PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2608.18397v1cs.AI

TL;DR

Wearable stress classifiers can achieve high average accuracy while concealing severe failures for specific individuals. This paper introduces ICCM, a subject-specific monitor that routes ambiguous windows before inference, and finds that coupling divergence is significantly associated with lower accuracy across two datasets, though it does not establish a stand-alone safety mechanism.

  • Problem

    High mean wearable stress-classification accuracy can conceal severe failures for specific individuals, motivating methods that identify unsupported inter-signal relationships.

  • Method

    ICCM calibrates a subject-specific non-stress coupling reference and uses a classifier-external three-zone gate to classify, defer, or abstain before inference.

  • Results

    Across two datasets, significant full-cohort Pearson associations show that greater coupling divergence accompanies lower subject-level accuracy.

  • Takeaways & Limitations

    ICCM provides a transparent physiological routing signal for structurally unsupported inputs and severe individual classifier failure, but is not a validated stand-alone safety mechanism.

  • Takeaways & Limitations

    Reliability across repeated sessions, additional classifiers, and naturalistic cohorts remains untested, and abstention is not proof that a decision is harmless.

Abstract

from arXiv · show

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.

I. INTRODUCTION

Wearable stress classifiers can achieve high average performance while failing severely for particular individuals when plausible channels exhibit unsupported inter-signal coupling. ICCM addresses this pre-inference safety gap with a subject-specific, classifier-external monitor that routes windows to classify, defer, or abstain without retraining.

  • Motivation: 93% mean accuracy coexists with F1 = 0.000 for Subject 14 on WESAD, where EDA–BVP coupling weakens near stress onset.The channels contain no missing samples, but residual artifacts cannot be excluded.
  • Limitations of existing approaches: Existing architectures, augmentation, and robust optimization can improve average accuracy but do not determine whether current coupling structure is supported before inference.Confidence and output-level uncertainty are computed only after classifier processing and do not directly detect structurally invalid inputs.
  • Proposed approach: ICCM calibrates a subject-specific non-stress coupling reference and applies a three-zone gate to classify, defer, or abstain before inference.The monitor is classifier-external, requires no model retraining, and multiarchitecture performance remains untested.
  • Problem definition: Structural ambiguity denotes insufficiently supported inter-signal coupling despite individually plausible physiological channels.It captures a subject-specific pattern that may be poorly supported by the person’s non-stress reference.

II. RELATED WORK

Prior wearable stress-detection and abstention methods address multimodal classification, signal quality, behavioral coupling, and uncertainty. ICCM instead monitors subject-specific inter-signal coupling for classifier-external routing, without a formal conformal-coverage guarantee here.

  • Wearable Stress Detection: Wearable stress detection commonly uses BVP, EDA, and skin temperature with Random Forests and other models achieving high average LOSO performance on WESAD.ICCM instead checks whether each window is supported by a subject-specific non-stress reference; its Stress-Predict extension also uses stress labels from LOSO training subjects.
  • Signal Quality and Abstention: Signal-quality methods detect hardware degradation, motion artifacts, or poor recordings, whereas ICCM checks inter-signal coupling and provides a classifier-external physiological routing reason.Behavior-adaptive models show interpretable coupling changes across behavioral phases, while selective prediction and clinical AI use abstention under high risk.
  • Signal Quality and Abstention: ICCM’s overlapping baseline and rank windows preclude a formal conformal-coverage guarantee here.Conformal and Bayesian uncertainty methods support abstention, but ICCM is framed as a physiological, classifier-external routing approach.

III. METHOD · A. External Sources

ICCM is a deterministic physiological filter that combines hybrid coupling nonconformity with empirical rank calibration. Its external inputs are Empatica E4 BVP, EDA, and TEMP signals, with HR derived from BVP and processed in overlapping windows.

  • III. METHOD: ICCM is deterministic and uses physiological measurements as a filter.The method is described as a deterministic physiological filter.
  • III. METHOD: Hybrid coupling nonconformity is one basis of ICCM’s filtering procedure.ICCM combines hybrid coupling nonconformity with empirical rank calibration.
  • III. METHOD: Empirical rank calibration is the second stated basis of ICCM’s filtering procedure.The method explicitly uses empirical rank calibration alongside hybrid coupling nonconformity.
  • III. METHOD: ICCM has three components.The passage states that the method consists of three components, shown in Fig. 2.
  • A. External Sources: 64 Hz BVP and 4 Hz EDA and TEMP are provided by Empatica E4.The external sources specify the device and sampling rates for all three signals.
  • A. External Sources: HR is derived from BVP using sliding 5-s peak detection, while signals use 60-s windows with a 30-s step.The external-source processing combines BVP-derived heart rate with overlapping signal windows.

