Risk of Postoperative Recurrence in Patients With Unusual Smooth Muscle Neoplasia and Construction of a Prediction Model

Shiying Wu: data curation. Weiwei Hu: conceptualization, methodology, writing – original draft. Xiangyang Zeng: visualization. Yueran Li: supervision. Songshu Xiao: writing – review and editing. Jingxi Zhang: data curation.

This study was approved by the Ethics Committee of the Third Xiangya Hospital of Central South University (approval no. Kuai 26266; approved on 27 March 2026) and was conducted in accordance with the Declaration of Helsinki. Owing to the retrospective design of the study, the requirement for informed consent was waived by the committee.

The authors have nothing to report.

Of the 248 patients, 87 (35.1%) recurred within 12 months and 161 did not. The recurrence and non‐recurrence groups differed in BMI, coexisting endometriosis and adenomyosis, maximum tumor diameter, tumor number, tumor location, surgical approach and T 2 WI signal (all p  < 0.05); age, menarche, gravidity, parity, infertility, previous caesarean, abnormal uterine bleeding, benign adnexal masses, morcellator use, preoperative hemoglobin, CA125 and LDH did not differ significantly (all p  > 0.05). Notably, histological subtype was similarly distributed in the two groups ( p  = 0.793). Details are shown in Table  1 . Baseline characteristics of the recurrence and non‐recurrence groups. Abbreviations: BMI, body mass index; IQR, interquartile range; LDH, lactate dehydrogenase; T 2 WI, T 2 ‐weighted imaging. Special site refers to broad‐ligament and cervical leiomyoma. Data are presented as mean ± SD, median (interquartile range) or n (%). The 248 patients were divided 7:3 into a training set ( n  = 173) and a validation set ( n  = 75). No baseline variable differed significantly between the two sets (all p  > 0.05), indicating a balanced split for model development and validation (Table  S1 ). LASSO regression was applied to the 21 preoperative variables in the training set. Cross‐validated deviance was minimized at log( λ ) = −3.3399, at which nine variables retained non‐zero coefficients (Figure  1 ). Entering these into multivariable logistic regression, intramural location, T 2 WI high signal, maximum diameter ≥ 6 cm, multiple tumors, BMI and coexisting adenomyosis were each independently associated with recurrence (all p  < 0.05; Table  2 ). The VIF was below 5 for every variable (highest 1.08 for maximum diameter), indicating no meaningful collinearity. These six variables were carried forward to model building. LASSO coefficient paths (left) and cross‐validation curve (right) for the 21 preoperative variables. Multivariable logistic regression for postoperative recurrence. Abbreviations: B , regression coefficient; BMI, body mass index; CI, confidence interval; OR, odds ratio; S.E., standard error; T 2 WI, T 2 ‐weighted imaging. Because the cohort combined three biologically distinct variants, we compared recurrence rates across subtypes as a sensitivity analysis. Recurrence occurred in 64/180 cellular (35.6%), 10/26 atypical (38.5%) and 13/42 vascular (31.0%) tumors, with no significant difference between subtypes ( χ 2  = 0.463, p  = 0.793). Subtype was likewise not retained by LASSO. Pooling the three variants for modeling therefore appears reasonable, although the atypical and vascular groups were small and the analysis is underpowered to exclude modest subtype‐specific differences. Before weighting, the three surgical groups were markedly unbalanced (8–11 of 13 covariates with SMD > 0.1 across pairwise comparisons); after stabilized‐weight IPTW the number of covariates with SMD > 0.1 fell to three in every comparison. In the weighted, covariate‐adjusted model (laparoscopy as reference), neither hysteroscopy (OR 1.835, 95% CI: 0.799–4.214, p  = 0.153) nor laparotomy (OR 2.663, 95% CI: 0.952–7.448, p  = 0.062) was significantly associated with recurrence, and the overall effect of approach did not reach significance (Wald χ 2  = 4.793, p  = 0.091). Pairwise 1:1 matching gave concordant results, with no significant between‐group difference in any comparison (all p  > 0.05). Surgical approach was therefore not entered into the prediction model. All eight algorithms were trainable, and several fitted the training data closely (RF AUC 0.952, GBM 0.910, AdaBoost 0.905; Table  S2 , Figure  S1 ). Performance in the held‐out validation set was more modest and is the basis for model selection (Table  3 , Figure  2 ). GBM achieved the highest validation AUC (0.785, 95% CI: 0.672–0.898) together with the best sensitivity (0.750) and F1 (0.679); LR and AdaBoost followed closely on AUC (0.765 and 0.764) but with substantially lower sensitivity. RF, which had the highest training AUC, fell to 0.757 in validation, consistent with overfitting; model choice was therefore based on validation performance rather than training fit. Performance of the eight machine‐learning models in the validation set. Abbreviations: AdaBoost, adaptive boosting; AUC, area under the receiver operating characteristic curve; F1, F1 score; GBM, gradient boosting machine; KNN, k‐nearest neighbors; LR, logistic regression; NN, neural network; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting. ROC curves of the eight machine‐learning models in the validation set. Calibration curves showed acceptable agreement between predicted and observed risk for most models, with GBM among the closest to the diagonal in the validation set (Figure  3 ). On decision‐curve analysis, the models generally provided net benefit over the treat‐all and treat‐none strategies across a range of thresholds, and GBM gave stable net benefit in the intermediate‐threshold range relevant to clinical decision‐making (Figure  4 ). Taken together (validation AUC, F1, calibration and DCA), GBM was selected as the best model. Calibration curves of the models in the validation set. Decision‐curve analysis of the models in the validation set. At a probability threshold of 0.40, the GBM model correctly identified 18 of 24 recurrences in the validation set (sensitivity 0.750) while misclassifying 11 non‐recurrent patients as high risk (Figure  S2 ). The comparatively high sensitivity is clinically desirable, since missing a high‐risk patient carries a greater cost than an additional follow‐up visit. SHAP analysis of the GBM model ranked tumor location, T 2 WI signal, maximum diameter, tumor number, BMI and adenomyosis as the leading contributors to prediction (Figure  5 ). The beeswarm plot showed that intramural location, T 2 WI high signal, larger diameter, multiple tumors, higher BMI and coexisting adenomyosis all pushed the prediction towards recurrence (Figure  6 ). These directions agree with the multivariable regression, while SHAP additionally displays how the strength of each feature varies across patients. A representative individual explanation is provided in Figure  S3 . SHAP global feature importance for the GBM model. SHAP beeswarm plot showing the distribution and direction of each feature's effect (GBM).

