Query-Conditioned Graph Retrieval for Contextualized LLM Reasoning in Personalized Wearable Data

Zhenyu Lu; Mahyar Abbasian; Amir M. Rahmani

cs.IRApr 10, 2026

Query-Conditioned Graph Retrieval for Contextualized LLM Reasoning in Personalized Wearable Data

Authors: Zhenyu Lu, Mahyar Abbasian, Amir M. Rahmani

Organizations: Carnegie Mellon University · University of California, Irvine

Abstract

Large language models (LLMs) are increasingly applied to analyzing wearable sensing data, which are long-term, multimodal, and highly personalized. A key challenge is context selection: providing insufficient context limits reasoning, while including all available data leads to inefficiency and degraded generation quality. We propose Wearable As Graph (WAG), a graph-based context retrieval framework that enables query-adaptive reasoning over wearable data with LLMs. WAG organizes wearable metrics and user-specific signals into a personalized knowledge graph, and retrieves a query-conditioned subgraph to support downstream generation. The retrieval process integrates global relationships, capturing prior knowledge and population- and individual-level patterns via hierarchical Bayesian modeling, with local relationships that reflect short-term signal deviations. A query openness signal further controls retrieval breadth. We evaluate WAG on over 10,000 data-grounded queries from real-world wearable datasets. Across LLM-based and human evaluations, WAG achieves an approximately 70% win rate over baseline and standard RAG methods, demonstrating the effectiveness of structured, query-adaptive context retrieval for LLM-driven analysis of wearable data.

Figures & tables

Figure 1: Overview of the WAG framework for query-adaptive context retrieval and reasoning over wearable data.

Figure 2: Visualization of the generated WAG personal knowledge graph.

Dataset	$\mathcal{W}^{global}$	$\mathcal{W}^{ind}$	$\mathcal{W}^{pop}$	$\mathcal{W}^{prior}$
Globem	2.14	2.62	2.50	2.22
IFH Affect	2.08	2.42	2.40	2.39
Lifesnap	2.14	2.41	2.32	2.34
Pmdata	2.28	2.42	2.40	2.32
Average	2.16	2.47	2.40	2.32

Table 3: Exp-G - Comparison of different weighting within global modeling across datasets.

Dataset	$\mathcal{W}^{final}$	$\mathcal{W}^{global}$	$\mathcal{W}^{local}$
Globem	1.90	1.98	2.01
IFH Affect	1.88	1.98	2.02
Lifesnap	1.85	1.94	2.10
Pmdata	1.89	2.01	1.99
Average	1.88	1.98	2.03

Table 4: Exp-L - Evaluation of the effectiveness of local modeling across datasets.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

	General Knowledge	Data Retrieval	Trend Analysis	Comparative Insight	Anomaly Detection	Actionable Advice	Exploratory Analysis	Metric Relationships	Contextual Queries
Main experiment
Globem	104	531	219	196	199	125	387	147	53
IFH Affect	164	749	348	330	305	256	569	141	59
Lifesnap	188	800	335	315	307	254	573	143	57
Pmdata	126	654	287	265	263	188	504	128	70
Total	582	2734	1189	1106	1074	823	2033	559	239

Appendix

Table 8: Query counts per query type

Dimension	Description
Insightfulness	Similar to Interestingness in SSI ( Thoppilan et al., 2022 ) . Captures whether incorporating high-quality, context-aware information leads to more insightful responses.
Relevance	Derived from Specificity and Sensibility in SSI . Assesses whether the system retrieves highly relevant content tailored to the user’s context.
Groundedness	Evaluates whether responses are supported by factual or retrievable content.
Personalization	Measures how accurately responses reflect the user’s specific data and context.
Clarity	Related to Conciseness . Judges whether responses are clear, and accessible, even for complex queries.
Safety & Security	Ensures responses avoid unsafe, harmful, or misleading content.

Appendix

Table 9: Evaluation dimensions of response quality.

Figure 3: Comparison of weights encoded by spearman correlation and mutual information.

