PALM: Scaling Physiologically-Aware AI Tutoring Through Consumer Wearables and Large Language Models

Abstract

AI tutors usually respond without regard to the learner's physiological state, yet learning depends heavily on it. PALM (Physiologically-Aware Language Model) is a three-stage automated pipeline: ingest wearable health data, compute personal baselines, and inject abstracted state summaries into the AI tutor's system prompt — enabling “invisible care” without requiring students to disclose their condition.

Problem & Motivation

AI tutoring systems typically respond to students without considering their current physiological state, yet learning is highly dependent on physiological conditions — insufficient sleep, elevated stress, or depleted energy reserves all impair cognition. How can an AI tutor automatically sense learner physiological state and adjust response strategy without requiring the student to disclose it?

Method

PALM (Physiologically-Aware Language Model) is a three-stage automated framework: (1) health-data abstraction — ingest sleep, stress, and activity data from Garmin wearables, Apple HealthKit, and Google Health Connect; (2) personal baseline computation — establish a personalized baseline with 7-day and 30-day rolling windows, detecting deviations via Z-score into normal, elevated, and significantly elevated states; (3) context construction — automatically inject the abstracted physiological state into the AI tutor's system prompt, modulating tone, response length, information density, and scaffolding strategy. The process is transparent to students, realizing “invisible care”.

Findings

• Scalable: the three-stage pipeline is fully automated, requires no clinical-grade hardware, and deploys on any learner with a consumer wearable.
• Proof-of-concept: end-to-end pipeline tests show the AI tutor appropriately adjusts response style (simplified language, increased encouragement, lower information density) when poor sleep or elevated stress is detected.
• Privacy-preserving: raw physiological data is not transmitted to the LLM — only abstracted state summaries — protecting learner privacy.

Implications

PALM demonstrates a viable path for combining physiological sensing with AI-in-education: AI tutors can adapt interaction style based on wearable data without forcing students to disclose their state. This “invisible care” mode lets every student receive support calibrated to their current physical and mental state, particularly valuable in large-scale online learning where instructors cannot perceive learner state in real time.

Citation

@inproceedings{chang2026palm,
  author    = {Chia-Kai Chang and Kuei-Hao Li},
  title     = {{PALM}: Scaling Physiologically-Aware {AI} Tutoring Through Consumer Wearables and Large Language Models},
  booktitle = {Proc. ACM Conf. on Learning at Scale (L@S '26)},
  year      = {2026},
  month     = jul,
  doi       = {10.1145/3774398.3811576},
  note      = {Work-in-Progress},
}