EarStreAM: A Closed-Loop Earable System for Personalized Stress-Adaptive Meditation

It proposes and implements an end-to-end closed-loop earable stress intervention that integrates in-ear physiological sensing, stress detection, LLM-generated meditation, and audio playback within a single earable device plus a smartphone. The authors note that prior earable work has largely addressed stress detection or conceptualized the potential of closed-loop interventions, while existing closed-loop stress interventions have mainly relied on biofeedback or virtual environments; to their knowledge, no earable system has yet combined continuous physiological stress detection with automatic initiation and real-time adaptation of interventions within a single device. This is a system-and-demo contribution; evidence comes from the architecture description (Figure 2), the hardware modalities used (Figure 3), and the demo flow (Figure 4), with no reported participant counts or controlled experimental data.

It builds a signal-processing and context-aware detection pipeline from in-ear PPG to stress determination, including artifact removal, signal-quality gating, individualized baseline estimation, and an activity-classification heuristic. The system uses 200 Hz in-ear PPG and a 100 Hz accelerometer, applying green/red/IR channel selection, ambient-light cancellation, accelerometer-referenced motion artifact suppression, band-pass filtering, and adaptive envelope normalization, with a composite signal-quality score from waveform quality, motion, peak yield, and RR regularity; the baseline fuses demographic priors with real-time data at 30% prior and 70% data-driven, and stress is by default detected when heart rate rises and heart rate variability falls beyond a 10% relative threshold. Method details are given with concrete parameters (e.g., 0.45-5.5 Hz band-pass, RR validation at 250-2000 ms, Kalman smoothing with Q=0.02, R=5.0), but thresholds and heuristics largely cite prior literature and no detection performance metrics are reported here.

It implements LLM-based personalized meditation generation and speech synthesis so that content adapts to both user preferences and current physiological state. Users can configure voice characteristics, thematic elements, and meditation style, and may provide free-form preferences; these inputs, together with current physiological measurements, are integrated into a prompt that generates meditation in approximately 280-word segments, later converted to audio via TTS. The authors describe this as enabling real-time personalization that varies with physiological state and personal preferences within each session. The implementation is explicit (DeepSeek V4 Pro with ElevenLabs Flash v2.5 or Tencent Cloud TTS), but no evaluation of generation quality or user preference is reported.

It offers an interactive demo with two modes and describes its privacy handling. The first mode uses a shortened 30 s baseline, an ultra-sensitive 3% threshold, and roughly 30 s meditation segments, with an optional stress induction inspired by the Maastricht Acute Stress Test (cold water exposure plus a timed mental arithmetic task); the second mode explores personalized meditation only. Physiological processing stays local to the smartphone, and prompts sent to the LLM and TTS services exclude personal and physiological information. Evidence consists of the demo setup and flow, a feasibility demonstration without controlled or statistical results.

The system targets office-like, predominantly sedentary work settings, depends on stable in-ear PPG fit and signal-quality gating, and does not initiate meditation when sustained movement is detected via the activity-classification heuristic; it is intended for adult users seeking timely, personalized stress support during everyday work, with closed-loop logic using heart rate as the recovery indicator and relative changes in heart rate and heart rate variability from an individualized baseline as the stress trigger.

A careful reader may still watch: the robustness of the stress-detection threshold and activity-classification heuristic across individuals and office settings; how using heart rate as the sole recovery indicator and heart rate variability as the trigger performs in real long-duration use; the quality and user acceptance of LLM-generated meditation content; and the practical boundaries of local processing and prompt de-identification for privacy. As a demo paper, it provides no quantitative results on detection accuracy, intervention effects, or user experience, which remain open questions for follow-up research.