Ensemble machine learning for predicting breast cancer recurrence and mortality using clinical and hemogram data

Abstract

Accurate prediction of breast cancer recurrence remains difficult because prognosis varies significantly across molecular subtypes. This underscores the need for scalable, affordable, and interpretable prognostic tools. We developed machine learning models integrating hematological indices with clinicopathologic data to predict 2- and 10-year recurrence or all-case death. We retrospectively analyzed 4277 women with primary breast cancer (2008–2022). The cohort included hormone receptor-positive (HR + ), HER2-positive, and triple-negative (TNBC) subtypes. We trained multiple classifiers and integrated them into a stacked ensemble using logistic regression as the final learner. Class imbalance was addressed with SMOTE applied to training sets. The ensemble achieved strong discrimination: general cohort AUC 0.859 (2-year) and 0.811 (10-year), with specificity 84–82% and sensitivity 68–61%. Subtype-specific performance remained robust across time horizons: HR + AUC 0.862/0.804, HER2 + AUC 0.877/0.831, and TNBC AUC 0.826/0.826 (2-year/10-year respectively). SHAP analysis identified advanced tumor stage, elevated inflammatory ratios (NLR, PLR, MLR), elevated red cell distribution width, and age as key adverse predictors with stronger effects on early recurrence. This interpretable tool uses routine blood tests and clinico-pathological data at diagnosis. It flaggs high-risk patients for intensified therapy, useful where genomic testing is unavailable.

Introduction

Breast cancer is the most frequently diagnosed cancer among women worldwide and remains a leading cause of cancer-related mortality¹. Despite advances in screening and therapy, outcomes vary widely which underscores the need for prognostic tools beyond traditional clinical factors. Prognosis and treatment decisions are guided by tumor subtype and stage at diagnosis. Subtype classification based on hormone receptor (estrogen receptor - ER, progesterone receptor - PR) delineates luminal tumors (ER⁺ and/or PR⁺) and human epidermal growth factor receptor 2 HER2-enriched tumors (HER2/C-erbB2 overexpressing) or tumors without expression are classified as triple-negative breast cancers (TNBC; ER⁻/PR⁻/HER2⁻)². Each subtype follows a distinct therapeutic approach including endocrine treatment for hormone receptor-positive cancers, HER2-targeted therapies, and chemotherapy. While these markers are invaluable and define therapy, patients in the same subtype and stage may experience very different outcomes. Additional prognostic biomarkers could refine risk stratification. New therapeutic approaches include immune checkpoint inhibitors such as anti-PD1/PDL1 combined with chemotherapy as first-line treatment for metastatic TNBC in 2020, then for early stage TNBC in 2021^3,4. Several other trials are in progress with response rate ranging from 15 to 60%, corroborating with the relevance of tumor immune microenvironment in the disease⁵. Consistent with this, the presence of tumor-infiltrating lymphocytes has been linked to better outcomes in breast cancer, especially when cytotoxic CD8⁺ T cells are abundant. In contrast, an immunosuppressive tumor microenvironment, characterized by high levels of neutrophils or tumor-associated macrophages, often correlates with poorer prognosis and resistance to therapy⁶. These observations suggest that systemic measures of inflammation and immunity might carry prognostic significance. In light of the tumor–immune interplay, peripheral blood leukocyte counts and their derived ratios have emerged as convenient prognostic markers in oncology^7–9. Routine complete blood count indices such as the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and monocyte-to-lymphocyte ratio (MLR) reflect systemic inflammatory status and immune balance. A meta-analysis of 15 cohorts (> 8500 patients) found that elevated NLR (cutoff ~3) is linked to poorer survival, with a hazard ratio of 2.6⁷. Thus, inexpensive blood-based metrics may represent a low-cost opportunity to enhance risk assessment beyond classical clinicopathological features. Recently, we published a statistical model for disease-free survival based on blood-derived indices, indicating that PLR and MLR had an unfavorable prognosis in our cohort⁸. In the triple-negative cases in our cohort, NLR > 2 was associated with poorer EFS and OS¹⁰.

Recent studies have successfully integrated routine blood biomarkers into prognostic models applying machine learning (ML) methods; for example, Zhang et al.¹¹ constructed a LightGBM ensemble using MRI imaging features and clinicopathologic characteristics to predict breast cancer metastasis from 574 non-metastatic patients. Notably, one recent analysis applied ML algorithms to structured electronic health record data to predict breast cancer recurrence from 342 patients using 25 features^12,13. Similarly, another group built an explainable ML model for long-term outcomes using standard clinical data, with 1704 breast cancer cases with 64 features¹³. Here, we address this challenge by developing a multimodal ML model with only 16 features extracted from hemogram with clinical parameters at diagnosis to predict breast cancer recurrence or death in a cohort of 4277 breast cancer-only cases to predict early (< 2 years recurrence) and 3675 cases for late recurrence (> 10 years), without any cost increment, and potentially aid risk-stratification for health care professionals. This finding is especially novel, as prior prognostic models have rarely incorporated red blood cell parameters; our work appears to be among the first to highlight that elevated RDW and low hemoglobin (anemia) are independently associated with higher 10-year recurrence risk in breast cancer patients^14,15. This design renders the model globally accessible and suitable for implementation across diverse resource settings. Furthermore, because RDW and systemic inflammatory markers reflect potentially modifiable physiological states such as chronic inflammation, nutritional deficiencies, and anemia, these findings establish a foundation for future interventional studies investigating whether optimization of these hematologic parameters during treatment and surveillance improves recurrence-free survival^16,17.

Results

Hemogram at diagnosis has a prognostic value by supervised machine learning models

Hemograms are common and cheap exams made by cancer patients at diagnosis, before any treatment, as a benchmark. As NLR, PLR, and other hemogram values may have prognostic values, we evaluated by supervised machine learning algorithms if the whole hemogram and clinical factors obtained at diagnosis, such as clinical stage in TNM, molecular subtype, and age at diagnosis, could be used to predict early (2-years) and late (10-years) recurrence or death. We grouped recurrence or death because patients that progress during treatment did not register a recurrence, but died early from cancer progression even in palliative care. We tested if the clinical factors alone could predict 2-years recurrence/death, and we obtained an AUC-ROC = 0.824. Only hematologic features AUC-ROC was 0.625, which was worse, but when combined (Clinical+Hematological) AUC-ROC was 0.85, which demonstrated a gain to use the cheap and available hemogram (Suppl. Fig. S1 and Suppl. Table S1). In the late event scenario (10-years), clinical only was AUC-ROC 0.78 while combined AUC-ROC = 0.81.

The machine learning models achieved AUC > 0.80 for both short-term (2-year) and long-term (10-year) survival prediction on the original class-imbalanced data. In the overall cohort, five algorithms (SVM, Random Forest, XGBoost, LightGBM, and Stacking ensemble) were evaluated, with the top-performing models selected based on validation ROC-AUC. For 2-year prediction, SVM ranked first on the validation set (AUC 0.8524) with test AUC 0.8512 (sensitivity 74.1%, specificity 79.1%), while Stacking followed as a near-equivalent second (validation AUC 0.85, PR-AUC 0.66, Brier score 0.13, sensitivity 72.8%, specificity 87.2%) (Fig. 1, Table 2). The near-identical validation AUCs confirm the equivalence of both models, with Stacking offering a more specificity-balanced profile. This profile may be preferred in clinical settings requiring high certainty before intervention.

Fig. 1. Receiver Operating Characteristic (ROC) Curves for the 2-year Predictive Models.

ROC curves depict the performance of different machine learning algorithms in predicting recurrence or death within two years across the overall cohort and breast cancer molecular subtypes (HR + , HER2 + , and TNBC). The upper panels correspond to the independent test set, while the lower panels present results from the validation set. Among the classifiers evaluated on the original, class-imbalanced data, the Stacking ensemble, Random Forest, and SVM models achieved the highest discriminative ability, with AUCs ranging from 0.80 to 0.88.

Table 2.