Figure 6: Visualization of hierarchical Bayesian modeling (HBM) updates for all nodes related to the metric “Circadian rhythm patterns” .

\boldsymbol{\mu^{\text{prior}}}

denotes the prior,

\boldsymbol{\mu^{\text{pop}}}

denotes the posterior after population data update, and

\boldsymbol{\mu^{s}}(\mathcal{W}^{\text{global}})

denotes the final posterior after incorporating individual-specific data.

Figure 8: Distributions of different edge weight components in the personal knowledge graph. The eight subplots display: (a) prior weights (

\boldsymbol{\mu^{\text{prior}}}

), representing the initial LLM-assigned weight; (b) empirical relationships from the population (

R^{\text{pop}}

), representing correlations observed across all subjects; (c) empirical relationships from the individual (

R^{s}

), representing correlations computed from the subject’s personal data; (d) posterior after population data update (

\boldsymbol{\mu^{\text{pop}}}

); (e) posterior after incorporating individual-specific data (

\boldsymbol{\mu^{s}}(\mathcal{W}^{\text{global}})

); (f) local weights (

\mathcal{W}^{\text{local}}

); (g) final weights (

\mathcal{W}^{\text{final}}

), obtained via Eq. 1 . Mean (red) and median (green) values are marked for reference.

Dataset	$\mathcal{W}^{global}$		$\mathcal{W}^{ind}$		$\mathcal{W}^{pop}$		$\mathcal{W}^{prior}$
	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate
Globem	2.14	0.30	2.62	0.17	2.50	0.22	2.22	0.36
IFH Affect	2.08	0.34	2.42	0.23	2.40	0.25	2.39	0.27
Lifesnap	2.14	0.28	2.41	0.24	2.32	0.30	2.34	0.28
Pmdata	2.28	0.30	2.42	0.23	2.40	0.23	2.32	0.32
Average	2.16	0.31	2.47	0.22	2.40	0.25	2.32	0.31

Appendix

Table 13: Experiment-G - Comparison of different weighting within global modeling across datasets.

Dataset	$\mathcal{W}^{global}$		$\mathcal{W}^{ind}$		$\mathcal{W}^{pop}$		$\mathcal{W}^{prior}$
	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate
1	2.15	0.31	2.49	0.21	2.35	0.28	2.29	0.31
7	2.20	0.29	2.40	0.24	2.25	0.28	2.42	0.28
14	2.08	0.29	2.52	0.21	2.44	0.24	2.29	0.31
30	2.18	0.32	2.45	0.22	2.50	0.21	2.31	0.32
Average	2.16	0.30	2.47	0.22	2.39	0.26	2.33	0.30

Appendix

Table 15: Experiment-G - Comparison of different weighting within global modeling across time periods.

Dataset	$\mathcal{W}^{global}$		$\mathcal{W}^{ind}$		$\mathcal{W}^{pop}$		$\mathcal{W}^{prior}$
	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate
low	2.03	0.38	2.57	0.21	2.54	0.16	2.31	0.31
medium	2.19	0.31	2.47	0.20	2.43	0.25	2.35	0.31
high	2.16	0.30	2.45	0.24	2.36	0.26	2.29	0.31
Average	2.12	0.33	2.50	0.22	2.45	0.23	2.32	0.31

Appendix

Table 16: Experiment-G - Comparison of different weighting within global modeling across abnormal levels.

Dataset	$\mathcal{W}^{final}$		$\mathcal{W}^{global}$		$\mathcal{W}^{local}$
	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate
Globem	1.90	0.36	1.98	0.33	2.01	0.35
IFH Affect	1.88	0.38	1.98	0.33	2.02	0.34
Lifesnap	1.85	0.38	1.94	0.37	2.10	0.29
Pmdata	1.89	0.36	2.01	0.36	1.99	0.33
Average	1.88	0.37	1.98	0.35	2.03	0.33

Appendix

Table 17: Experiment-L - Evaluation of the effectiveness of local modeling across datasets

Query Type	$\mathcal{W}^{final}$		$\mathcal{W}^{global}$		$\mathcal{W}^{local}$
	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate
General Knowledge	1.48	0.63	2.04	0.33	2.30	0.11
Data Retrieval	1.90	0.31	1.75	0.44	2.23	0.25
Trend Analysis	1.94	0.40	1.88	0.34	2.06	0.32
Comparative Insight	1.95	0.34	1.91	0.40	2.09	0.28
Anomaly Detection	1.91	0.34	2.02	0.32	1.93	0.40

Appendix

Table 18: Experiment-L - Evaluation of the effectiveness of local modeling across query types.

