Development of a risk prediction model for second primary thyroid cancer in female breast cancer patients based on machine learning algorithms

Abstract

Female breast cancer patients remain at risk of developing additional primary cancers even following surgical treatment. Second primary thyroid cancer (SPTC) is the most common type of multiple primary cancer (MPC) in this demographic during their survivorship period, significantly burdening patients’ quality of life. This study aimed to develop a risk prediction model for the occurrence of SPTC in postoperative female breast cancer patients using various machine learning algorithms. Key risk factors were progressively identified utilizing a literature review, the Delphi method, and the LASSO algorithm. A retrospective cohort of 684 breast cancer patients (228 in the SPTC group and 456 in the non-SPTC group) from 2014 to 2019 was used to construct and internally validate the optimal prediction model for postoperative SPTC risk. An independent cohort of 110 breast cancer patients treated from 2021 to 2024 was collected for temporal-split validation. Five machine learning algorithms—Logistic Regression, K-Nearest Neighbors (KNN), Decision Tree, Support Vector Machine (SVM), and Naive Bayes—were compared. The Decision Tree model demonstrated superior performance on both the training and test sets, with AUC values of 0.915 (95% CI: 0.891–0.939) and 0.909 (95% CI: 0.879–0.939), alongside recall rates of 0.80 for both. On the temporal-split validation set, it achieved an AUC of 0.861 (95% CI: 0.802–0.920) and a recall of 0.77. Calibration curve analysis revealed that the Decision Tree model achieved the lowest Log loss value, indicating high consistency between the predicted and observed probabilities of SPTC. Interpretability analysis using the SHAP framework revealed that BMI ≥ 30 kg/m², anxious or depressive psychological state, history of thyroid disease, ER positivity, and PR positivity positively influenced the model output. Decision tree models can effectively assist in evaluating the risk of SPTC in postoperative female breast cancer patients, providing a valuable reference for the formulation of personalized survivorship monitoring strategies.

Introduction

Recent epidemiological data indicate that breast cancer incidence remains consistently high worldwide, with the number of new cases rising each year¹. It remains the most frequently diagnosed malignancy and a leading cause of cancer-related mortality among women globally. In China, breast cancer is the most commonly diagnosed malignancy in women, accounting for approximately 19.9% of all new female cancer cases². With continuous advances and innovations in medical technology, the range of effective treatment options for female breast cancer patients has expanded considerably, resulting in markedly prolonged survival and improved clinical outcomes³. However, this progress also elevates the likelihood of developing multiple primary cancers (MPC), also known as second primary cancers (SPC). Studies report that among female breast cancer patients who have undergone surgery, the cumulative incidence of MPC is approximately 7.43% at 10 years, 14.41% at 15 years, and 20.08% at 20 years⁴. The incidence of MPC increases over time, adversely affecting patients physically, psychologically, socially, and economically, thereby placing a serious burden on their quality of life.

Second primary thyroid cancer (SPTC) is the most prevalent type of MPC occurring among female breast cancer survivors⁵. While there is no universally accepted definition of SPTC, it generally refers to the sequential or simultaneous development of primary thyroid cancer in an individual already diagnosed with a primary malignancy. Each cancer must be histologically confirmed as malignant, and the second tumor must not be a metastasis of the original malignancy⁶. Research indicates that female breast cancer patients face a 1.29-fold increased risk of developing SPTC compared to the general population⁷. Following an SPTC diagnosis, the 5-, 10-, and 15-year overall survival rates decrease to 88.34%, 64.42%, and 54.66%, respectively⁸. This not only impacts patient prognosis but also escalates economic and family caregiving burdens⁹. At present, numerous studies have shown a bidirectional association between breast cancer and thyroid cancer, although the specific mechanisms remain unclear^10,11. The potential risk factors associated with the co-occurrence of breast and thyroid cancers include pathological grade and tumor stage, postoperative radiotherapy, estrogen receptor (ER) status, progesterone receptor (PR) status, human epidermal growth factor receptor-2 (HER-2) status, family history of malignancy, and age at diagnosis^12–14.

Given the profound clinical implications of a second malignancy, the early identification of risk factors for SPTC in breast cancer patients is of crucial importance. A review of the literature reveals that current clinical assessment tools primarily evaluate the risk of a single primary cancer. There remains a lack of specialized assessment tools and corresponding predictive models for MPC or SPTC in female breast cancer patients. Among the earlier developed and more widely used breast cancer risk prediction models in the United States are the Gail model¹⁵ and the Claus model¹⁶. It is evident that the Claus model demonstrates significantly lower predictive accuracy than the Gail model. Although the Gail model performs relatively well, it still has certain limitations, such as incomplete inclusion of risk factors and limited generalizability across different ethnic groups. To address these issues, Chinese researchers have developed a breast cancer risk prediction model tailored to the characteristics of the Chinese population¹⁷. Compared to the Gail model, this model shows better predictive capability and accuracy¹⁸. “Nevertheless, models designed solely for initial breast cancer risk are insufficient to accurately assess a patient’s risk of developing SPTC after surgery. Therefore, there is a need to develop a specific model to support more refined risk assessment in clinical practice.