Model performance metrics on the validation and test sets, for each cohort, prediction horizon, and ranked model configuration

Cohort	Horizon	Rank	Model	Split	ROC-AUC	PR-AUC	Brier	Sensitivity [%]	Specificity [%]	PPV [%]	NPV [%]
General	2-year	1	SVM	VALIDATION	0.8524	0.6604	0.0973	75.9	83.1	50.4	93.8
				TEST	0.8512	0.6348	0.0989	74.1	79.1	44.5	93.1
General	2-year	2	Stacking	VALIDATION	0.8523	0.6593	0.1340	72.8	87.2	56.4	93.4
				TEST	0.8588	0.6233	0.1415	68.3	83.8	48.9	92.1
General	10-year	1	Stacking	VALIDATION	0.7853	0.7722	0.1837	61.4	83.7	72.9	75.3
				TEST	0.8109	0.7878	0.1761	61.4	82.0	70.9	74.9
General	10-year	2	SVM	VALIDATION	0.7848	0.7689	0.1769	62.1	83.2	72.5	75.5
				TEST	0.8140	0.7911	0.1692	63.4	81.3	70.8	75.7
HR +	2-year	1	SVM	VALIDATION	0.8531	0.5871	0.0800	73.6	87.9	49.5	95.4
				TEST	0.8260	0.5237	0.0830	69.4	87.0	46.3	94.6
HR +	2-year	2	Stacking	VALIDATION	0.8338	0.5484	0.1430	76.4	78.9	36.9	95.4
				TEST	0.8614	0.5669	0.1404	79.2	77.8	36.5	95.9
HR +	10-year	1	Stacking	VALIDATION	0.7953	0.7571	0.1790	77.8	66.1	58.9	82.7
				TEST	0.8043	0.7580	0.1823	80.8	58.6	55.2	82.9
HR +	10-year	2	RandomForest	VALIDATION	0.7934	0.7557	0.1747	55.0	88.0	74.0	75.8
				TEST	0.8029	0.7438	0.1748	59.9	83.9	70.1	76.8
HER2 +	2-year	1	RandomForest	VALIDATION	0.8728	0.6444	0.1099	69.4	93.8	73.5	92.5
				TEST	0.8592	0.7135	0.1062	77.8	91.0	68.3	94.3
HER2 +	2-year	2	Stacking	VALIDATION	0.8711	0.6573	0.1408	72.2	94.5	76.5	93.2
				TEST	0.8774	0.7625	0.1423	77.8	92.4	71.8	94.4
HER2 +	10-year	1	SVM	VALIDATION	0.7723	0.7744	0.1807	50.8	95.7	88.9	74.0
				TEST	0.8179	0.7799	0.1597	50.8	95.7	88.9	74.2
HER2 +	10-year	2	Stacking	VALIDATION	0.7567	0.7646	0.1912	54.0	94.6	87.2	75.0
				TEST	0.8312	0.7717	0.1649	57.1	91.4	81.8	75.9
TNBC	2-year	1	Stacking	VALIDATION	0.8421	0.7411	0.1510	76.0	82.2	66.7	88.0
				TEST	0.8256	0.7286	0.1707	78.0	70.1	54.9	87.2
TNBC	2-year	2	RandomForest	VALIDATION	0.8327	0.7240	0.1496	76.0	80.4	64.4	87.8
				TEST	0.8009	0.6877	0.1634	74.0	67.3	51.4	84.7
TNBC	10-year	1	SVM	VALIDATION	0.7321	0.8121	0.2044	51.4	90.5	86.1	62.0
				TEST	0.8294	0.8672	0.1693	58.3	89.1	85.7	65.5
TNBC	10-year	2	Stacking	VALIDATION	0.7291	0.8035	0.2088	50.0	92.1	87.8	61.7
				TEST	0.8264	0.8687	0.1723	58.3	93.8	91.3	66.7

Top-2 machine learning models on original class-imbalanced data, evaluated by Youden J threshold. Performance of the top-2 machine learning models on the original, class-imbalanced data (no oversampling applied). For each breast cancer molecular subtype (General, HR + , HER2 + , TNBC) and time horizon (2-year, 10-year), the two best-performing models were selected by ROC-AUC on the validation set among five candidates (Random Forest, XGBoost, LightGBM, SVM, and Stacking ensemble). Validation metrics reflect model performance during the selection phase, where models were trained on 60% of the data and evaluated on the held-out validation set (20%). The selected models and their validation-derived Youden J thresholds were applied unchanged to the test set (no data leakage). Test metrics reflect the same models applied unchanged to the held-out test set (20%), which was never used during training or selection. Metrics reported: ROC-AUC, PR-AUC, Brier score (lower = better calibration), Sensitivity, Specificity, PPV, and NPV. All values are point estimates; 95% bootstrap confidence intervals are provided in Suppl. Table S2.

Rows in bold indicate test-set performance. ROC-AUC area under the receiver operating characteristic curve, PR-AUC area under the precision–recall curve, PPV positive predictive value, NPV negative predictive value.

For 10-year combined recurrence/death prediction, Stacking ranked first on the validation set (AUC 0.78, PR-AUC 0.77), with SVM following closely (validation AUC 0.78) (Fig. 2, Table 2).

Fig. 2. Receiver Operating Characteristic (ROC) Curves for the 10-year Predictive Models.

ROC curves depict the discriminative performance of different algorithms in predicting recurrence or death within ten years across the entire cohort and breast cancer molecular subtypes (HR + , HER2 + , and TNBC). The upper panels correspond to the independent test set, and the lower panels to the validation set. Among the evaluated classifiers on the class-imbalanced dataset, SVM, Stacking ensemble, and Random Forest models consistently achieved the highest AUCs (≈0.80-0.83) across both datasets. The overall cohort achieved AUC 0.814 (SVM, test), HR+ reached 0.804 (Stacking, test), HER2+ achieved 0.831 (Stacking, test), and TNBC showed 0.829 (SVM, test).

Compared to the 2-year models, the 10-year predictions demonstrated higher PR-AUC values (~0.77–0.79 versus ~0.62–0.66), reflecting the greater cumulative event incidence over a decade. The operating characteristics differed between time horizons: 2-year models achieved higher specificity (70–91%) with moderate sensitivity (74–78%), whereas 10-year models captured more events with sensitivity around 51–81% at specificity of 59–96%. These results indicate that short-term models were calibrated more conservatively to reduce false alarms, while long-term models prioritized identifying a larger proportion of patients who would eventually experience recurrence or death. Consistent performance across independent test and validation sets confirms model generalizability (Table 2). Pairwise DeLong tests confirmed that performance differences between top-ranked models (SVM and Stacking) were non-significant, while both significantly outperformed LightGBM in most cohort–horizon combinations (Suppl. Table S5).

Subtype-specific model performance

Model performance remained robust across major breast cancer molecular subtypes, though with notable variation in discrimination and calibration metrics. The hormone receptor-positive (HR + ) subgroup demonstrated excellent performance, particularly for short-term prediction. The 2-year HR+ model showed SVM as the top validator (AUC 0.85, sensitivity 73.6%, specificity 87.9%), while Stacking achieved the highest test AUC (0.86, sensitivity 79.2%, specificity 77.8%), indicating complementary performance profiles between the two leading models (Fig. 1, Table 2). For 10-year prediction, HR+ models showed test AUCs of 0.804 (Stacking) and 0.802 (Random Forest), with validation AUCs of 0.795 and 0.793, respectively (Fig. 2, Table 2). The consistency between test and validation performance (difference <1%) demonstrates that the models did not present overfitting. Calibration was strong, with Brier scores ranging from 0.080 (2-year validation) to 0.183 (10-year test), reflecting the increased uncertainty inherent in longer-horizon predictions.

Patients with HER2-positive tumors exhibited the strongest overall predictive performance across both time horizons. The 2-year HER2+ model achieved exceptional discrimination, with Random Forest ranking first on the validation set (AUC 0.8728, PR-AUC 0.6444, Brier 0.1099, sensitivity 69.4%, specificity 93.8%) and Stacking following closely (validation AUC 0.8711) (Fig. 1, Table 2). On the test set, Stacking achieved the highest AUC (0.8774, sensitivity 77.8%, specificity 92.4%), with Random Forest also performing strongly (test AUC 0.8592, sensitivity 77.8%, specificity 91.0%), confirming robust generalization across both evaluation splits. For 10-year prediction, SVM led on the validation set (AUC 0.7723), while Stacking achieved a test AUC of 0.8312 (sensitivity 57.1%, specificity 91.4%), maintaining high specificity in this subtype (Fig. 2, Table 2). The consistently high PR-AUC values in HER2+ disease (0.64–0.78) reflect favorable event prediction characteristics, likely attributable to the more homogeneous biological behavior and treatment responsiveness of HER2-positive tumors.

