Beyond Latent Personalization: Integrating HRV-Conditioned Personalization via LLM for Interpretable Stress Detection

Wei-Hsiang Ma; Ya-Tse Wu; Hsing-Hang Chou; Chi-Chun Lee

eess.SPOct 16, 2026

Beyond Latent Personalization: Integrating HRV-Conditioned Personalization via LLM for Interpretable Stress Detection

Authors: Wei-Hsiang Ma, Ya-Tse Wu, Hsing-Hang Chou, Chi-Chun Lee

Organizations: Institute of Communication Engineering, National Tsing Hua University, Taiwan · Department of Electrical Engineering, National Tsing Hua University, Taiwan

Abstract

Stress detection based on heart rate variability (HRV) has received increasing attention. However, existing personalized approaches integrate personal attributes with HRV through latent representations, limiting the interpretability of their interactions. In this work, we propose an LLM-enhanced personalized stress detection framework that integrates HRV features and personal attributes through LLM-based semantic interpretation. The proposed framework achieves a 29.9% relative MCC improvement over the strongest baseline on TILES-2018 while providing human-readable descriptions of individual physiological responses. The results highlight the potential of LLM-based semantic interpretation for achieving both effective and interpretable personalized stress detection.

Figures & tables

Figure 1: Overview of the proposed BLP framework.

Attribute Integration

[Neutral Baseline Profile] This individual is a { gender } participant in the { age } age group, and { student_desc } . Professionally, { subject } works as a(n) { position } , with { nurseyears } of professional experience and { duration } of tenure at the current organization. Working the { shift } , { subject } averages { hours } per week with approximately { overtime } of monthly overtime, and supervises { quantsup } . Regarding personal background, { subject } identifies as { race } and has { englyrs } of English language experience. { possessive } current relationship status is { relationship } , with an annual household income of { income } . { Subject } has { children } child(ren) under 18 in the household and { housing_desc } . { possessive } typical daily commute is { commute_time } .

HRV Integration (Take time domain as example)

[Data Interpretation Context] Note: The following HRV metrics are not instantaneous readings. They are strictly calculated based on rigorous daily aggregation: Feature Group: ’Time’: ”Time-domain HRV features (e.g., meanNN, sdNN, RMSSD, pNN50, cvNN)”; Base calculation: Extracted from 5-minute rolling windows of Normal-to-Normal (NN) intervals. Time-domain differences and Frequency domains were normalized; Suffixes meaning: Metric ending in ‘_mean’ is the daily average; ‘_std’ is the daily standard deviation; ‘_cv’ (Coefficient of Variation) shows daily volatility; ‘_skew’/‘_kurt’ show distribution shape; ‘_1quartile’/‘_3quartile’ are the 25th/75th percentiles. Please interpret these metrics as the overall autonomic stability and allostatic load for the entire day.

[Current Physiological Markers] { feature_1 } : { value_1 } { feature_2 } : { value_2 } … { feature_N } : { value_N }

General

[Task] Based on the individual’s baseline profile and today’s HRV metrics, write a daily analytical log assessing their current stress state.

Table 1: Physiological Interpretation Prompt Template

Table 2: Overview of the 61 HRV Features

Methods						TILES-2018			WESAD
	HRV	Semantic Representation				UAR	Acc	MCC	UAR	Acc	MCC
	HRV	LLM Interp.	Extractor	HRV Integ.	Attr. Integ.	UAR	Acc	MCC	UAR	Acc	MCC
LLM-Only	$\times$	$\surd$	$\surd$	$\surd$	$\surd$	0.516	0.504	0.040	0.616	0.683	0.260
HRV-Only	$\surd$	$\times$	$\times$	$\times$	$\times$	0.600	0.588	0.201	0.799	0.821	0.599
w/o Attribute Integration	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	0.594	0.600	0.189	0.798	0.820	0.595
w/o HRV Integration	$\surd$	$\surd$	$\surd$	$\times$	$\surd$	0.560	0.557	0.119	0.787	0.809	0.576

Table 3: Ablation Results on the TILES-2018 and WESAD Datasets. Interp. = Interpretation; Integ. = Integration; Attr. = Attribute; Extractor = Semantic Embedding Extractor.

