clusters around about negative 0.2 and about 0.15 to 0.2; VAS_score spans from about negative 0.2 to about 0.2, with a dense band near 0.0 and another band around about 0.1; Number_of_pregnancy spans from about negative 0.2 to about 0.15, with many points near 0.0; Number_of_abortions spans from about negative 0.12 to about 0.1, concentrated near 0.0; CA125_level spans from about negative 0.03 to about 0.08; GnRH_a_course spans from about negative 0.03 to about 0.08; Age spans from about negative 0.03 to about 0.03; Painful_nodule spans from about negative 0.03 to about 0.06; Duration spans from about negative 0.03 to about 0.03. Two plots showing feature importance ranking and shap value distribution for model input features. Interpretability analysis of the SVM model. ( A ) Global feature importance ranking based on the mean absolute shap value; ( B ) SHAP summary plot illustrating the impact of each feature on the model output. Based on the global SHAP analysis in optimal SVM model, the seven most influential predictors were identified, including age, CA125 level, number of pregnancies, number of abortions, VAS score, hysterectomy, and GnRH-a course. In the multivariate Cox regression analyses ( Table S2 ), number of pregnancies (HR=0.49, 95% CI: 0.25–0.96, P = 0.03) and hysterectomy (HR=0.22, 95% CI: 0.12–0.39, P < 0.01) were associated with IFS. In the fertility-sparing surgery subgroup, the number of pregnancy (HR=0.42, 95% CI: 0.20–0.89, P = 0.02) remained independently associated with IFS ( Table S2 ). These results indicate that the retained risk factors capture recurrence risk among uterine-preserving patients independently of the hysterectomy-related treatment effect. Subsequently, the final clinical prediction model was a Cox-based nomogram based on the seven selected variables for predicting IFS ( Figure 4A ). The nomogram was constructed using the proportional transformation of each regression coefficient from the multivariate Cox proportional hazards regression model to a scale ranging from 0 to 100 point. Points were summed up for all independent variables in order to obtain a total point which was subsequently converted into predicted probabilities for 3-year, 4-year, and 5-year IFS. Time-dependent ROC analysis demonstrated good discrimination performance. In the training set, the AUCs (95% CI) for predicting 3-, 4-, and 5-year IFS were 0.78 (0.73–0.83), 0.79 (0.74–0.84), and 0.80 (0.75–0.85), respectively ( Figure 4B ). Consistent results were observed in the test set, with AUCs (95% CI) of 0.72 (0.62–0.81), 0.73 (0.64–0.81), and 0.73 (0.65–0.81) for the three timepoints ( Figure 4C ). The calibration curves demonstrated good agreement between predicted and observed survival probabilities across all timepoints (3-year, 4-year, and 5-year) for IFS in the training ( Figure 4D ) and test set ( Figure 4E ). The calibration slopes were close to 1.0 (ranging from 1.02 to 1.08) and intercepts approached 0 (ranging from −0.03 to 0.02), indicating minimal deviation between predicted and actual risks. Decision curve analysis revealed a substantial net benefit associated with the nomogram for predicting 3-year, 4-year, and 5-year IFS in the training ( Figure 4F ) and test set ( Figure 4G ). The nomogram demonstrated a favorable net benefit across a broad range of threshold probabilities (approximately 5% to 55%). Figure 4 Development and validation of nomogram for predicting IFS. ( A ) A nomogram for predicting 3-year, 4-year, and 5-year IFS; the ROC curve of nomogram for predicting 3-year, 4-year, and 5-year IFS in the training set ( B ) and testing set ( C ); the calibration curves of nomogram for predicting 3-year, 4-year, and 5-year IFS in the training set ( D ) and testing set ( E ); decision curve analysis for 3-year, 4-year, and 5-year IFS in the training set ( F ) and testing set ( G ). A) Nomogram with a top scale labeled, Points, from 0 to 100 and a title, Nomogram. Rows list variables with category labels and a marked selection box on each row: Hysterectomy with Yes and No; Numberofabortions