Tamoxifen (TAM) is one of the most important drugs in breast cancer treatment. For premenopausal women with estrogen receptor-positive breast cancer, it is a cornerstone of hormone therapy that dramatically reduces recurrence risk and saves lives. Millions of women take it for 5-10 years as part of their breast cancer management.
However, tamoxifen has a well-documented side effect: it acts as a weak estrogen in the uterus. While it blocks estrogen receptors in breast tissue, it stimulates endometrial (uterine lining) growth in the uterus. Long-term use is associated with a 2-5 fold increased risk of endometrial abnormalities ranging from benign polyps to precancerous hyperplasia to endometrial carcinoma. This creates a genuine clinical dilemma: how frequently should patients be monitored for uterine changes, and in whom?
This study analyzed 224 premenopausal breast cancer patients on tamoxifen to identify which factors predict development of endometrial lesions (abnormal uterine changes detectable on ultrasound), with the goal of building a machine learning prediction model that clinicians could use to personalize monitoring intensity.
224 premenopausal breast cancer patients who had received tamoxifen were enrolled in this retrospective study. All patients had regular gynecologic ultrasound follow-up. The outcome was whether any endometrial lesion (polyp, hyperplasia, or carcinoma) had developed during follow-up. 79 of the 224 patients (35%) developed endometrial lesions.
Candidate predictor variables included demographic factors (age, BMI), tamoxifen duration, endometrial thickness on ultrasound, specific ultrasound characteristics of the endometrium, menstrual changes (colporrhagia - abnormal vaginal bleeding), and reproductive history. Three machine learning methods were tested: LASSO logistic regression, stepwise logistic regression, and a random forest model.
Each model was evaluated on a randomly held-out test set using AUC, C-index, calibration analysis, and decision curve analysis. The final model was presented as a nomogram - a visual tool that converts individual patient measurements into a probability score, formatted for practical bedside use without requiring a computer.
The LASSO logistic regression approach outperformed both stepwise regression and random forest, achieving a C-index of 0.874 and AUC of 0.891 in the test set. This is a strong performance for a clinical prediction model built from routinely available variables - comparable to some validated tools used in other cancer settings.
Four predictors were selected by the model as the most informative: ultrasound characteristics of the endometrium (specific texture or structure features visible on imaging), duration of tamoxifen use, colporrhagia (abnormal vaginal bleeding or discharge), and endometrial thickness. A threshold of 0.825 cm for endometrial thickness emerged as a key cutoff - above this value, risk increases substantially.
Calibration analysis showed good agreement between the model's predicted probabilities and actual observed lesion rates. Decision curve analysis confirmed the nomogram provides real clinical benefit compared to either monitoring all patients or monitoring none - it correctly identifies the high-risk group that benefits most from intensive surveillance.
The practical value of this nomogram is stratifying patients by risk before deciding on monitoring frequency. Currently, many guidelines recommend either routine annual ultrasound for all tamoxifen users or monitoring only when symptoms arise. Both approaches have limitations: routine screening in all patients leads to anxiety and costly procedures in low-risk women, while symptom-only monitoring misses asymptomatic endometrial lesions.
With this nomogram, clinicians could potentially identify the highest-risk patients - those with abnormal ultrasound features, longer tamoxifen use, bleeding symptoms, and thicker endometrium - and offer them more intensive monitoring (more frequent ultrasound, or earlier referral for uterine biopsy). Lower-risk patients could be reassured and monitored less frequently, reducing unnecessary interventions.
The 0.825 cm endometrial thickness threshold deserves specific attention. Current clinical guidelines often use 5 mm as the threshold for further investigation, but this study suggests that in premenopausal tamoxifen users, 8.25 mm may be a more relevant clinical cutoff when combined with other risk factors. Prospective validation of this specific threshold would be clinically important.
This is a relatively small single-center retrospective study from China. The patient population, tamoxifen regimens, and gynecologic monitoring protocols may differ from those in other countries, affecting how broadly the results generalize. External validation in European or North American breast cancer cohorts would be needed before clinical adoption.
The 35% rate of endometrial lesions in this cohort is somewhat higher than rates reported in some Western studies, possibly due to differences in patient age distribution, tamoxifen dose, or duration of follow-up. This higher event rate may make the model appear more discriminating than it would be in a population with lower baseline risk.
Despite these limitations, the study addresses a genuinely important clinical problem. The four-factor model is simple enough for practical use, the performance is strong, and the nomogram format makes it accessible to clinicians without specialized AI tools. A prospective multicenter validation study is the logical next step, ideally enrolling patients at the start of tamoxifen therapy and following them longitudinally.