Introduction
Cancer treatment delays and modifications can significantly impact patient survival and quality of life. Research has consistently shown that marginalized populations, including Black, Hispanic, Asian, and American Indian/Alaska Native (AIAN) patients, experience higher rates of late-stage cancer diagnoses and lower rates of timely treatment.
For my project, I used machine learning models to examine how race, socioeconomic status, tumor characteristics, cancer stage, grade, subtype, age, clinical trial participation, and time to treatment initiation predict treatment interruptions in breast cancer patients.
This project highlights both the potential and limitations of using machine learning for predicting healthcare disparities in cancer treatment.
Dataset:
This project utilized Simulacrum v2.1.0, a synthetic dataset derived from the National Disease Registration Service (NDRS) Cancer Analysis System at NHS England. While this dataset mimics real-world cancer data, it ensures patient anonymity.
After data cleaning and removing incomplete observations, the dataset contained 69,367 patients, all diagnosed with breast cancer.
Demographics Overview
- Gender Distribution:
- Women: 98.7%
- Men: 1.3% (Male breast cancer cases were retained for analysis.)
- Age Distribution:
- Mean Age: 61.06 years
- Median Age: 61 years
- Racial Distribution:
- White: 85.9%
- Asian: 3.7%
- Black: 1.9%
- Other: 1.7%
- Mixed Race: 0.6%
- Unknown: 6.1% (Reweighting was applied to mitigate algorithmic bias.)
- Neighborhood Deprivation (Socioeconomic Status):
- Scored from 1 (most deprived) to 5 (least deprived).
- The dataset was fairly balanced across deprivation levels.
Average Time to Treatment Initiation (by Race)
- Overall: 61 days
- Other Racial Groups: 71 days
- Black Patients: 63 days
- White Patients: 62 days
- Asian Patients: 54 days
- Mixed Race Patients: 45 days
- Unknown Race: 57 days
Clinical Trial Participation
- 84.4% of patients were enrolled in a clinical trial.
Data Processing & Feature Engineering
1. Cleaning & Standardizing Data
- Removed inconsistent staging classifications.
- Imputed missing values for key variables such as comorbidity scores and time to treatment initiation.
2. Encoding Variables
- One-Hot Encoding:
- Race, tumor subtype, estrogen receptor (ER), progesterone receptor (PR), and HER2 status.
- Ordinal Encoding:
- Tumor stage, node stage, metastasis stage, overall stage, grade, and deprivation index.
3. Feature Engineering
- Created a new feature,
ANY_REGIMEN_MOD- Combined dose reduction, time delay, and early termination variables into one binary target variable.
- Grouped tumor biomarkers into cancer subtypes:
- Luminal A: ER+, PR+, HER2- (Least aggressive)
- Luminal B: ER+, PR+, HER2+ (Slightly more aggressive)
- HER2-Enriched: ER-, PR-, HER2+
- Triple Negative (Basal-like): ER-, PR-, HER2- (Most aggressive)
Bias Mitigation Strategies
1. Gender Bias
- While 98% of patients were female, male breast cancer cases were retained for rare case analysis.
- Applied weighting techniques to balance gender representation during model training.
2. Racial Bias
- Since 85% of patients in the dataset were White, inverse weighting was applied to ensure fair contributions across racial groups.
Machine Learning Models & Performance
I trained Random Forest and XGBoost models to predict treatment modifications.
1. Initial Model Performance (Random Forest & XGBoost)
Results were poor. The models performed only slightly better than random guessing.
2. Hyperparameter Tuning & Feature Selection
- Used GridSearchCV to optimize parameters.
- Dropped the least important features and retrained the models.
- Results did not improve, and performance worsened in some cases.
3. Model Performance Issues
- Models failed to generalize to testing data.
- ROC AUC scores hovered around 0.5, meaning models were barely better than random guessing.
- Models overfitted the ‘no treatment modification’ group, inaccurately predicting treatment delays.
4. Alternative Models Attempted
To address these issues, I tested additional models:
- Support Vector Machines (SVM)
- Neural Networks
- Logistic Regression
- K-Nearest Neighbors (KNN)
- Naïve Bayes
Findings from Alternative Models:
- SVM and Logistic Regression performed best (Accuracy: 0.65), but logistic regression completely failed at predicting treatment modifications.
- Neural Networks, KNN, and Naïve Bayes were slightly better balanced but still had low accuracy (~0.60–0.62).
5. Last Attempt – Oversampling with SMOTE
- Used Synthetic Minority Oversampling (SMOTE) to balance the dataset.
- SVM improved slightly, but performance gains were minimal.
Uncovering Insights Through Clustering
Since predictive modeling was unsuccessful, I conducted clustering analysis to find patterns in the data.
Three Distinct Patient Clusters Emerged:
Cluster 0: Moderate-Stage Cancer (35.4% Treatment Modifications)
- Average Age: 60.8
- Tumor Stage: T2, N0.7, M0.03 (Localized)
- Time to Treatment: 32.6 days
Cluster 1: Early-Stage Cancer with Delayed Treatment (36.7% Treatment Modifications)
- Average Age: 61.0
- Tumor Stage: T1, N0.03, M0.0 (Localized, very early-stage)
- Time to Treatment: 46.8 days (longest)
Cluster 2: Older Patients with Shortest Time to Treatment (36.8% Treatment Modifications)
- Average Age: 69.4 years
- Higher Comorbidities (Charlson Score = 0.63)
- Time to Treatment: 14.8 days (shortest)
Key Takeaways from Clustering:
- Cluster 1 had the longest time to treatment despite being early-stage. Further investigation needed.
- Older patients (Cluster 2) received faster treatment but had more comorbidities.
Conclusion & Next Steps
Machine learning models failed to predict treatment modifications accurately.
Clustering analysis revealed patterns in treatment delays and patient subgroups.
Future Steps:
- Explore more sophisticated models (Deep Learning, Bayesian Networks).
- Use real-world data instead of synthetic datasets.
- Investigate non-quantifiable factors, like patient-provider interactions and healthcare policies.
Final Thoughts
This project highlighted the complexity of predicting cancer treatment interruptions and the importance of interdisciplinary approaches in health equity research.
If you’re interested in machine learning for healthcare, data-driven health equity research, or predictive modeling, let’s connect!
Sydney Lash 