with less than 3 and greater than or equal to 3; CA125level with less than or equal to 35 U mL and greater than 35 U mL; GnRHacourse with 0 and greater than or equal to 1; VASscore with greater than or equal to 1 and 0; Age with greater than 38 years and less than or equal to 38 years; Numberofpregnancy with greater than or equal to 2 and less than or equal to 2. Below is a scale labeled, Total points, with tick labels 400, 420, 440, 460, 480, 500, 520, 540, 560 and 580 and a wavy line drawn above it. Three probability scales follow: Pr(Time <’ 60) with tick labels 0.03, 0.05, 0.07, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6 and 0.7; Pr(Time <’ 48) with tick labels 0.03, 0.05, 0.07, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6 and 0.65; Pr(Time <’ 36) with tick labels 0.02, 0.04, 0.06, 0.08, 0.15, 0.25, 0.35, 0.45 and 0.55. B) Receiver operating characteristic plot labeled, Train. X axis label, 1 minus Specificity, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. Y axis label, Sensitivity, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. Three receiver operating characteristic curves are drawn along with a diagonal reference line. Text box values: 3 year AUC (95 percent CI) equals 0.78 (0.73 to 0.83); 4 year AUC (95 percent CI) equals 0.79 (0.74 to 0.84); 5 year AUC (95 percent CI) equals 0.80 (0.75 to 0.85). C) Receiver operating characteristic plot labeled, Test. X axis label, 1 minus Specificity, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. Y axis label, Sensitivity, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. Three receiver operating characteristic curves are drawn along with a diagonal reference line. Text box values: 3 year AUC (95 percent CI) equals 0.72 (0.62 to 0.81); 4 year AUC (95 percent CI) equals 0.73 (0.64 to 0.81); 5 year AUC (95 percent CI) equals 0.73 (0.65 to 0.81). D) Calibration plot titled, Calibration Curves at 3-, 4- and 5-Year. X axis label, Nomogram Predicted Probability, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. Y axis label, Actual Observed Probability, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. A diagonal reference line runs from the lower left to the upper right. Three calibration lines with vertical error bars are shown at several probability positions. E) Calibration plot titled, Calibration Curves at 3-, 4- and 5-Year. X axis label, Nomogram Predicted Probability, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. Y axis label, Actual Observed Probability, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. A diagonal reference line runs from the lower left to the upper right. Three calibration lines with vertical error bars are shown at several probability positions. F) Decision curve analysis plot titled, Decision Curve Analysis at 3-, 4- and 5-Year. X axis label, Threshold Probability, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. Y axis label, Net Benefit, with tick labels approximately 0.0, 0.1, 0.2, 0.3 and 0.4. Multiple lines are shown, including three solid curves and horizontal reference lines labeled, Treat all and a baseline labeled, Treat none. G) Decision curve analysis plot titled, Decision Curve Analysis at 3-, 4- and 5-Year. X axis label, Threshold Probability, with ticks 0.0, 0.2, 0.4, 0.6, 0.8 and 1.0. Y axis label, Net Benefit, with tick labels approximately 0.0, 0.1, 0.2, 0.3 and 0.4. Multiple lines are shown, including three solid curves and horizontal reference lines labeled, Treat all and a baseline labeled, Treat none. A nomogram with six plots for training and test performance at 3, 4 and 5 years. Development and validation of nomogram for predicting IFS. ( A ) A nomogram for predicting 3-year, 4-year, and 5-year IFS; the ROC curve of nomogram for predicting 3-year, 4-year, and 5-year IFS in the training set ( B ) and testing set ( C ); the calibration curves of nomogram for predicting 3-year, 4-year, and 5-year IFS in the training set ( D ) and testing set ( E ); decision curve analysis for 3-year, 4-year, and 5-year IFS in the training set ( F ) and testing set ( G ).