In this cohort of 248 patients with special‐type USMN, 35.1% recurred within 12 months—a high early rate for a formally benign group of tumors, and in keeping with reports that these variants are not entirely indolent [ 6 , 15 ]. Because recurrence was assessed only after a negative 3‐month ultrasound and at a fixed 12‐month landmark, this figure is unlikely to be inflated by residual disease. We identified six independent, preoperatively available correlates of recurrence and built an interpretable model around them. Intramural location was associated with recurrence, echoing analyses of ordinary leiomyoma in which intramural tumors predicted recurrence [ 16 ]. Similar risk factors have been reported for special‐type leiomyomas that recur after laparoscopic surgery [ 17 ]. Intramural lesions lie within the myometrium and are less easily exposed and enucleated; when the margin with surrounding muscle is indistinct, complete removal is harder, and this may be accentuated in special‐type variants whose texture and vascularity differ from ordinary myoma. Residual microscopic foci in a hormonally responsive myometrium may then regain proliferative activity. Subserosal tumors, which project outward and are more readily excised, behave oppositely. Intramural location is thus best read as a marker of the difficulty of complete clearance rather than a direct cause of recurrence. High T 2 WI signal was also independently associated with recurrence. Cellular and vascular variants, and mesenchymal tumors that are difficult to distinguish from sarcoma, frequently show high or heterogeneous T 2 signal [ 18 ], so in this population a high T 2 WI signal may flag a more cellular, biologically active lesion [ 19 ]. As with the other variables, we interpret this as a risk marker rather than a causal factor: the imaging phenotype reflects the underlying tissue, not a mechanism of recurrence. A maximum diameter ≥ 6 cm and the presence of multiple tumors were both associated with recurrence, consistent with previous work [ 20 ]. Larger tumors are more difficult to enucleate, control for bleeding and repair, raising the chance of residual tissue, and each additional tumor carries its own risk of incomplete removal or of a small lesion being missed. A broader myometrial predisposition—reflected molecularly in changes such as MED12 mutation or HMGA2 rearrangement across independent clones—may also underlie multiplicity, so that myomectomy removes existing lesions without altering the tendency to form new ones. Elevated BMI was independently associated with recurrence. Adipose aromatase activity raises circulating estrogen, and obesity‐related low‐grade inflammation and insulin resistance increase IGF‐1, all of which can stimulate smooth‐muscle proliferation and favor regrowth of residual or new lesions [ 21 , 22 ]. A metabolic contribution is further suggested by reports linking elevated serum lipids to fibroid recurrence [ 23 ]. Coexisting adenomyosis was likewise associated with recurrence; although a distinct disease, it shares an estrogen‐dependent basis with leiomyoma, and the thickened, heterogeneous myometrium it produces both complicates complete enucleation and may provide a microenvironment more permissive to regrowth [ 24 ]. Endometriosis, significant on univariate analysis, did not remain in the model, plausibly because it frequently coexists with adenomyosis and its effect was partly absorbed by the latter. Crude recurrence rates differed by surgical approach, but this difference did not persist after propensity‐score adjustment. Approach is selected largely on tumor burden and exposure, so the groups differ at baseline; after IPTW and matching, no approach was independently associated with recurrence, although laparotomy retained a non‐significant trend towards higher risk (OR 2.663, p  = 0.062) that likely reflects the larger, more complex tumors selected for open surgery and the small numbers involved. Importantly, propensity methods balance only recorded covariates; surgeon experience, the extent of enucleation and the quality of repair were not captured and could not be adjusted for, and these factors may themselves influence recurrence. Surgical approach is therefore better regarded as a correlate of tumor characteristics than as an independent, modifiable determinant of recurrence. Recurrence rates were similar across the cellular, atypical and vascular subtypes, and subtype was not selected as an independent predictor. This supports analyzing the variants together for the purpose of recurrence prediction, but the atypical ( n  = 26) and vascular ( n  = 42) groups are small, and the rarer