1 of 1

Breast Cancer Prediction Using Supervised Machine Learning

NIcole Adelson, Abbot Tubeine, Salah Mohammed, Azeezat Akinboro, Alexia Rendon

This project uses supervised machine learning to improve the early detection of breast cancer. Traditional diagnostic methods like mammograms have had limitations with approximately 1 in 8 cases of breast cancer going undetected.��This gap leaves many patients vulnerable to late-stage diagnosis and fewer treatment options. Factors such as tumor complexity and individual health variations can make accurate detection challenging which highlights a need for more innovative and reliable diagnostic tools. Our project seeks to address this issue by providing a modern solution grounded in machine learning.

We mainly focused on a Neural Network due to its ability to identify complex patterns in high-dimensional data, making it well suited for detecting subtle indicators of breast cancer. We also use other algorithms such as Random Forest Classification and the XGboost algorithm for passes of robustness and comparison.��This approach gave us precise predictions, helping us to gauge the accuracy of our neural network. Adopting this model in health care will help increase survival rates as it will aid in detecting breast cancer at its early stages.

Next Steps:

Image Data Integration: In future iterations, the model can be expanded to classify breast cancer using mammogram or ultrasound images, incorporating convolutional neural networks (CNNs) for image recognition.

Background

Dataset

Models Utilized

https://www.geeksforgeeks.org/pros-and-cons-of-decision-tree-regression-in-machine-learning/
Random Forest https://www.geeksforgeeks.org/what-are-the-advantages-and-disadvantages-of-random-forest/
City of Hope. (2022, September 20). Breast cancer tumor size chart & measurement explained. City of Hope. https://www.cancercenter.com/cancer-types/breast-cancer/stages/tumor-size-chart
Teng, J. (2019, January 18). Seer Breast Cancer Data. IEEE DataPort. https://ieee-dataport.org/open-access/seer-breast-cancer-data

Github Repository

Citation and Documentation

Dataset Visualizations

Model Evaluations

Decision Trees	XGBoost	Random Forest	Neural Networks
Uses a tree-like structure to make predictions by inquiring about the different features and dividing the dataset into smaller subsets according to feature homogeneity until it reaches the leaf nodes, which is when it makes the final prediction.	Also called “Extreme Gradient Boosting,” this technique uses the ensemble “boosting” approach by connecting weak models (or decision trees) in series to build a stronger classifier. In this case, each model corrects the errors of their predecessors, which improves the overall performance of the model.	Several decision trees are created using random subsets of the data and features (also called bagging). Each decision tree makes a prediction according to these random values. Whichever prediction is more frequent among all decision trees is the result of the algorithm.	Neural networks consist of artificial neurons, an input layer, one or more hidden layers, and an output layer. The neurons are interconnected and have their own linear regression models, each consisting of weights, a bias/threshold, and output layer. This method simulates processes of the human brain and enables computers to recognize patterns in non-linear, complex data.

Decision Trees	XGBoost	Random Forest	Neural Networks
Pros: Interpretability: Due to the tree structure, it is easy to interpret how decisions are made by the algorithm. Great with non-linear data: When handling data with complex patterns, decision trees are able to capture relationships between features and the target. Cons: Overfitting: Prone to overfitting when the dataset is too noisy and tends to make generalizations about trends in data. Instability: Small variations in data can lead to different subsets and different trees and, in turn, varying results after each iteration.	Pros: Regularization: Incorporates L1 (Lasso Regression) and L2 (Ridge Regression) in order to prevent overfitting Handles missing values: When encountering a missing node, it samples the left and right branches in order to learn the trend. Cons: Overfitting: XGBoost can be prone to overfitting, especially when too many trees are used. Memory Requirements: It can also be memory intensive when working with large datasets, making it not suitable when memory resources are limited	Pros: Feature Importance: Random forest models is good for calculating the importance of features, determining what features should be used. Flexibility: Random forest allows flexibility by handling multiple tasks with high accuracy. Cons: Prediction Time: Random forest algorithms can take longer than other algorithms, even though they are more effective. Overfitting: Random forest models can be prone to overfitting if the model captures noise from the training data.	Pros: Complex data: Neural networks are capable of processing complex data, making them suitable for more complex works. Feature extraction: Neural networks are capable of extracting features from raw data, reducing the need of manual feature engineering. Cons: Complexity: Neural networks can be a bit complex to master, and require a lot of expertise. Biases: Neural networks can be more prone to bias especially in fields like healthcare.

We utilized the Wisconsin breast cancer diagnostic dataset. Some key features it contains include the mean texture, perimeter, and area which describe the tumor size and surface, and the concavity, concave points, and symmetry which highlight shape irregularities.�

Impact & Bias

In the process of creating our AI/ML solutions, we mitigated bias by investigating the availability of reliable data to create our model, as we had to pivot once due to the unavailability of reliable data.

We also worked to determine the exact features that had a significant effect on the model to reduce the notice within the neural network.