Patients

This study was conducted and reported in accordance with the STROCSS guideline 21 and Declaration of Helsinki. This retrospective cohort study consecutively enrolled patients who underwent surgical treatment and received pathological confirmation of AM at our institution between March 2016 and December 2020. This study was approved by the Institutional Review Committee of our hospital (TJ-IRB20221157, dated November 2022), which authorized both the retrospective review of medical records and the subsequent outcome follow-up conducted through September 2025. As this retrospective study involved only the review of existing medical records and posed minimal risk to patients, the IRB waived the requirement for written informed consent. All patient data were anonymized prior to analysis and accessed exclusively by the study investigators, thereby safeguarding patient confidentiality and privacy. The inclusion criteria were as follows: (1) received a confirmed pathological diagnosis of AM; (2) aged 18 years or older; (3) underwent open or laparoscopic surgery for AM; and (4) availability of complete preoperative baseline and postoperative follow-up data. The exclusion criteria were: (1) presence of severe organ damage; (2) had surgical pathology confirming precancerous or malignant lesions; and (3) loss to follow-up or incomplete clinical data. The final follow-up was completed in September 2025, ensuring a minimum postoperative follow-up duration of 12 months. Clinical data were retrieved from the hospital’s electronic medical record system, including age, height, weight, disease duration, age at menarche, regular menstruation, menorrhagia, marital history, pregnancy history, number of pregnancies, number of deliveries, number of abortions, history of cesarean section, dysmenorrhea, VAS score, history of painkiller use, hormone therapy, intrauterine device use, preoperative GnRH-a treatment, and surgical history of EMS. Blood samples were obtained prior to treatment on the same day. The comprehensive hematologic and metabolic parameters included: leukocyte, neutrophil, lymphocyte, and monocyte counts; hemoglobin and platelet levels; fibrinogen; albumin, globulin, glutamyltransferase, creatinine; and CA125 and CA199 antigen levels. The comprehensive assessment of the lesions was conducted based on preoperative ultrasound imaging and direct intraoperative exploration. The assessed parameters included uterine diameter, lesion location, and the presence or absence of uterine adenomyoma, uterine fibroids, endometrial polyps, ovarian EMS, pelvic peritoneal lesions, painful nodules, and pelvic adhesions. Missing values in candidate predictors were handled using multiple imputation by chained equations (MICE). Continuous predictors were imputed using predictive mean matching, and categorical predictors were imputed using logistic regression. To avoid information leakage during internal validation, the imputation model was fitted within each resampling fold using only the training data and then applied to the internal split-sample validation data. The surgeries were performed by senior gynecologists with at least 5 years of laparoscopic experience. All hysterectomy procedures were conducted in accordance with the standard protocol at our institution. Under general anesthesia, a standard laparoscopic procedure was performed, involving ligation of uterine ligaments, vessel hemostasis, and dissection of the cervix and vaginal fornix. The intact uterus was removed via the vaginal or abdominal route based on size, and bilateral ovaries were preserved in all patients. Regarding fertility-sparing surgery, hemostatic measures were implemented, including the placement of a tourniquet around the uterine isthmus and myometrial injection of vasopressin. Subsequently, the uterus was bisected to expose the endometrial cavity, facilitating the meticulous excision of adenomyotic lesions. Ultimately, uterine integrity was restored, followed by the re-establishment of uterine perfusion. Postoperative follow-up data, including the occurrence of IFS, were obtained from medical records and structured telephone interviews. IFS status was evaluated through a comprehensive assessment comprising medical history inquiry, gynecological examination, and pelvic ultrasound. IFS is defined as the duration from the completion of surgery (and any adjuvant therapy, if applicable) to the first occurrence of any additional therapeutic intervention for adenomyosis. An IFS event included initiation of new or escalated medical therapy specifically for adenomyosis-related symptoms, repeat uterine-sparing procedures, or hysterectomy. Isolated symptom recurrence, imaging progression, or CA125 elevation without subsequent therapeutic intervention were not counted as IFS events. The decision to initiate additional therapy was made by the treating clinicians in routine practice based on persistent or recurrent symptoms, patient preference, and clinical assessment, rather than by imaging findings or biomarker changes alone. In this study, seven commonly used machine learning algorithms were employed to model the training set based on all included feature variables. The primary target was the individual risk of experiencing an IFS event over time, expressed as a Cox-based linear predictor (risk score). This risk score was derived from a Cox proportional hazards model fitted to the observed follow-up times and IFS events, so that right censoring was accounted for in its estimation. To assess their predictive validity for the actual endpoint, model performance was evaluated on the true time-to-event outcome using time-dependent AUC, with the 5-year time point used as the primary reference for model selection because it represented the longest prespecified follow-up horizon with adequate event numbers, and RMSE used as an auxiliary metric to assess the agreement between predicted and observed risks. The model selection progressed a gradient from simple to complex, to balance predictive performance with interpretability. In addition, this study utilizes SHapley Additive exPlanations (SHAP) to interpret the association between feature variables and prediction outcomes. This model is a non-parametric supervised learning method known for its ease of understanding and interpretation, which predicts target variables by inferring a series of simple decision rules. 