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 nine-head frozen-encoder ensemble is evaluated against expert video annotation on an independent cohort — ICC(%TF) = 0.899 [0.700, 0.970], using one lower-back IMU and no target-cohort training.

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 [0.863, 0.950]	0.812 [0.550, 0.923]	0.876 [0.810, 0.920]
Stanford (7) — site / device / med state	0.734 [0.624, 0.846]	0.400	-0.119 [-0.920, 0.670]
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 the DeFOG validation folds. Full definitions and confidence intervals are in manifest.yaml under results:.

The tDCS-FOG AP of 0.812 [0.550, 0.923] above is retained as a historical fold-safe analysis result, not the released nine-head reproduction target. The existing public code manifest records AP 0.866 [0.661, 0.946] for the all-71-participant nine-head assembly used by the reproduction workflow. No result has been recomputed for this documentation update. Use the public code's release/manifest.yaml for reproduction targets; this model repository's older manifest.yaml also contains historical analysis entries.

Controlled comparison (a different model set)

The paper's headline effect is a separate, deliberately constrained experiment: two arms differing only in encoder initialization, under matched downstream training. Its numbers are lower than the released detector's on the same cohort because it is a different model set — not a worse estimate of the same thing.

Cohort	Metric	Self-supervised	Supervised from scratch	Difference [95% CI]
FogAtHome-provoking	AUROC	0.861	0.752	+0.109 [0.029, 0.182]
FogAtHome-provoking	AP	0.784	0.592	+0.192 [0.064, 0.338]
tDCS-FOG	AUROC	0.869	0.657	+0.212 [0.126, 0.280]
tDCS-FOG	AP	0.811	0.501	+0.310 [0.133, 0.397]

Each result in manifest.yaml names its model set and the manuscript table it comes from, so the two sets stay distinguishable.

Released weights

Pretrained FORGE encoders (the backbones)

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

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

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

What this release contains

The released nine-head frozen-encoder ensemble contains all nine BiGRU heads: three participant-level DeFOG folds (0, 1, 2) × seeds 42, 43, and 44, over one shared pretrained frozen MC encoder (encoders/mc.safetensors). External evaluation averages all nine heads; the DeFOG reference remains the seed-42 out-of-fold result. The 27 seed-42 classification files listed above are joined by:

Seed	Fold 0	Fold 1	Fold 2
43	classification/mc_probe_s43_fold0.safetensors	classification/mc_probe_s43_fold1.safetensors	classification/mc_probe_s43_fold2.safetensors
44	classification/mc_probe_s44_fold0.safetensors	classification/mc_probe_s44_fold1.safetensors	classification/mc_probe_s44_fold2.safetensors

Usage

This release is weights only: each file is a .safetensors tensor set with small string metadata (name, context, phase, fold, seed, and the experiment config that rebuilds the model). No training configuration, optimizer state or local file path is included, and loading executes no pickled code.

Rebuild a model from the companion repo, which composes the architecture from the experiment config named in the file's metadata and in manifest.yaml:

from utils.released_weights import load_released_model

model, config = load_released_model(
    "release/forge-fog/classification/mc_probe_fold0.safetensors",
)

Or read the tensors directly:

from safetensors.torch import load_file
from safetensors import safe_open

state_dict = load_file("classification/mc_probe_fold0.safetensors")
with safe_open("classification/mc_probe_fold0.safetensors", framework="pt") as f:
    meta = f.metadata()   # name / kind / context / phase / fold / seed / experiment / splits

Each classification file already contains its encoder, so encoders/*.safetensors are needed only to train new heads.

Citation

Lior Nisimov, Amit Salomon, Eran Gazit, Talia Herman, Lior Rokach, Jeffrey M. Hausdorff, Nathaniel Shimoni. Self-supervised learning improves cross-cohort freezing-of-gait detection from a single lower-back accelerometer. Submitted to npj Digital Medicine, 2026.

Reproduce the released detector on the four external cohorts using the public REPRODUCE.md and ./reproduce.sh (or reproduce.bat on Windows). The public code's release/manifest.yaml pins the weights and datasets. The workflow downloads public safetensors and data, builds the evaluation inputs, and generates its own predictions; private checkpoints, precomputed predictions, and private paths are not required. This workflow does not reproduce every manuscript analysis. Code: github.com/Lior-Nis/forge-public. Data: Liornis/fog-dataset.

Intended use and limitations

For research use only. FORGE is not clinically validated and is not a medical device. It must not be used to diagnose, monitor, or make treatment decisions for an individual. Performance and burden calibration vary across cohorts, devices, and recording conditions.

License: MIT.