variants excluded here were fewer still; subtype‐specific effects can therefore not be ruled out, and this remains a limitation. Among the eight algorithms, GBM gave the best validation performance and was chosen as the final model. By fitting successive residuals, gradient boosting accommodates non‐linear effects and interactions, and it proved more stable than RF, whose high training AUC was not reproduced in validation. GBM's advantage over logistic regression lay mainly in sensitivity (0.750 vs. 0.542): for a recurrence‐prediction task, correctly flagging high‐risk patients matters more than the slightly higher specificity of LR. All six inputs are obtainable before surgery from routine ultrasound or MRI and clinical assessment, so the model can be used at the counseling and planning stage rather than requiring postoperative pathology. Machine‐learning models have recently been developed to predict uterine fibroid occurrence and recurrence more broadly [ 25 ]; our study extends this approach to the special‐type variants using only preoperative variables. Coupling the model with SHAP addresses the interpretability that often limits clinical uptake of machine learning. SHAP confirmed that the same six variables drove prediction and that their directions matched the regression, while adding patient‐level explanations. It should be read as showing association, not causation: a high T 2 WI signal, for example, marks a higher‐risk phenotype rather than a change that itself produces recurrence, and SHAP output is most useful when interpreted alongside clinical and pathological reasoning. These findings have practical implications. When the model indicates high risk, uterine preservation versus hysterectomy can be weighed against the patient's age and fertility wishes; the myometrium can be explored carefully and, where helpful, with intraoperative ultrasound, to maximize clearance; and follow‐up can be intensified—for instance transvaginal ultrasound every 3 months in the first 2 years. BMI is a modifiable factor, so weight management can be included in follow‐up, and coexisting adenomyosis may prompt consideration of adjuvant medical measures. The aim of closer surveillance is not to prevent recurrence outright but to detect lesion change early, when management options are broadest. A note of caution is warranted regarding malignant transformation. The manuscript does not treat transformation as a study outcome, and no such events were analyzed here; reports linking repeated recurrence of atypical cellular tumors with dedifferentiation come from the wider literature and remain uncertain for benign variants [ 6 , 8 ]. We therefore raise it only as a reason for long‐term follow‐up in selected patients, not as an established consequence of the recurrences observed in this study. This study has several limitations. It is a single‐centre, retrospective analysis, and although 248 cases is a reasonable size for an uncommon disease, selection and information bias cannot be excluded and the validation set ( n  = 75) is modest. Excluding patients with residual disease at 3 months, adopted to separate recurrence from residual tumor, may remove some incompletely resected and potentially higher‐risk cases and could bias the observed association between tumor number and recurrence; this should be borne in mind when interpreting the results. Recurrence was ascertained by ultrasound, which is operator‐dependent and less sensitive for subserosal lesions, so some events may have been missed. The model uses a fixed 12‐month binary outcome rather than a time‐to‐event analysis; a Kaplan–Meier/Cox approach would use follow‐up duration more fully and is a natural next step. Potentially informative variables—surgical margin status, mitotic activity and proliferation markers such as Ki‐67—were not included because they depend on postoperative pathology and are unavailable preoperatively. Finally, there was no external validation. Multi‐centre data, longer follow‐up with time‐to‐event modeling and the addition of radiomic or molecular features are needed to confirm and extend these findings. Special‐type USMN carries a substantial early recurrence risk. Intramural location, high T 2 WI signal, multiple tumors, a maximum diameter ≥ 6 cm, elevated BMI and coexisting adenomyosis were independently associated with recurrence, and a GBM model built on these six preoperative variables discriminated recurrence reasonably well in an independent validation set. Interpreted with SHAP, the model may help identify high‐risk patients before surgery and support individualized counseling and follow‐up, pending external validation.

