Deep Neural Network Classifies Metastatic Breast Cancer Using Four Predictors

A deep feedforward neural network can predict metastatic breast cancer occurrence across multi-year clinical horizons using only four selected predictors, matching the predictive performance of models reliant on dozens of clinical variables. In a retrospective computational study evaluating clinical records from the Lynn Sage Metastasis database, researchers developed a specialized deep learning framework designed to systematically shrink input complexity without sacrificing discriminative power. By coupling hyperparameter optimization with demographic-aware feature selection and structured architectural pruning, the investigative team compressed an initial thirty-one-variable clinical profile down to four essential features. Across five-, ten-, and fifteen-year follow-up horizons, the resulting compact neural architectures achieved high overall classification accuracy and strong area under the receiver operating characteristic curve metrics. This structural reduction demonstrates that high-dimensional clinical datasets often harbor significant statistical redundancy, confirming that compact deep learning architectures can retain critical prognostic signals when guided by disciplined pruning rules.

Metastatic progression accounts for the vast majority of breast cancer fatalities worldwide, driving an ongoing clinical requirement for accurate, interpretable, and computationally accessible risk-stratification instruments. Clinical outcomes and mortality patterns diverge sharply across racial and ethnic cohorts, with epidemiological surveys consistently demonstrating higher incidence and mortality rates among non-Hispanic Black women compared with other groups, while non-Hispanic Asian women exhibit lower observed rates. These recorded disparities reflect complex interactions between social determinants of health, differential access to early diagnostic screening, treatment variations, and biological characteristics, including the higher prevalence of aggressive triple-negative breast cancer among Black women. However, electronic health records frequently suffer from inconsistent, incomplete, or missing demographic documentation stemming from administrative variations, institutional privacy policies, and time constraints during patient intake. To address these systemic gaps without introducing algorithmic bias, the investigators sought to examine how clinical predictors correlate with demographic categories and whether models can maintain diagnostic accuracy when restricted to compact feature sets.

The experimental investigation utilized retrospective patient cohorts derived from the Lynn Sage Metastasis database, partitioned across three independent observation intervals designated as the five-year, ten-year, and fifteen-year datasets. Each dataset cataloged thirty-two total clinical variables, encompassing thirty-one prognostic candidate predictors alongside a single binary endpoint indicating whether distant metastasis emerged within the assigned follow-up timeframe. In the five-year cohort, patients were categorized as metastatic if secondary disease manifested within five years of the primary diagnosis and non-metastatic otherwise. To resolve missing values without discarding partially documented patient histories, the data preparation pipeline implemented a five-nearest-neighbor imputation protocol before initiating computational partitioning. Categorical predictors underwent standardized encoding across all datasets, after which records were partitioned into separate training and testing subsets under a strict data isolation protocol that prevented testing information from informing either preprocessing parameters or training cycles.

The methodological strategy centered on quantifying the statistical association between each available prognostic factor and recorded demographic categories of race and ethnicity. For every candidate predictor, researchers calculated mathematical correlation coefficients by dividing the empirical covariance between the predictor and the demographic indicator by the product of their respective standard deviations. Predictors were subsequently ranked in descending order according to their absolute correlation magnitudes, producing distinct prioritized candidate lists for race and ethnicity. To synthesize these distinct demographic relationships into unified input subsets, the investigators formulated mathematical union and intersection operators. The union set aggregated variables demonstrating dominant statistical alignments with either racial or ethnic groupings, thereby ensuring that all prominent demographic-associated prognostic indicators were captured. Conversely, the intersection set isolated only those features sharing strong simultaneous correlations across both categories, providing an alternative analytical baseline to assess how demographic associations affect predictive classification.

