Paperdoi:10.3390/make8110379
Partition Choice Accounts for Fifty-Two Accuracy Points in Wearable Driving Context Classification
Machine Learning and Knowledge Extraction (MDPI), published 2026-11-16
TL;DR
With windows, features and classifier fixed, wrist EDA/BVP/ACC driving-context classification scored 0.909 under a stratified random 5-fold split and 0.388 when grouped by driver: a 52-point gap (55.7 for a CNN). Shuffling route labels between sessions left random-split accuracy at 0.910, and one window identified its session with 0.890 accuracy, so the random split was learning sessions, not routes. On AffectiveROAD the gap was small.
Abstract
Accuracies above ninety percent for driving context from wearable physiology usually come from random window splits. Electrodermal activity, blood volume pulse, and acceleration were recorded with an Empatica EmbracePlus over eighteen real road journeys by eleven drivers on rural, expressway, and urban routes (1520 non-overlapping 60 s windows). With windows, features, and classifier fixed, a stratified fivefold split returned 0.909, and a split grouped by driver 0.388—a gap of 52.0 accuracy points; a convolutional network showed a gap of 55.7 points. Each session covered one route. With route labels shuffled between sessions, the random split averaged 0.910 (p = 0.53), and a single window identified its session with 0.890 accuracy. On the seven drivers with more than one class, no configuration significantly exceeded a within-driver majority rule, and a two-class contrast free of the cohort confound reached 0.631 across five drivers, not significantly above chance. On public AffectiveROAD data, where every drive contains city and highway, the gap was 2.2 balanced accuracy points with all sensors and 9.1 with physiology alone, and the permutation test detected a signal (p = 0.005) mostly from acceleration. This study establishes the partitioning effect, not a classification result.
Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://creativecommons.org/licenses/by/4.0/
Results on ChatPPG leaderboards
All leaderboardsNo results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.
Tasks & signals
Signals
Keywords
How it works
Cite this paper
Em, P. P., & Teoh, T. S. (2026). Partition Choice Accounts for Fifty-Two Accuracy Points in Wearable Driving Context Classification. Machine Learning and Knowledge Extraction. https://doi.org/10.3390/make8110379
Em, Poh Ping and Tai Shie Teoh. 2026. "Partition Choice Accounts for Fifty-Two Accuracy Points in Wearable Driving Context Classification." Machine Learning and Knowledge Extraction. https://doi.org/10.3390/make8110379.
Em, P. P. and Teoh, T. S. (2026) 'Partition Choice Accounts for Fifty-Two Accuracy Points in Wearable Driving Context Classification', Machine Learning and Knowledge Extraction. doi: 10.3390/make8110379.
Em, Poh Ping, and Tai Shie Teoh "Partition Choice Accounts for Fifty-Two Accuracy Points in Wearable Driving Context Classification." Machine Learning and Knowledge Extraction, 2026. https://doi.org/10.3390/make8110379.
P. P. Em and T. S. Teoh, "Partition Choice Accounts for Fifty-Two Accuracy Points in Wearable Driving Context Classification," Machine Learning and Knowledge Extraction, 2026, doi: 10.3390/make8110379.
@article{em2026partition,
title = {{Partition Choice Accounts for Fifty-Two Accuracy Points in Wearable Driving Context Classification}},
author = {Em, Poh Ping and Teoh, Tai Shie},
year = {2026},
journal = {Machine Learning and Knowledge Extraction},
publisher = {MDPI},
doi = {10.3390/make8110379}
}Added to the ChatPPG research hub