Breast cancer is the most common cancer among women globally, and survival rates vary dramatically between wealthy and low-income countries. The gap is largely explained by the absence of advanced early detection systems in underdeveloped regions, where late-stage diagnoses are far more frequent and outcomes are correspondingly worse.
Earlier treatment is directly linked to lower mortality, making early and accurate diagnosis the single most important factor in breast cancer control. Conventional detection methods -- primarily imaging -- can be costly, inaccessible, and cause patient distress. Researchers have therefore explored whether routine blood tests and simple body measurements could serve as a low-cost alternative screening pathway.
Prior work identified several candidate biomarkers from blood chemistry and anthropometric measurements, including adipokines like leptin, adiponectin, and resistin, as well as insulin, glucose, and body mass index (BMI). These markers reflect the metabolic and hormonal environment that is known to influence breast cancer risk.
This study builds on that foundation by asking whether a modern artificial neural network architecture, specifically a pattern recognition network, can combine a small set of these easily collected features into a clinically useful prediction tool.
The study used a publicly available dataset of 116 participants -- 52 healthy controls and 64 breast cancer patients -- originally assembled by Patricio et al. Each participant had nine measured variables: age, BMI, glucose, insulin, HOMA (homeostatic model assessment of insulin resistance), leptin, adiponectin, resistin, and MCP-1 (monocyte chemoattractant protein-1).
A pattern recognition network (PRN), a type of feedforward neural network, was the core machine learning architecture. In these networks, data flows in one direction from input neurons through one or more hidden layers to an output layer, with each connection assigned a numerical weight that is adjusted during training to minimize prediction errors.
Data were split into 70% for training, 15% for validation, and 15% for final testing. The Levenberg-Marquardt algorithm was used as the training function because of its demonstrated efficiency, with tan-sigmoid activation functions and mean square error (MSE) as the loss metric. Training was allowed to run for up to 1,000 epochs.
Seven distinct model configurations (M1 through M5, plus M5-9 and M5-4) were compared, varying the number of input features (4, 5, or 9), the number of hidden layers (one or two), and the number of neurons per hidden layer (ranging from 1 to 100). This systematic comparison was designed to identify the configuration that best balances accuracy and simplicity.
The winning configuration, Model M5, used five input features -- age, BMI, glucose, resistin, and adiponectin -- and two hidden layers containing 11 and 100 neurons respectively. On the held-out test data, M5 achieved 94.1% accuracy, 100% sensitivity, and 88.9% specificity.
Sensitivity of 100% means the model correctly identified every actual breast cancer case in the test set -- a critical property for a screening tool where missing a cancer has serious consequences. The Matthews correlation coefficient (MCC) of 88.9% and Fowlkes-Mallows (FM) index of 94.3% further confirm that performance was strong even given the slight class imbalance between healthy and patient groups.
Across all 116 samples (training, validation, and test combined), M5 correctly classified 114 out of 116 participants -- an overall accuracy of 98.3%. The area under the ROC curve (AUC) plots confirmed these strong results across each data partition.
Interestingly, adding more features did not help. The nine-feature model (M5-9) achieved only 70.6% accuracy on the test set, substantially worse than the five-feature M5. This finding highlights the risk of overfitting when too many weakly predictive features are included, and demonstrates that feature selection is as important as model architecture.
A feedforward neural network (FFNN) consists of an input layer, one or more hidden layers, and an output layer. Each artificial neuron in a hidden layer receives inputs from the previous layer, multiplies them by learned weights, adds a bias term, and applies an activation function to produce its output signal. The tan-sigmoid function used here maps any input to a value between -1 and 1, introducing the non-linearity that allows the network to learn complex patterns.
The network is trained by adjusting weights to minimize the loss function, which measures the difference between predicted and actual class labels. The Levenberg-Marquardt algorithm combines the fast convergence of gradient descent with the stability of Gauss-Newton optimization, making it particularly effective for small-to-medium-sized networks.
Model performance was evaluated using a confusion matrix that tracks true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). From these four counts, accuracy, sensitivity, specificity, MCC, and FM index are all derived. This multi-metric approach gives a fuller picture of performance than accuracy alone, especially when class sizes differ.
Because no established rule governs the optimal number of hidden layers or neurons, the researchers used a trial-and-error approach, systematically testing seven architectures. This empirical strategy, while computationally modest for small datasets, is standard practice in applied neural network research.
Several earlier studies used the same dataset, allowing direct comparison. Saritas and Yasar (2019) applied a standard ANN and a Naive Bayes classifier to all nine features, achieving accuracies of 86.95% and 83.54% respectively. The current M5 model surpassed both using only five features, suggesting that careful feature selection matters more than the raw number of inputs.
Patricio et al., the original creators of the dataset, used a support vector machine (SVM) with four features (age, resistin, BMI, glucose), obtaining sensitivity of 82-88% and specificity of 85-90%. The current M5-4 model (same four features, PRN architecture) achieved 77.8% sensitivity and 87.5% specificity -- slightly lower -- while the five-feature M5 substantially improved sensitivity to 100%.
In the author's own prior work using K-nearest neighbor (KNN) and SVM algorithms with the same five features as M5, the best accuracy was 85.3% with 80.8% sensitivity and 89.1% specificity. The PRN-based M5 outperformed these across all three metrics, reinforcing the value of neural network approaches for this particular classification task.
One additional finding worth noting: the fifth feature, adiponectin, appears to carry meaningful predictive signal beyond what the four-feature model captures. Its inclusion alongside age, BMI, glucose, and resistin is what drives the jump from good to excellent sensitivity.
The five features powering the best model -- age, BMI, glucose, resistin, and adiponectin -- are all obtainable from standard clinical visits without specialized equipment. Age is recorded on intake; BMI is measured with a scale and height gauge; glucose, resistin, and adiponectin are drawn from a routine blood panel. No mammography, ultrasound, or biopsy is required to generate these inputs.
This accessibility is clinically significant. In low- and middle-income countries where imaging infrastructure is limited, a validated blood-based model could function as a first-line triage tool, directing limited imaging resources toward higher-risk individuals. Even in well-resourced settings, it could complement existing screening programs.
The paper carefully avoids claiming that this approach should replace imaging-based diagnosis. Rather, it positions the model as a decision-support aid: a tool that flags elevated risk and prompts further investigation. The authors acknowledge that the dataset is small (116 participants) and that external validation in diverse populations is needed before clinical deployment.
Beyond practical accessibility, the authors note that a blood-based screening approach may cause less patient anxiety, pain, and radiation exposure than mammographic screening -- a non-trivial benefit given that psychological distress is a known barrier to cancer screening participation.
This study demonstrates that a pattern recognition network trained on just five easily collected features -- age, BMI, glucose, resistin, and adiponectin -- can distinguish breast cancer patients from healthy controls with 94.1% accuracy and 100% sensitivity on the test set. The result is competitive with or superior to prior machine learning approaches on the same dataset.
The finding that the five-feature model outperforms the nine-feature model is a reminder that dimensionality reduction and feature selection are not merely computational conveniences -- they actively improve generalization by removing noise that would otherwise mislead the network.
The most critical limitation is dataset size. With only 116 participants, the model may not capture the full diversity of real-world breast cancer presentations. Large-scale prospective studies incorporating this feature set, ideally across multiple ethnicities and clinical settings, would be needed to establish clinical validity.
If validated, the approach represents a meaningful advance toward equitable early cancer detection -- a model that works with measurements available in virtually any clinical setting, reducing the global disparity in breast cancer outcomes driven by unequal access to imaging technology.