Time Period	$\mathcal{W}^{final}$		$\mathcal{W}^{global}$		$\mathcal{W}^{local}$
	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate
1	1.89	0.36	1.99	0.33	1.99	0.36
7	1.91	0.35	1.89	0.40	2.08	0.30
14	1.85	0.36	1.96	0.37	2.11	0.30
30	1.87	0.40	2.02	0.33	2.01	0.32
Average	1.88	0.37	1.96	0.36	2.05	0.32

Appendix

Table 19: Experiment-L - Evaluation of the effectiveness of local modeling across time periods.

Abnormal Level	$\mathcal{W}^{final}$		$\mathcal{W}^{global}$		$\mathcal{W}^{local}$
	Mean	Win Rate	Mean	Win Rate	Mean	Win Rate
low	1.74	0.45	1.92	0.36	2.20	0.22
medium	1.90	0.36	1.95	0.36	2.04	0.32
high	1.89	0.36	2.01	0.33	1.99	0.35
Average	1.84	0.39	1.96	0.35	2.08	0.30

Appendix

Table 20: Experiment-L - Evaluation of the effectiveness of local modeling across abnormal levels.

Table 18

Table 19

Category	Openness	Description	Example Questions
Single-Entity
General Knowledge	0.2-0.4	Definition and basic understanding of metrics, optimal ranges, and benchmarks	What is HRV and its optimal range? What is resting heart rate?
Data Retrieval	0.1-0.3	Specific time-bound numerical queries	What was my step count this week? Average deep sleep minutes past 14 days?
Trend Analysis	0.4-0.6	Trend identification and behavioral patterns	Identify trends in my daily movement last 30 days. When do I typically have my most active days?
Comparative Insight	0.5-0.7	Time-based comparisons between periods	How does this week’s activity compare to last week? Has my sleep duration improved this month?
Anomaly Detection	0.6-0.8	Outlier identification and unusual deviations	Any unusual sleep metrics this month? Was there an abnormal recovery time this week?

Appendix

Table 23: Question Categories for Health Data Analysis

Notation	Description
$D$	a dataset
$s\in\mathcal{S}_{sel}$	participants selected
$m\in\mathcal{M}$	a concept of health metric, eg. heart rate, mood
$t\in\mathcal{T}$	all timestamps (day) of data
$v_{t}$	data value at timestamp(day) $t$
$\text{MD}_{s}$	missing data percentage for a subject $s$ ’s data

Appendix

Table 24: Notations and Descriptions

Field	Description
ID	Id of the node
Name	Name of the node
Description	Description of the node
Range	Range of values with units
Recommendation	Recommendation for improvement
DataSource	Data source specification including dataset name, feature name, description, range, unit, and type, path to data

Appendix

Table 25: Node Field Descriptions

Field	Description
ID	Unique relationship identifier
Node_1_name	Name of first node in relationship
Node_1_description	Description of first node
Node_1_id	Unique identifier of first node
Node_2_name	Name of second node in relationship
Node_2_description	Description of second node