“In recent years, the landscape of artificial intelligence in bioinformatics and clinical prediction modeling has evolved rapidly. Beyond traditional machine learning, deep learning models have demonstrated strong capabilities in medical image analysis and genomic sequence interpretation by automatically learning hierarchical features of high-dimensional data^19,20. Although these advanced models demonstrate significant advantages in discriminative performance, their internal decision-making processes are often complex and difficult to interpret (i.e., the ‘black box’ characteristic), which poses challenges in clinical decision-making environments that require transparent decisions and traceable logic. This study aims to establish an evaluative framework that assists clinicians in rapid initial risk screening. Therefore, in model selection, we prioritize interpretability, computational efficiency, and robustness on structured clinical data, so that clinicians can intuitively understand the logic of risk decision-making, thereby building trust and guiding individualized monitoring. Accordingly, we selected five well-established, widely validated traditional machine learning algorithms with clear theoretical foundations for systematic development and comparison: logistic regression, K-nearest neighbor (KNN), decision tree, support vector machine (SVM), and Naive Bayes. Combining machine learning with SHAP model interpretability analysis helps us identify key risk factors for the occurrence of SPTC in postoperative female breast cancer patients, providing scientific guidance for early risk monitoring of patients, as well as a scientific basis for constructing a predictive model directly applicable to clinical evaluation.

Methods

Design and participants

This retrospective study included breast cancer patients diagnosed at the Affiliated Cancer Hospital of Xinjiang Medical University between January 2014 and January 2019. The inclusion criteria were as follows: (1) age ≥ 18 years at diagnosis; (2) female sex; (3) histopathological confirmation of breast cancer as the first primary malignancy; and (4) underwent surgical resection after diagnosis. The exclusion criteria were: (1) development of second primary cancers other than thyroid cancer; (2) thyroid metastases originating from breast cancer; (3) an interval of less than 6 months between the diagnoses of the two primary tumors; and (4) incomplete clinical data or an ambiguous pathological diagnosis. Based on the time interval between the diagnoses of two primary malignancies, it is generally accepted that an interval of ≤ 6 months between diagnoses defines synchronous multiple primary tumors, while an interval of > 6 months defines metachronous multiple primary tumors²¹. Because this study focuses on metachronous SPTC following an initial breast cancer diagnosis, we restricted our cohort to cases with an interval of > 6 months between the diagnoses of the two primary tumors.

We conducted a retrospective case-control study, dividing participants into a case group (SPTC group) and a control group (non-SPTC group), based on whether they developed SPTC within 6 months to 3 years after their initial breast cancer diagnosis. The specific patient selection process is detailed in Fig. 1. Ultimately, the study included 684 breast cancer patients, with 228 in the SPTC group and 456 in the non-SPTC group. The cohort of 684 patients was randomly partitioned into a training set (n = 547) and a test set (n = 137) at an 8:2 ratio using Python. This study strictly adhered to the TRIPOD reporting guidelines for multivariable prediction models for individual prognosis or diagnosis. The study protocol was approved by the Institutional Review Board and exempted from the requirement for informed consent (approval No. K-2023048), and all procedures complied with relevant ethical guidelines.

Fig. 1.

Flowchart of the study inclusion process.

Determine variables and data collection

Our research team initially identified potential risk factors for second primary thyroid cancer (SPTC) in female breast cancer patients through a systematic literature search. Subsequently, we designed an expert survey questionnaire and conducted two rounds of Delphi correspondence to refine the variable set. Following group discussions to finalize the study variables, a standardized data collection form was developed. Clinical data were then extracted by reviewing electronic medical records. The collected variables were organized into two sections. Section  1: Basic Information, consisting of eight variables: age at diagnosis (categorized as 18–40, 41–60, or > 60 years), body mass index (BMI < 25 kg/m²: normal/underweight; 25 ≤ BMI < 30 kg/m²: overweight; BMI ≥ 30 kg/m²: obese), age at menarche (≤ 12 years, > 12 years), menopausal status, family history of malignancy, history of diabetes, anxiety/depression, and pre-existing thyroid diseases (including hyperthyroidism, hypothyroidism, thyroiditis, and thyroid nodular lesions). Section  2: Tumor Pathological Information, comprising nine variables: location of the largest tumor (left or right breast), quadrant of the largest lesion (outer, inner, or other), maximum tumor diameter (< 2 cm, 2–5 cm, or ≥ 5 cm), pathological grade (Ⅰ, Ⅱ, or Ⅲ), pathological type (invasive ductal carcinoma or non-invasive ductal carcinoma), estrogen receptor (ER) status (positive/negative), progesterone receptor (PR) status (positive/negative), human epidermal growth factor receptor‑2 (HER‑2) status (positive/negative), and lymph node metastasis (yes/no). Pre-existing thyroid disease was defined as the presence of thyroid dysfunction (e.g., hyperthyroidism or hypothyroidism indicated by abnormal laboratory hormone levels) or solid thyroid lesions (e.g., nodules or goiter identified via imaging) detected between the initial breast cancer diagnosis and surgical resection. In the basic information section, anxiety/depression were assessed using the Generalized Anxiety Disorder-7 (GAD-7) scale and the Patient Health Questionnaire-9 (PHQ-9).