In triple-negative breast cancer (TNBC), models demonstrated acceptable but comparatively lower performance, consistent with the biological heterogeneity and aggressive nature of this subtype. For 10-year prediction, SVM achieved test AUC 0.8294 (PR-AUC 0.8672, Brier 0.1693), with sensitivity of 58.3% and specificity of 89.1%, reflecting a specificity-favoring threshold that limits false positive risk escalation in this clinically aggressive subtype (Fig. 2). Validation performance showed AUC 0.7321 (SVM, PR-AUC 0.8121); the divergence between ROC-AUC and PR-AUC reflects the high event rate in TNBC, where the model retains strong positive predictive value even when overall discrimination is more modest, a clinically relevant distinction given the priority of identifying true high-risk patients in this aggressive subtype. For 2-year prediction, Stacking led on the validation set (AUC 0.8421, PR-AUC 0.7411, sensitivity 76.0%, specificity 82.2%), with Random Forest as the second-ranked model (validation AUC 0.8327), confirming robust generalization (Fig. 1, Table 2). Brier scores in TNBC were higher than other subtypes (0.147–0.209), particularly for 10-year predictions, reflecting the inherent difficulty in long-term prognostication for this heterogeneous disease. Importantly, no subgroup showed dramatic performance degradation (AUC > 0.73), underscoring that our approach retains utility across diverse biological subtypes. To assess potential differential performance across age strata, we stratified General cohort results by age group (< 40, 40–60, > 60 years). The models performed better in 40–60 years old group, and maintained useful discrimination across all age groups, with AUCs generally above 0.75 even in patients under 40 (Suppl. Fig. S12, Table S6), suggesting consistent performance regardless of age-related biological differences.

Impact of feature engineering: embedding augmentation analysis

To evaluate the impact of data augmentation strategies, we conducted a systematic comparison of models trained with and without latent embedding features (Table 3). The comparison involved 480 model configurations across three oversampling techniques (SMOTE, Borderline-SMOTE, ADASYN), two feature engineering pipelines (With EMBED: oversampling + PCA/UMAP embeddings + GMM cluster indicators; No EMBED: oversampling only), and all five algorithms. The benefit of embedding augmentation proved context-dependent rather than uniformly advantageous.

Table 3.

Model performance metrics across oversampling and embedding configurations, for each cohort, prediction horizon, and top-ranked scenario

Cohort	Horizon	Sc. Rank	Oversampler	Embed	Model	Split	ROC-AUC	PR-AUC	Brier	Sens (%)	Spec (%)	PPV (%)	NPV (%)
General	2-year	1	ADASYN	Yes	Stacking	VALIDATION	0.8552	0.6609	0.0963	78.5	79.9	47.0	94.3
						TEST	0.8497	0.6124	0.1012	78.5	75.6	42.2	94.0
General	2-year	2	BL-SMOTE	Yes	Stacking	VALIDATION	0.8532	0.6568	0.0962	71.5	87.1	55.7	93.1
						TEST	0.8490	0.6148	0.1011	69.6	84.8	50.9	92.5
General	2-year	3	BL-SMOTE	No	Stacking	VALIDATION	0.8520	0.6587	0.0962	67.7	91.0	62.9	92.6
						TEST	0.8495	0.6167	0.1000	66.5	88.0	55.6	92.0
General	10-year	1	SMOTE	No	Stacking	VALIDATION	0.7856	0.7704	0.1784	62.1	83.7	73.1	75.6
						TEST	0.8162	0.7916	0.1689	61.4	82.3	71.2	75.0
General	10-year	2	SMOTE	Yes	SVM	VALIDATION	0.7840	0.7763	0.1798	58.8	87.6	77.2	74.9
						TEST	0.8133	0.7923	0.1731	56.9	86.5	75.0	73.8
General	10-year	3	BL-SMOTE	Yes	SVM	VALIDATION	0.7823	0.7718	0.1837	59.8	86.7	76.2	75.1
						TEST	0.8131	0.7916	0.1753	58.2	84.9	73.2	74.0
HR +	2-year	1	SMOTE	Yes	Stacking	VALIDATION	0.8509	0.5553	0.0827	76.4	81.2	39.6	95.5
						TEST	0.8392	0.5353	0.0828	75.0	82.3	40.6	95.3
HR +	2-year	2	SMOTE	No	SVM	VALIDATION	0.8411	0.5652	0.1217	68.1	90.6	53.8	94.6
						TEST	0.8312	0.5216	0.1239	66.7	90.4	52.8	94.4
HR +	2-year	3	BL-SMOTE	Yes	Stacking	VALIDATION	0.8375	0.5239	0.0874	77.8	77.8	36.1	95.6
						TEST	0.8297	0.5177	0.0858	79.2	75.6	34.3	95.7
HR +	10-year	1	SMOTE	No	Stacking	VALIDATION	0.7851	0.7502	0.1738	60.2	82.1	67.8	76.8
						TEST	0.8028	0.7554	0.1706	65.7	78.0	65.3	78.3
HR +	10-year	2	ADASYN	Yes	Stacking	VALIDATION	0.7850	0.7540	0.1724	59.1	85.0	71.1	76.9
						TEST	0.7997	0.7457	0.1709	60.5	79.8	65.4	76.2
HR +	10-year	3	SMOTE	Yes	Rand. Forest	VALIDATION	0.7841	0.7516	0.1758	75.4	67.9	59.5	81.6
						TEST	0.7966	0.7355	0.1761	76.7	64.8	57.9	81.6
HER2 +	2-year	1	SMOTE	No	SVM	VALIDATION	0.8705	0.6671	0.1478	75.0	93.1	73.0	93.8
						TEST	0.8908	0.7647	0.1445	77.8	92.4	71.8	94.4
HER2 +	2-year	2	SMOTE	Yes	XGBoost	VALIDATION	0.8657	0.6210	0.1190	80.6	83.5	54.7	94.5
						TEST	0.8310	0.6007	0.1178	77.8	81.4	50.9	93.7
HER2 +	2-year	3	ADASYN	No	SVM	VALIDATION	0.8638	0.6481	0.1581	77.8	86.2	58.3	94.0
						TEST	0.8833	0.7537	0.1552	80.6	85.5	58.0	94.7
HER2 +	10-year	1	BL-SMOTE	No	SVM	VALIDATION	0.7771	0.7750	0.1900	57.1	87.0	75.0	74.8
						TEST	0.8114	0.7812	0.1792	60.3	92.5	84.4	77.5
HER2 +	10-year	2	ADASYN	No	SVM	VALIDATION	0.7724	0.7700	0.1912	54.0	90.2	79.1	74.1
						TEST	0.8097	0.7779	0.1804	58.7	92.5	84.1	76.8
HER2 +	10-year	3	SMOTE	No	SVM	VALIDATION	0.7724	0.7700	0.1858	49.2	95.7	88.6	73.3
						TEST	0.8099	0.7848	0.1748	58.7	95.7	90.2	77.4
TNBC	2-year	1	BL-SMOTE	Yes	Stacking	VALIDATION	0.8495	0.7756	0.1335	82.0	79.4	65.1	90.4
						TEST	0.8385	0.7167	0.1483	82.0	65.4	52.6	88.6
TNBC	2-year	2	BL-SMOTE	No	SVM	VALIDATION	0.8477	0.7728	0.1709	80.0	79.4	64.5	89.5
						TEST	0.8318	0.6772	0.1843	82.0	67.3	53.9	88.9
TNBC	2-year	3	SMOTE	Yes	Stacking	VALIDATION	0.8471	0.7655	0.1361	78.0	79.4	63.9	88.5
						TEST	0.8381	0.7168	0.1485	80.0	67.3	53.3	87.8
TNBC	10-year	1 †	ADASYN	No	Stacking	VALIDATION	0.7266	0.8035	0.2098	54.2	85.7	81.2	62.1
						TEST	0.8253	0.8705	0.1714	65.3	89.1	87.0	69.5
TNBC	10-year	2 †	BL-SMOTE	No	Stacking	VALIDATION	0.7266	0.8035	0.2098	54.2	85.7	81.2	62.1
						TEST	0.8253	0.8705	0.1714	65.3	89.1	87.0	69.5
TNBC	10-year	3 †	SMOTE	No	Stacking	VALIDATION	0.7266	0.8035	0.2098	54.2	85.7	81.2	62.1
						TEST	0.8253	0.8705	0.1714	65.3	89.1	87.0	69.5

Performance of the top-3 oversampling scenarios per cohort and time horizon. Six oversampling pipelines were evaluated: SMOTE, BorderlineSMOTE, and ADASYN, each with or without embedding augmentation (PCA + UMAP + GMM-derived cluster indicators). For each breast cancer molecular subtype (General, HR + , HER2 + , TNBC) and time horizon (2-year, 10-year), scenarios were ranked by the best validation ROC-AUC achieved across five candidate models (Random Forest, XGBoost, LightGBM, SVM, Stacking ensemble); the top-3 scenarios are shown with their best-performing model. Oversampling was applied exclusively within training folds. Validation-derived Youden J thresholds were applied unchanged to the test set (no data leakage). Metrics reported: ROC-AUC, PR-AUC, Brier score (lower = better calibration), Sensitivity, Specificity, PPV, and NPV. Point estimates only; 95% bootstrap confidence intervals are provided in Supplementary Table S3.

