FM–USP
Turning clinical data into understanding.
FM–USP / MACHINE LEARNINGPatterns of possibility.SYSTEM STUDY / 01
A machine learning predictive model that analyzes breast cancer evolution variability and treatment responses to estimate patient risk levels and survival rates. Using electronic medical records from 3,769 patients across four breast cancer subtypes (Luminal-A, Luminal-B, HER-2, and TNBC), the model helps healthcare professionals make more informed treatment decisions and optimize resource allocation in Brazil's public healthcare system (SUS).
The intention
- Predict breast cancer evolution variability and treatment responses
- Estimate patient survival rates categorized by quartiles
- Enable efficient resource allocation in public healthcare system
- Support personalized treatment planning based on data-driven insights
- Improve quality of life through optimized consultation scheduling
The complexity
- Processing 61,684 records with 104 columns from dispersed medical data
- Handling biased data from tertiary-level hospital specializing in complex cases
- Addressing data quality issues including missing values and inconsistencies
- Preventing overfitting while maintaining model generalization capability
- Ensuring clinical relevance while meeting privacy regulations
What took shape
- Achieved 74.70% accuracy using K-Nearest Neighbors model
- Successfully compared Decision Tree, KNN, Neural Networks, and SVM models
- Identified key predictive features: tumor duration, menarche age, patient age
- Implemented CRISP-DM methodology for structured data mining
- Developed comprehensive data treatment pipeline for medical records
Tools & disciplines
- Python
- Scikit-learn
- Pandas
- NumPy
- Matplotlib
- Google Colab
- Jupyter Notebook