17 Its construction process relies on information gain for feature selection, recursively splitting branches from the root node based on the feature with the maximum gain until the gain is negligible or features are exhausted. This study employs the Classification and Regression Tree (CART) algorithm based on the Gini coefficient and sets the maximum number of leaf nodes to 10 to optimize the model structure. RF is an ensemble learning algorithm composed of a large number of independent DT. Its construction combines bootstrap sampling with random feature selection at node splits, outputting the final conclusion by aggregating the predictions of all trees. 22 In this study, the number of trees is set to 500, with other parameters kept at default settings. This study employs the XGBoost algorithm for feature importance ranking and selection. The model iteratively generates DT weak classifiers, optimizing based on the bias of previous trees. 23 A grid search was used to explore hyperparameter combinations. To prevent overfitting, training was capped at 500 rounds, and an early stopping mechanism was applied using a validation subset derived from the training data. Training was stopped if the validation performance did not improve for 30 consecutive rounds. The independent test set was not used for model training, hyperparameter tuning, or early stopping, and was reserved exclusively for final evaluation. The SVM aims to find the optimal hyperplane in the feature hyperspace to maximize the classification margin. For non-linearly separable data, this study uses a Gaussian kernel function to map samples into a high-dimensional space for linear separation. 24 Data standardization was performed prior to modeling to accommodate distance calculations. The regularization parameter C is set to 0.5. Neural network models simulate human brain neuronal connections to construct a network structure comprising an input layer, hidden layers, and an output layer. This model applies activation functions to linear models and nests them layer by layer, transforming simple linear relationships into complex non-linear mappings to fit the data. 17 Specifically, the MLP is formed by multiple hidden layers and can extract deeper and higher-dimensional features from input information. In this study, the MLP comprised three hidden layers with 64, 32, and 16 neurons, respectively. The ReLU activation function was used, and the model was trained using the Adam optimizer with a learning rate of 0.001, L2 regularization, and a mini-batch size of 32. LightGBM is a highly efficient gradient boosting algorithm specifically developed to handle massive, high-dimensional datasets. As a cutting-edge ensemble learning technique analogous to XGBoost, it represents a significant advancement over the traditional framework, offering superior computational efficiency and reduced memory consumption. 25 KNN is a simple, instance-based classification algorithm. Its core logic involves finding the k nearest neighbors to a target sample in the feature space and determining the sample’s category based on a majority voting principle. 26 Due to its reliance on distance calculations, the algorithm requires data standardization. This study uses Euclidean distance as the metric and sets the key hyperparameter k to 9. A unified 5-fold cross-validation framework was employed to ensure stability and generalizability, with hyperparameters optimized using Bayesian optimization methods. In this process, the dataset was randomly partitioned into five equal-sized subspaces. The model was trained on four subspaces and validated on the remaining one, repeating this process five times so that each subspace was used for validation exactly once. The predictive accuracy and robustness of the models were compared and evaluated using the root mean squared error (RMSE). Furthermore, regression plots and residual analysis were utilized to visually inspect the goodness-of-fit and bias of the candidate models. The algorithm demonstrating the highest accuracy (lowest RMSE) and the most compact error distribution was selected as the final model for subsequent analysis. A nomogram was developed based on a multivariate Cox proportional hazards regression to predict 3-year, 4-year, and 5-year IFS. The discriminatory capacity of the nomogram was assessed using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. The calibration curves were used to assess the agreement between predicted IFS probabilities and observed outcomes. The clinical net benefit and decision-making value of the nomogram were determined using decision curve analysis (DCA). Predictor selection followed a multistage procedure to balance model stability, interpretability, and clinical applicability. LASSO regression with 20-fold cross-validation was first used to reduce the candidate predictors. Seven machine-learning algorithms were then evaluated using 5-fold cross-validation, and SHAP values were used to rank variables by their predictive contribution to the IFS risk score. The final predictors were selected for multivariable Cox regression and nomogram construction based on SHAP-based ranking, clinical relevance, and Cox model estimates. Cross-validation was applied within each stage to reduce information leakage. All the statistical results were done quantitatively with R software (Version 4.4.1: http://www.r-project.org/ ). Given that most variables deviated from a normal distribution, they are presented as medians with interquartile ranges (Q1, Q3). Between-group comparisons were performed using the two-sided Mann–Whitney U -test. Categorical variables are expressed as numbers (percentages) and were compared using the χ 2 test. The selection of candidate variables was employed by the Least Absolute Shrinkage and Selection Operator (LASSO) analysis using “glmnet” R package. To explore the prognostic factors, seven machine learning algorithms, including DT, XGBoost, RF, LSVM, MLP, LightGBM, and KNN, were compared and evaluated using the “tidymodels” R packages. The hazard ratio (HR) and its corresponding 95% confidence interval (CI) was computed with Cox proportional hazards regression model using the “survival” R packages. Then, a nomogram was developed based on the multivariate Cox proportional hazards regression model using the “regplot” R packages. Next, the discriminatory ability of nomogram model was evaluated by the area under the ROC curve (AUC), calibration curve and decision curve analysis (DCA) using the “timeROC”, “rms” and “dcurves” R package. A P -value < 0.05 was deemed to be statistically significant.