† TNBC 10-year: all three scenarios tied at ROC-AUC = 0.7266 on the validation set.

Rows in bold indicate test-set performance. Sc. Rank scenario rank within the cohort and horizon, BL-SMOTE borderline SMOTE, Embed whether clinical-feature embeddings were used, ROC-AUC area under the receiver operating characteristic curve, PR-AUC area under the precision–recall curve, Sens sensitivity, Spec specificity, PPV positive predictive value, NPV negative predictive value.

In the overall cohort, embedding provided minimal advantage for 10-year prediction: the no-embed pipeline achieved AUC 0.77 versus embed’s 0.78, with the former offering slightly higher sensitivity (0.622 vs. 0.615) and the latter higher specificity (0.837 vs 0.876), differences too small to be clinically meaningful. For 2-year prediction, both pipelines performed comparably (AUC ~ 0.85–0.86), suggesting that the baseline clinical features captured sufficient prognostic information for the general population.

However, subtype-specific patterns emerged with clear clinical implications. In HR-positive disease, embedding provided meaningful improvements for long-term prediction: Random Forest with embedding achieved test AUC 0.8201 (PR-AUC 0.775, sensitivity 59.4%, specificity 84.3%) compared to no-embed’s 0.8015, a gain of 2 percentage points in AUC (Table 3). This suggests that unsupervised latent structure captures clinically relevant heterogeneity within luminal tumors, potentially reflecting underlying biological diversity in HR+ disease that manifests over longer follow-up. For 2-year HR+ prediction, the embed pipeline presented a little improvement: Stacking Classifier achieved AUC 0.85 (PR-AUC 0.55, sensitivity 76.4%, specificity 81.2%), indicating that near-term risk is adequately characterized by conventional features.

Conversely, in HER2-positive disease, the simpler no-embed pipeline consistently prevailed across both time horizons. At 2 years, SVM without embedding (SMOTE) achieved test AUC 0.89 (PR-AUC 0.76, sensitivity 77%, specificity 92%), outperforming the best embedded variant (XGBoost+embed, AUC 0.8310). At 10 years, SVM without embedding (BL-SMOTE) led with test AUC 0.8114 (PR-AUC 0.7812), exceeding all embedded alternatives (Table 3). This pattern suggests that HER2+ tumors possess inherently high baseline clinicopathologic separability, potentially due to the dominant prognostic signal from HER2 status and treatment response, making additional latent features redundant or even introducing noise.

For triple-negative breast cancer, embedding provided meaningful gains at 2 years but was neutral at 10 years. At 2 years, Stacking with embedding (BL-SMOTE) achieved test AUC 0.8385 (PR-AUC 0.7167, sensitivity 82.0%, specificity 65.4%), outperforming the best no-embed variant (SVM, BL-SMOTE, AUC 0.8318), suggesting that latent features capture subtle biological variation within the diverse TNBC landscape. At 10 years, all three oversampling strategies produced equivalent results regardless of embedding (test AUC 0.8253), reflecting the limited cohort size constraining differentiation between configurations (Table 3).

Validation set performance consistently tracked test set rankings within each subtype-horizon combination, confirming that the observed embedding effects represent genuine signal rather than overfitting. In summary, embedding augmentation demonstrated subtype-specific utility: most beneficial for HR+ disease at 10 years and TNBC at both horizons; unnecessary or counterproductive for HER2+ disease; and neutral for the general cohort. These findings suggest that feature engineering strategies should be tailored to biological context, with embedding particularly valuable in heterogeneous subtypes lacking dominant single-feature prognostic signals. Pairwise DeLong comparisons across all oversampling scenarios confirmed consistent model rankings, with SVM and Stacking achieving statistically superior discrimination relative to LightGBM in most configurations, particularly for 10-year predictions (Suppl. Fig. S13a–d).

Feature importance and risk stratification

To interpret model predictions and enhance clinical transparency, we applied SHAP (Shapley Additive Explanations) analysis¹⁷. This game theory-based framework quantifies the contribution of each feature to individual predictions. SHAP provides both global feature importance rankings and local explanations, making the model’s decision-making process interpretable to clinicians^17,18. Aggregate feature importance rankings across all five algorithms and molecular subtypes are shown in Suppl. Fig. S2 and Table S2. SHAP summary plots for the best-performing models on the test set (Fig. 3) revealed consistent feature hierarchies across all cohorts and time horizons, with distinct patterns in both magnitude and directionality of feature contributions. In these plots, each point represents a patient, colored by feature value (red = high; blue = low), with horizontal position indicating impact on predicted risk.

Fig. 3. SHAP summary plots for best-performing models (test set).

Beeswarm plots display feature contributions (x-axis: SHAP value) ranked by importance (top to bottom). Points represent individual patients colored by feature value (red = high; blue = low). Tumor stage variables (T_stage, N_stage, M_stage) dominate risk prediction across all cohorts, with advanced stages showing large positive SHAP values. CBC-derived markers (NLR, MLR, PLR, RDW-SD, RDW-CV, MCHC, hemoglobin, hematocrit) contribute as secondary features: higher inflammatory ratios and RDW increase risk, while higher hemoglobin confers protection. SHAP magnitudes are larger in 2-year models, indicating stronger near-term effects. TNBC models show broader feature dispersion, reflecting subtype heterogeneity.

Tumor stage variables (T_stage, N_stage, M_stage) emerged as the most influential predictors across all models. In the general cohort 10-year SVM model, advanced T-stages showed large positive SHAP values (> 1.0), substantially increasing predicted risk, while early T-stages exhibited protective effects (negative SHAP values). M_stage displayed particularly concentrated risk effects, with metastatic disease showing SHAP values between 1.5 and 2.0. Notably, in the 2-year general stacking model, M_stage ranked first in feature importance, surpassing T_stage, indicating that distant metastasis is especially critical for near-term recurrence prediction. N_stage consistently ranked among the top three features, with higher nodal involvement driving predictions toward increased risk.

This hierarchy remained stable across breast cancer subtypes. In HR+ models, both 2-year Random Forest and 10-year Stacking algorithms showed M_stage, T_stage, and N_stage with the widest SHAP value distributions. HER2+ models similarly demonstrated stage variable dominance at both time horizons, with M_stage effects particularly pronounced in the 10-year model (SHAP values reaching 1.5–2.0). TNBC models preserved this general ordering, though with notably broader dispersion of SHAP values across multiple features, consistent with greater biological and clinical heterogeneity in triple-negative disease¹⁷.

Beyond classical staging variables, complete blood count-derived inflammatory and hematologic markers provided meaningful secondary contributions to risk prediction. Inflammatory cell ratios consistently appeared in the upper tier of feature importance, with higher values associated with positive SHAP values (increased predicted risk) and lower values conferring protective effects. NLR ranked particularly high across multiple models: appearing as the 4th–5th most important feature in the general 10-year SVM model, among the top 6 features in HR+ models at both time horizons, and prominently in TNBC models. MLR and PLR showed similar patterns, with consistent directional effects: higher ratios shifted predictions toward increased recurrence probability. Red blood cell indices also contributed substantially to model predictions. RDW (both RDW-SD and RDW-CV) appeared consistently across all models, with elevated values associated with higher predicted risk. Supplementary Fig. S3 displays distributions of hematologic indices stratified by predicted risk category in the general cohort, visually demonstrating the biomarker differences underlying model predictions. Conversely, higher hemoglobin and hematocrit levels were associated with negative SHAP values (protective effects), indicating that patients without pre-treatment anemia had lower predicted recurrence risk. Pre-treatment anemia is prevalent in breast cancer. Additional erythrocyte parameters (MCHC, MCH, MCV) showed more variable effects depending on model and subtype, but generally contributed to risk stratification in biologically plausible directions.

SHAP value magnitudes were generally larger in 2-year models compared to 10-year models, particularly for inflammatory markers and RDW. This suggests that hematologic biomarkers exert a stronger influence on near-term recurrence, while their effects attenuate over longer follow-up periods. The same feature hierarchy persisted across both time horizons, but with reduced absolute SHAP values at 10 years, indicating that while these factors remain prognostically relevant long-term, their relative contribution diminishes as other biological processes and treatment effects accumulate¹⁷.

