GitHub Repository My Methods

Overview

This is an ongoing machine learning research project focused on decision support for fixation strategy in total hip arthroplasty (THA): cemented versus non-cemented implants. The pipeline uses CT-derived quantitative imaging features and compares multiple modeling approaches while explicitly addressing surgeon disagreement, cohort heterogeneity, and data quality constraints.

The current repository reflects active development and iteration; results and conclusions will continue to evolve as additional analyses are completed.

What I Accomplished

  • Built an end-to-end, script-based ML workflow from data audit through model evaluation (01_data_audit.py to 05_compare_all_experiments.py)
  • Created leakage-safe preprocessing with patient-level grouped splitting, train-only fitting for imputation/scaling, and same-split reuse for controlled sensitivity analyses
  • Implemented multiple label strategies to manage clinical uncertainty:
    • original-surgeon labels
    • majority-vote labels
    • vote-fraction (soft/probabilistic) labels
  • Engineered clinically motivated features from Gruen zone BMD and cortical geometry, expanding from 13 base CT features to 22 total model inputs
  • Trained and benchmarked Logistic Regression, Random Forest, Gradient Boosting, and SVM models with grouped cross-validation and held-out test evaluation
  • Added cohort-specific reporting to evaluate domain shift across scanner eras/manufacturers (Toshiba vs Philips)

Technical Approach

Data and Labeling

  • Retrospective THA cohort from preoperative CT scans
  • Unit of analysis: hip-level observations, grouped by patient to prevent leakage
  • Harmonized surgeon labels and quantified inter-rater agreement before defining target labels

Preprocessing and Feature Engineering

  • Quality filtering for unusable scans and implausible segmentation-derived values
  • Cohort-aware KNN imputation (fit on training set only)
  • Standardized scaling fit on train data, applied to both train and test
  • Derived BMD summary statistics, paired-zone averages, and ratio features tied to cortical measurements

Modeling and Evaluation

  • Grouped 5-fold cross-validation by patient ID
  • Test metrics: ROC-AUC, PR-AUC, accuracy, precision, recall, F1, confusion matrix
  • Additional cohort-specific performance breakdowns to assess robustness

Current Findings (Interim)

  • The project demonstrates meaningful predictive signal from CT-derived features, but outcomes are sensitive to label definition and cohort differences.
  • Logistic Regression currently provides a strong baseline ROC-AUC in majority-vote experiments.
  • Comparative experiments suggest combined CT + demographic models may outperform CT-only models in this dataset, while demographics-only models retain moderate signal.

Majority-vote ROC curves

These findings are preliminary and intended to guide the next research iterations rather than serve as final clinical conclusions.

Ongoing Work

Because this project is not yet published, current priorities are:

  • sensitivity analyses across alternative ground truth definitions
  • deeper unanimous-only versus majority-vote comparisons
  • quality-flag impact analysis (including/excluding questionable scans)
  • cohort transfer/generalization testing and potential harmonization strategies
  • refinement of feature interpretation and model calibration

Research Context

This work is intended as clinical decision-support research, not autonomous decision-making. It is being developed with de-identified data, documented preprocessing safeguards, and reproducibility-focused scripting to support transparent, iterative validation.