To identify the most informative variables, we utilized LASSO regression with L1 regularization. This approach is particularly efficient for variable selection, as it automatically drives the coefficients of non-contributing variables to zero, retaining only the key risk factors in the final model.

Diagnostic criteria for SPTC

SPTC was diagnosed based on histopathological confirmation of distinct thyroid cancer characteristics, differing entirely from the primary breast tumor in both location and nature. Furthermore, metastasis or recurrence of the primary breast malignancy within the thyroid gland had to be definitively ruled out. To ensure diagnostic accuracy, all cases of SPTC underwent independent double-review by specialized pathologists within the study team.

Data preprocessing

To address missing values inherent in real-world clinical data, records with significant missing data were excluded, and variables with fewer missing values were imputed using the mode or replaced with fixed values. Based on reference ranges, guidelines, and consensus, all continuous variables were converted to categorical variables to eliminate the influence of varying feature scales. To facilitate the model’s ability to learn class-specific patterns and enhance overall interpretability, one-hot encoding was applied to all categorical features.

Model development and validation

The dataset was randomly split into a training set and a test set at an 8:2 ratio. These data were used to build and evaluate five machine learning models: Logistic Regression, k-Nearest Neighbors (KNN), Decision Tree, Support Vector Machine (SVM), and Naive Bayes. To mitigate overfitting, ten-fold cross-validation was applied on the training set. Given the constrained parameter space in this study, a grid search strategy was employed to systematically traverse predefined candidate ranges and identify the optimal hyperparameters for each model. For temporal-split validation, a temporal cohort of 110 female breast cancer patients meeting the same inclusion and exclusion criteria was collected from January 2021 to January 2024. This independent dataset was used to assess the model’s generalizability and predictive performance on unseen future data.

Model performance was evaluated using accuracy, recall (sensitivity), the area under the receiver operating characteristic curve (AUC), and calibration curves. The AUC and recall rate were designated as the primary evaluation metrics. An AUC value closer to 1 indicates superior model performance, signifying that the model’s predicted probabilities for positive cases are consistently higher than for negative cases. Recall measures the proportion of actual positive cases correctly identified by the model. We prioritized screening sensitivity to maximize the detection of high-risk patients who might develop postoperative SPTC, thereby minimizing clinically significant missed diagnoses. A calibration curve plots the actual observed event rate against the model’s predicted probability. A curve closer to the ideal diagonal, accompanied by a lower Log loss value (approaching 0), indicates better calibration and higher predictive accuracy. Finally, we evaluated the robustness of the models across different breast cancer molecular subtypes and employed the SHAP framework to analyze model interpretability.

Statistical analysis

Basic data analysis was performed with R software (version 4.3.1), and machine learning modeling was conducted using Python (version 3.8.1). Categorical variables were expressed as frequencies and percentages (%), and differences between groups were compared using the Chi-square (X²) test was used for inter-group comparisons. P-values < 0.05 were considered statistically significant.

Results

Baseline characteristics

The study included 228 patients in the SPTC group and 456 in the non-SPTC group. In the overall cohort, the majority of breast cancer patients (72.08%) were aged 41–60 years. Most patients were classified as overweight (30.56%) or obese (38.01%), experienced menarche after age 12, were premenopausal (65.06%), and had no family history of malignancy (79.82%) or personal history of diabetes (95.32%). Among all the baseline and clinicopathological variables evaluated, no significant differences (p > 0.05) were observed between the training and testing sets, except for age at diagnosis (p = 0.013), as shown in Table 1.

Table 1.

Comparison between groups of training set and test set.