Among subtypes, TNBC models displayed the broadest dispersion of SHAP values across features, with less dominance by any single predictor. This reflects triple-negative breast cancer heterogeneity, where multiple features contribute moderate predictive signals rather than a single factor dominating outcome prediction^14,19. In contrast, HR+ and HER2+ models showed more concentrated SHAP distributions around top-ranked stage variables, suggesting more predictable risk stratification. Age contributed moderately but consistently across all models (generally ranking 5th-8th), with directional effects varying by model and subtype, reflecting the complex relationship between age and breast cancer biology across different molecular contexts.

These SHAP-derived insights demonstrate that while tumor stage remains the cornerstone of prognostic assessment, readily available complete blood count parameters provide clinically meaningful augmentation of risk stratification¹⁷. The integration of inflammatory and hematologic biomarkers captures systemic host factors, inflammation, immune status, and physiologic reserve, that complement anatomic disease extent and contribute to personalized recurrence risk estimation^15,20,21. The transparency afforded by SHAP analysis enhances clinical trust, enabling physicians to understand not only which patients are at high risk, but also why specific biological and clinical factors drive that risk¹⁸.

Finally, we conducted survival analyses to assess the prognostic relevance of the model-derived risk estimates. For each algorithm and prediction horizon, patients were stratified into high- and low-risk groups using the median predicted probability within the respective test set, thereby avoiding any risk of information leakage. Univariate Cox proportional hazards analysis confirmed the prognostic significance of hematologic biomarkers and clinical variables (Suppl. Table S4). Kaplan–Meier curves consistently demonstrated substantial and statistically significant separation between risk strata (log-rank p < 0.001), indicating that higher model-predicted risk was associated with markedly worse recurrence-free survival across the 2- and 10-year follow-up periods. The principal results for the best-performing model are presented in the main text. Comprehensive Kaplan–Meier analyses encompassing all five algorithms, both time horizons, and all tumor subtypes (HR + , HER2 + , and TNBC) are provided in Suppl. Figs. S5–S11. Across all settings, patients classified as high-risk exhibited significantly reduced event-free survival compared with those classified as low-risk, with hazard ratios ranging from 3.02 to 10.47. This reproducible pattern of risk discrimination across models and clinical subgroups supports the robustness of the predictive framework and highlights its potential utility for individualized prognostic assessment in breast cancer.

Discussion

In this study, we developed and validated a multimodal machine learning model to predict breast cancer recurrence or death at two clinically relevant time points (2 and 10 years post-treatment) in a single-center cohort from a public health care hospital, part of the Brazilian universal health system (SUS). Our results demonstrate that even without genomic profiling, an integrated model combining clinicopathologic features with routine blood biomarkers can achieve robust prognostic performance. The 2-year models achieved strong discrimination (general cohort AUC 0.859), with HER2+ disease demonstrating exceptional performance (AUC 0.877), the highest across all subtypes. The 10-year models maintained good discrimination (general cohort AUC 0.814, HER2 + 0.831, TNBC 0.829). The superior performance in HER2+ disease likely reflects the biological homogeneity and predictable treatment response patterns of HER2-amplified tumors following targeted therapy with trastuzumab, whereas the comparatively lower (though still clinically useful) performance in TNBC reflects this subtype’s well-documented biological heterogeneity and variable clinical trajectories²². The HR+ subtype demonstrated intermediate performance, consistent with its diverse spectrum ranging from highly endocrine-sensitive luminal A-like tumors to more aggressive luminal B-like disease. Our model was designed to predict overall survival and any disease recurrence (including local, regional, and distant sites). While this comprehensive approach captures a broader clinical picture, it does not distinguish between cancer-specific mortality and other causes of death.

The heterogeneity in optimal algorithm choice across subtypes and time horizons (SVM excelling in HER2 + 2-year and TNBC, Random Forest in HR + 2-year, Stacking in long-term predictions) underscores the importance of algorithm-agnostic evaluation rather than assuming a single method will universally perform best. This pattern suggests that biological context influences not only which features are prognostically relevant, but also which algorithmic approaches best capture the underlying risk structure.

Our models achieve performance consistent with published machine learning approaches for breast cancer prognosis, though direct numerical comparisons should be interpreted cautiously, as differences in AUC may reflect dataset characteristics, event rates, and cohort composition rather than true model superiority. A recent systematic review by Silveira et al.²³ found that machine learning methods, especially support vector machines and neural networks combining clinical and imaging data, improve recurrence prediction accuracy, though dataset diversity and evaluation standardization remain challenges. Similarly, a meta-analysis of 34 studies (67,560 patients) by Lu et al.²⁴ reported moderate predictive performance for ML-based recurrence models (pooled validation C-index ~0.77), emphasizing the need for larger multicenter studies and external validation to improve long-term prognostic accuracy. Our models perform favorably against these benchmarks: the 2-year model achieved AUC 0.859 (general cohort) and the 10-year model AUC 0.814, demonstrating sustained discrimination over an extended horizon where predictive accuracy typically declines.

While some studies report remarkably high AUCs, Zuo et al.¹² (AUC 0.987, 5-year, Chinese cohort) and Lou et al.²⁵ (AUC 0.98, 10-year external validation in over 1,100 Taiwanese patients), these results often reflect differences in endpoints, favorable data distributions, or limited cohort sizes, and may not generalize broadly. Lou et al. identified surgeon volume, hospital volume, and tumor stage as the strongest predictors²⁵, highlighting the importance of contextual factors that may vary across healthcare systems. Modern multimodal approaches incorporating imaging and multi-omics data have achieved higher discrimination (AUC 0.90–0.96)^11,26, yet require specialized infrastructure and high-cost tests. Our work distinguishes itself by attaining competitive accuracy using only baseline clinical data and routine blood counts available at diagnosis worldwide; genomic assays, advanced imaging, and longitudinal electronic health records are not required. This balance of accuracy and accessibility represents a key pragmatic advantage for real-world implementation across diverse healthcare settings.

A distinctive feature of our approach is the systematic integration of CBC-derived inflammatory and hematologic markers (NLR, PLR, MLR, RDW, MCHC) with standard clinical variables. Traditional breast cancer prognostic models focus on tumor characteristics and patient demographics, rarely incorporating routine blood metrics. Our SHAP analysis revealed that while tumor stage variables dominate prediction, these blood biomarkers provide meaningful additive prognostic value. Notably, RDW (a measure of red blood cell size variability) emerged as a top contributor to 10-year risk. Elevated RDW ( > 13.5%) has been associated with ~1.7-fold higher recurrence risk in breast cancer¹², consistent with our findings. This is further corroborated by a meta-analysis of 4884 breast cancer patients demonstrating that high RDW is significantly associated with worse overall survival (HR = 2.12, 95% CI 1.47–3.08) and disease-free survival (HR = 1.77, 95% CI 1.32–2.37), with particularly strong effects in surgery-only patients (HR = 2.41). Elevated RDW reflects red cell size variability arising from inflammation, oxidative stress, or nutritional deficiencies, and is associated with larger tumor size, worse differentiation, deeper invasion, and lymph node metastasis^27,28. Similarly, low MCHC and elevated inflammatory ratios (NLR, PLR, MLR) increased predicted risk, reflecting the prognostic impact of anemia and systemic inflammation on cancer outcomes^7,14. Higher hemoglobin levels were associated with lower predicted risk, consistent with evidence that pre-treatment anemia is prevalent in breast cancer and that hemoglobin trajectories during early treatment predict long-term 10-year survival outcomes²⁹. These findings are further supported by meta-analyses demonstrating that elevated PLR is associated with lower pathologic complete response rates and poor survival (HR 0.77, 95% CI 0.67–0.88) in patients treated with neoadjuvant chemotherapy^20,21, and that high LMR confers improved outcomes, particularly in luminal subtypes¹⁹. Since RDW reflects potentially modifiable states (inflammation, nutritional deficiencies, anemia), future interventional studies could investigate whether optimizing these parameters during follow-up improves recurrence outcomes, transforming RDW from a passive prognostic marker into an active therapeutic target^16,20. Importantly, our model joins a growing trend toward explainable AI in oncology. Like Zuo et al., who used 11 algorithms on clinicopathological and laboratory data and identified an AdaBoost model as optimal¹², we employed SHAP interpretability to demonstrate that biologically plausible factors drive predictions. Their model achieved high AUC and was the first to employ SHAP for interpretability in breast cancer recurrence prediction, confirming the prominence of features like tumor size and CA-125 for breast cancer¹². This interpretability is crucial for clinical adoption, allowing clinicians to understand and verify which factors (e.g., advanced stage, high NLR) contribute to each patient’s risk score. The ability to provide patient-specific explanations distinguishes modern AI tools from “black box” algorithms and facilitates integration into clinical decision-making. However, while SHAP quantifies which features drive predictions, it does not elucidate the underlying biological mechanisms linking these markers to disease progression. Mechanistic studies integrating longitudinal biomarker trajectories with molecular profiling are needed to determine whether elevated inflammatory markers represent causal pathways or proxies for unmeasured biological processes.

