99% Sure? Gabor Begs to Differ

Code and paper for "99% Sure? Gabor Begs to Differ: A Time-Frequency Look at Parkinson's Handwriting and a Structure-Preserving Image Encoding."

We study Parkinson's disease (PD) detection from online handwriting on the public PaHaW benchmark (72 subjects: 37 PD, 35 healthy controls). The paper makes three points:

1. Literature audit — reported PaHaW accuracies (often >90%, sometimes 98–100%) are inflated by subject leakage, feature-selection bias, feature explosion, and the absence of variance/uncertainty in the reported numbers.
2. A time-frequency limit — the PD signature (a 4–6 Hz tremor interleaved with sub-decisecond pen arrests) is bounded by the Heisenberg–Gabor uncertainty principle (σ_t·σ_f ≥ 1/4π) upstream of any classifier, so no fixed representation can resolve both regimes on a short 150 Hz signal.
3. A structure-preserving encoding — instead of separating PD from healthy (ill-posed), we re-encode the kinematics of the Archimedean spiral (Task 1) as a colour image read by a frozen ImageNet backbone (ViT-B/16, Swin-T, EfficientNet-B0) and classified by a per-fold-tuned head.

Under repeated nested cross-validation with the Nadeau–Bengio corrected t-test, the best encoding (effort, ViT, random forest) clears chance and a clinical baseline (F1 0.7457 vs. 0.5080, p = 0.0187) — but no encoding is statistically separable from its nearest rivals. The honest number on 72 subjects is a modest one.

The full paper is in paper/handwriting_uncertainty.pdf.

	velocity	effort (v·p)	rgb (v·p·e)
Healthy
Parkinson's

Structure-preserving colour encoding of the Task-1 Archimedean spiral (PaHaW). Each pen sample is coloured by a kinematic channel; the Parkinsonian spiral reads slower, tighter and lower-effort.

---

Project Structure

.
├── data/
│   ├── raw/                     # PaHaW source data (git-ignored)
│   └── processed/               # Cached renders + embeddings (git-ignored)
├── paper/
│   ├── handwriting_uncertainty.tex   # Paper source (self-contained: TikZ + tables)
│   ├── handwriting_uncertainty.pdf   # Compiled paper
│   ├── references.bib
│   ├── neurips_2023.sty / natbib.sty # Style files
└── src/
    ├── data/
    │   ├── pahaw.py             # load_subjects(), read_svc(), task_path()
    │   ├── features.py          # 33-dim handcrafted clinical feature vector
    │   └── trajectory_image.py  # spiral → colour image (10 modalities)
    ├── models/
    │   ├── image/vit_encoder.py # build_vit_encoder(), encode_image() (timm, frozen)
    │   └── heads/mlp.py         # FixedMLPClassifier, SmallMLPClassifier
    ├── experiments/grid_sweep/
    │   ├── generate_embeddings.py   # render images + cache frozen-backbone embeddings
    │   ├── param_grids.py           # head registry + grids (single-seed sweep)
    │   ├── param_grids_robust.py    # head registry + grids (repeated sweep)
    │   ├── run_all_heads.py         # input registry, subject-level metrics
    │   ├── run_repeated.py          # repeated nested 10×3 CV (seeds 42,43,44) ★
    │   ├── summarize_repeated.py    # aggregate repeated results
    │   └── results/                 # per-fold result CSVs (committed)
    ├── analysis/
    │   ├── nadeau_bengio.py     # corrected resampled paired t-test ★
    │   └── full_pairwise.py     # all-pairs significance table ★
    └── utils/
        ├── metrics.py           # classification_metrics(), aggregate_fold_scores()
        └── results.py           # save_results_csv()

★ = the three reproducibility scripts cited in the paper.

---

Setup

uv sync

PaHaW is not redistributed here. Place the dataset under data/raw/ (one folder per subject, <id>__<task>_1.svc files) so that src/data/pahaw.py:load_subjects() can find it.

---

Reproducing the Paper

# 1. Render the 10 spiral image modalities and cache frozen-backbone embeddings
uv run python src/experiments/grid_sweep/generate_embeddings.py

# 2. Repeated nested 10×3 StratifiedGroupKFold CV (seeds 42, 43, 44) over all cells
uv run python src/experiments/grid_sweep/run_repeated.py

# 3. Aggregate the per-cell fold CSVs
uv run python src/experiments/grid_sweep/summarize_repeated.py

# 4. Significance testing (Nadeau–Bengio corrected t-test, all pairs)
uv run python src/analysis/full_pairwise.py

Results are written to src/experiments/grid_sweep/results/all_heads_repeated/.

---

Evaluation Protocol

Loop	Folds	Splitter	Purpose
Outer	10	StratifiedGroupKFold (≡ StratifiedKFold here)	Unbiased test-set evaluation
Inner	3	StratifiedKFold	Per-fold head hyperparameter tuning

Because we use only Task 1, every subject contributes exactly one recording, so subject = sample: each group has size 1 and StratifiedGroupKFold reduces to an ordinary StratifiedKFold. There is no cross-task subject leakage to prevent — it is excluded by construction (one recording per subject), not by the splitter. The group-aware splitter is kept in code only as a harmless guard (a no-op when groups are singletons). The protocol is repeated over 3 seeds (42, 43, 44) → 30 test folds per cell. Model comparisons use the Nadeau–Bengio corrected resampled paired t-test, which repairs the optimistic variance of the naive fold-wise test.

---

Representations & Models

• 10 image modalities of the spiral: 6 single-signal colour maps (velocity, acceleration, pressure, azimuth, altitude, effort), 3 RGB triplets (rgb_vpa, rgb_vpe, rgb_vpz), and 1 raw uncoloured stroke (geometry control).
• 3 frozen backbones: ViT-B/16, Swin-T, EfficientNet-B0.
• 7 heads: logistic regression, 3 SVMs, random forest, 2 MLPs.
• Baseline: 33-dim handcrafted clinical feature vector (Drotář et al., Impedovo).

In all: 31 inputs (10 modalities × 3 backbones + baseline) × 7 heads = 217 cells.

---

Slides

A talk deck built with open-slide lives in slides/pahaw-uncertainty/ (deck: slides/pahaw-gabor/, 20 pages).

cd slides/pahaw-uncertainty
npm install       # first time only
npm run dev       # live preview at localhost — press F to present, export PDF from the browser
npm run build     # static site → dist/

---

Citation

@misc{porcelli2025gabor,
  title  = {99\% Sure? Gabor Begs to Differ: A Time-Frequency Look at
            Parkinson's Handwriting and a Structure-Preserving Image Encoding},
  author = {Porcelli, Andrea},
  year   = {2025},
}

---

License

Released under the MIT License. PaHaW is the property of its original authors and is not redistributed in this repository.