The Individual Conformal Coupling Monitor is introduced, 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.
Abstract
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.
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.
Jaden Moon, Y. Wu, Arvind Pillai et al.· 0 citations
Large language models provide a promising framework for wearable-based health prediction by converting structured physiological and behavioral measurements into natural-language prompts. In this paper, we investigate whether pre-trained lightweight open-weight LLMs can predict depression and anxiety symptoms from short-horizon consumer wearable data. Using 4-8 days of Oura Ring data from 1,285 participants in the Northern Finland Birth Cohort 1986, we convert activity, sleep, heart rate, heart rate variability, demographic, and anthropometric measurements into structured prompts. We evaluate Llama 3.1, BioMistral, and Qwen 2.5 under zero-shot, rule-based, and few-shot in-context learning settings. To contextualize LLM performance, we compare them against machine learning models and recurrent neural networks. Our results show that prompt design is critical for LLM-based wearable inference. Zero-shot LLMs achieve high accuracy but largely predict the majority class, failing to identify participants with depression and anxiety symptoms. In contrast, few-shot prompting substantially improves positiveclass detection. Llama 3.1 with four in-context examples achieves the strongest performance, with 0.92 accuracy, 0.82 macro-F1, and 0.69 F1 for the positive class, among evaluated models. These findings suggest that lightweight LLMs can use in-context examples to better interpret structured wearable summaries and possibly provide a scalable direction for mental health prediction from consumer wearable data in combination with pre-trained LLMs.
S. Azadifar, A. Sameh, M. Niemela et al.· medRxiv· 0 citations
Removing wearables from physiological monitoring also removes their supervision: the signal indicating where and when a stress response occurred. Contactless stress sensing therefore becomes a weakly supervised evidence-localization problem, where a clip-level label must be traced to the body regions and moments that produced it. We address this with FABLE-Therm, a weakly supervised architecture that preserves localized evidence across body regions, time, and encoder-specific representations until the final decision. FABLE-Therm fuses frozen foundation-model encoders at the embedding level, with theory explaining why localized fusion can outperform feature concatenation and prediction averaging. We study this problem in opioid use disorder (OUD), where stress is a major relapse trigger and sustained wearable use can be difficult during early recovery. Using fixed thermal video, FABLE-Therm achieves 0.938 AUROC on held-out participants, and its learned representation transfers to self-reported craving, providing, to our knowledge, the first evidence that craving can be recovered from contactless thermal video. Localized evidence also enables participant-level analysis of deployment failure. We find that improving representation alone is insufficient for equitable deployment: additional data from the underserved group would recover only about half of the cohort gap, while the remainder reflects person-to-person heterogeneity. This modality-agnostic decomposition applies to models with identifiable subpopulations. Together with the first cohort-structured contactless thermal OUD benchmark, our results show that preserving localized evidence supports both accurate sensing and principled analysis of who a model fails and why.
Sachin Deb, Harshit Sharma, A. Salekin· 0 citations
This paper presents a two-stage pipeline for implicit feature engineering in time series-based physiological stress detection using electrodermal activity (EDA) signals. In the first stage, we forecast three descriptive statistics of future EDA signals over short horizons (3, 5, and 10 seconds) based on a 60-second context window. In the second stage, a lightweight linear classifier detects stress from these predicted statistics. We evaluate three forecasting architectures spanning the domain expertise spectrum: a domain-specific bidirectional long short-term memory (BiLSTM) recurrent neural network, zero-shot and fine-tuned variants of Amazon Chronos T5 time series foundation model, and the Tabular Prior-data Fitted Network (TabPFN) applied to engineered physiological features. Experiments on the publicly available Wearable Stress and Affect Detection (WESAD) dataset, comprising chest-worn multimodal physiological signals from 15 subjects under baseline and stress conditions, use subject-independent 5-fold cross-validation and show that the domain-specific BiLSTM and TabPFN achieve comparable classification performance, with mean area under the receiver operating characteristic curve (AUC) values of 0.859–0.882 and 0.863–0.883 respectively. Both remain well ahead of the Chronos variants, which yield 0.629–0.777. Chronos models quickly reach performance saturation regardless of training depth, highlighting challenges in tokenizing continuous physiological time series. The proposed approach advances implicit feature engineering for wearable stress monitoring by leveraging forecasting as a powerful inductive bias, thereby improving robustness and providing insights into the limitations of the foundation model for physiological signals.
J. G. Gonzalez Nunez, Soheil Sabri, Parham M. Kebria et al.· bioRxiv· 0 citations
Stress remains a significant health concern, yet commercial monitoring devices are often too expensive and complex for populations in developing countries. This research presents Spondon, a low-cost, multi-sensor wearable designed to detect stress in resource-constrained environments. Utilizing an ESP32 microcontroller with MAX30102, GSR, and temperature sensors, the device collects physiological data. We evaluated the system with 15 participants using the Stroop test to induce stress across baseline, task, and recovery phases. Statistical analysis confirmed a distinct 7.95% drop in HRV (RMSSD) during stress. The system achieved a 75.0% accuracy in binary acute stress detection (Baseline vs. Stress) using a Random Forest classifier with personalized baseline normalization. In addition, when evaluating continuous real-world monitoring across three physiological states (Baseline, Stress, and Recovery), Spondon achieved a 58.4% multi-class accuracy, successfully outperforming the 33.3% random-chance baseline. This work contributes a validated, accessible hardware platform and methodology, democratizing stress research for low-income communities and highlighting the importance of personalized, context-aware design in HCI.
Shahariar Ifti, R. Rony, Shuvashish Chakraborty et al.· The Compass· 0 citations
Mental health conditions are widespread, yet scalable assessment remains limited, as clinical interviews are time-intensive and self-report screeners are episodic and prone to bias, motivating passive, data-driven approaches. This study investigates whether short-term physical activity patterns captured by consumer wearables can classify self-rated mental health in a large, diverse cohort. Using the All of Us Research Program’s Registered Tier Dataset v7, we aligned Fitbit activity data with self-rated mental health, binarized into positive (good/very good/excellent) and negative (poor/fair) classes. Addressing class imbalance (88.5% positive, 11.5% negative) issue through undersampling resulted in a final cohort of 2,092 participants (837 negative, 1,255 positive). In the cohort, each participant was represented by eleven seven-day physical activity features together with age. Five models, logistic regression (LR), support vector machine (SVM), decision tree (DT), random forest (RF), and XGBoost, were trained and evaluated using AUC, accuracy, precision, recall, and F1. Result indicated that XGBoost performed best (AUC = 0.77, F1 = 0.774, recall = 0.883), followed by RF, while LR and SVM showed similar performance (AUC ≈ 0.75) and DT performed worst (AUC = 0.63). These outcomes suggest that short-term wearable activity data can inform mental health classification, with ensemble methods performing best. In practice, such systems could enable low-burden, continuous monitoring to support early identification of at-risk individuals, but should be designed as decision-support tools that complement, but not replace, clinical assessment, with careful attention to interpretability, user trust, and appropriate handling of uncertainty.
Guannan Liu, Jae Yoo, Gaojian Huang et al.· Proceedings of the Internati...· 0 citations
