PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.23555v1eess.SPcs.LG

TL;DR

Personalized HRV stress detection has used latent representations that limit insight into how personal attributes contextualize physiological responses. BLP integrates HRV and personal attributes through LLM-generated semantic descriptions, achieving a 29.9% relative MCC improvement over the strongest baseline on TILES-2018 while providing human-readable descriptions.

  • Problem

    Existing personalized HRV approaches integrate personal attributes through latent representations, limiting insight into how those attributes contextualize changing HRV responses.

    Source: §1 · p. 1

  • Method

    BLP combines HRV features with LLM-generated domain-specific descriptions that integrate personal attributes and HRV information for stress classification.

    Source: §1 · p. 1

  • Results

    29.9% relative MCC improvement over the strongest baseline on TILES-2018, while performance on WESAD supports generalizability across datasets.

    Source: §1 · p. 1

  • Takeaways & Limitations

    BLP provides human-readable descriptions of individual physiological responses alongside improved stress classification.

    Source: §1 · p. 1

    LLM-generated semantic representations alone are insufficiently discriminative, so the framework remains reliant on HRV representations.

    Source: §5 · p. 4

Abstract

from arXiv · show

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.

1. INTRODUCTION

HRV responses vary across people, making personal context relevant to stress detection. BLP uses LLM-generated descriptions to integrate personal attributes with HRV, addressing the limited interpretability of latent feature-based approaches.

  • Prior personalized approaches integrate HRV and personal attributes through latent representations, limiting insight into how attributes contextualize changing HRV responses.Examples include contrastive learning, attribute-similarity regularization, and hypergraph learning.Source: §1 · p. 1
  • BLP combines HRV features with LLM-generated descriptions integrating personal attributes and domain-specific HRV, then uses orthogonality and learnable source weights to aggregate domains.Descriptions expose how personal context relates to physiology, while source weights quantify each HRV domain’s contribution.Source: §1 · p. 1
  • 29.9% relative MCC improvement over the strongest baseline on TILES-2018; performance on WESAD supports generalizability across datasets.Source: §1 · p. 1

2. METHODOLOGY

BLP combines temporally encoded HRV with domain-specific LLM descriptions of physiological state in the context of personal baselines. It learns semantic representations and fuses them with HRV representations for stress classification.

2. METHODOLOGY

  • BLP processes learned HRV and semantic representations through bidirectional cross-attention, then concatenates them for MLP stress classification.The framework comprises HRV representation learning, LLM-based semantic interpretation, semantic representation learning, and multimodal fusion.Source: §2 · p. 1

2.1. HRV Representation Learning

  • A Transformer encodes temporally ordered HRV feature vectors, with positional encodings, into embeddings that an HRV projector maps for cross-attention.Source: §2.1 · p. 2

2.2. LLM-based Semantic Interpretation

  • The LLM prompt combines a shared personal baseline profile with HRV markers and interpretation context for each of four physiological domains.The generated description situates physiological state relative to individual baseline characteristics without predicting the stress label.Source: §2.2 · p. 2

2.3. Semantic Representation Learning

  • Domain-specific descriptions are embedded and projected, then orthogonally regularized and adaptively combined with learnable weights into a shared semantic representation.The shared representation is 100-dimensional and is used for subsequent cross-modal fusion.Source: §2.3 · p. 2

2.4. Training Objective

  • Training combines intra-subject Multi-Similarity Loss on HRV and semantic representations with an orthogonal loss to reduce redundancy across domain-specific semantic representations.The total objective also includes weighted cross-entropy; auxiliary-loss contributions are controlled by λ_intra and λ_ortho.Source: §2.4 · pp. 2–3

3. EXPERIMENT SETUP

The experiments evaluate personalized stress detection on longitudinal naturalistic and smaller controlled datasets using subject-independent five-fold cross-validation. BLP is compared with three personalized baselines and six ablation models.

3.1. Dataset and Feature Extraction

  • TILES-2018 includes 212 hospital workers followed for 10 weeks, while WESAD includes 15 participants in controlled Baseline, Stress, and Amusement conditions.The datasets share 61 HRV features across four domains, but use different extraction and aggregation procedures.Source: §3.1 · p. 3
  • TILES-2018 includes 17 self-reported attributes and WESAD includes 11 attributes spanning demographics and personal or day-of-collection characteristics.Source: §3.1 · p. 3
  • The evaluation uses subject-independent five-fold cross-validation with UAR, accuracy, and MCC, selecting MCC as the primary metric because of class imbalance.Source: §3.1 · p. 3

3.2. Comparison Models

  • Three baselines represent multi-task learning, graph-based attribute integration, and dynamic hypergraph learning; six ablations remove or alter model branches and components.The ablations test semantic-only and HRV-only inputs, integration of attributes or HRV, the embedding extractor, and LLM interpretation.Source: §3.2 · p. 3

4. RESULTS ANALYSIS AND ABLATION

BLP outperforms the compared baselines on TILES-2018, and case studies illustrate how it connects physiological patterns with personal attributes. Ablations assess the contributions of HRV, personal attributes, and semantic interpretation.

4.1. Performance Comparison

  • 0.630/0.631/0.261 UAR/Acc/MCC are BLP’s highest results, with statistically significant MCC improvements of 51.7%, 33.2%, and 29.9% over three baselines.The comparison uses the same subject-independent setting.Source: §4.1 · p. 3, Table 4 · p. 4
  • Removing HRV or personal attributes from the prompt reduces TILES-2018 MCC by 54.4% or 27.6%, respectively.Removing LLM interpretation reduces MCC by 8.0% on TILES-2018 and 3.1% on WESAD; removing the semantic embedding extractor reduces it by 17.6% and 2.3%, respectively.Source: §4.1 · p. 4, Table 3 · p. 4

4.2. Case Study

  • Case studies show BLP correcting errors from the setting without HRV integration by relating physiological patterns to personal attributes.The examples associate VLF variability with occupational demands in a stress case, and elevated pNN50 and RMSSD with parasympathetic activity and recovery capacity in a non-stress case.Source: Table 5 · p. 4, §4.2 · p. 4

5. CONCLUSION AND FUTURE WORK

The conclusion presents BLP as an LLM-based approach that integrates HRV and personal attributes for personalized stress classification. It reports stronger TILES-2018 performance than compared baselines and identifies reliance on HRV representations as a limitation.

  • BLP integrates HRV and personal attributes through LLM-based semantic interpretation, generating human-readable descriptions fused with temporal HRV representations for stress classification.Source: §5 · p. 4
  • BLP outperforms all compared baselines on TILES-2018, while WESAD performance supports generalizability across datasets.Source: §5 · p. 4
  • LLM-generated semantic representations alone are insufficiently discriminative; future work aims to improve task-relevant information while balancing interpretability and predictive performance.Source: §5 · p. 4

Related papers

Loading 2610.23555v1…