Lightweight edge AI-based home vestibular rehabilitation monitoring method

Abstract

The invention discloses a household vestibular rehabilitation monitoring method based on lightweight edge AI, and belongs to the field of medical artificial intelligence and edge computing. Eye movement and head movement data are collected through intelligent glasses and an IMU bracelet, a neural network model smaller than 100 KB is operated at an equipment end, and head movement deviation, eye closing over-limit and falling risks are recognized in real time; the original data is destroyed immediately, and only the structured event abstract is uploaded. The system adopts a special home gateway repeater to support the old to complete automatic pairing by pressing a physical key for 3 seconds for the first time, and subsequent training does not need a mobile phone or App. Event reporting supports multi-mode dynamic switching of Wi-Fi / 5G / 4G / SMS and Beidou / low orbit satellites, and a satellite channel is started when no ground network exists. Doctors can remotely issue prescriptions or initiate video follow-up visit, and patients can answer the prescriptions by voice. The method is compatible with future 6G communication evolution, realizes a home rehabilitation closed loop of zero interaction, high privacy and full-scene communication, and is suitable for urban and rural areas and remote areas.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical artificial intelligence, edge computing and medical Internet of Things. Specifically, it relates to a home-based vestibular rehabilitation monitoring method based on lightweight edge AI, and more particularly to a non-intrusive rehabilitation behavior recognition and event reporting method for elderly patients that integrates 5G / satellite multi-mode communication and national cryptographic security mechanisms. Background Technology

[0002] Vestibular rehabilitation training has a significant therapeutic effect on improving dizziness and balance disorders in the elderly. Existing remote monitoring solutions mainly rely on smartphone apps (such as CN113456987A) or cloud-based video analytics systems (such as US20210386345A1), which have three major drawbacks:

[0003] High barrier to entry: Users need to frequently operate the mobile phone and enter account passwords, making it difficult for the elderly to use independently;

[0004] Significant privacy risks: Original videos are often uploaded to the cloud, violating GDPR, HIPAA, and China's Personal Information Protection Law;

[0005] Unreliable communication: It only supports Wi-Fi or 4G, and is prone to losing contact in rural areas, islands or emergency scenarios, lacking emergency channels such as satellite.

[0006] While some studies have attempted to deploy lightweight models at the edge (such as EP3566692B1), these have not been linked to clinical prescriptions and have not addressed the issues of trusted device access and multi-generational communication evolution. Currently, there is no solution that integrates "physical one-click pairing + edge AI privacy computing + 5G / satellite full-scenario communication + 6G scalable architecture" into a unified medical closed loop.

[0007] Therefore, there is an urgent need for a new monitoring method that can meet the ease of use needs of elderly users, ensure the security of medical data, and be oriented towards the future evolution of communication for home-based vestibular rehabilitation monitoring. Summary of the Invention

[0008] The purpose of this invention is to provide a home-based vestibular rehabilitation monitoring method based on lightweight edge AI, which realizes a closed-loop system of "zero-interaction pairing, local privacy inference, multi-modal reliable reporting, and dynamic prescription linkage", supports full-scenario coverage from cities to border areas, and is compatible with future 6G communication evolution.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] The method described in this invention is applied to a home-based vestibular rehabilitation system consisting of smart glasses, an IMU wristband, and a dedicated home gateway repeater. The home gateway repeater is a medical-grade device with a built-in main control chip, a 5G / 4G / Wi-Fi multi-mode communication module, a national cryptographic security chip, and a physical pairing button.

[0011] When using it for the first time, elderly patients only need to press and hold the physical pairing button on the home gateway repeater for at least 3 seconds, and the gateway will enter pairing mode and start Bluetooth 5.0 broadcast, which carries a temporary pairing identifier.

[0012] After the smart glasses and IMU wristband detect the broadcast, they automatically send a binding request containing the device type and public key to the gateway. The gateway verifies whether the device belongs to the whitelist pre-registered by the doctor. If the verification is successful, it establishes an encrypted BLE connection with the smart glasses and IMU wristband and stores the device ID in the secure storage area.

[0013] Once pairing is complete, the gateway's LED indicator will turn solid green, eliminating the need for further user intervention during all subsequent training processes and completely eliminating reliance on accounts, passwords, or apps.

[0014] The doctor generates a structured rehabilitation prescription (including parameters such as training type, direction, angular velocity, and duration), and after digitally signing the prescription with SM2, pushes it to the home gateway repeater via the cloud using the MQTT over TLS protocol. The home gateway repeater uses its built-in security chip to verify and decrypt the prescription, and broadcasts it to the smart glasses via Bluetooth 5.0. After receiving the prescription, the smart glasses automatically update their local task parameters without requiring user confirmation or clicking, achieving seamless prescription synchronization.

[0015] During training, the smart glasses collect video of the eye area at a frequency of no less than 2 frames per second, and the IMU wristband collects three-axis acceleration and angular velocity data of the wrist at a sampling rate of no less than 100 Hz.

[0016] The video data and IMU data are synchronously transmitted to the local processing unit of the smart glasses via Bluetooth 5.0, with an end-to-end latency of less than 20 milliseconds. A quantized MobileNetV2-Tiny neural network model (deployed based on the TensorFlow Lite Micro framework) with fewer than 100KB of parameters is run on the local processing unit to identify at least three types of vestibular rehabilitation-specific events in real time:

[0017] The direction or amplitude of head movement deviates from the prescription's permissible range;

[0018] The cumulative time spent with eyes closed within the 5-second sliding window exceeded 40%.

[0019] A trunk tilt angle greater than 60° and an acceleration mutation exceeding 2.5g, estimated by combining wrist height changes, are used as indicators of fall risk.

[0020] When the identification confidence of any event is greater than 0.8, a structured event summary is generated. The event summary includes JSON format and contains event type, duration, UTC timestamp, prescription ID, and de-identified patient hash value.

[0021] The original video frames and IMU raw data streams are immediately overwritten and destroyed in the device memory after inference is completed. They are not stored locally or uploaded to any external devices, thus protecting patient privacy from the source.

[0022] The event summary is sent to the home gateway repeater via Bluetooth 5.0; the home gateway repeater dynamically selects the communication channel based on the current network status.

[0023] Prioritize uploading to the cloud server via Wi-Fi;

[0024] If Wi-Fi is unavailable and the 5G or 4G signal strength is greater than -100 dBm, switch to the 5G / 4G Cat.1 channel;

[0025] If there is no terrestrial cellular network coverage for more than 2 hours, the event summary will be compressed into SMS format (e.g., [VH]EYE: 2.1s; TS: 1734684000) and sent via SMS to the designated SMS number of the medical center through the built-in SIM card using AT commands.

[0026] When an external BeiDou-3 short message module, TianTong-1 terminal, or low-orbit satellite communication equipment is detected, the satellite channel is automatically set as the highest priority reporting path, and the event summary is encoded and reported through the L / S band link.

[0027] Meanwhile, the home gateway repeater encrypts and caches the event digests that failed to upload in the eMMC storage, and automatically resumes the upload once the network is restored.

[0028] The cloud server receives and aggregates event summaries from multiple patients, and displays statistical trends of abnormal events on the doctor's web management platform;

[0029] Doctors can use the platform to push new rehabilitation prescriptions or initiate video follow-up invitations with a single click.

[0030] When a video follow-up invitation arrives, the smart glasses automatically wake up the camera and microphone, and prompt the patient with a voice message to "say 'connect' to start the call." The patient only needs to respond with a voice message to establish an end-to-end encrypted video session.

[0031] Furthermore, the home gateway repeater is compatible with future 6G communication standards, and its communication protocol stack reserves a 6G New Radio (NR) lightweight access layer interface, supporting the loading of 6G NR-Light protocol modules through secure firmware over-the-air (FOTA) upgrades, thus enabling a smooth evolution to 6G.

[0032] In summary, this invention, through the collaborative design of a dedicated home gateway and lightweight edge AI, has for the first time achieved a closed-loop home vestibular rehabilitation monitoring system for elderly patients that features "zero interaction, high privacy, full-scenario communication, and evolvability."

[0033] Beneficial effects:

[0034] Senior-friendly: No need to operate a mobile phone or app throughout the entire process. Only a 3-second long press of the physical button is required for the first time. Subsequent processes are completely hands-free, eliminating the digital divide for the elderly and achieving "zero-interaction" rehabilitation monitoring, which conforms to the "zero-interaction" design principle.

[0035] Strong privacy protection: Raw physiological data does not leave the device; only structured summaries are uploaded.

[0036] Full communication coverage: Supports five-mode redundancy including Wi-Fi, 5G, 4G, SMS, and satellite, suitable for urban and rural areas, border regions, islands, and emergency scenarios;

[0037] A strong clinical closed loop: automation of the entire process from prescription to execution to feedback to intervention;

[0038] Technology is evolvable: Hardware and protocol stacks support future 6G upgrades, extending product lifecycles;

[0039] High compliance: Meets the requirements of YY / T 1708-2020, GDPR, HIPAA and the Personal Information Protection Act. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;

[0041] Figure 2 Flowchart for device pairing and prescription dispensing;

[0042] Figure 3 A logic diagram for edge AI event recognition;

[0043] Figure 4 Flowchart for multi-mode communication switching decision-making for home gateways;

[0044] Figure 5 The timing diagram for automatic wake-up of video follow-up. Detailed Implementation

[0045] To enable those skilled in the art to clearly and completely implement the present invention, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 5 Four specific implementation examples are provided, and the hardware configuration, software process and technical implementation details of each module are described in detail.

[0046] Example 1: System Composition

[0047] This system includes the following three types of terminal devices:

[0048] Smart glasses (100)

[0049] Hardware: OV7670 camera (1080P@30fps), ESP32-S3 AI chip (including Tensilica HiFi 4DSP), BLE 5.0 module;

[0050] Features: Real-time video capture, edge AI inference, Bluetooth communication, audio and video encoding / decoding;

[0051] Software: Deploy a lightweight model (MobileNetV2-Tiny) based on TensorFlow Lite Micro, running on the ESP32-S3 coprocessor (COPROC).

[0052] IMU wristband (200)

[0053] Hardware: BN0055 six-axis sensor (accelerometer + gyroscope), nRF52840 BLE chip;

[0054] Functions: Inertial data acquisition (100Hz sampling rate), data preprocessing (FIR filtering), Bluetooth pass-through;

[0055] Software: Data synchronization and low-power mode management are implemented based on the nRF Connect SDK.

[0056] Home gateway repeater (300)

[0057] hardware:

[0058] Main controller: RK3308B (ARM Cortex-A35, dual-core, 1.8GHz);

[0059] Communication modules: RG500U 5G / 4G dual-mode module (PCIe / USB 3.0 interface), SMA antenna interface (shared for 5G / satellite), Beidou DM22 module (UART expansion port);

[0060] Security chip: CIU9832 (supports national cryptographic standards SM2 / SM3 / SM4);

[0061] Storage: 8GB eMMC (AES-256 encrypted cache).

[0062] Function:

[0063] Network status detection and communication protocol switching;

[0064] Whitelist matching and national cryptographic verification;

[0065] Multi-mode communication (Wi-Fi / MQTT, 5G / HTTPS, 4G / CoAP, satellite / SMS);

[0066] Firmware over-the-air (FOTA) updates and security audit logs.

[0067] Topology:

[0068] The smart glasses (100) and the IMU wristband (200) are directly connected to the home gateway (300) via BLE 5.0, forming a star topology;

[0069] The home gateway (300) communicates with the cloud server via Wi-Fi / 5G / 4G / satellite links.

[0070] Example 2: Implementation of multimode communication (see appendix) Figure 4 )

[0071] The communication protocol stack of the home gateway repeater (300) dynamically selects the communication path through the following steps:

[0072] -Network status detection

[0073] Wi-Fi Priority: Check Wi-Fi connection status every 5 minutes (by querying WPA_STATE=COMPLETED using the wpa_cli command-line tool);

[0074] Cellular network assessment:

[0075] G RSSI (signal strength is obtained via AT+CSQ command, >-100dBm is considered usable);

[0076] G RSSI (same as above);

[0077] Satellite channel detection:

[0078] The L-band link status of the Beidou DM22 module (polled via the UART interface to the BEIDOU_STATUS register).

[0079] - Communication protocol selection logic

[0080] Wi-Fi available: Event digests are uploaded to the cloud using the MQTT over TLS protocol (port 1883);

[0081] G / 4G available:

[0082] In GSA mode, the HTTPS protocol is used (port number 443);

[0083] In G NSA mode, the CoAP protocol is used (port number 5683);

[0084] No terrestrial network:

[0085] If no valid network is detected for two consecutive hours, the SMS channel will be activated.

[0086] The event summary is compressed into SMS format ([VH]EYE: 2.1s; TS: 1734684000; PID: B2E8) and sent to the medical center's SMS number via AT commands;

[0087] If a BeiDou-3 module is detected, it will be uploaded via L-band short message first.

[0088] -Data caching and resume download

[0089] All unsuccessfully uploaded event digests are encrypted and stored in eMMC (AES-256 encryption, key generated by CIU9832 security chip);

[0090] After the network is restored, the gateway automatically reads the cache and resumes transmission according to the original communication protocol.

[0091] Example 3: Complete Usage Process Demonstration (Demonstrating "Zero Interaction")

[0092] Scene: 78-year-old Mr. Zhang uses the system for gaze stability training for the first time.

[0093] -First pairing

[0094] The community nurse instructed Mr. Zhang to press and hold the physical pairing button on his home gateway (300) for ≥3 seconds to trigger BLE broadcast;

[0095] The smart glasses (100) and the IMU wristband (200) automatically send a binding request (including device type, public key, and unique ID);

[0096] The gateway queries the local doctor pre-registration whitelist (stored in the CIU9832 security chip). If a match is found, an SM2 key is generated, the LED turns green, and the pairing is complete.

[0097] -Daily Training

[0098] After Mr. Zhang put on the device, the smart glasses automatically loaded today's prescription (gaze_stabilization, horizontal);

[0099] During training:

[0100] The glasses capture video at 2 frames per second, and the IMU wristband captures inertial data at 100Hz;

[0101] Edge AI models can identify events in real time (such as eyes closed >40%, torso tilt angle >60°);

[0102] The event summary is sent to the gateway via BLE 5.0 and uploaded according to the multi-mode communication strategy.

[0103] - Doctor intervention

[0104] After the doctor platform detects an abnormal event, it pushes a new prescription (SM2 signature);

[0105] Once the gateway verifies the signature, it broadcasts the information to the smart glasses, which then automatically adjust the training parameters.

[0106] During video follow-up, the doctor initiates an invitation, and the glasses' voice prompt says "Say 'connect'," after which the patient responds and an SRTP encrypted call is established.

[0107] Example 4: Communication backup in remote areas

[0108] In a mountain village in Yunnan with no broadband and weak 4G signal:

[0109] The gateway detected that there had been no MQTT connection for two consecutive hours.

[0110] Automatically enable SMS channel, and send event summaries compressed into SMS format;

[0111] After the medical center analyzes the text messages, manual intervention is provided to ensure the continuity of rehabilitation.

[0112] Feasibility description

[0113] Those skilled in the art can implement the method of the present invention through the following means:

[0114] Edge AI processing module: Employs an edge AI chip with an NPU (such as the ESP32-S3 series) to run a neural network model based on a lightweight AI framework (parameters < 100KB) to achieve local event recognition of eye-tracking videos (such as eye closure exceeding limits, head movement deviation).

[0115] Device pairing and communication: Automatic binding and encrypted data transmission between smart glasses, IMU wristbands and home gateways are achieved through the central / peripheral mode of the Bluetooth Low Energy protocol stack (such as the standard implementation in Nordic nRFConnect SDK).

[0116] Privacy protection mechanism: The event digest is encrypted using national cryptographic algorithms (such as SM4), and key derivation and storage security are achieved by combining hardware acceleration security chips (such as BGI CIU9832).

[0117] Multi-mode communication integration: Based on the main control chip that supports Wi-Fi and 5G / 4G dual-mode communication (such as Rockchip RK3308), expand the Beidou short message module or low-orbit satellite terminal to achieve signal coverage in urban and rural areas and remote areas.

[0118] Technological integration and innovation of this invention:

[0119] Privacy and performance balance: Reduce the risk of privacy breaches by destroying raw data (video and IMU streams) locally and uploading only event summaries;

[0120] Communication redundancy design: dynamically switches between Wi-Fi / 5G / 4G / SMS / satellite channels to ensure reliable data upload in extreme environments;

[0121] Age-friendly interaction: Device binding can be completed with a single physical button, requiring no further manual intervention, and is adapted to the operating habits of elderly users.

[0122] All technical solutions are based on existing AI development kits and open-source frameworks (such as lightweight AI frameworks, Bluetooth protocol stacks, and national cryptographic algorithm libraries), but have achieved technological breakthroughs in the field of home vestibular rehabilitation monitoring through specific integration logic (such as multi-mode communication + privacy destruction + zero-interaction pairing).

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. Automatically wakes up the audio and video module and prompts the patient with a voice prompt to "say 'connect' to start the call". The patient only needs to respond with a voice to establish an end-to-end encrypted video session; in, The communication protocol stack of the home gateway repeater reserves a 6G New Radio (NR) lightweight access interface, supporting the loading of the 6G NR-Light protocol module via over-the-air firmware upgrade.

2. The method according to claim 1, characterized in that, The quantized neural network model is MobileNetV2-Tiny deployed on TensorFlowLite Micro, with a model parameter size of 96KB and an inference latency of less than 80 milliseconds.

3. The method according to claim 1, characterized in that, Data synchronization between the smart glasses and the IMU wristband is achieved via Bluetooth 5.0 direct connection, with an end-to-end latency of less than 20 milliseconds.

4. The method according to claim 1, characterized in that, The SMS compression format for the event summary is: [VH] Event Type Abbreviation: Duration; TS: Unix timestamp; PID: patient hash, each SMS message cannot exceed 160 bytes.

5. The method according to claim 1, characterized in that, The cached data in the eMMC memory is encrypted using AES-256, and the key is derived from the national cryptographic security chip, so the main control chip cannot access the plaintext key.

6. The method according to claim 1, characterized in that, The home gateway repeater automatically identifies the 5G SA / NSA or 4G network type by querying the communication module's capabilities, and prioritizes using the 5G standalone networking mode for data transmission.

7. The method according to claim 1, characterized in that, The satellite communication equipment includes a BeiDou-3 regional short message terminal, a TianTong-1 satellite phone module, or a Starlink-type low-orbit broadband satellite modem.

8. The method according to claim 1, characterized in that, The video follow-up sessions are encrypted using the SRTP protocol, and the key is negotiated using the SM4 algorithm to ensure end-to-end security.

9. The method according to claim 1, characterized in that, After pairing is complete, the home gateway repeater automatically rejects connection requests from new devices not on the whitelist to prevent unauthorized devices from accessing the network.

10. The method according to claim 1, characterized in that, The 5G or 4G signal strength threshold of -100dBm can be configured remotely by the doctor, and the configuration parameters are sent after being signed by SM2.

11. The method according to claim 1, characterized in that, The home gateway repeater connects to the 5G / 4G dual-mode communication module via USB 3.0 or PCIe bus, and supports future direct replacement with a 6G communication module.

12. The method according to claim 1, characterized in that, The de-identified patient hash value is generated by performing an SM3 hash operation on the patient ID, and the original identity information cannot be reversed.

Images