Eye Tracking for Assembly Fatigue: SSM
Complete fitted models accompanying Estimating Operator Fatigue During Manual Assembly with a State-Space Model of Wearable Eye-Tracking Signals.
The state-space model (SSM) learns shared parameters from training participants and uses individual ocular references and initial SSS ratings. Evaluation is offline.
Files and intended use
Final SSM for inference: reference_all_participants.joblib, trained on
all 22 participants and 44 sessions with the paper's 150 s / 300 s reference
configuration. Its matching .manifest.json and .verification.json identify
the full-data fit and record serialization and fresh-process checks. No
independent test-performance score is assigned to this final fit.
Models for numerical reproduction:
reference_fold_1.joblib through reference_fold_5.joblib reproduce five fixed
participant-held-out folds. Each contains fitted Ridge, Extra Trees and logistic
pipelines, imputation/scaling, Q/R, target mean/SD and training/test provenance.
manifest.json records hashes, configuration, feature ordering and library
versions. Those five fold files are not an ensemble; the all-participant fit is
provided separately as the primary inference artifact.
The configuration is 150-s trailing ocular features / 300-s nominal state cadence. It is separate from the 150-s/30-s realtime prototype.
Loading
The repository includes paper_ssm.py, paper_bundle.py and requirements.txt.
After downloading its files, install the pinned dependencies with
python3.11 -m pip install -r requirements.txt and run Python from that directory:
from paper_bundle import load_model
model = load_model("reference_all_participants.joblib")
output = model.predict(prepared_chain, feedback="none")
prepared_chain is constructed from the extracted feature CSV with
load_dataset(..., windows=[150]), prepare_window(..., 150), and
build_sss_chain(..., 300). It includes initial subjective/ocular references.
For cross-validation reproduction, load a reference_fold_N.joblib file and use
the matching held-out participant subset. These custom joblib
bundles require paper_ssm.py and paper_bundle.py from the companion package;
they are not standalone executables. Load only trusted serialized files.
The same bundle supports feedback="both": reveal the first, fifth and ninth
eligible session ratings after their predictions, update state and session logit
offset, and keep the trained SSM components fixed.
The dataset repository contains the extracted input features. The code is available on GitHub.
Cross-validation performance
F1 is the primary measure because it balances precision and recall for detecting elevated sleepiness (41.3% of the ratings). Their harmonic mean summarizes positive-class detection through true positives, false positives and missed detections. Comparisons use pooled binary F1 for SSS โฅ 4; fold-mean F1 and SD describe between-fold variability.
Across 504 held-out endpoints, the reference SSM achieves pooled F1 0.713, ROC-AUC 0.787 and Brier score 0.188. The SSM with subjective feedback achieves F1 0.817, ROC-AUC 0.894 and Brier score 0.123. These scores belong to the five held-out fits, not an independent test of the final all-participant model. Last-rating persistence, which carries forward the last available SSS with the same feedback schedule, gives F1 0.819, accuracy 0.853 and ROC-AUC 0.919. The SSM with subjective feedback and this comparator achieved similar observed pooled F1 (paired difference -0.002, 95% interval [-0.039, 0.033]), highlighting the predictive value of recent subjective assessments. Serialization/reloading was checked for exact prediction equality, including loading in a fresh Python process.
The outcome is SSS-defined self-reported sleepiness. The reference SSM had higher observed pooled F1 than the contextual controls, with paired intervals including zero. The 504 assessments come from 22 participant resampling units, which limits precision for modest performance differences. The seven geometry inputs describe eyelid opening, binocular alignment and ocular variability. The paper documents their literature rationale and the exploratory temporal comparison supporting the 150-s/300-s configuration. Larger independent cohorts and prospective evaluation can assess use in workplace monitoring.
Distribution and citation
The trained model artifacts and accompanying code are distributed under the MIT
License; see LICENSE and NOTICE.md for the license text and attribution. Cite
the paper and specific model revision used. Publication identifiers will be
added when available.