Method	UAR	Acc	MCC	MCC $p$ -value
Trans.+MTL+GRL [ 11 ]	0.584	0.603	0.172	0.0019*
Trans.+MGCN+MS [ 6 ]	0.597	0.613	0.196	0.0085*
DESHM [ 15 ]	0.598	0.619	0.201	0.0496*
Proposed	0.630	0.631	0.261	–

Table 4: Baseline Comparison on TILES-2018 (

{}^{*}p<0.05

w/o HRV Integration (Pred: 0, GT: 1) : …the cumulative effect of prolonged overtime, leadership duties, and a lengthy commute creates a state of sustained activation where the body and mind are likely operating under continuous demand, potentially leading to physiological weariness and psychological strain… Proposed (Pred: 1, GT: 1) : …This pattern is further underscored by the significant variability in very-low-frequency power , which reflects a disrupted circadian rhythm or allostatic load, likely exacerbated by the cumulative demands of his supervisory role, extended overtime hours, and long commute …

w/o HRV Integration (Pred: 1, GT: 0) : …With over 15 years of experience in clinical support and a tenure of 6 to 9 years at the current organization , she possesses significant institutional knowledge and likely established coping mechanisms … Furthermore, her stable marital status and the presence of a child under 18 indicate strong social anchoring…

Proposed (Pred: 0, GT: 0) : … significantly elevated parasympathetic activity , as evidenced by the exceptionally high pNN50 and RMSSD values , which suggest a strong capacity for recovery and stress buffering … Given her extensive professional tenure and established home environment , this pattern aligns with a resilient allostatic load profile where the body effectively manages daily demands without exhibiting signs of chronic depletion or dysregulation…

Table 5: Case Study of HRV-Conditioned LLM Interpretation

Interpretable ML

Interpretable machine learning develops models and explanations that make predictions, evidence, and decision processes understandable to people.

See all 310 papers

Retrieval-Augmented Personalization with Foundation Models for Wearable Stress Detection
Jun 23, 2026Louis Simon, Mohamed ChetouaniWearable Health Monitoring
Institut des Systèmes Intelligents et de Robotique Sorbonne University Paris, France
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
Explainable graph attention network for stress recognition (StressGAT) via differential action units
Jul 23, 2026Thomas Kassiotis, Stefanos Gkikas, Nikolaos Smyrnis +1Graph Attention Networks Facial Expression Recognition
Department of Electronic Engineering Hellenic Mediterranean University Chania, Greece · Honda Research Institute Japan Wako City, Japan · University Mental Health, Neurosciences and Precision Medicine, Research Institute “COSTAS STEFANIS” 2nd Department of Psychiatry, Medical School, National and Kapodistrian University of Athens Athens, Greece +1

Jun 23, 2026J/K move · Enter open · S save

Retrieval-Augmented Personalization with Foundation Models for Wearable Stress Detection

Louis Simon, Mohamed Chetouani

Institut des Systèmes Intelligents et de Robotique Sorbonne University Paris, France

Wearable Health Monitoring

Open paper AskPDF arXiv

Personalization in wearable-based stress detection remains challenging due to substantial inter-individual variability in physiological and behavioral responses. While traditional approaches rely on user-specific fine-tuning or costly self-supervised pre-training on large datasets, we propose a lightweight alternative based on retrieval-augmented personalization. Our method leverages frozen, out-of-domain foundation models to retrieve similar patterns from a target user's history and encode them into a compact personalized embedding that modulates representations extracted by a lightweight transformer network. We evaluate our approach on the WESAD stress detection dataset with N=15 users, comprising wrist-worn physiological (EDA, BVP, temperature) and activity (accelerometer) signals, and report gains of +3.92% in accuracy and +4.76% in macro F1-score over a non-personalized transformer baseline, approaching supervised fine-tuning performance without requiring any labeled user data. We further show that temporal retrieval, where only prior user samples are available, achieves performance close to full intra-user retrieval, demonstrating robustness to limited user history. Finally, we explore personalization in a cross-dataset retrieval setting, leveraging embeddings from the K-Emocon dataset to personalize representations for stress detection on the WESAD dataset.

Abstract

Interpretable ML

Explore similar work

Retrieval-Augmented Personalization with Foundation Models for Wearable Stress Detection

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

Explainable graph attention network for stress recognition (StressGAT) via differential action units