Using these ordered feature subsets, the researchers constructed a novel deep feedforward neural network and conducted systematic hyperparameter optimization to maximize generalization capacity. The training framework employed the Adam optimization algorithm configured with an initial learning rate of 0.001 to govern adaptive weight adjustments while minimizing potential overfitting. To resolve gradient saturation and eliminate the dead-neuron vulnerabilities commonly observed in traditional rectified linear units, hidden layers incorporated Leaky Rectified Linear Unit activation functions that permit controlled, non-zero gradients for negative inputs. Network regularization combined L1 and L2 penalty constraints on weight matrices, preventing mathematical divergence while promoting parameter compression across intermediate layers. The optimization protocol also integrated batch normalization to accelerate training stability and self-attention mechanisms to dynamically assign focus weights to the most informative input representations, systematically evaluating mini-batch sizes of thirty-two and sixty-four to ensure consistent model convergence.

To transition the model toward minimal feature reliance, the research team instituted dynamic linear and exponential network pruning schemes designed to adapt network topology to varying input dimensions. The mathematical formulation bounded all pruning operations within a predefined operational corridor, establishing a minimum pruning rate of 0.1 and a maximum ceiling of 0.6. Under the linear pruning rule, structural reduction scaled gradually as a direct linear function of input feature counts, providing stable and predictable parameter elimination when processing wider predictor subsets. For exponential pruning, the framework applied an exponential decay function governed by an initial scaling factor of 0.5 to enforce more aggressive compression as feature counts dropped. In sequential pruning configurations, the framework evaluated an empirical threshold of ten features, executing linear reduction when inputs met or exceeded ten and transitioning to exponential pruning below that threshold to strip away redundant connections.

Experimental evaluation demonstrated that combining hyperparameter tuning with structured pruning enabled minimal feature subsets to match or exceed baseline models utilizing all thirty-one clinical predictors. When evaluating the union of race- and ethnicity-correlated predictors, the pruned deep neural network required only four features to equal the predictive benchmark in both the five-year and ten-year metastatic cohorts. In the five-year follow-up dataset, the four-feature union architecture reached a maximum receiver operating characteristic score of 0.89, outperforming conventional benchmark configurations reliant on full-dimensional clinical inputs. For the ten-year cohort, four union predictors similarly matched established multi-variable performance levels, achieving a peak metric of 0.81. In the fifteen-year dataset, relying on ethnicity-correlated predictors alone required seven features to match traditional performance levels, while sixteen union features achieved a maximum receiver operating characteristic score of 0.85, confirming that strategic feature selection preserves long-term prognostic validity.

To verify that these prognostic outcomes did not reflect stochastic initialization anomalies or data partitioning artifacts, the authors performed rigorous statistical stability assessments using repeated experimental trials and bootstrap resampling. Each optimized model architecture underwent ten independent experimental training runs governed by fixed pseudo-random generator seeds, from which the researchers calculated mean performance metrics and standard deviations. Over these repeated runs, the deep neural network registered mean area under the curve values of 0.7592 with a standard deviation of 0.0313 for the five-year horizon, 0.7842 with a standard deviation of 0.0303 for the ten-year horizon, and 0.8182 with a standard deviation of 0.0431 for the fifteen-year horizon. Subsequent non-parametric bootstrap analysis evaluating 5,000 resamples established 95% confidence intervals of 0.7249 to 0.8217 for the five-year test set, 0.7656 to 0.8579 for the ten-year test set, and 0.7812 to 0.9430 for the fifteen-year test set, quantifying the statistical bounds of model stability.

Complementing global discriminative metrics, the study integrated Shapley Additive Explanations to establish granular interpretability by quantifying how retained predictors influence individual classification outputs. Drawing on background reference samples of one hundred observations per cohort, the interpretability framework confirmed that the compressed four-predictor models maintained coherent biological and clinical attribution profiles. On final test partitions, the four-feature neural network achieved area under the precision-recall curve values of 0.2721 in the five-year cohort, 0.6828 in the ten-year cohort, and 0.9593 in the fifteen-year cohort, alongside overall classification accuracies of 0.8961, 0.7213, and 0.8079, respectively. The pronounced variation in precision-recall performance across time horizons highlighted how metastatic prevalence within the cohorts influences classification thresholds, underscoring the necessity of evaluating precision-recall dynamics alongside receiver operating metrics when deploying minimal feature algorithms in unbalanced clinical settings.

