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

🕒 Published (v1): 2026-06-16 16:45 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

WEQA is a query-adaptive agentic framework for answering natural-language questions about wearable sensor data (activity, sleep, stress, cardiovascular, respiratory). An LLM controller dynamically routes each query through sensor-native analytical tools or specialized predictive models, then audits the response against sensor evidence and external medical knowledge. It outperforms LLM-only and prior agentic baselines by at least 24% across a new four-family benchmark and rates higher in blinded expert evaluation.

Problem

Existing approaches to wearable health QA either serialize sensor data as text (losing fine-grained temporal/morphological structure) or apply fixed reasoning pipelines and single foundation models. Neither handles the heterogeneity of real-world health queries — which range from lightweight statistical lookups to longitudinal trend analysis to complex predictive inference from raw physiological signals — nor do they align continuous, high-dimensional sensor data with text-centric LLM pretraining distributions.

Method

WEQA decomposes every query into three stages:

1. Query-Aware Planning (\(a = \pi(q, m, e, S)\)): A prompted LLM controller reads the user query \(q\), sensor metadata \(m\), a retrieved few-shot history bank \(e\), and the available toolset \(S\) to produce a structured execution plan. The plan identifies query objective, required modalities, temporal scope, reasoning type, risk level, and selected tools/models. After each execution step \(t\), the controller may replan based on accumulated evidence \(s_t = \{q, a_t, Z_t\}\).

2. Evidence Construction (\(Z = \Phi(X, H, q, a)\)): Two complementary pathways execute the plan:

3. Sensor Analytical Reasoning (\(z_{\text{ana}} = \phi_{\text{ana}}(X, q, a)\)): runs code and invokes analytical tools (stats calculator, temporal profiler, trend fitter, correlation checker) for descriptive/longitudinal questions.

4. Adaptive Predictive Inference (\(z_{\text{pred}} = \phi_{\text{pred}}(X, H, q, a)\)): selects specialized models (audio FM OPERA-CT/GT, PPG-based hypertension classifier, stress detection heads) conditioned on query type; performs uncertainty-aware orchestration (querying fallback models on low-confidence outputs); applies few-shot personalization when labeled user history is available, or history-aware baseline normalization when only unlabeled history exists.

5. Grounded Response Auditing (\(y = g(q, z_{\text{ana}}, z_{\text{pred}})\)): Verifies claims against sensor evidence, calibrates uncertainty communication, retrieves external medical knowledge for clinical context, and enforces risk-sensitive language.

Key Contributions

• Identification of two core challenges: sensor-to-text misalignment and query heterogeneity requiring adaptive computational pathways.
• WEQA framework: training-free, query-adaptive agent with a reusable wearable-health toolset spanning statistical, temporal, and predictive modalities.
• A benchmark of 1,123 QA pairs across 358 users, 6 sensing modalities, 4 datasets (TILES, UK COVID-19 Sounds, PPG-BP), and 4 task families (Short-Horizon Analytical, Long-Horizon Analytical, Predictive Reasoning, Open-Ended Insight).
• Automated + human evaluation showing WEQA scores 3.9/5 vs. 3.1 (ReAct) and 2.9 (Multi-Agent) from 12 medical experts, with largest gains in personalization and usefulness.

Results

• Short-Horizon Analytical QA: WEQA 95.6 exact match / 9.2 MAE vs. ReAct 64.8 / 31.3 and Multi-Agent 72.0 / 157.2.
• Long-Horizon Analytical QA: WEQA 94.0% numerical accuracy, 95.1% trend & correlation accuracy (vs. ReAct 80.9% / 56.6%).
• Predictive Reasoning QA: WEQA 83.9% UAR (classification) / 10.9 MAE (regression) vs. ReAct 59.2% / 15.4.
• Token efficiency: ~10k tokens/query vs. 31k–42k for agentic baselines; lowest end-to-end latency among all methods.
• Human evaluation: WEQA avg. 3.9/5 (experts), highest across correctness/grounding, personalization, usefulness, and clinical plausibility; user ratings preserve the same ordering.
• Overall improvement: ≥24% over LLM-only and existing agentic baselines across all task families.

Limitations

• Benchmark is synthetically constructed from existing datasets; real-world free-form user queries may differ in distribution and ambiguity.
• Relies on Gemini-3.0-Flash as default backbone; cross-LLM robustness tested only with Qwen3-Max.
• Predictive models in the toolset were trained on specific datasets/modalities; generalization to unseen device types or novel health conditions is not evaluated.
• Open-Ended Insight QA results are evaluated primarily through human ratings (small panel of 20 raters), not a scalable automated metric.
• Personalization requires either labeled or unlabeled personal history; cold-start performance for new users is not characterized.

Relevance to Harnesses / Meta-Harnesses

WEQA is a concrete instance of a query-adaptive meta-harness: rather than committing to a fixed workflow at design time, the LLM controller dynamically composes a computation graph — selecting analytical tools, predictive models, and auditing steps — conditioned on the incoming query's intent and sensor context at inference time. This is directly analogous to meta-harness design patterns where a coordinator layer selects and sequences specialized sub-agents or tools rather than hard-coding a pipeline. The replanning loop (revising \(a_t\) after each execution step based on accumulated evidence) and the grounded response auditing phase — which acts as a post-execution verifier — are reusable architectural primitives relevant to any harness that must handle heterogeneous task types safely. The efficiency gains (3–4× fewer tokens than fixed agentic baselines) also demonstrate the practical value of adaptive routing over monolithic pipelines.