flap-prediction-rf

Model Card — Flap Circulation Prediction (RF2)

Overview

A RandomForest classifier predicting free-flap circulation status from objective skin measurements. Reproduces a 2022 published / conference-presented research prototype (CGMH Plastic & Reconstructive Surgery; Huang CJ, Tsai CY, et al.).

Intended use

• Research / methodology demonstration only.
• NOT a medical device. NOT prospectively validated. NOT for clinical decisions.

Inputs / outputs

• Inputs (4 objective features): [temp_diff, ΔR, ΔG, ΔB]
  • temp_diff = flap surface temperature − control (normal skin) temperature (°C)
  • ΔR/ΔG/ΔB = mean(flap region RGB) − mean(control region RGB)
• Output: one of {Normal Flap, Arterial Insufficiency, Venous Insufficiency} + class probabilities.

Training data

• Free-flap monitoring photographs + surface temperatures, CGMH.
• Region-of-interest (ROI) sampling per flap; multiple ROIs per patient.
• Class imbalance handled with SMOTE-Tomek.

Reported performance (with caveat)

• Reported test accuracy (RandomForest, objective features): ~97.9%.
• ⚠️ Known limitation — patient-level data leakage: SMOTE oversampling was applied before the train/test split, and a single patient contributed multiple ROI rows that ended up on both sides of the split. The reported figure therefore overstates real-world (patient-level) performance. A rigorous leakage-free re-evaluation (patient-level aggregation, ~225 patients) is the correct next step and has not been completed.

Limitations

• Single-institution data; external validity unverified.
• Severe class imbalance (Arterial insufficiency is rare, ~10 patients aggregated).
• Colour features are sensitive to lighting / white balance / camera.
• The honest leakage-free accuracy is expected to be substantially lower.

License

Model weights CC-BY-NC-4.0. Code MIT.