Beyond predictive discrimination, computational profiling confirmed that aggressive feature reduction substantially curtailed algorithmic execution overhead and computational hardware requirements. Across all three follow-up datasets, each final four-predictor neural network comprised exactly 6,771 trainable parameters, requiring an estimated computational complexity of approximately 0.012 million floating-point operations per evaluated patient sample. Training completion times measured 24.46 seconds for the five-year model, 16.40 seconds for the ten-year model, and 11.50 seconds for the fifteen-year model within a standard computing environment possessing 5.81 gigabytes of available random-access memory. Single-sample inference latency remained exceptionally low, logging processing speeds of 0.196 seconds, 0.197 seconds, and 0.176 seconds across the five-, ten-, and fifteen-year models, while total memory expansion during operational training remained restricted within a modest range of forty-eight to fifty megabytes, highlighting practical feasibility for resource-constrained clinical infrastructure.

Comparative benchmarking against established machine learning algorithms illustrated practical performance trade-offs, demonstrating that the compact deep feedforward architecture functions as a computationally lightweight alternative rather than a universally superior classifier. When benchmarked against ensemble gradient boosting frameworks on identical datasets, the optimized neural network achieved accuracy and receiver operating characteristic values of 0.896 and 0.537 for five-year prediction, 0.685 and 0.719 for ten-year prediction, and 0.807 and 0.657 for fifteen-year prediction. In contrast, XGBoost, LightGBM, and CatBoost recorded higher discriminative scores on the ten-year dataset, achieving values of 0.775, 0.774, and 0.780, respectively. Similarly, on the fifteen-year cohort, XGBoost and LightGBM attained 0.796 while CatBoost reached 0.768, outperforming the compressed neural network. These comparative outcomes emphasize that while gradient boosting models deliver superior discrimination on complex tabular matrices, the pruned neural network framework provides extreme topological compactness, interpretable feature attributions, and negligible computational overhead.

The study authors highlighted several methodological limitations that temper immediate clinical translation, foremost among them being the reliance on retrospective electronic records from a single database without independent external validation across distinct healthcare systems. Because complementary oncology datasets containing structured race and ethnicity metadata were unavailable, the generalizability of these specific four-feature sets across diverse global populations remains unconfirmed. Furthermore, the underlying datasets exhibited inherent class imbalances and variable demographic subgroup sizes, meaning small sample counts within certain racial categories could destabilize subgroup-specific performance metrics. The authors noted that removing explicit demographic variables does not guarantee algorithmic neutrality, as remaining clinical and socioeconomic indicators can function as unmeasured demographic proxies. Consequently, prospective clinical trials, broader demographic representation, and extensive external validation remain essential before feature-pruned deep neural networks can enter clinical decision workflows.

Subject of Research: Deep learning and feature pruning for metastatic breast cancer prediction across multi-year follow-up horizons

Article Title: High-performance deep learning for metastatic breast cancer classification using a minimal feature representation

Article References: Youssef, K., H. Abd El-Malek, A., Kanaya, H., & Abo-Zahhad, M. (2026). High-performance deep learning for metastatic breast cancer classification using a minimal feature representation. Discover Artificial Intelligence, 6(1), Article 1854. https://doi.org/10.1007/s44163-026-02442-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02442-2

Keywords: deep learning, metastatic breast cancer, feature reduction, neural network pruning, Lynn Sage Metastasis database, deep feedforward neural network, hyperparameter optimization, ROC-AUC evaluation, SHAP interpretability, gradient boosting comparison, health disparities, demographic predictors, bootstrap confidence intervals, computational efficiency, medical oncology AI, nearest neighbor imputation