Dae Hyeon Kim RESEARCH 한국어

NeuroTruth

Connected wearable craving estimates and rule-based risk decisions to app notifications and conversational support.

GitHub: neurotruth
Role
Technical lead and project manager
Contribution
Designed the system and implemented risk logic, backend services, notifications, speech recognition and speech synthesis.
Period & affiliation
2026.03 ~ Present NeuroTruth

01 / RESEARCH DETAILNeuroTruth: sensing, state checks and dialogue

Challenge
Wearable predictions must trigger state checks and notifications while keeping dialogue separate from risk decisions.
Approach
Backend rules determine risk levels and notification timing. The app supports state confirmation, questionnaires and speech interaction.
Contribution
Served as technical lead and PM; implemented risk logic, backend, notifications, Korean STT and TTS integration.
Prototype display of received PPG and EDA signals
Prototype display of received PPG and EDA signals
View full-size figure
State-check decision using craving level and previous notifications
State-check decision using craving level and previous notifications
View full-size figure
Prototype dialogue view: user speech and system response
Prototype dialogue view: user speech and system response
View full-size figure
Implemented NeuroTruth prototype functions
Current function Implementation scope
Physiological input 4-level craving estimation from PPG and EDA
Dialogue support AWS Bedrock LLM, STT, Android TTS
Records AUQ, state and notification history
Evaluation conditions and scopeFour-level craving-estimation prototype using PPG and EDA. MBID and clinical intervention evaluation are planned.

02 / RESEARCH DETAILNeuroTruth: integration and notification latency

Integration test
Confirmed the connection between classification, risk decisions, notifications and dialogue in 21 of 21 repeated integration tests.
Camera test
A separate camera test processed 10 of 10 static videos. Mean processing time for a 20-second video was 14.8 seconds.
Next steps
Next steps are MBID scoring and intervention evaluation, with clinical assessment of approximately 80 participants at two hospitals planned.
Prototype view of state changes over the previous hour
Prototype view of state changes over the previous hour
View full-size figure
Craving classification, LOSO (%) Classification performance: Balanced acc. 71.0, Macro F1 70.0 0 22 43 65 87 71 Balanced acc. 70 Macro F1 Classification performance
Craving classification, LOSO (%)
Notification delivery latency (ms) Mean: Wi-Fi 78.0, LTE 121.1; P95: Wi-Fi 89.0, LTE 133.0 0 41 81 122 162 78 89 Wi-Fi 121.1 133 LTE Mean P95
Notification delivery latency (ms)
Integration and latency measurement scope
Measurement scope Interpretation
Integration tests Repeated functional tests excluding MBID
Latency Notification delivery; excludes LLM response
Evaluation conditions and scopeNotification latency excludes LLM response time. The camera test measured processing success and duration.