You can deploy this model and start earning money today!

ZeroSync/
GaitAnalysis-fixes

Views

No views yet

GaitAnalysis — Bug Fixes & EDA Overhaul

Fixed version of ZeroDiscord/GaitAnalysis

Drop these 8 files into the original repo to apply all fixes. No other files are modified.

Files Changed (8)

File	Fix
`dataset.py`	Moved dead code (methods after `create_dataloaders` return) back into `GaitPathologyDataset` class; removed duplicate `__len__`/`__getitem__`
`train.py`	Changed all 4 model init branches from `input_dim=5` → `input_dim=input_dim` (uses dynamic feature count)
`evaluate.py`	Added missing `input_dim` argument to `load_model()` call (was passing 3 args to 4-param function)
`inspect_data.py`	Removed hardcoded `C:\Users\ASUS\...` path; added argparse; fixed `pd.read_csv` to use `header=None`
`gait_phase.py`	Changed fallback values from hardcoded healthy-gait numbers to `float('nan')` when cycle detection fails
`EDA/eda_features.py`	Added `gp_extraction_failed` column to track which windows had failed gait phase extraction
`EDA/eda_visualization.py`	PCA/LDA/ICA/correlation now use patient-level aggregation; waveform plots pick median-representative patient
`EDA/permutation_importance.py`	Complete rewrite: patient-level GroupKFold CV, no StandardScaler, cross-validated importance, diagnostics

Critical Bugs Fixed

1. Dead Code in `dataset.py` (P0)

Methods _extract_legacy_features, _get_cache_path, _load_or_compute_features, etc. were defined after create_dataloaders() return — outside the class body. They were dead code that never executed.

2. Hardcoded `input_dim=5` in `train.py` (P0)

All model constructors hardcoded input_dim=5, ignoring the dynamic input_dim variable computed from the dataset. Enhanced features (17+) would crash or silently truncate.

3. Wrong Argument Order in `evaluate.py` (P0)

load_model(args, num_classes, device) called with 3 args but function expects 4: load_model(args, input_dim, num_classes, device).

4. EDA Subject Signal Mixing (P1)

All EDA analyses (PCA, LDA, ICA, correlation, permutation importance) treated 961 overlapping windows as independent samples, but they come from only 38 patients with consecutive-window Pearson r ≈ 0.97.

5. Permutation Importance Data Leakage (P1)

Window-level train_test_split guaranteed overlapping windows from the same patient in both train and test. Test accuracy was inflated and importance rankings were wrong.

6. Gait Phase Fallback Pollution (P2)

When cycle detection failed (15.7% of windows), hardcoded "healthy gait" values (e.g., stance_duration=62.0) were injected regardless of actual pathology class.

Permutation Importance Overhaul

The original PI had 6 compounding flaws:

Flaw	Before	After
Split level	Window-level (leaks patient data)	Patient-level GroupKFold
Effective n	961 windows (fake)	38 patients (real)
Window correlation	r ≈ 0.97 between adjacent windows	Eliminated by grouping
Feature scaling	StandardScaler (distorts RF importance)	None (RF is scale-invariant)
Stability	Single 70/30 split	K-fold CV averaged
Fallback features	151 rows with identical healthy-gait values	NaN → 0 with failure flag

ZeroSync/GaitAnalysis-fixes