Variables	Total sample size (N = 684)	The training set (n = 547)	The test set (n = 137)	χ2-value	P-value
1. Basic information
Age at diagnosis				8.74	0.013
18–40 years	96 (14.04)	25 (18.25)	71 (12.98)
41–60 years	493 (72.08)	85 (62.04)	408 (74.59)
> 60 years	95 (13.89)	27 (19.71)	68 (12.43)
BMI				3.10	0.212
< 25 kg/m2	215 (31.43)	35 (25.55)	180 (32.91)
≥ 25 kg/m2,< 30 kg/m2	209 (30.56)	48 (35.04)	161 (29.43)
≥ 30 kg/m2	260 (38.01)	54 (39.42)	206 (37.66)
Age at menarche				0.41	0.520
≤ 12 years	146 (21.35)	32 (23.36)	114 (20.84)
> 12 years	538 (78.65)	105 (76.64)	433 (79.16)
Menopause				0.18	0.670
Yes	239 (34.94)	50 (36.50)	189 (34.55)
No	445 (65.06)	87 (63.50)	358 (65.45)
Family history of malignant tumors				χ²=0.40	0.530
Yes	138 (20.18)	25 (18.25)	113 (20.66)
No	546 (79.82)	112 (81.75)	434 (79.34)
History of diabetes				0.07	0.789
Yes	32 (4.68)	7 (5.11)	25 (4.57)
No	652 (95.32)	130 (94.89)	522 (95.43)
Pre-existing thyroid diseases				2.73	0.098
Yes	402 (58.77)	72 (52.55)	330 (60.33)
No	282 (41.23)	65 (47.45)	217 (39.67)
Anxiety/depression				0.40	0.528
Yes	358 (52.34)	75 (54.74)	283 (51.74)
No	326 (47.66)	62 (45.26)	264 (48.26)
2. Tumor pathological information
The largest tumor site				0.75	0.386
Left breast	377 (55.12)	71 (51.82)	306 (55.94)
Right breast	307 (44.88)	66 (48.18)	241 (44.06)
Tumor lesion quadrant				0.89	0.642
Outer quadrant	447 (65.35)	90 (65.69)	357 (65.27)
Inner quadrant	163 (23.83)	35 (25.55)	128 (23.40)
Others	74 (10.82)	12 (8.76)	62 (11.33)
Tumor maximum diameter				3.26	0.196
< 2 cm	207 (30.26)	38 (27.74)	169 (30.90)
2–5 cm	446 (65.20)	89 (64.96)	357 (65.27)
≥ 5 cm	31 (4.53)	10 (7.30)	21 (3.84)
Pathological grade				0.97	0.616
Grade I	36 (5.26)	6 (4.38)	30 (5.48)
Grade II	469 (68.57)	91 (66.42)	378 (69.10)
Grade III	179 (26.17)	40 (29.20)	139 (25.41)
Tumor pathological type				1.64	0.200
Invasive ductal carcinoma	601 (87.87)	116 (84.67)	485 (88.67)
Non-invasive ductal carcinoma	83 (12.13)	21 (15.33)	62 (11.33)
ER status				0.01	0.925
Positive	447 (65.35)	90 (65.69)	357 (65.27)
Negative	237 (34.65)	47 (34.31)	190 (34.73)
PR status				0.17	0.676
Positive	454 (66.37)	93 (67.88)	361 (66.00)
Negative	230 (33.63)	44 (32.12)	186 (34.00)
Her−2 status				0.09	0.762
Positive	163 (23.83)	34 (24.82)	129 (23.58)
Negative	521 (76.17)	103 (75.18)	418 (76.42)
Lymph node metastasis				2.06	0.151
Yes	256 (37.43)	44 (32.12)	212 (38.76)
No	428 (62.57)	93 (67.88)	335 (61.24)

Feature screening

The initial variable screening was conducted through two rounds of Delphi expert consultation, administered on August 20, 2024, and September 15, 2024, respectively. A total of 15 experts and professors from leading hospitals and medical schools across multiple regions participated. All experts had extensive experience in oncology, including eight nursing specialists in breast tumor care, six clinical oncologists, and one researcher specializing specializing in genetic research on breast cancer immunotherapy. The panel had a mean age of 46.45 ± 4.93 years (range: 38–54 years) and a mean professional experience of 22.33 ± 6.87 years (range: 14–32 years). The expert response rates of the experts in the two rounds were 86.67% and 100%, respectively. The average expert authority coefficients were 0.86 and 0.87, respectively. The coefficient of concordance for expert opinions exceeded 0.7 in both rounds, with details regarding the degree of expert consensus provided in Table 2.

Table 2.

Expert opinion concentration in two rounds.

