ReliaGate: Reliability Routing for Low-Stakes Wearable Stress Prediction

Jaden Moon; Yu Wu; Arvind Pillai; Andrew Campbell

cs.LGAug 16, 2026

ReliaGate: Reliability Routing for Low-Stakes Wearable Stress Prediction

Authors: Jaden Moon, Yu Wu, Arvind Pillai, Andrew Campbell

Organizations: Dartmouth College Hanover, New Hampshire, USA

Abstract

We study when a wearable stress system should surface a prediction rather than change it. In low-stakes reflection and summary settings, aggregate accuracy is insufficient because withholding can reduce error while leaving some people with little or no information. We formulate fixed-label reliability routing: after a locked classifier emits a protocol-defined stress/non-stress label, a post-hoc gate surfaces that unchanged label or withholds it as unavailable. ReliaGate assembles established confidence, signal-quality/trust, agreement, train-standardized atypicality, and train-fitted geometry cues into a post-hoc correctness score. We evaluate four wearable datasets using subject-disjoint folds, validation-selected routing, paired held-out-subject intervals, and pooled and per-subject analyses. WESAD point estimates favored ReliaGate, UBFC-Phys primary coverage/risk intervals favored ReliaGate, and E4 checks were mixed. ReliaGate provides an operational framework for studying surfaced-label error, output availability, and accepted-output distribution across subjects, without revising labels or providing clinical or finite-sample risk guarantees.

Figures & tables

Figure 1. ReliaGate routes an unchanged label using confidence, quality/trust, agreement, train-standardized atypicality, and train-fitted geometry. Outer-training subjects fit models and references; validation selects checkpoint, temperature, scorer, and threshold; test data are reporting-only. Canonical feature-model predictions are subject-wise out of fold (OOF), while primary geometry may retain same-subject windows. A left-to-right five-stage pipeline. Wearable BVP, EDA, TEMP, and optional ACC enter a locked predictor that emits a fixed stress/non-stress label and learned representations. Reliability evidence combines prediction confidence, signal quality and modality trust, cross-stream agreement, train-standardized atypicality, and embedding geometry relative to outer-training references. A correctness scorer assigns a score, and a validation-selected threshold either surfaces the unchanged label or withholds it; no replacement label is produced. A lower panel shows subject-disjoint roles: training fits models and references, validation selects the checkpoint, temperature, scorer, and threshold, and held-out test subjects are used only for reporting.

Dataset	Evaluation role	Model inputs	Non-stress periods	Stress periods	Folds	Windows	Counts N/S
WESAD ( Schmidt et al., 2018 )	Primary mechanism and robustness analysis	BVP, EDA, TEMP	Baseline; amusement	TSST	15	2,986	2,099 / 887
UBFC-Phys ( Meziati et al., 2021 ; Sabour et al., 2023 )	Controlled BVP/EDA evaluation	BVP, EDA	T1 rest	T2 speech; T3 arithmetic	56	1,848	616 / 1,232
EmpaticaE4Stress ( Campanella et al., 2023a ; Campanella et al., 2024 )	E4 protocol check	BVP, EDA, TEMP; ACC quality	Rest blocks	Task blocks	29	4,301	1,044 / 3,257
PhysioNetE4 ( Hongn et al., 2025a ; Hongn et al., 2025c )	Acute-stressor protocol check	BVP, EDA, TEMP; ACC quality	Baseline/rest blocks	Stressor blocks	34	4,784	3,973 / 811

Table 1. Analyzed post-filter samples, manuscript-defined binary tasks, and held-out-subject folds.

System	$N_{\mathrm{acc}}$	Cov.	Risk	SF	mF1	BAcc	Reach
Always	2,986	1.000	0.325	0.325	0.638	0.652	1.000
Conf.	619	0.207	0.288	0.060	0.625	0.668	0.533
NG	1,594	0.534	0.233	0.124	0.719	0.719	0.933
RG	1,908	0.639	0.175	0.112	0.774	0.767	0.933

Table 2. WESAD routing at

\alpha=0.10

N_{\mathrm{acc}}

: accepted windows; Conf.: confidence-only; Cov.: coverage; Risk: accepted risk; SF: silent-failure rate; mF1: accepted macro-F1; BAcc: accepted balanced accuracy; Reach: subject reach.

