Decentralized Tele-Rehabilitation via Edge AI-Oracle Architecture for Spatiotemporal Pain Assessment

Area of Science:

• Computer Science
• Artificial Intelligence
• Medical Informatics

Background:

• Remote tele-rehabilitation necessitates objective pain assessment, a challenge unmet by current self-report or cloud-based computer vision methods due to falsification risks or privacy violations.
• Existing methods like Visual Analog Scale (VAS) and Numeric Pain Rating Scale (NPRS) are susceptible to falsification, while cloud-based systems compromise patient privacy by transmitting raw biometric data.

Purpose of the Study:

• To propose a decentralized Edge AI-Oracle architecture for privacy-preserving, objective pain assessment in tele-rehabilitation.
• To develop a system that maps facial Action-Unit sequences to pain scores without transmitting raw video data, ensuring compliance with privacy regulations like GDPR and HIPAA.

Main Methods:

• Utilized MediaPipe Face Mesh for landmark extraction and a Long Short-Term Memory (LSTM) recurrent neural network for classifying Action-Unit sequences into pain scores.
• Implemented an Edge AI-Oracle architecture with inference and Elliptic Curve Digital Signature Algorithm (ECDSA) signing within a Trusted Execution Environment (TEE).
• Employed off-chain storage on the InterPlanetary File System (IPFS) and on-chain anchoring via smart contracts with an optimistic verification window for an off-chain auditor.

Main Results:

• The LSTM model achieved an F1 score of 0.683 on the SynPAIN dataset and 0.519 on the BioVid Heat Pain Database, demonstrating robust pain recognition.
• Inference latency in the TEE was significantly below the per-frame budget (0.45 ms for INT8 at T=120), and dual-layer storage reduced gas costs by over 23x.
• The system demonstrated detection of gross score tampering, though it bounds rather than eliminates subtle patient-side fraud, necessitating further research into zero-knowledge proofs.

Conclusions:

• The proposed Edge AI-Oracle framework offers a cryptographically verifiable, privacy-preserving, and economically viable solution for tele-rehabilitation pain assessment.
• The architecture aligns with GDPR and HIPAA requirements through its Zero-Video Transmission principle and Layer-2 economic feasibility.
• Further improvements in recognition accuracy and robustness to tampering are needed for clinical deployment, with zero-knowledge proofs suggested for enhanced fraud detection.