Variables	First round of expert consultation				Second round of expert consultation
	M	SD (±)	CV	Full score ratio (%)	M	SD (±)	CV	Full score ratio (%)
1. Basic information
Age at diagnosis	5.00	0.00	0.00	100.00%	5.00	0.00	0.00	100.00%
Educational level	1.67	0.49	0.29	0.00%	-	-	-	-
Occupation types	3.07	0.80	0.26	0.00%	-	-	-	-
Marital Status	1.67	0.49	0.29	0.00%	-	-	-	-
BMI	4.20	0.56	0.13	26.67%	4.40	0.51	0.12	40.00%
Age at menarche	3.73	0.60	0.16	6.67%	3.87	0.52	0.13	6.67%
Menopause	3.20	0.56	0.18	26.67%	4.07	0.70	0.17	26.67%
Family history of malignant tumors	5.00	0.00	0.00	100.00%	5.00	0.00	0.00	100.00%
Diabetes	4.07	0.70	0.17	26.67%	4.07	0.70	0.17	26.67%
Pre-existing thyroid diseases	5.00	0.00	0.00	100.00%	5.00	0.00	0.00	100.00%
Anxiety/Depression	4.07	0.70	0.17	100.00%	4.47	0.52	0.12	46.67%
2. Tumor pathological information
The largest tumor site	3.60	0.51	0.14	6.67%	3.73	0.74	0.20	13.33%
Tumor lesion quadrant	3.60	0.51	0.14	6.67%	3.73	0.74	0.20	13.33%
Tumor maximum diameter	4.07	0.70	0.17	26.67%	4.27	0.46	0.11	26.67%
Pathological grade	4.27	0.59	0.14	33.33%	4.47	0.52	0.12	46.67%
Tumor pathological type	4.67	0.49	0.10	66.67%	4.73	0.458	0.10	73.33%
ER status	5.00	0.00	0.00	100.00%	5.00	0.00	0.00	100.00%
PR status	5.00	0.00	0.00	100.00%	5.00	0.00	0.00	100.00%
Her-2 status	5.00	0.00	0.00	100.00%	5.00	0.00	0.00	100.00%
Lymph node metastasis	3.93	0.59	0.15	13.33%	4.07	0.46	0.11	13.33%

M: mean; SD: standard deviation; CV: Coefficient of Variation.

Secondary variable screening was performed using the LASSO algorithm, which applies L1 regularization to shrink coefficients toward zero. At the optimal λ value, five variables with non-zero coefficients were selected from 17 variables as the influencing factors of the occurrence of SPTC: BMI, pre-existing thyroid disease, anxiety/depression, ER status, and PR status. As shown in Fig. 2.

Fig. 2.

Lasso Algorithm for Variable Selection. A Select the tuning parameter (λ) in the Lasso model using 10-fold cross-validation based on one standard error of the minimum criterion, and the λ value that gives the minimum binomial average deviation is used to select features; B draw a vertical line at the value selected by 5-fold cross-validation, where the optimal result is 5 non-zero coefficients.

Machine learning model

For the development and evaluation of the machine learning models, 547 female breast cancer patients were allocated to the training set and 137 to the testing set. In the training set, the AUC values for the training set in logistic regression, KNN, decision tree, SVM, and Naive Bayes models were 0.853 (95%CI: 0.821–0.885), 0.733 (95%CI: 0.693–0.773), 0.915 (95%CI: 0.891–0.939), 0.793 (95%CI: 0.758–0.828), and 0.743 (95%CI:0.704–0.782), respectively (Fig. 3A). Correspondingly, the AUC values in the test set were 0.844 (95%CI: 0.801–0.887), 0.734 (95%CI: 0.679–0.789), 0.909 (95%CI: 0.879–0.939), 0.762 (95%CI: 0.713–0.811), and 0.736 (95%CI: 0.683–0.789), respectively (Fig. 3B). The calibration curves for the five models revealed that the decision tree model yielded the lowest log-loss value (0.314), with its curve aligning most closely with the ideal diagonal, indicating excellent calibration and model fit (Fig. 3C); The recall rates for the decision tree model in the training set, test set, and temporal-split validation set were 0.8, 0.8, and 0.77, respectively (Table 3). Based on a comprehensive evaluation of the AUC values, recall rates, and calibration metrics across the five models, the decision tree model demonstrated superior overall performance.

Fig. 3.

The effectiveness of building models with five types of machine learning. A ROC curve of the training set; B ROC curve of the test set; C Calibration curve of the model; D Analysis of important features in the Decision Tree.

Table 3.

Evaluation of the efficacy of SPTC risk prediction models under five machine learning algorithms.

Machine learning algorithms	Data set	AUC value	Accuracy	Precision	Recall rate	F1 Score
Logistic regression	Training set	0.853 (0.821–0.885)	0.78	0.76	0.75	0.755
	Test set	0.844 (0.801–0.887)	0.77	0.76	0.74	0.750
KNN	Training set	0.733 (0.693–0.773)	0.68	0.65	0.66	0.655
	Test set	0.734 (0.679–0.789)	0.67	0.64	0.65	0.645
Decision tree	Training set	0.915(0.891–0.939)	0.83	0.82	0.80	0.810
	Test set	0.909(0.879–0.939)	0.82	0.81	0.80	0.805
	Temporal-split validation set	0.861 (0.802–0.920)	0.78	0.76	0.77	0.765
SVM	Training set	0.793 (0.758–0.828)	0.72	0.70	0.69	0.695
	Test set	0.762 (0.713–0.811)	0.70	0.68	0.66	0.670
Naive Bayes	Training set	0.743 (0.704–0.782)	0.66	0.67	0.66	0.665
	Test set	0.736(0.683–0.789)	0.68	0.65	0.66	0.655