Our systematic evaluation of feature engineering strategies revealed subtype-specific patterns that provide insights into biological heterogeneity. Embedding augmentation (PCA/UMAP-derived features with GMM clustering) improved performance specifically in HR+ disease at 10 years (AUC gain: 0.8201 vs 0.8015) and across both horizons in TNBC, suggesting these subtypes harbor latent biological complexity not fully captured by conventional clinical features¹⁷. Conversely, HER2+ models performed best without embedding (10-year AUC 0.8773 vs. 0.8701 with embedding), indicating that this subtype’s prognostic signal is dominated by readily observable clinical characteristics, likely reflecting the strong predictive power of HER2 status and treatment responsiveness. These findings suggest that unsupervised dimensionality reduction and clustering capture meaningful biological variation in heterogeneous subtypes (HR + , TNBC) where no single marker dominates, but may introduce noise when applied to more homogeneous disease profiles (HER2 + ). This has practical implications: simpler models may suffice for HER2+ disease, while more complex feature engineering may be warranted for subtypes with greater intrinsic heterogeneity.

The practical advantage of our model lies in its exclusive reliance on universally available data collected during routine breast cancer workup. This design enables deployment in resource-constrained settings where genomic profiling and advanced imaging analytics are financially or logistically prohibitive. While genomic assays such as Oncotype DX and MammaPrint have demonstrated incremental prognostic value, with tools like RSClin combining the 21-gene score with clinical-pathological features to provide enhanced 10-year distant recurrence prediction, such tests remain unavailable for most patients in public healthcare cohorts globally due to cost and access barriers³⁰. Our model could serve as a first-tier risk assessment tool in settings lacking access to genomic profiling, or as a complementary approach where such tests are available. Future research integrating blood-based biomarkers with genomic data could create tiered models adaptable to varying resource availability, though validation of such hybrid approaches remains to be established.

The model could be integrated into electronic health record systems to automatically calculate personalized recurrence risk upon completion of baseline diagnostics, helping clinicians tailor treatment intensity and surveillance schedules. This approach aligns with emerging evidence that AI-based risk tools can improve clinical decision-making when properly implemented. Arasu et al.³¹ demonstrated that AI models using screening mammograms outperformed traditional clinical risk tools in predicting 5-year breast cancer incidence (AUC ~ 0.63–0.67 vs. 0.61), with consistent performance across diverse patient subgroups, highlighting the potential for EHR-integrated AI to broaden access to accurate and equitable risk assessment.

By requiring no additional testing beyond standard care, our approach is inherently equitable and scalable, potentially reducing care disparities between high-resource academic centers and community hospitals lacking facilities for multi-gene assays. This focus on accessible data echoes recent efforts to leverage EHR-derived features for prognosis without costly specialized tests³². Our work extends this paradigm by demonstrating that even long-term outcomes (10 years post-diagnosis) can be predicted with clinically useful accuracy from routine inputs alone. This represents a pragmatic application of machine learning aligned with value-based medicine principles by extracting maximum prognostic information from data already being collected.

Several limitations warrant acknowledgment. First, our single-center retrospective design may introduce selection bias and limit generalizability. External validation in diverse geographic regions and healthcare systems is essential; we are establishing collaborations for multicenter validation studies. Second, our feature set excluded genomic assay data, which provide proven incremental value but were unavailable for most patients in our public healthcare cohort. Third, we predict a composite outcome (recurrence or all-cause mortality) rather than cancer-specific events; future work with larger datasets could distinguish these endpoints for more granular risk stratification.

Despite these limitations, our findings carry important implications. The model’s inputs, CBC indices, and standard pathology are routinely collected worldwide at minimal cost. Unlike genomic panels, our approach can serve as a low-cost decision support tool in settings where expensive assays are infeasible, helping identify high-risk patients who might benefit from intensified therapy or surveillance. With appropriate validation, such algorithms could improve clinical judgment: high-risk patients (e.g., young TNBC patients with elevated inflammatory markers) might be candidates for extended treatment or preventive trial enrollment, while low-risk patients could potentially forgo unnecessary interventions. Integration into clinical workflows via EHR-embedded calculators could facilitate widespread adoption, bridging the gap between high-resource academic centers and community settings in delivering personalized survivorship care.

Methods

Cohort description

Our study retrospectively analyzed a cohort of 4277 breast cancer patients treated at the Instituto do Câncer do Estado de São Paulo (ICESP, Brazil) between 2008 and 2022.

Inclusion criteria

Female patients with a single primary tumor of the breast and at least one pre-treatment Complete Blood Count (CBC) data point with at least 2 years of follow-up for the 2-year models or 5 years of follow-up for the 10-year model (or recurrence or death before these periods). The minimum 5-year follow-up requirement for the 10-year outcome analysis was established because patients without evidence of recurrence are discharged from ICESP within 5 years of diagnosis per institutional protocol, consistent with standard Brazilian cancer center surveillance practices, where follow-up is most intensive in the first 5 years, when recurrence risk is highest³³. Patients without events by year 5 who were lost to follow-up were assumed to remain disease-free through year 10. We checked the vital status of the cohort at the Brazilian CPF (Individual Taxpayer Registry) Status Certificate (equivalent to Social Security Number) and date of birth for the register status as “Regular” (valid) or deceased. In deceased status, the year of death is available, and we apply 07/01/yyyy date as the date of death. Figure 4 depicts a flowchart of inclusion and exclusion criterias. We considered death by any cause as an event. Suppl. Fig. 7 shows more detailed recurrence in local, regions, and distance site, from the cohort.

Fig. 4. Data Pipeline and Model Workflow.

Starting from 4867 breast cancer patients at ICESP (São Paulo, Brazil; 2008–2022), 590 were excluded (incomplete hemogram, < 2 years follow-up, or DCIS/IS/PAGET/LCIS), yielding cohorts of 4277 (2-year) and 3675 (10-year) patients, with subtype-specific models for HR⁺, HER2⁺, and TNBC. Outcomes included recurrence (local, regional, or distant) or death within the specified timeframe. Sixteen features were engineered from clinical variables and hematologic biomarkers, with dimensionality reduction via PCA, UMAP, and GMM clustering. Random Forest, XGBoost, LightGBM, and SVM classifiers were combined through a stacking meta-learner, trained on a stratified 60/20/20 split with hyperparameter tuning via GridSearchCV (5-fold StratifiedKFold), and evaluated after retraining on the combined train+validation set (80%). Performance was assessed by ROC-AUC, PR-AUC, Brier Score, sensitivity, specificity, PPV, and NPV (95% CI via bootstrap, n = 1000). Classification used two thresholds: Youden’s J and a fixed 90% sensitivity cutoff. The incremental value of hematologic features was assessed by comparing clinical-only (A), hematologic-only (B), and combined (C) models via DeLong’s test.

Exclusion criteria

Patients with missing key features or stage 0 at diagnosis. Also, the following morphologies were excluded if alone: ductal carcinoma in situ (DCIS), lobular carcinoma in situ (LCIS), and PAGET disease. From the initial 4867 patients, 590 were excluded due to incomplete data, resulting in final analytical cohorts of 4,277 patients for the general model. Subtype-specific final cohorts included: HR+ (2588 patients), HER2+ (904 patients), and TNBC (785 patients) (Fig. 4). Patient demographics, tumor details, treatment timelines, and outcomes were extracted directly from ICESP’s electronic medical records (EMR) and maintained at our internal REDCap database. Laboratory values, including CBC and tumor markers, were routinely recorded pre-treatment. To protect patient confidentiality, all data were anonymized before analysis. Institutional Ethical approval CEP N° 0068 Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo - HCFMUSP for this study was granted under CAAE: 87092224.4.0000.0068. This study was conducted in accordance with the Declaration of Helsinki. Given the retrospective nature of the analysis and the use of fully anonymized data, the requirement for written informed consent was waived by the Ethics Committee of Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo (HCFMUSP).

This study follows a Type 1b development design according to TRIPOD guidelines, developing a multivariable prediction model without external validation³⁴. Sample size adequacy was evaluated beyond the traditional “10 events-per-variable” rule, which recent guidance cautions lacks theoretical support²⁴. With 4277 patients and 1087 outcome events (25% event rate), our dataset provided adequate statistical power for stable model development, accounting for the larger sample requirements of ensemble machine learning methods compared to traditional regression³⁵. For outcome definition, patients who experienced breast cancer recurrence or death within 2 or 10 years after diagnosis were classified as “high-risk of recurrence”, while those without recurrence or death during the respective follow-up period were classified as low-risk.