Conclusion

This study developed and internally validated a prognostic model for IFS in patients with AM. Among the algorithms evaluated using 541 patient records, the SVM showed the best discrimination in internal split-sample validation. Furthermore, SHAP analysis identified seven key predictors: age, CA125 level, number of pregnancies, number of abortions, VAS score, history of hysterectomy, and GnRH-a courses. The nomogram showed moderate discrimination and good calibration in the internal split-sample validation cohorts, suggesting it may serve as a hypothesis-generating aid to risk stratification. Specifically, it may help identify patients at higher risk of postoperative recurrence or intervention, including younger women seeking fertility preservation. These risk estimates may help inform the intensity of postoperative follow-up. Accordingly, the model requires external validation in independent cohorts and prospective evaluation before it can inform clinical practice.

Discussion

AM is an estrogen-dependent gynecological disorder affecting mainly reproductive-aged women, which contributes to conditions like chronic pelvic pain, dyspareunia, and infertility. 27 Characterized by uterine wall thickening, this condition affects 20%–40% of reproductive-age women, 5 and imposes an annual economic burden estimated between $15.58 billion and $78.05 billion. 6 The etiology of AM is associated with several risk factors, including increased estrogen exposure (eg, increased parity, early menarche, high BMI, and the use of tamoxifen) and a history of uterine surgical procedures (eg, D&C, caesarean section, and myomectomy). 28 AM should be conceptualized not as a singular clinical entity, but as a heterogeneous condition characterized by variations in lesion localization, extent, inflammatory activity, and coexisting disorders. 3 Therefore, AM is recognized as a chronic condition necessitating individualized, long-term management. This study included 541 AM patients, comprising 203 who underwent hysterectomy and 338 who received fertility-sparing surgery. Among the 198 recorded IFS events, severe clinical manifestations accounted for nearly half (47.0%), followed by imaging-confirmed disease progression (42.4%) and significant serum CA125 elevation (10.6%). Over a median follow-up of 77 months, the incidence of IFS events was significantly lower in the hysterectomy group compared to the fertility-sparing surgery group ( Table 1 ). We developed a suite of seven machine learning models, including DT, RF, XGBoost, SVM, MLP, LightGBM, and KNN, to identify key predictive factors. Based on a comprehensive comparison of mean RMSE values and 5-fold cross-validation results, the SVM model demonstrated superior overall performance. In contrast, the KNN model exhibited the poorest performance, primarily due to significant instability observed during validation. Both KNN and SVM models rely on distance metrics within the feature hyperspace, yet their classification mechanisms differ fundamentally. KNN determines class probabilities by evaluating the class distribution of neighboring samples, whereas SVM seeks to identify the optimal hyperplane for sample separation. Significant data overlap regarding IFS within the feature hyperspace was observed in this study ( Figure S1A ), which rendered the KNN model ineffective in distinguishing between the two classes. Consequently, it yielded the lowest prediction accuracy among the seven machine learning models. 26 In contrast, the SVM model effectively overcame the classification barrier posed by data overlap by employing a Gaussian kernel function, which maps linearly inseparable data into a higher-dimensional space where it becomes linearly separable. 24 Based on SHAP analysis of the optimal SVM model, a nomogram incorporating seven key predictors was developed to estimate 3-year, 4-year, and 5-year IFS probabilities. This nomogram demonstrated good predictive accuracy, with AUCs ranging from 0.72 to 0.80 across both the training and test sets. In addition, DCA showed good calibration and modest net benefit on decision curve analysis; its reliability and clinical utility remain to be confirmed in external cohorts. The novelty of this study lies not in identifying previously unknown prognostic factors, but in integrating clinically established predictors into a unified framework for predicting IFS after surgery treatment for adenomyosis. AM can manifest with a symptom profile comparable in spectrum and intensity to both superficial peritoneal EMS and DIE. 29 Evidence indicates a strong link between AM and EMS, as the concurrent presence of both conditions is reported in 13% to 89% of patients with EMS. In our cohort of 541 patients with AM, 159 (29.4%) were found to have concomitant ovarian EMS ( Table 1 ). However, these estimates are significantly constrained by the absence of histologic confirmation for AM and/or EMS, often relying instead on imaging findings merely suggestive of these conditions. The coexistence of both disorders is associated with more severe clinical manifestations, including heavy menstrual bleeding, dysmenorrhea, and chronic pelvic pain. 15 In addition, the patients with EMS were significantly more likely to present with severe dysmenorrhea compared to those without the condition. 