Subgroup analysis

To evaluate the robustness of the prediction model across different molecular subtypes of breast cancer, we conducted a pre-specified subgroup analysis in the external validation cohort. The results demonstrated that the model achieved the highest discriminative ability in patients with luminal-type breast cancer (AUC: 0.751, 95% CI: 0.661–0.841), whereas its performance declined in the triple-negative subgroup (AUC: 0.610, 95% CI: 0.508–0.712) (Table 4). This trend may be attributable to the absence of hormone receptor expression rates and the distinct tumor biological behavior characteristic of triple‑negative breast cancer. However, owing to the limited sample size of the triple‑negative subgroup (n = 16), these findings should be regarded as exploratory. Future studies with larger, molecularly balanced cohorts are warranted to further validate the generalizability of the model across diverse subtypes.

Table 4.

Evaluation of model prediction performance across different molecular subtypes of breast cancer.

Subgroup	n	AUC value (95% CI)	Accuracy	Precision	Recall rate	F1 score
Luminal-type	83	0.751 (0.661, 0.841)	0.771	0.795	0.722	0.757
HER-2-positive	11	0.648 (0.530, 0.766)	0.636	0.625	0.833	0.714
Triple-negative	16	0.610 (0.508, 0.712)	0.625	0.636	0.700	0.667
Overall	110	0.719 (0.650, 0.788)	0.718	0.731	0.709	0.720

SHAP interpretability analysis

The SHAP plot (Fig. 4A) illustrates that the top five features contributing to the model, ranked in descending order of importance, are BMI, anxiety/depression, pre-existing thyroid diseases, ER status, and PR status. Specifically, a BMI ≥ 30 kg/m², the presence of depressive or anxious psychological states, a history of thyroid disease, and positive ER/PR status positively drive the model’s risk predictions. In other words, the presence or positive encoding of these five features correlates directly with an increased risk of SPTC. Notably, feature importance analysis in the decision tree model confirms that BMI is the dominant predictor, making the greatest overall contribution (see Fig. 3D). Figure 4B presents a local SHAP plot for an individual patient who developed SPTC, demonstrating that a BMI ≥ 30 kg/m², pre-existing thyroid disease, anxiety/depression, and PR positivity are the primary factors driving the positive risk prediction. Conversely, Fig. 4C displays the SHAP plot for a patient who did not develop SPTC. For this individual, the absence of anxiety or depression, the lack of pre-existing thyroid disease, and ER negativity yield negative SHAP values, effectively lowering the predicted risk and serving as key contributors to the patient remaining SPTC-free.

Fig. 4.

SHAP visualizes machine learning. A Global model output based on feature density; B Influence plot of single-sample characteristics of patients who develop SPTC; C Influence plot of single-sample characteristics of patients without developing SPTC.

Discussion

This study identified BMI, anxiety/depression, pre‑existing thyroid disease, ER status, and PR status as key risk factors associated with the development of SPTC in female breast cancer patients after surgery. Previous clinical studies examining the pathogenic links and prognosis between breast cancer and thyroid cancer have highlighted several influencing factors, including pathological grade and stage of breast cancer, receipt of postoperative radiotherapy, ER status, and family history of malignancy, among others¹². Subsequent studies by Allen et al.¹³, Bakos et al.¹⁴, and Li et al.²² expanded the list of potential risk factors contributing to SPTC, incorporating PR status, HER‑2 status, Ki‑67 proliferation index, and age at diagnosis. Building upon this existing evidence, the present study comprehensively integrated a range of potential risk factors and employed machine learning to identify those with the strongest influence on SPTC development. This approach holds important clinical value for improving the screening of high‑risk patients.

Weight management is of particular importance for overweight women with breast cancer. This study confirms that elevated BMI is a significant risk factor for the development of second primary thyroid cancer (SPTC) in female breast cancer patients, with most surgically treated patients in our cohort classified as overweight or obese. Multiple studies have further established a close association between BMI and tumor progression^23,24. One plausible explanation lies in the metabolic symbiosis between adipocytes and tumor cells in obesity, which can alter the tumor microenvironment and suppress anti‑tumor immune responses^25,26. Additionally, excess weight may enhance interactions between white adipose tissue in the breast and tumor cells, while also influencing thyroid hormone secretion, thereby potentially promoting the development of both breast and thyroid cancers^27,28. Therefore, comprehensive management of BMI—especially in those already overweight or obese—through integrated strategies such as dietary guidance, functional exercise, and targeted health education is clinically warranted. For postoperative breast cancer patients, we recommend combining limb functional rehabilitation with a supervised weight-control plan during recovery and long-term survivorship care, aiming to mitigate adverse outcomes and improve overall prognosis²⁹.