Clinical features

The final analytical cohort comprised 4277 breast cancer patients (Table 1). The mean age at diagnosis was 53 years (SD 13), with similar distributions across subtypes. Tumor subtype classification revealed HR+ as the most prevalent (60%), followed by HER2+ (21%) and TNBC (18%). Most patients presented with intermediate disease burden: 43% had node-negative disease (N0), while clinical stage distribution showed 23% stage I, 39% stage II, 20% stage III, and 12% stage IV. Metastatic disease at diagnosis (M1) was present in 12% of cases overall, with higher rates in HER2+ (15%) and TNBC (13%) compared to HR+ disease (11%). During follow-up, 1,087 patients (25%) experienced recurrence, with a median follow-up of 3.28 years (IQR 1.87) for the overall cohort. TNBC subtypes presented more recurrences and lower times to event than the other subtypes. Baseline complete blood counts showed mean values within normal ranges for most parameters (Table 1). However, inflammatory ratios demonstrated notable variation: NLR ranged from 0.3 to 44.2 (mean 2.20), PLR from 18 to 1,100 (mean 136), and MLR from 0.04 to 3.8 (mean 0.28). RDW-CV (Red Blood Cell Distribution Width - coefficient of variation) averaged 13.74% (SD 1.44), with subtle differences across subtypes. These inflammatory and erythrocyte indices formed key predictive features alongside clinical staging variables.

Table 1.

Clinical and laboratory characteristics of the study cohort, stratified by molecular subtype

Characteristic	Overall N = 4277	Luminal N = 2589	HER2-pos N = 903	Triple-negative N = 785	p-value¹
Age at diagnosis, Mean (SD)	53 (13)	55 (13)	51 (13)	52 (14)	< 0.001
Clinical Stage T, n (%)					< 0.001
1	1,005 (23%)	742 (29%)	169 (19%)	94 (12%)
2	1,667 (39%)	1,082 (42%)	332 (37%)	253 (32%)
3	851 (20%)	429 (17%)	202 (22%)	220 (28%)
4	754 (18%)	336 (13%)	200 (22%)	218 (28%)
Clinical Stage N, n (%)					< 0.001
0	1,820 (43%)	1,207 (47%)	306 (34%)	307 (39%)
1	1,441 (34%)	859 (33%)	326 (36%)	256 (33%)
2	703 (16%)	378 (15%)	185 (20%)	140 (18%)
3	313 (7.3%)	145 (5.6%)	86 (9.5%)	82 (10%)
Clinical Stage M, n (%)					0.003
0	3,752 (88%)	2,304 (89%)	766 (85%)	682 (87%)
1	525 (12%)	285 (11%)	137 (15%)	103 (13%)
Estrogen receptor, n (%)					< 0.001
Negative	1,074 (26%)	32 (1.2%)	257 (34%)	785 (100%)
Positive	3,044 (74%)	2,547 (99%)	497 (66%)	0 (0%)
Progesterone receptor, n (%)					< 0.001
Negative	1,344 (33%)	216 (8.5%)	343 (46%)	785 (100%)
Positive	2,743 (67%)	2,339 (92%)	404 (54%)	0 (0%)
HER-2, n (%)					< 0.001
0 (Negative)	2,930 (70%)	2,218 (88%)	0 (0%)	712 (92%)
+ (Negative)	97 (2.3%)	79 (3.1%)	1 (0.1%)	17 (2.2%)
++ (Inconclusive)	265 (6.3%)	212 (8.4%)	11 (1.2%)	42 (5.4%)
+++ (Positive)	890 (21%)	0 (0%)	890 (99%)	0 (0%)
Recurrence, n (%)					< 0.001
Recurrence	1,087 (25%)	567 (22%)	248 (27%)	272 (35%)
Recurrence Free	3,190 (75%)	2,022 (78%)	655 (73%)	513 (65%)
Follow up days to recurrence, Mean (SD)	828 (747)	1,010 (871)	731 (570)	537 (439)	< 0.001
Local recurrence, n (%)	311 (7.3%)	143 (5.5%)	69 (7.6%)	99 (13%)	<0.001
Regional recurrence, n (%)	258 (6.0%)	113 (4.4%)	69 (7.6%)	76 (9.7%)	<0.001
Distant recurrence, n (%)	754 (18%)	407 (16%)	164 (18%)	183 (23%)	< 0.001
Vital status, n (%)					< 0.001
Alive with cancer	182 (4.3%)	107 (4.1%)	44 (4.9%)	31 (3.9%)
Alive, NOS	2,697 (63%)	1,696 (66%)	583 (65%)	418 (53%)
Death by cancer	1,239 (29%)	690 (27%)	246 (27%)	303 (39%)
Death, NOS	159 (3.7%)	96 (3.7%)	30 (3.3%)	33 (4.2%)
Follow up days, Mean (SD)	2,313 (1,220)	2,422 (1,171)	2,346 (1,238)	1,918 (1,280)	< 0.001
Erythrocyte, Mean (SD)	4.54 (0.44)	4.55 (0.44)	4.56 (0.42)	4.48 (0.46)	< 0.001
Hematocrit, Mean (SD)	39.5 (3.7)	39.7 (3.7)	39.8 (3.4)	38.8 (3.8)	< 0.001
Hemoglobin, Mean (SD)	13.15 (1.37)	13.19 (1.38)	13.25 (1.26)	12.88 (1.44)	< 0.001
Leucocyte, Mean (SD)	7.30 (2.31)	7.20 (2.25)	7.40 (2.29)	7.49 (2.54)	0.010
Eosinophils, Mean (SD)	0.19 (0.19)	0.18 (0.19)	0.19 (0.18)	0.19 (0.20)	0.71
Lymphocyte, Mean (SD)	2.18 (0.81)	2.19 (0.86)	2.21 (0.71)	2.14 (0.74)	0.059
Neutrophil, Mean (SD)	4.34 (1.91)	4.25 (1.82)	4.40 (1.94)	4.54 (2.17)	0.011
Platelets, Mean (SD)	267 (72)	262 (69)	270 (69)	280 (84)	< 0.001
NLR, Mean (SD)	2.20 (1.44)	2.14 (1.31)	2.22 (1.52)	2.39 (1.69)	0.006
PLR, Mean (SD)	136 (66)	132 (59)	136 (64)	148 (88)	< 0.001
MLR, Mean (SD)	0.28 (0.15)	0.27 (0.14)	0.28 (0.14)	0.31 (0.17)	< 0.001
RDW.CV, Mean (SD)	13.74 (1.44)	13.79 (1.51)	13.59 (1.19)	13.75 (1.48)	0.038
RDW.SD, Mean (SD)	43.2 (4.0)	43.4 (4.1)	42.8 (3.4)	43.0 (4.0)	< 0.001
CHCM, Mean (SD)	33.24 (1.20)	33.25 (1.22)	33.31 (1.13)	33.14 (1.22)	0.026
HCM, Mean (SD)	29.04 (2.33)	29.08 (2.43)	29.12 (2.08)	28.82 (2.26)	0.001

Clinical and Hematologic Features Included in the Predictive Models. The table displays the set of baseline variables incorporated into model development, encompassing patient demographics, tumor stage, receptor status, and hematologic indices derived from pre-treatment complete blood counts. In addition to standard erythrocyte, leukocyte, and platelet parameters, we explicitly calculated inflammatory ratios such as neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and monocyte-to-lymphocyte ratio (MLR), along with red cell distribution indices. These features integrate established clinical predictors with readily available blood biomarkers, forming the 16 candidate variables subjected to feature selection for model training. NOS Not Otherwise Specified. p-values indicate the difference between molecular subtypes of each parameter evaluated. p < 0.05 were considered significant. Death and recurrence events are not censored by time in this table.

¹Kruskal-Wallis rank sum test; Pearson’s Chi-squared test.

Data storage and access

We securely stored the clinical and laboratory data on institutional servers managed by ICESP’s IT department. Clinical data from our REDCap system were exported as CSV files and integrated with records from the hospital EMR. Access to these datasets was strictly controlled, requiring authorized credentials to protect patient privacy. All analyses were performed in Python (version 3.10). Prior to analysis, all patient-identifying information was removed and replaced with anonymized identifiers, ensuring that no personal data was accessible during the computational workflow. Throughout the study, we adhered to the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) guidelines to ensure transparency, reproducibility, and methodological rigor³⁴.