14 EMS predominantly affects women of reproductive age, with a global prevalence of 10%–15%. Although histologically benign, EMS exhibits malignant-like behaviors, such as implantation, invasion, distant metastasis, and recurrent hemorrhage. 30 Imaging studies have further revealed that the proportion of patients with specific AM subtypes (eg, anterior and posterior wall types) co-occurring with EMS can exceed 80%. 31 AM and EMS can provoke diffuse peritoneal inflammation and uterine hypercontractility, which may account for the increased burden of chronic pain and deep dyspareunia. 29 Notably, patients with concomitant AM and EMS may be at risk of inadequate therapeutic intervention if treated with hysterectomy alone. This underscores the critical importance of laparoscopic pelvic assessment during hysterectomy for patients with suspected EMS, particularly in those presenting with severe dysmenorrhea. 32 Total hysterectomy is widely recognized as the most effective surgical intervention for AM, as it entails the complete removal of the affected uterus, thereby ensuring symptom resolution. 33 Hysterectomy emerged as an important predictor of IFS in our analysis. This finding should be interpreted cautiously because hysterectomy is itself a definitive treatment that directly reduces the likelihood of requiring additional therapeutic intervention. Its strong predictive contribution may therefore reflect treatment effect, clinical indication, and treatment-selection bias rather than an independent biological risk factor. However, this procedure results in the irreversible loss of reproductive potential and carries inherent risks of pelvic floor disfunction. 34 Furthermore, hysterectomy may cause a potential decline in sexual function postoperatively attributable to anatomical and neurovascular alterations. 35 Therefore, identifying these patients prior to surgery can guide preoperative counseling and ensure surgical preparedness for the concurrent management of EMS at the time of hysterectomy. In addition to hysterectomy, advanced age has been identified as a crucial protective factor for IFS. A retrospective study involving 420 patients demonstrated that the risk of recurrence after conservative surgery without adjuvant medication significantly decreased with advancing age. The cumulative recurrence rate for patients with endometrioma aged 40–45 was only 10.2%, far lower than that of younger patients. 36 Furthermore, another study retrospectively analyzed 134 patients who underwent surgery and long-term medication, reporting an overall recurrence rate of 26.1%. The postoperative dysmenorrhea and younger age were identified as independent risk factors for endometrioma recurrence. 37 Typically, GnRH-a treatment could significantly reduce uterine volume and adenomyotic foci but induced menopausal symptoms like hot flushes. 38 Moreover, while the drug helps prevent pain recurrence, patients require close surveillance for abnormal uterine bleeding. 39 Although the number of GnRH-a courses was identified as a predictor of IFS in our model, this association should not be interpreted as evidence that GnRH-a therapy reduces recurrence or improves IFS. In addition, the efficacy of add-back therapy combined with GnRH-a in preventing induced menopausal symptoms remains to be fully elucidated and warrants further investigation. AM is an estrogen-dependent gynecological disorder defined by the diffuse infiltration and proliferation of endometrial tissue within the myometrium. 1 It is hypothesized that the disruption of the endometrial-myometrial interface serves as the initiating event for AM. 40 Such disruption, often precipitated by induced abortion or uterine dilatation and curettage (D&C), may predispose individuals to the development of this disease. 41 Serum CA125 is a glycoprotein predominantly expressed by coelomic epithelial derivatives within an inflammatory environment. 42 It may facilitate ectopic endometrial migration and enhance myometrial adhesion capacity, clinically manifesting as increased pain severity. 43 Typically, elevated preoperative CA125 levels were correlated with larger adenomyotic lesion volumes, particularly in external-type AM, which is characterized by higher rates of pelvic adhesions and concurrent EMS. 44 Therefore, serum CA125 levels are significantly elevated in patients with AM and also serve as a predictive marker for symptom recurrence. 45 In conclusion, this study evaluated seven machine learning algorithms (DT, RF, XGBoost, SVM, MLP, LightGBM, and KNN) to identify key predictive factors for IFS. The SVM model demonstrated superior overall performance. Based on the seven selected variables, age (CA125 level, number of pregnancies, number of abortions, VAS score, hysterectomy, and GnRH-a course), a nomogram was developed to estimate the probability of 3-year, 4-year, and 5-year IFS. This model showed moderate discrimination with acceptable calibration and net benefit, suggesting it may assist individualized clinical management.