Informed consent was waived by the Ethics Committee of the Third Xiangya Hospital of Central South University because of the retrospective design of the study, and no patient‐identifiable data are included in this article.

Unusual smooth muscle neoplasia (USMN), also referred to as special‐type uterine leiomyoma, denotes a group of smooth muscle tumors whose histological appearance or growth pattern differs from that of the ordinary leiomyoma. Within the 2020 WHO classification of female genital tumors these lesions are recognized as variants of leiomyoma, and the commonly encountered types are cellular, atypical (bizarre/symplastic) and vascular leiomyoma, with epithelioid leiomyoma, lipoleiomyoma, intravenous leiomyomatosis and benign metastasizing leiomyoma being far less frequent [ 1 ]. They are reported to account for roughly 0.03% to 5% of all uterine leiomyomas [ 2 , 3 ]. Although these variants are formally benign, their cells proliferate actively and may show mild atypia and variable mitotic activity [ 4 ], and the transcriptional profile of some types overlaps with that of leiomyosarcoma [ 5 ]. Local recurrence after conservative surgery is therefore a recognized concern [ 6 , 7 ], and for cellular leiomyoma with marked atypia repeated recurrence has occasionally been linked with dedifferentiation over long follow‐up [ 6 , 8 ]. With the growing adoption of minimally invasive techniques and shifts in national fertility policy, a rising number of reproductive‐age women seek to preserve their uterus and maintain fertility [ 9 ]. Myomectomy—laparoscopic myomectomy in particular—has emerged as a mainstay of treatment, yet it does not eliminate the risk of postoperative recurrence; this concern is especially pronounced for special‐type variants, whose biological behavior is less predictable and for which uterine‐sparing surgery may leave residual lesions. Currently, there are few tools specifically designed to predict recurrence in this patient population, as most published studies focus on ordinary leiomyoma, and the limited research on single variants such as cellular leiomyoma rarely goes beyond describing associated risk factors [ 10 ], or relies on postoperative pathological markers (e.g., Ki‐67) that are unavailable during preoperative surgical planning. Conventional logistic and Cox models perform well at capturing linear associations, but are less adept at handling non‐linear effects and variable interactions. To address this gap, we analyzed 248 patients with special‐type USMN, screened for preoperative risk factors, compared the performance of eight machine learning algorithms, and applied SHAP analysis to add interpretability to the best‐performing model, with the aim of providing a practical reference for preoperative counseling and postoperative follow‐up.

The authors declare no conflicts of interest.