Data processing and feature engineering

We utilized Python (Pandas) for data cleaning and preparation. Demographics, clinical records, therapies, and laboratory tables were linked via unique patient IDs. This threshold aligns with pragmatic approaches in clinical prediction studies, though we acknowledge that multiple imputation methods could potentially retain more patients and reduce bias³⁶. For transparency per TRIPOD guidelines, we report that key clinical variables (tumor stage, receptor status) and laboratory-derived ratios were excluded on a per-patient basis when source values were unavailable. For the 4277 patients in the final cohort, continuous variables were standardized (z-scores) after trimming outliers beyond 1.5× interquartile range. Guided by prior clinical evidence, we derived inflammatory indices from the complete blood count: neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), monocyte-to-lymphocyte ratio (MLR), and basophil-to-lymphocyte ratio (BLR). We also included red cell distribution width (RDW) and all available hemogram features. Key clinical variables (e.g., tumor stage, diagnosis date, recurrence, death) were sourced from the hospital registry/EMR, with categorical fields encoded for modeling. For survival analyses, time from diagnosis to recurrence or last contact was recorded in months; patients without events were censored at last follow-up. We plotted Kaplan–Meier curves with log-rank tests and fitted Cox proportional hazards models to estimate hazard ratios with 95% confidence intervals. All tests were two-tailed with α = 0.05.

Model development and validation

The General Model was trained with all breast cancer cases. Each molecular subtype was trained separately from each other and from the General model. We tested several machine learning algorithms to predict patient recurrence or death within 2 or 10 years, including Random Forest, XGBoost, and a non-linear SVM. These algorithms are frequently chosen in clinical prediction studies for their robust performance and algorithmic diversity, random forests excel in handling nonlinear interactions, XGBoost in optimizing the bias-variance tradeoff, and SVM in capturing complex decision boundaries³⁷. Initially, we noticed the XGBoost model tended to overfit the rare-event cases, leading us to explore an ensemble approach. We implemented a stacked generalization (stacking) ensemble³⁸, which combines multiple base learners by training a meta-learner on their outputs. This technique has been shown to perform at least as well as the single best base model, often outperforming individual classifiers by exploiting their complementary strengths³⁹. For the meta-learner, we chose logistic regression to optimally weight base model predictions, training it specifically on validation predictions.

During model training, class imbalance was a significant challenge, as patients without events greatly outnumbered those with recurrence or death. We addressed this through two complementary strategies evaluated in parallel. In the primary pipeline (Table 2), all classifiers were configured with cost-sensitive learning: Random Forest, LightGBM, and SVM used class_weight = ‘balanced’, which automatically adjusts the loss function to penalize misclassification of minority-class samples in proportion to their underrepresentation, while XGBoost used scale_pos_weight calibrated to the inverse class ratio in the training set. This approach achieves similar objectives to resampling, improving model sensitivity to rare events²⁹, without introducing synthetic observations.

In a secondary pipeline (Table 3), we additionally evaluated three oversampling variants. SMOTE (Synthetic Minority Over-sampling Technique)²⁸ generates synthetic minority samples by interpolating between a minority instance and its k-nearest neighbors, uniformly across the feature space. Borderline-SMOTE restricts synthesis to minority samples near the decision boundary, those whose nearest neighbors include majority class instances, thereby focusing augmentation on the most ambiguous cases. ADASYN (Adaptive Synthetic Sampling)^30,40 extends this concept by adaptively weighting harder-to-learn minority examples, generating more synthetic samples in regions where the local class ratio is most imbalanced. All three techniques were applied exclusively within training folds using imbalanced-learn’s ImbPipeline, ensuring that synthetic samples were never generated from or applied to validation or test data.

Model development for the primary pipeline (Table 2) followed a structured sequence: (1) Data were split into training (60%), validation (20%), and test (20%) sets using stratified random sampling to preserve outcome proportions (train_test_split, scikit-learn, random_state = 42). (2) For each base classifier (Random Forest, XGBoost, LightGBM, SVM), hyperparameters were optimized via exhaustive grid search with 5-fold stratified cross-validation on the training set only (GridSearchCV, scikit-learn, scoring by ROC-AUC). (3) The best hyperparameter configuration for each algorithm was selected based on validation set ROC-AUC. (4) A stacking ensemble was constructed by combining the four optimized base learners with a logistic regression meta-learner (StackingClassifier, scikit-learn), using 5-fold cross-validation internally to generate out-of-fold predictions for meta-learner training. The ensemble was fitted on the training set (60%). (5) Final performance was evaluated on the held-out test set (20%), which was never used during any training, tuning, or model selection step. The models applied to the test set are the same models trained on 60% of the data, with hyperparameters and thresholds fixed during the selection phase.

To capture nonlinear structure beyond the raw input variables, we also evaluated models incorporating dimensionality reduction features (referred to as “embed” models in Table 3). In this pipeline, Principal Component Analysis (PCA) was applied for linear variance-preserving projection (6 components), Uniform Manifold Approximation and Projection (UMAP) for nonlinear embedding (3 components), and Gaussian Mixture Model (GMM) clustering for probabilistic latent group assignment (2 components). These derived features were appended to the original 16 clinical and hematologic variables via FeatureUnion (scikit-learn), and models were trained with and without these representations (Table 3) to assess their incremental value. All dimensionality reduction transformations were fitted on the training data only and applied to validation and test sets to prevent data leakage. Each combination of oversampling strategy (SMOTE, Borderline-SMOTE, ADASYN) and feature set (with and without embed) was evaluated, yielding six configurations per model and time horizon.

Models were evaluated by several metrics, including area under the curve (AUC) from the receiver operating curve (ROC) (AUC ROC), which shows a balance between specificity and sensitivity. Area Under the Precision-Recall Curve (PR AUC), which summarizes a trade-off between precision (positive predictive values) and recall, good for models with less positive events as in our case. Model calibration was evaluated using the Brier score⁴¹, a proper scoring rule that measures the mean squared difference between predicted probabilities and observed outcomes, with values closer to 0 indicating better calibration.

For clinical interpretation, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were computed at two operating points. The first was the Youden’s J statistic threshold⁴², which maximizes the sum of sensitivity and specificity and provides a balanced classification cutoff, though it implicitly assumes equal weight for false positives and false negatives. The second was a clinically motivated threshold targeting ≥90% sensitivity, selected to minimize missed high-risk patients in a potential treatment de-escalation setting, consistent with approaches used in recent breast cancer prediction studies⁴³. At this high-sensitivity threshold, specificity is traded for detection rate, resulting in more false positives but fewer missed recurrences. All metrics were reported with 95% confidence intervals derived from 1,000-iteration bootstrap resampling on the test set⁴⁴. Pairwise comparisons of ROC-AUC between models were performed using DeLong’s test⁴⁵, with p < 0.05 considered statistically significant.

Web-based clinical decision support tool

The best-performing models were deployed as an interactive web platform (BreastOncoPredict), publicly accessible at https://breastoncopredict.com. The tool accepts the 16 clinicopathologic and hemogram variables used during training and returns individualized risk estimates with SHAP-based feature attributions. The platform is intended for research and educational purposes only and is not approved for clinical use.

Author contributions

P.H.M. and L.R.C.B. conceived and designed the study. P.H.M. developed the machine learning models, performed statistical analyses, and created visualizations. A.S.O.C., R.O., J.G., and W.A.L.L. contributed to data curation, clinical data extraction, and validation. R.C.B. and B.S.M. provided clinical expertise and interpretation of results. A.F.R. contributed to methodological design and computational infrastructure. F.S. provided guidance on machine learning methodology. T.A.R. developed the web platform. R.C. and L.R.C.B. supervised the project and provided critical revision of the manuscript. P.H.M. drafted the initial manuscript. All authors contributed to manuscript revision, read, and approved the final version.

Data availability

The datasets generated and/or analysed during the current study are not publicly available due to patient privacy and institutional data protection policies but are available from the corresponding author on reasonable request and with permission of ICESP.

Code availability

The underlying code for machine learning model development, training, and evaluation used in this study is available in the GitHub repository breast-cancer-recurrence-predictive-model at https://github.com/ipatriciahonorato/breast-cancer-recurrence-predictive-model. The algorithm is under evaluation for registration at Brazilian National Institute of Industrial Property (INPI).

Competing interests

R.C.B. declares speaker fees and/or honoraria for consulting or advisory functions: Daiichi-Sankyo, Nestle Health Science, Addium, Gilead, MSD, BMS, AstraZeneca, Ache, Pfizer, Roche, Novartis, Libbs, Lilly, GSK; financial support for educational programs and symposia: AstraZeneca, Daiichi-Sankyo, MSD, Lilly, Roche, Novartis; institutional Research grant: Novartis, AstraZeneca, Gilead. The other authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41523-026-01005-z.