Breast cancer is the most common cancer affecting women worldwide and a leading cause of cancer death. Despite advances in treatment, approximately 30% of breast cancer patients experience disease recurrence during follow-up, creating an urgent need for precise, individualized prognosis tools.
Traditional recurrence prediction relies on clinicopathologic factors such as tumor size, axillary nodal status, histologic grade, and receptor status (ER, PR, HER2). Online tools like Adjuvant! Online and PREDICT estimate survival from these baseline characteristics, but they do not update their predictions as new clinical information becomes available over time.
A critical limitation of existing models is that they capture only a snapshot at diagnosis. They cannot reflect changes such as developing comorbidities, evolving laboratory values, or imaging findings that emerge during the surveillance period and may themselves affect recurrence risk.
This study addressed this gap by developing a deep learning recurrence prediction model that incorporates both baseline clinical data at diagnosis and serial follow-up measurements, enabling real-time personalized recurrence probability estimates throughout post-treatment surveillance.
This retrospective analysis drew from a registry of 13,117 breast cancer patients who received curative surgery at Samsung Medical Center between January 2000 and December 2016. All patients had undergone adjuvant treatment and were followed prospectively as part of standard surveillance.
The dataset began with 325 clinical data elements covering demographics, baseline pathologic characteristics, receptor status, treatment details, and serial follow-up measurements including laboratory tests and imaging results. This breadth of longitudinal data distinguishes this dataset from those used in prior machine learning models.
Patients who received neoadjuvant chemotherapy before surgery, had ductal carcinoma in situ, were male, or had evidence of distant metastases at baseline were excluded, ensuring the cohort represented a surgical adjuvant population with curative intent.
The dataset was split 60/20/20 into training, validation, and test sets, preserving the proportion of recurrence events (9.2% overall) in each set to ensure balanced model training and evaluation.
The team implemented a Weibull Time To Event Recurrent Neural Network (WTTE-RNN), a deep learning framework designed for survival analysis with time-stamped sequential data. Unlike conventional logistic regression models, WTTE-RNN models recurrence as a probability that evolves with each new patient encounter.
Recurrent neural networks (RNNs) are uniquely suited for longitudinal clinical data because they maintain a hidden state that accumulates information from prior time points. At each follow-up visit, the model uses both the current measurements and the patient's complete history to update its recurrence probability estimate.
The model incorporated both time-independent features -- such as age at surgery, pathologic stage, and receptor status -- and time-dependent features -- such as serial laboratory values (complete blood count, liver enzymes, tumor markers), mammography findings, and breast sonography results. A grid search identified a 12-month lookback window as optimal for incorporating recent visit data.
The network architecture used two hidden layers (32 and 20 cells), gated recurrent units, hyperbolic tangent activation functions, and a dropout rate of 0.25 to prevent overfitting. The Weibull distribution was used to model the hazard function because it allows the risk of recurrence to vary non-linearly over time, reflecting the biological reality of breast cancer relapse patterns.
After 200 training epochs, the model achieved a Harrell's C-index of 0.92 on the training set and 0.89 on both the validation and test sets. The C-index measures how well the model ranks patients by predicted recurrence time, with 1.0 representing perfect discrimination and 0.5 representing random chance.
Time-specific performance was strong across all evaluated windows: AUC of 0.90 at 2 years, 0.91 at 5 years, and 0.91 at 7 years. This consistency across time points suggests the model captures enduring biological features rather than transient patterns specific to short-term follow-up.
When compared against three alternative machine learning approaches, the WTTE-RNN substantially outperformed all competitors. Logistic regression achieved AUCs of only 0.69 to 0.72. Random forest and gradient boosting models produced AUCs of 0.80 to 0.83. Only the deep learning model exceeded an AUC of 0.90 at any time point.
Subgroup analysis stratifying patients by pathologic T stage, N stage, receptor subtype, EGFR status, and CK5/6 status showed that the model's predicted recurrence proportions closely matched observed rates, with median absolute error (MAE) as low as 3.5% across most subgroups.
Feature selection using time-dependent Cox regression and backward stepwise selection identified 32 variables significantly associated with breast cancer recurrence. These spanned baseline clinicopathologic factors (13 features), adjuvant treatment characteristics (4 features), and serial follow-up measurements (15 features).
At the time of surgery, pathologic T stage, N stage, and lymphovascular invasion emerged as the strongest stationary predictors. Notably, individual receptor status variables (ER, PR, HER2) did not independently predict recurrence in the multivariate model, consistent with the hypothesis that appropriate subtype-matched treatment neutralizes their prognostic effect.
Among follow-up laboratory measurements, elevated serum glucose, absolute neutrophil count, and CA-15-3 tumor marker increased recurrence risk. Conversely, higher white blood cell count, hemoglobin, and total protein levels were associated with a protective effect -- findings that align with the role of systemic inflammation and nutritional status in cancer biology.
Mammography and breast ultrasound findings during surveillance predictably influenced recurrence estimates. However, the model's strength lies in integrating all routine test data that would be available at each standard follow-up visit, without requiring additional specialized or costly molecular testing.
Most prior machine learning prognosis models for breast cancer used only baseline data from the time of surgery. This study is notable as one of the first to incorporate longitudinal surveillance data -- serial lab results, imaging findings, and treatment outcomes -- into a dynamic prediction framework.
The use of the Weibull distribution within the RNN framework allows for a non-constant hazard rate over time. This is biologically appropriate because the risk of recurrence for different breast cancer subtypes peaks at different intervals: hormone receptor-positive cancers may recur years after initial treatment, while triple-negative cancers tend to recur earlier.
The model is designed for integration into electronic medical records (EMR) and could display updated recurrence risk probabilities at each scheduled follow-up visit. This would allow clinicians to identify patients whose risk has meaningfully increased, potentially warranting more intensive surveillance or earlier therapeutic intervention.
An important finding is that the model's predictions remained accurate across diverse patient subgroups defined by tumor subtype and pathologic stage, with errors rarely exceeding 3.5%. This subgroup consistency is a prerequisite for clinical deployment in a heterogeneous patient population.
The primary limitation of this study is its single-institution design. Although the cohort of 13,117 patients is large, the patient population, treatment protocols, and follow-up practices at Samsung Medical Center in South Korea may not fully represent breast cancer populations at other centers worldwide.
External validation using datasets from multiple institutions with diverse patient demographics, treatment histories, and follow-up protocols is an essential prerequisite before the model can be deployed more broadly. Prospective validation studies are also needed to confirm that real-time recurrence estimates translate into actionable clinical benefits.
The model relies on routine data already collected in standard clinical care, which is both a strength (no additional testing required) and a potential limitation: patients with fewer follow-up visits or missing data provide less temporal information for the model to learn from.
Future refinements could explore incorporating genomic data, imaging biomarkers, and patient-reported outcomes to further improve prediction accuracy. The concept of a real-time, continuously updating recurrence risk score could ultimately transform how post-treatment breast cancer surveillance is personalized and prioritized.