Chronic adverse psychological states represent a potential risk factor for driving oncogenic processes. This study indicates that depressive or anxious psychological states in female breast cancer patients are associated with an increased risk of developing SPTC. Such adverse emotional states, present in some patients from the initial admission assessment, can persist through follow‑up at 6 months post-discharge and may continue to affect patients throughout the entire recovery period. Studies have reported that the prevalence of anxiety and depression among female breast cancer patients can reach 30.7%³⁰ and 38%³¹, respectively. These adverse psychological conditions may accelerate malignant progression, growth, and metastasis, contributing to a poorer prognosis. Therefore, early psychological intervention for breast cancer patients experiencing anxiety or depression is particularly important³². We recommend implementing comprehensive psychological health management, strengthening psychosocial support systems, and enhancing psychological resilience in these patients^33,34. Furthermore, previous studies on risk factors have seldom focused on patients’ psychological status. While this study considered anxiety or depression as key features of adverse mental states, it only followed patients who presented with these conditions at admission. Changes in mental state among patients without initial psychological distress during treatment or recovery remain unknown. Additionally, this study focused solely on depression and anxiety; future research should also consider other adverse emotional states, such as self‑doubt, guilt, and anger³⁵.

Pre-existing thyroid disease increases the risk of developing SPTC in breast cancer patients. A Mendelian randomization study supports a positive causal association between breast cancer and thyroid cancer³⁶, indicating that the risk of SPTC in breast cancer survivors rises by 17%³⁷. The present study identifies pre-existing thyroid disease as a key risk factor for SPTC³⁸. Both abnormalities in thyroid hormones and structural thyroid lesions can serve as risk factors that promote the occurrence of SPTC. Research indicates that, given that both the breast and thyroid are endocrine organs regulated by the hypothalamic-pituitary axis, their comorbidity may be explained by hormonal crosstalk³⁹. Notably, compared with the general healthy population, breast cancer patients show higher levels of thyroid hormones⁴⁰, with particularly high rates of abnormality in thyroid-stimulating hormone (TSH) and free thyroxine (FT4) levels⁴¹. Therefore, these findings underscore the importance of prioritizing thyroid screening in female breast cancer patients—for example, by incorporating thyroid ultrasound and thyroid function tests into their care—while also recommending routine breast screening for patients with thyroid cancer⁴². Unfortunately, this study did not analyze or compare different degrees of thyroid hormone level changes or types of thyroid nodules in breast cancer patients. Consequently, the variable “pre-existing thyroid disease” has a clinical interpretability limitation. Its broad definition (combining dysfunction with benign nodules) may reduce clinical actionability, as a positive result cannot distinguish between these two distinct pathological entities. Although model performance (area under the curve, AUC) remained robust in the sensitivity analysis conducted earlier in this study, future research should collect data on thyroid dysfunction and nodules separately to enhance clinical utility.

Hormone receptor‑positive breast cancer patients face a higher risk of SPTC compared to those with receptor‑negative disease. In our study, both ER and PR positivity contributed positively to the SPTC risk‑prediction model. It is noteworthy that, compared with the general population, breast cancer survivors with hormone receptor-positive status have been reported to have a 20% higher risk of any second primary cancer and a 29% higher risk specifically for SPTC; if receptor positivity persists after treatment, the SPTC risk rises further⁴³. A Mendelian randomization study suggested that, from a genetic perspective, ER positivity is an important causal factor increasing the risk of SPTC in breast cancer patients^44,45. However, clinically, luminal-type breast cancers typically express ER and/or PR, while triple-negative breast cancers lack hormone receptor expression. Consequently, tailored screening strategies should be adopted for different subtypes of breast cancer.

Other potential risk factors also warrant attention. Several variables considered potentially predictive based on univariate analyses or prior literature—such as HER2 status—had their coefficients shrunk to zero during the LASSO selection process in this study. This finding does not negate their biological relevance; rather, it more likely reflects limited statistical power due to small effect sizes or an insufficient number of events within our cohort. Therefore, these variables should be regarded as candidate predictors requiring reevaluation in larger, more statistically robust studies, rather than as definitively refuted risk factors. In addition, previous studies have demonstrated significant associations between radiotherapy, chemotherapy, and the occurrence of SPTC in postoperative breast cancer patients⁴⁶. In the present study, these variables were not included in the modeling process due to considerations regarding heterogeneous treatment regimens and incomplete documentation of treatment details.

To date, few studies have employed predictive modeling to assess the risk of multiple primary cancers in this population, and there remains no consensus on which risk factors should be routinely included in such assessments⁴⁷.A study by Jin et al. included 6,864 participants and used logistic regression and random forest algorithms to explore risk factors for developing MPC in female breast cancer patients, ultimately incorporating 12 variables into model construction⁴⁸. As the authors noted, a larger number of variables can increase the risk of overfitting. Compared with the present study, the research by Jin et al. was based on a larger sample size; however, their prediction model was not subjected to external validation, thus limiting its generalizability. By employing clearer feature categorization and selection, this study effectively mitigates the risk of overfitting while maintaining focused predictive utility for SPTC. However, as this is a single-center study, although temporal-split validation was employed to assess the model’s out-of-sample generalization ability, providing evidence of temporal stability, our findings further validation in geographically diverse cohorts to ensure broader generalizability. In other words, this research is still in the exploratory phase. We aim to emphasize that the value of a clinically useful predictive model lies not only in algorithmic performance but also in how it is cautiously and contextually integrated into clinical workflows.