This was a single‐centre, retrospective study. Using the gynecology database of the Third Xiangya Hospital of Central South University, we identified patients who underwent myomectomy for a uterine mass between January 2015 and December 2024 and whose paraffin sections confirmed a special‐type uterine smooth muscle tumor. Inclusion criteria: (1) surgical treatment in our department with a postoperative paraffin diagnosis of special‐type USMN; (2) complete clinical data; and (3) no residual myoma on early postoperative imaging. Exclusion criteria: (1) other malignancy; (2) severe concurrent medical disease; (3) no surgical treatment; and (4) residual myoma still visible on the early (within 3 months) postoperative ultrasound. The last criterion was applied so that the baseline for follow‐up was a uterus without demonstrable residual disease; consequently, a lesion found later was more likely to represent genuine recurrence than persistent residual tumor. We recognize that this choice may exclude some incompletely resected, potentially higher‐risk cases, and its possible effect is considered in the Discussion. A total of 285 patients were screened. Thirty‐seven were excluded for a follow‐up shorter than 12 months or for more than 30% missing data, leaving 248 patients for analysis. The study was approved by the Ethics Committee of the Third Xiangya Hospital of Central South University (approval no. Kuai 26266); given the retrospective design, the requirement for informed consent was waived. The diagnosis rested on the histological report of the postoperative paraffin sections, interpreted in line with the 2020 WHO classification of leiomyoma variants. Three subtypes were included: (1) cellular leiomyoma—densely packed, relatively uniform smooth muscle cells without significant atypia and 0–4 mitoses per 10 high‐power fields (HPF); (2) vascular leiomyoma—tumor cells of variable shape with enlarged, hyperchromatic nuclei and occasional multinucleated or lobulated forms but preserved differentiation, and fewer than 2 mitoses/10 HPF; and (3) atypical (bizarre) leiomyoma—pleomorphic cells with moderate to severe atypia, enlarged hyperchromatic nuclei and occasional multinucleation but overall good differentiation, and fewer than 2 mitoses/10 HPF. Epithelioid leiomyoma ( n  = 2), lipoleiomyoma ( n  = 2) and intravenous leiomyomatosis ( n  = 1) were too few for meaningful analysis and were not included; benign metastasizing leiomyoma was not present in this cohort. Recurrence was defined as a uterine or extrauterine smooth‐muscle lesion ≥ 1 cm detected by ultrasound within 12 months of surgery, in a patient whose transvaginal ultrasound at 3 months had been negative [ 11 ]. All patients had at least 12 months of follow‐up, so recurrence status was assessed at a common, fixed 12‐month landmark; this uniform observation window was adopted to keep the binary outcome comparable across patients despite the retrospective design. Twenty‐one preoperative clinical and imaging variables were retrieved, including age, age at menarche, gravidity, parity, infertility, previous caesarean section, abnormal uterine bleeding, BMI, coexisting endometriosis, adenomyosis and benign adnexal masses, maximum tumor diameter, tumor number, tumor location, surgical approach, use of a morcellator, preoperative hemoglobin, CA125, LDH, T 2 WI signal and histological subtype. Patients were followed by telephone and outpatient review; the follow‐up cut‐off was 31 May 2025. Analyses were performed in SPSS 26.0 and R 4.5.1. Categorical variables were compared with the chi‐square or Fisher exact test and continuous variables with the appropriate parametric or non‐parametric test; p  < 0.05 was considered significant. As a sensitivity analysis of histological heterogeneity, recurrence rates were compared across the three subtypes. Because surgical approach is chosen largely on the basis of tumor characteristics, its crude comparison is confounded by indication. We therefore estimated a propensity score from 13 covariates and applied inverse probability of treatment weighting (IPTW), truncating extreme weights at the 99th percentile and checking balance with the standardized mean difference (SMD < 0.1 considered balanced); a weighted logistic model with laparoscopy as reference was then fitted. Pairwise 1:1 nearest‐neighbor matching was performed as an additional check. For the prediction model, the cohort was split 7:3 into training ( n  = 173) and validation ( n  = 75) sets by stratified sampling. To avoid information leakage, imputation, standardization, cut‐off determination and feature selection were carried out in the training set only and then applied to the validation set. Variables with more than 30% missingness were dropped and the remainder imputed by random forest [ 12 ]. Continuous variables were standardized ( z ‐score) using training‐set parameters. Candidate predictors were screened by LASSO regression with 10‐fold cross‐validation, and the variables retained at the penalty minimizing cross‐validated deviance were entered into a multivariable logistic regression to define the independent factors; collinearity was checked by the variance inflation factor (VIF). Using the selected features, eight algorithms (LR, SVM, RF, GBM [ 13 ], XGBoost, KNN, AdaBoost, NN) were trained with grid‐search hyper‐parameter tuning and internal cross‐validation. Discrimination (AUC, accuracy, precision, sensitivity, specificity, F1), calibration and decision‐curve analysis (DCA) were assessed in the validation set, and the best model was interpreted globally and individually with SHAP [ 14 ].

Table S1: Comparison of baseline characteristics between the training and validation sets. Table S2: Performance of the eight machine‐learning models in the training set. Figure S1: ROC and calibration curves of the eight machine‐learning models in the training set. Figure S2: Confusion matrices of the GBM model in the training and validation sets (probability threshold 0.40). Figure S3: SHAP waterfall plot for a representative recurrent case (GBM).