B. Interface · C. Orchestrator

The Interface returns classifications only when physiological coupling is structurally supported, while the Orchestrator calibrates subject-specific coupling and monitors deviations for routing. Its conformal-style monitoring remains constrained by limited calibration requirements and unestablished exchangeability.

  • B. Interface: The Interface returns the classifier output when structural support is present and reports insufficient physiological evidence otherwise.
  • C. Orchestrator: Two phases—calibration and real-time monitoring—define the Orchestrator’s coupling analysis and routing process.
  • C. Orchestrator: Once per individual, calibration computes hybrid coupling vectors over EDA–HR, EDA–TEMP, and HR–TEMP signal pairs using known non-stress windows.
  • C. Orchestrator: Coupling features combine absolute Pearson correlation, maximum absolute cross-correlation over physiological delays, and normalized Granger-style directed coupling.The delay range is τ ∈[1s, 10s], and gx→y = min(−log pGC, 10)/10 is normalized to [0, 1].
  • C. Orchestrator: The calibration distance set D = {D1, . . . , DK} represents the individual’s normal coupling variation.
  • C. Orchestrator: At each window t, real-time monitoring computes v(t) and evaluates an empirical conformal-style rank score against the calibration distribution.
  • C. Orchestrator: A low p(t) flags coupling divergence from the individual baseline as evidence of structural ambiguity, but overlapping windows do not establish exchangeability or split-conformal independence.
  • C. Orchestrator: At α = 0.05, Zone 3 requires K ≥40, while 60-s windows with a 30-s step require at least 20.5 minutes of contiguous calibration.For WESAD, K = 72–76 because all labeled non-stress periods are used.

D. Output: 3-Zone Safety Gate

The Orchestrator applies a three-zone gate at α = 0.05 to route each window by p(t), forwarding supported windows, withholding borderline windows, and abstaining on unsupported windows. Abstention is a routing action, not a clinical safety guarantee.

  • Zone Definitions: α = 0.05 defines the three-zone routing threshold applied to each window using p(t).The Orchestrator routes windows according to their p(t) values.
  • Zone 1: Zone 1 (p(t) ≥ α) forwards structurally supported windows to the downstream Random Forest classifier.This zone represents windows meeting or exceeding the threshold.
  • Zone 2: Zone 2 (α/2 ≤ p(t) < α) withholds borderline windows from classification and issues no prediction.Borderline windows are neither forwarded nor assigned a classifier output.
  • Zone 3: Zone 3 (p(t) < α/2) triggers abstention for structurally unsupported windows, but this routing action is not itself a clinical safety guarantee.The gate therefore identifies unsupported structure without claiming clinical safety from abstention alone.

E. Protocol-Aware Coupling Selection

ICCM uses protocol-specific coupling features: magnitude-based divergence for single-protocol WESAD and direction-aware coupling for multi-protocol Stress-Predict. These configurations rely on LOSO training-subject information, while the test subject’s reference remains label-free, and are exploratory rather than universally transferable.

  • Single-protocol configuration: WESAD uses six features, ρ and ℓonly, because directed coupling adds noise when protocol variability is low.Magnitude-based divergence from the individual baseline is sufficient to detect coupling collapse in the single-protocol setting.
  • Multi-protocol configuration: Stress-Predict uses all nine features, ρ, ℓ, and g, with a direction-aware score to accommodate multi-protocol variability.The population mean coupling-change direction is estimated from training subjects under LOSO.
  • Data-use assumptions: Both the population mean coupling-change direction and empirical routing distribution use labeled stress windows from LOSO training subjects, while only the test subject’s reference is label-free.This separates training-subject supervision from test-subject reference construction.
  • Configuration limitations: The two dataset-specific configurations were selected after ablation and remain exploratory; a fixed unsupervised configuration does not transfer to Stress-Predict (r = 0.474).The supplied passage also begins to describe a fixed direction-aware configuration but does not provide its complete result.

F. Signal-Quality Diagnostic

Subject 14 passed basic signal-validity checks, but quantized EDA and temperature and non-minimal motion left residual motion/contact artifact as an alternative explanation. The diagnostic is not a validated device-specific quality index.

  • Basic signal checks: All samples were finite, and detected BVP inter-beat intervals were plausible within 0.3–2.0 s.Screening also covered channel ranges, constant runs, and wrist-acceleration magnitude.
  • Motion and quantization: 80th percentile: stress-window motion was within the cohort range, but motion was not minimal.EDA and temperature were highly quantized, limiting confidence that the signals were artifact-free.
  • Limitations: The diagnostic is not a validated device-specific quality index, and residual motion/contact artifact remains an alternative explanation.These caveats qualify interpretation of Subject 14’s signal-quality findings.

