Explainable Prediction from Mobile Sensing Data through LLM-guided Concept Integration
Introduces Concept-Integrated Transformer (CIT), a novel architecture combining transformer-based prediction with LLM-guided concept supervision for mobile sensing data Uses pretrained LLMs to generate baseline-aware concept abnormality targets with confidence weights, eliminating the need for manual concept annotation Achieves state-of-the-art F1 scores on two longitudinal health datasets: 0.756 on AFFECT and 0.765 (tied) on PHQ-9 Learned concept scores yield interpretable behavioral patterns,
Analysis
TL;DR
- Introduces Concept-Integrated Transformer (CIT), a novel architecture combining transformer-based prediction with LLM-guided concept supervision for mobile sensing data
- Uses pretrained LLMs to generate baseline-aware concept abnormality targets with confidence weights, eliminating the need for manual concept annotation
- Achieves state-of-the-art F1 scores on two longitudinal health datasets: 0.756 on AFFECT and 0.765 (tied) on PHQ-9
- Learned concept scores yield interpretable behavioral patterns, with sleep quantity and quality showing clearest differentiation between high and low negative affect groups
- Addresses the dual challenge of accuracy and interpretability in small-cohort health-sensing studies where outcome supervision is limited relative to heterogeneous data
Why It Matters
This work bridges a critical gap in digital health AI by demonstrating that LLMs can serve as effective concept generators for interpretable machine learning without costly manual annotation. For practitioners building health monitoring systems, it provides a scalable pathway to models that are both accurate and clinically meaningful, which is essential for real-world deployment where stakeholder trust depends on understanding model reasoning.
Technical Details
- Concept-Integrated Transformer (CIT): A transformer-based architecture that integrates concept-level supervision into prediction tasks, where concepts are derived from mobile sensing features (e.g., sleep, activity, physiological signals)
- LLM-Guided Concept Supervision: A pretrained LLM generates concept abnormality targets with confidence weights, conditioned on individual baselines, enabling zero-shot concept extraction without manual labeling
- Datasets: Evaluated on two longitudinal mobile sensing datasets — AFFECT (negative affect prediction) and PHQ-9 (depression severity assessment)
- Performance: Highest F1 of 0.756 on AFFECT and tied highest F1 of 0.765 on PHQ-9, demonstrating strong generalization across different mental health prediction tasks
- Interpretability: Concept scores reveal meaningful patterns; sleep metrics emerged as the most discriminative features between high and low negative affect groups, aligning with established clinical knowledge
Industry Insight
- LLMs can serve as powerful priors for concept discovery in domains with limited labeled data, reducing annotation costs while improving model interpretability — a strategy applicable beyond health sensing to any domain with heterogeneous sensor or time-series data
- The baseline-aware concept generation approach addresses a key limitation in small-cohort studies, suggesting that personalization at the concept level (rather than only at the model level) could become a standard practice in digital health
- As regulatory bodies increasingly demand explainability in health AI, methods like CIT that natively produce interpretable concept scores may accelerate adoption of mobile sensing-based diagnostics in clinical workflows
Disclaimer: The above content is generated by AI and is for reference only.