System	$N_{\mathrm{acc}}$	Cov.	Risk	SF	mF1	BAcc	Reach
Always	1,848	1.000	0.366	0.366	0.607	0.614	1.000
Conf.	534	0.289	0.324	0.094	0.649	0.652	0.554
NG	1,057	0.572	0.311	0.178	0.614	0.614	0.893
RG	1,244	0.673	0.256	0.172	0.696	0.691	0.946

Table 3. UBFC-Phys routing at

\alpha=0.10

; abbreviations follow Table 2 .

Dataset/system	Cov.	Risk	SF	mF1	BAcc	Reach
Empatica–NG	0.681	0.178	0.121	0.731	0.756	1.000
Empatica–RG	0.742	0.175	0.130	0.757	0.788	1.000
PhysioNet–NG	0.536	0.254	0.136	0.554	0.562	0.912
PhysioNet–RG	0.545	0.266	0.145	0.566	0.574	0.882

Table 4. Independent within-dataset E4 outcomes at

\alpha=0.10

; abbreviations follow Table 2 .

Explore similar work

When Clean Signals Are Not Enough: Detecting Structural Ambiguity for Safe Wearable Stress Classification
Aug 19, 2026Saba A. Farahani, Hung Cao, Amir M. RahmaniWearable Health Monitoring Selective Prediction
University of California, Irvine Irvine, CA, USA
ReliableNet: A Chance-Constrained Approach to Trustworthy Classification in Deep Learning
Aug 10, 2026Ange-Clément Akazan, Ineza Remy Mugenga, Abebe Geletu +2Chance-Constrained Optimization Classification
School of Agriculture and Science, University of KwaZulu-Natal, Pietermaritzburg, KwaZulu-Natal, South Africa · 2AIMS Research and Innovation Centre, African Institute for Mathematical Sciences, Kigali, Rwanda · 3African Institute for Mathematical Sciences, Kigali, Rwanda
AdaptStress: Online Adaptive Learning for Interpretable and Personalized Stress Prediction Using Multivariate and Sparse Physiological Signals
Feb 19, 2026Xueyi Wang, Claudine J. C. Lamoth, Elisabeth WilhelmWearable Health Monitoring Explainable Time Series Forecasting
Engineering and Technology institute Groningen (ENTEG), University of Groningen, Groningen, 9742AG, the Netherlands · Department of Human Movement Sciences, University Medical Center Groningen, Groningen, Netherlands

Aug 19, 2026J/K move · Enter open · S save

When Clean Signals Are Not Enough: Detecting Structural Ambiguity for Safe Wearable Stress Classification

Saba A. Farahani, Hung Cao, Amir M. Rahmani

University of California, Irvine Irvine, CA, USA

Wearable Health Monitoring Selective Prediction Conformal Prediction

Open paper AskPDF arXiv

Wearable stress classifiers can achieve strong average performance while failing completely for a particular individual. On WESAD, a Random Forest reaches 93.0% mean accuracy yet yields F1 = 0 for Subject 14, whose cross-signal coupling weakens near stress onset. We call this structural ambiguity: individually plausible physiological channels form an inter-signal pattern that is poorly supported by the person's non-stress reference. We introduce the Individual Conformal Coupling Monitor (ICCM), a lightweight and transparent pre-inference monitor that quantifies subject-specific coupling divergence and routes each window to classify, defer, or abstain without retraining the downstream classifier. Across WESAD (N = 15) and Stress-Predict (N = 35), full-cohort Pearson associations between ambiguity and accuracy are negative (r = -0.607, p = 0.016; r = -0.412, p = 0.014). Robustness analyses temper this finding: rank correlations are not significant, and the WESAD association disappears when Subject 14 is removed. ICCM changes false-positive counts from 29 to 27 and 94 to 92, although neither paired change is significant. It withholds 3 of Subject 14's 21 stress windows but does not repair the missed-stress failure. These results position ICCM as an interpretable signal of unsupported physiology and individual failure, rather than a stand-alone safety guarantee.

Abstract

Explore similar work

When Clean Signals Are Not Enough: Detecting Structural Ambiguity for Safe Wearable Stress Classification

ReliableNet: A Chance-Constrained Approach to Trustworthy Classification in Deep Learning

AdaptStress: Online Adaptive Learning for Interpretable and Personalized Stress Prediction Using Multivariate and Sparse Physiological Signals