IV. EXPERIMENTS · A. Datasets · B. Experimental Setup

The experiments evaluate binary wearable-stress classification on WESAD and Stress-Predict using subject-wise validation, subject-specific ICCM calibration, and a Random Forest downstream classifier. Analyses compare association measures, abstention strategies, coverage-matched performance, and paired false-positive changes.

  • A. Datasets: WESAD includes Empatica E4 recordings from 15 subjects across baseline, amusement, meditation, and laboratory stress conditions, using BVP, EDA, and TEMP.The task is binary classification.
  • A. Datasets: Stress-Predict includes Empatica E4 recordings from 35 subjects performing Stroop and Interview stress tasks, with hyperventilation segments excluded.Remaining segments are treated as binary baseline-versus-stress data.
  • B. Experimental Setup: All experiments use LOSO cross-validation, with ICCM calibrated on each test subject’s labeled non-stress windows.This calibration is subject-specific and occurs within the test-subject evaluation process.
  • B. Experimental Setup: The downstream classifier is a Random Forest with 200 trees, trained on 14 time-domain features from the remaining subjects.The experimental significance level is α = 0.05.
  • B. Experimental Setup: The study reports Pearson and Spearman associations, Pearson correlation without Subject 14, and leave-one-subject-out influence.These analyses assess the relationship between ambiguity and classification outcomes under multiple correlation and influence procedures.
  • B. Experimental Setup: At matched coverage, ICCM is compared with random and confidence abstention, while confusion counts include only covered windows and abstention is not correct.Subject-paired false-positive changes are tested with a two-sided Wilcoxon signed-rank test.

C. Results

ICCM identifies subject-specific coupling divergence associated with lower accuracy, but robustness checks show that this association is high-leverage and rank-unstable. Its safety gate removes few false alerts and does not repair Subject 14’s missed-stress failure.

  • Structural Ambiguity Detection: Neither Spearman test is significant, and removing Subject 14 changes WESAD Pearson r from −0.607 to 0.185.Excluding Subject 14 is the sole influence-analysis exception and reverses the WESAD association’s sign.
  • Safety Gate Performance: 2 false alerts are removed in each dataset, changing counts from 29 to 27 and 94 to 92, but neither paired change is significant.Random and confidence baselines remove more Stress-Predict false alerts at matched coverage.
  • Safety Gate Performance: 0.129 to 0.131 and 0.951 to 0.949 are the covered sensitivity and specificity changes, respectively.Two Stress-Predict subjects lose more than two percentage points of covered-window accuracy.
  • Safety Gate Performance: 3 of 21 Subject 14 true-stress windows are withheld, while predictions for the remaining 18 are all false negatives.ICCM detects part of the anomalous interval but does not repair the motivating missed-stress failure.

V. DISCUSSION · VI. CONCLUSION

ICCM preserves the finding that individualized coupling divergence can expose structurally unsupported inputs and severe individual classifier failure beyond aggregate accuracy or classifier confidence. However, robustness limitations and unresolved validation needs position ICCM as a transparent candidate component, not a stand-alone safety guarantee.

  • V. DISCUSSION: Personalized coupling divergence exposes individual classifier failure that aggregate accuracy can conceal.The signal is external to classifier confidence and targets structurally unsupported inputs.
  • V. DISCUSSION: WESAD results are driven by a high-leverage case, and Stress-Predict has low mean F1.These dataset characteristics bound the contribution of the reported associations.
  • V. DISCUSSION: Confidence thresholding removes more false alerts, and configuration selection is post hoc.These analytic conditions constrain comparisons and practical interpretation.
  • V. DISCUSSION: Overlapping calibration windows preclude a formal coverage claim for ICCM.Accordingly, the monitor cannot yet support a formal conformal guarantee.
  • V. DISCUSSION: ICCM should complement classifier uncertainty and clinical escalation rather than establish that abstention or a non-stress decision is harmless.“Safe” describes a safety-oriented system objective, not proof of harmless outcomes.
  • V. DISCUSSION: Reliability across repeated sessions, window-length sensitivity, device-specific quality indices, selective-risk uncertainty, additional classifiers, and naturalistic cohorts remain future work.These studies are needed to assess generalizability and operational reliability.
  • VI. CONCLUSION: Across two datasets, negative Pearson associations motivate coupling divergence as an ambiguity signal, while robustness analyses prevent overinterpretation.The conclusion retains the cross-dataset finding but explicitly limits its interpretation.
  • VI. CONCLUSION: ICCM is a transparent candidate component for safer wearable stress systems, not yet a validated stand-alone safety mechanism.The conclusion frames ICCM as interpretable support for safer systems rather than an independent guarantee.
Loading 2608.18397v1…