Liornis/forge-fog

FORGE — FOG Representation via Generative Encoding

Self-supervised spectral-temporal encoders for Freezing of Gait (FOG) detection from a single lower-back accelerometer. Pretrained by masked autoencoding on 11,724 h (~21M windows) of unlabeled at-home recordings from 65 participants, then trained for FOG detection on the 57-participant DeFOG cohort only. Evaluated with no target-cohort training on four external cohorts: FogAtHome-provoking, tDCS-FOG, Stanford and FogAtHome daily living.

Headline: the released MC probe ensemble reaches clinical-grade agreement with expert video annotation on an independent cohort — ICC(%TF) = 0.899 [0.700, 0.970], zero-shot, one IMU.

External results (released detector)

Nine-head MC frozen-probe ensemble (3 participant folds x 3 seeds), evaluated with no target-cohort training and the unchanged DeFOG operating point of 0.35.

Cohort (N) — shift	AUROC	AP	ICC(%TF)
FogAtHome-provoking (12) — cross-study	0.887 [0.830, 0.922]	0.804 [0.573, 0.902]	0.899 [0.700, 0.970]
tDCS-FOG (71) — cross-protocol	0.917	0.812 [0.550, 0.923]	0.876
Stanford (7) — site / device / med state	0.734	0.400	-0.119
FogAtHome daily living (11) — naturalistic*	0.803 [0.737, 0.877]	0.105	0.656 [-0.129, 0.872]

* Daily living is scored inside a label-independent walking-and-standing domain (58.18 h of 301.8 h, 2.92% FOG); it is gait-conditioned burden, not whole-recording %TF. Stanford is negative evidence: discrimination survives the shift, the fixed threshold does not (its oracle-rule threshold is 0.18). In-distribution reference: window-level AP 0.730 on held-out DeFOG folds. Full definitions and confidence intervals are in manifest.yaml under results:.

Released weights

Pretrained FORGE encoders (the backbones)

Context	Window (frames)	File	Params
LC	1000	encoders/lc.ckpt	14,147,072
MC	500	encoders/mc.ckpt	14,147,072
SC	200	encoders/sc.ckpt	12,918,272

Downstream classification checkpoints (57-participant DeFOG, 3-fold participant-level CV)

File	Context	Phase	Fold
classification/lc_probe_fold0.ckpt	lc	probe	0
classification/lc_probe_fold1.ckpt	lc	probe	1
classification/lc_probe_fold2.ckpt	lc	probe	2
classification/mc_probe_fold0.ckpt	mc	probe	0
classification/mc_probe_fold1.ckpt	mc	probe	1
classification/mc_probe_fold2.ckpt	mc	probe	2
classification/sc_probe_fold0.ckpt	sc	probe	0
classification/sc_probe_fold1.ckpt	sc	probe	1
classification/sc_probe_fold2.ckpt	sc	probe	2
classification/lc_finetune_fold0.ckpt	lc	finetune	0
classification/lc_finetune_fold1.ckpt	lc	finetune	1
classification/lc_finetune_fold2.ckpt	lc	finetune	2
classification/mc_finetune_fold0.ckpt	mc	finetune	0
classification/mc_finetune_fold1.ckpt	mc	finetune	1
classification/mc_finetune_fold2.ckpt	mc	finetune	2
classification/sc_finetune_fold0.ckpt	sc	finetune	0
classification/sc_finetune_fold1.ckpt	sc	finetune	1
classification/sc_finetune_fold2.ckpt	sc	finetune	2
classification/lc_supervised_fold0.ckpt	lc	supervised	0
classification/lc_supervised_fold1.ckpt	lc	supervised	1
classification/lc_supervised_fold2.ckpt	lc	supervised	2
classification/mc_supervised_fold0.ckpt	mc	supervised	0
classification/mc_supervised_fold1.ckpt	mc	supervised	1
classification/mc_supervised_fold2.ckpt	mc	supervised	2
classification/sc_supervised_fold0.ckpt	sc	supervised	0
classification/sc_supervised_fold1.ckpt	sc	supervised	1
classification/sc_supervised_fold2.ckpt	sc	supervised	2

What this release contains

All 27 classification heads are the seed-42 runs. The manuscript's released detector averages nine heads (3 folds x 3 seeds) over one shared frozen encoder; the nine stored encoder parameter sets are bit-identical, so encoders/mc.ckpt + classification/mc_probe_fold(0, 1, 2).ckpt rebuild the seed-42 three-fold ensemble. That is the configuration this project's evaluation scripts run, and it lands within about 0.02 of the nine-head numbers tabled above.

Usage

Checkpoints are slimmed PyTorch Lightning checkpoints (weights + config; optimizer state stripped). Each keeps the state_dict and the hyper_parameters["config"] Pydantic config used to rebuild the model — the same fields the evaluation pipeline reads.

import torch
ckpt = torch.load("encoders/mc.ckpt", map_location="cpu", weights_only=False)
state_dict = ckpt["state_dict"]            # encoder weights
config = ckpt["hyper_parameters"]["config"]  # Config object to rebuild the model
meta = ckpt.get("forge_meta")              # name / context / phase / fold

Reproduce every paper number with the companion repo's reproduce-evaluations skill (see manifest.yaml, shipped in this repo). Code: github.com/Lior-Nis/forge. Data: Liornis/fog-dataset.

License: MIT.