DATASET COMPOSITION OF THE LOSO CONFUSION MATRIX
Description
Raw measurements and Python scripts to reproduce the Leave-One-Session-Out (LOSO) validation, the confusion matrix, and the leakage demonstrations in the manuscript: "Cross-Session Glucose Level Estimation with a CSRR-Loaded Microwave Sensor: A Baseline-Corrected Resonant Feature Outperforms Full-Spectrum Deep Learning"
Files
Steps to reproduce
Requirements: Python 3 with numpy, scipy, scikit-learn, matplotlib (TensorFlow is also required for the 1D-CNN). In each script, set the two session-folder paths near the top of the file to: raw_data/session1_blood and raw_data/session2_blood Then run, for example: python code/multi_notch_experiment.py -> single 4.3 GHz notch, LOSO accuracy = 82.5 % (also the multi-notch leakage: top-5 = 88.8% leaky -> 38% honest) python code/cnn_experiment.py -> full 101-point spectrum with LDA / MLP / 1D-CNN, augment-before-split (leaky) vs Leave-One-Session-Out (~98% -> 43-44%) python code/correct_split_experiment.py -> augment-before-split leakage (92.5% of test rows are duplicates) python code/gen_blood_figs_from_raw.py -> blood S11 spectrum and calibration figures, recomputed from raw .s2p