The value of this model lies in aggregating observable risk signals using routine clinical variables (BMI, anxiety/depression, pre-existing thyroid disease, ER/PR status) to provide a screening and stratification tool for SPTC. The associations between each variable and SPTC should be understood as statistical risk aggregation rather than established causal relationships. Clarifying the model’s operational premise as risk aggregation based on observable clinical signals helps clinicians correctly interpret its scope of application—namely, assisting in the early identification of high-risk patients and optimization of follow-up strategies, while avoiding overinterpretation of the predictors as direct causative factors. This positioning is precisely a key advantage of the model in facilitating clinical translation.

Currently, artificial intelligence is widely applied in the medical field, and its ability to handle complex data has shown great potential in disease diagnosis and prognosis⁴⁹. This study developed a risk-prediction model for SPTC in female breast cancer patients using five machine learning algorithms, among which the decision-tree model demonstrated the strongest performance. To enhance interpretability, we applied the SHAP framework, which integrates feature values and their interactions to provide both local and global explanations of the model’s decisions⁵⁰. Through SHAP analysis, this study offers detailed, individualized explanations of how key risk factors influence SPTC risk. Thus, we can understand the impact of different important features on individual samples, while revealing the functional details of the model, thereby facilitating the development of personalized preventive strategies for mitigating SPTC risk in postoperative female breast cancer patients.

Roadmap for multi-center external validation

To advance clinical translation, a multi‑center external validation should be conducted as the next step. Specifically, we propose: (1) Participating institutions should adopt standardized variable definitions and, where feasible, common data models to ensure consistent model input⁵¹. (2) At least three independent cohorts from geographically distinct regions and different tiers of healthcare facilities should be enrolled to assess generalizability across varied populations and practice settings. (3) SPTC outcomes should be collected according to a prespecified schedule (e.g., annual thyroid ultrasound and thyroid function tests) to minimize detection bias. (4) Model performance (reassessing model discrimination, calibration, and clinical net benefit) should be evaluated in each external cohort, with prespecified allowance for model recalibration or updating if substantial performance degradation is observed. Only after multi‑center validation can the model be considered for integration into clinical decision support systems or formal survivorship care guidelines.

Limitations

Several limitations warrant consideration. First, as a retrospective, single-center study, the relatively modest sample size constrains the broader generalizability of our findings. Future multi-center studies should incorporate independent cohorts from healthcare institutions across different geographic regions and levels of care, and use standardized data collection and variable definitions to evaluate the model’s generalizability across different populations and clinical settings. Second, the retrospective design inherently introduces temporal bias, limiting our ability to establish causal pathways; prospective longitudinal cohorts are needed. Finally, while the current model exhibits predictive utility, its performance could potentially be optimized by incorporating high-resolution biological features, such as genomic markers or comprehensive longitudinal biomarkers, to refine its precision and move toward a more comprehensive personalized risk assessment.

Conclusion

This study identified BMI, anxiety/depression, pre-existing thyroid diseases, and ER/PR status as pivotal clinical indicators significantly associated with the risk of developing SPTC in postoperative breast cancer patients. These factors can serve as critical clinical indicators for stratifying high-risk populations. To translate these associations into a clinical tool, we developed a prediction model using five machine learning algorithms. The decision tree model demonstrated superior performance, offering a valuable reference for informing personalized survivorship monitoring strategies.

Abbreviations

ML

Machine learning

BMI

Body mass index

MPC

Multiple primary cancer

SPC

Second primary cancers

SPTC

Second primary thyroid cancer

ER

Estrogen receptor

PR

Progesterone receptor

HER-2

Human epidermal growth factor receptor

KNN

K-nearest neighbors

SVM

Support vector machine

Author contributions

YQ, ZXH participated in the conception and design of the study; YQ, LZH performed the data analysis; YQ, ZYF, R Doktorzhan drafted the work and revised important intellectual content of the work; ZXH, GWJ ensured that issues related to the accuracy or completeness of any part of the work were properly investigated and resolved, with the approval of the final version.

Funding

The research support for this study was funded by the Key R&D Program Projects of Xinjiang Uygur Autonomous Region (No. 2022B03019-4); and Xinjiang Medical University Scientific Research and Innovation Team Project (XYD-2024C09).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Human ethics and consent to participate

The study was conducted in accordance with the principles of the Declaration of Helsinki. This study was reviewed by the Ethics Committee of the Cancer Hospital Affiliated with Xinjiang Medical University, and the requirement for written informed consent was waived (K-2023048).