Limitation

Several limitations should be acknowledged. First, the single-center retrospective design, particularly in a tertiary care setting, may predispose the study to selection bias. Furthermore, the absence of external validation cohorts limits the assessment of the model’s generalizability and restricts its external validity. Third, the frequent co-occurrence of AM with EMS hinders the ability to distinguish the independent impact of AM, suggesting that future multicenter studies should aim to stratify patients based on the presence or absence of combined lesions. Nevertheless, the utilization of prospectively collected follow-up data, adhering to standardized documentation practices, serves as a mitigating factor for data quality concerns inherent in retrospective analyses. Fourth, internal validation was performed using a single random 70/30 split-sample approach, which may yield unstable performance estimates in a dataset of modest sample size. Given the single-center design and the absence of external validation, the developed models should be regarded as hypothesis-generating and require external validation in independent cohorts before clinical implementation.

Text is read by the "Ask this paper" AI Q&A widget below. Extraction quality varies by source — PMC NXML preserves structure cleanly, OA-HTML may include some navigation residue, and OA-PDF can have broken hyphenation. The publisher copy (via DOI) is the canonical version.

My notes (saved in your browser only)

⚙ Ask this paper AI returns verbatim quotes from the full text · source: pmc-nxml ⓘ

Answers must be backed by verbatim quotes from this paper's full text. Hallucinated quotes are dropped automatically; if no verbatim passage answers the question, we say so. How this works

Citation neighborhood (no data yet)

We don't have any in-corpus citations linked to this paper yet. This is a recent paper (2026) — citers typically take a year or two to land, and the OpenAlex reference graph may still be filling in.

Source provenance

europepmc
last seen: 2026-11-18T06:10:28.346261+00:00
pubmed
last seen: 2026-11-18T06:04:25.516603+00:00
unpaywall
last seen: 2026-11-18T06:44:38.916964+00:00
License: CC-BY-NC-4.0 · commercial use OK · attribution required
Courtesy of the U.S. National Library of Medicine