Appendix

Table 26: Edge Field Descriptions

Category	Metrics
Physiological	Heart rate, Heart rate variability, Blood pressure, Blood oxygen saturation, Pulse wave velocity, Cardiac output, Peripheral blood flow, Respiratory rate, Oxygen uptake, Carbon dioxide exhalation rate, Lung volume, Breathing rhythm, Muscle activity, Blood glucose levels, Lactate levels, Basal metabolic rate, Core body temperature, Brain activity, Skin temperature, Sweat rate, Electrolyte concentration, Skin hydration levels, Skin conductance, pH of sweat
Sleep	Sleep stages (light, deep, REM), Sleep apnea events, Total sleep duration, Sleep onset latency, Wake after sleep onset, Sleep efficiency, Arousals per night, Breathing irregularities, Snoring patterns, Circadian rhythm patterns
Activity	Steps taken, Distance traveled, Active minutes, Energy expenditure, Activity level, Running dynamics, Balance and stability, Joint movement and flexibility
Mental	Mental stress, Mental fatigue, Mental workload, Engagement, Cognitive load, Memory performance, Decision-making speed, Mood
Environmental	Ambient temperature, Humidity, Barometric pressure, UV radiation exposure, Air quality, Noise levels, Electromagnetic field exposure, Altitude
Lifestyle	GPS location, Travel patterns, Proximity to other devices or people, Interaction frequency with social contacts, Time spent on specific activities, Screen interaction patterns, Daily routine adherence

Appendix

Table 27: Initial Health Metrics

Type	Metric	Globem	Ifh_affect	Lifesnap	Pmdata
Physiological	Blood oxygen saturation	✗	✗	✓	✗
	Skin conductance	✗	✗	✓	✗
	HRV RMSSD	✗	✓	✓	✗
	Resting heart rate	✗	✓	✓	✗
	Nightly skin temperature	✗	✗	✓	✗
	Sleep breathing rate	✗	✓	✓	✗

Appendix

Table 28: Metrics available across datasets. ✓ indicates presence, ✗ absence.

Graph Retrieval-Augmented Generation

Graph Retrieval-Augmented Generation connects language models to knowledge graphs to retrieve relational evidence for grounded, multi-hop answers.

See all 232 papers

WEQA: Wearable hEalth Question Answering with Query-Adaptive Agentic Reasoning
Jun 16, 2026Yuwei Zhang, Tong Xia, Bianca Emmerich +5Biomedical QA Wearable Sensing
University of Cambridge · Tsinghua University · University College London +2
WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data
Sep 4, 2026Ji Soo Lee, Xilun Chen, Pierce Chuang +5Medical QA Wearable Health Monitoring
Meta · KAIST · Korea University
ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents
Apr 24, 2026Yating Wu, Yuhao Zhang, Sayan Ghosh +4LLM Agent Memory Graph-Based Agent Memory
UT Austin · AWS AI Labs

Jun 16, 2026J/K move · Enter open · S save

WEQA: Wearable hEalth Question Answering with Query-Adaptive Agentic Reasoning

Yuwei Zhang, Tong Xia, Bianca Emmerich, Yu Yvonne Wu+4

University of Cambridge · Tsinghua University · University College London+2

Biomedical QA Wearable Sensing LLM Agents Wearable Health Monitoring

Open paper AskPDF arXiv

Language models are remarkably capable at medical question answering, in some cases surpassing the accuracy of general physicians. However, answering questions about wearable health data remains challenging and understudied, as these ubiquitous sensors produce continuous, high-dimensional, and longitudinal data, which is non-trivial to align with text-centric distributions in LLM pretraining. The diversity of sensor modalities and user intents cannot be effectively handled by a fixed reasoning workflow or a single pretrained foundation model. To address these challenges, we propose WEQA, a query-adaptive agent framework that unifies LLM reasoning with specialized wearable analytical and modeling tools. An LLM controller is employed to synthesize execution plans and dynamically route each query to the appropriate combination of sensor analysis and pretrained models, and perform grounded response auditing with external knowledge. We also curate a benchmark spanning four open wearable datasets comprising analytic and predictive tasks in three different health domains. Experiments show that our framework is 24% more accurate than LLM and agentic baselines, and a blinded study with 12 medical experts and 8 users shows substantial gains in usefulness and clinical soundness.

Abstract

Graph Retrieval-Augmented Generation

Explore similar work

WEQA: Wearable hEalth Question Answering with Query-Adaptive Agentic Reasoning

WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data

ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents