Comparative Study of Machine Learning Models in Breast Cancer Diagnosis: Clinical Decision Optimization Centered on Recall Rate

This article conducts a comparative study of six mainstream machine learning models (logistic regression, decision tree, random forest, gradient boosting, XGBoost, neural network) for breast cancer diagnosis, based on the UCI Wisconsin Breast Cancer Dataset, with a focus on recall rate (to reduce false negative risks). The study found that logistic regression and neural network had the highest recall rate (97.62%), providing a reference for optimizing clinical decisions.

Breast cancer is the most common malignant tumor among women globally, and early diagnosis directly affects survival rates. Machine learning has become an auxiliary tool in medical image diagnosis, but the consequences of missed diagnosis (false negatives) are far more serious than misdiagnosis. The team from the University of Toronto carried out a model comparison study with the core goal of maximizing recall rate to ensure the reduction of missed malignant cases.

The UCI Wisconsin Diagnostic Breast Cancer Dataset was used (569 samples: 212 malignant, 357 benign, with class imbalance). Stratified sampling was used to ensure consistent class distribution in the training/test sets; feature standardization was applied to logistic regression/neural network, while tree models leveraged their insensitivity to scaling.

The performance of the six models is as follows:

• Logistic regression: Recall rate 97.62%, precision 100%, F1 98.80%
• Neural network: Recall rate 97.62%, precision 100%, F1 98.80%
• XGBoost: Recall rate 92.86%, precision 97.50%
• Random forest/gradient boosting: Recall rate 90.48%, precision 100%
• Decision tree: Recall rate 88.10% (lowest). Logistic regression and neural network performed best in terms of recall rate.

Feature importance analysis identified two key features:

1. Perimeter Worst: Reflects the most severe morphology of the tumor
2. Concave Points: Describes the degree of indentation in the nuclear boundary, which is related to malignancy. These features are consistent with medical knowledge, proving that machine learning can learn clinically relevant features.

Implications for clinical application:

• High recall rate models can improve early detection rates, reduce missed diagnoses, and assist doctors in decision-making.
• Model selection needs to balance performance, computational cost, and interpretability: simple logistic regression and complex neural networks perform similarly, so comprehensive consideration should be given.
• Interpretability (such as feature importance of tree models) can enhance doctors' trust in AI and is key to clinical integration.

The project is built based on Python, relying on Scikit-learn, XGBoost, Pandas, NumPy, and Matplotlib. The complete code and experimental process have been open-sourced, providing a reproducible benchmark for subsequent research.

This study demonstrates the performance differences of different models in breast cancer diagnosis, emphasizing the importance of selecting evaluation metrics in medical AI. The recall rate priority strategy ensures model reliability and provides direction for the construction of intelligent diagnosis systems. In the future, with technological progress and data accumulation, machine learning will play a greater role in cancer screening.