Smart bracelet for monitoring neurological diseases

Abstract

The present application relates to the technical field of intelligent wearable medical equipment, in particular to an intelligent bracelet for monitoring neurological diseases, which comprises a wristband body and an electronic monitoring system, the electronic monitoring system comprising a multi-modal physiological signal acquisition module, a microprocessor control unit, a disease characteristic analysis module, an early warning feedback module, a wireless communication module and a power management module. By collecting multi-dimensional physiological signals, Parkinson's disease targeting features, epilepsy targeting features and double disease auxiliary features are extracted, then through the basic differentiated contribution weight of the convolution long short-term memory network model and the dynamic output distribution weight adjustment strategy triggered by the pathological threshold, the spatio-temporal fusion analysis is realized and the risk probability of the two diseases and the normal state probability are output, the early warning feedback module outputs differentiated prompts according to the risk probability, the wireless communication module synchronizes data or sends emergency alerts through dual-mode communication, and the power management module realizes power supply and low power reminder.

Description

Technical Field

[0001] This invention relates to the field of smart wearable medical devices, specifically to a smart bracelet for monitoring neurological diseases. Background Technology

[0002] Parkinson's disease and epilepsy are common chronic diseases in clinical practice. Parkinson's disease is characterized by resting tremor and bradykinesia, while epilepsy is characterized by myoclonus and loss of consciousness. Both diseases require long-term real-time monitoring to achieve early intervention and reduce the harm caused by disease attacks.

[0003] In existing technologies, wearable monitoring devices for neurological diseases are mostly designed for single-disease monitoring. For example, epilepsy monitoring bracelets only collect EEG, heart rate, and activity level signals, while Parkinson's disease monitoring bracelets only collect acceleration signals of hand tremors. This is insufficient to meet the clinical needs of monitoring patients with comorbidities or multiple symptoms. Furthermore, existing devices have limited signal acquisition dimensions and simple feature extraction methods, relying heavily on traditional machine learning algorithms or manual baseline comparisons for disease identification. This results in low detection accuracy and a high risk of false alarms and missed alarms. In addition, data processing in existing monitoring devices largely depends on cloud computing or server resources. The bracelet only completes signal acquisition; the data must be uploaded for analysis. For instance, patent application CN120000174A discloses an intelligent early warning bracelet for epilepsy monitoring. It collects EEG and other signals and uses a Hidden Markov Model for seizure pattern recognition on the server. This solution primarily targets a single epilepsy disease and relies on cloud processing, leading to data transmission delays and network dependency issues.

[0004] Therefore, there is an urgent need for a convenient wearable device that can simultaneously monitor Parkinson's disease and epilepsy in real time, and improve detection accuracy and early warning response speed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a smart bracelet for monitoring neurological diseases, enabling real-time, 24 / 7 monitoring and early warning of both Parkinson's disease and epilepsy. This solves the problems of existing technologies, such as single-disease monitoring, limited feature dimensions, reliance on cloud computing for response delays, high communication power consumption, and insufficient accuracy in early warning.

[0006] The basic solution provided by this invention is a smart bracelet for monitoring neurological diseases, comprising a wristband body and an electronic monitoring system embedded in the wristband body, wherein the electronic monitoring system includes: A multimodal physiological signal acquisition module is used to acquire multimodal physiological signals, including acceleration signals acquired by an accelerometer, angular velocity signals acquired by a gyroscope, electromyographic signals acquired by an electromyographic sensor, skin conductance signals acquired by a skin conductance response sensor, and photoelectric pulse wave signals acquired by a photoelectric pulse wave sensor. The microprocessor control unit is electrically connected to the multimodal physiological signal acquisition module and is used to control the synchronous sampling timing and analog-to-digital conversion of each sensor; The disease feature analysis module is electrically connected to the microprocessor control unit and includes a pathological targeted feature extraction algorithm unit and an embedded neural network processor. The pathological targeting feature extraction algorithm unit is used to extract multi-dimensional features from the multimodal physiological signals. These multi-dimensional features include Parkinson's disease targeting features, epilepsy targeting features, and dual-disease auxiliary features. The Parkinson's disease targeting features include a tremor index composed of 4-12Hz frequency band energy extracted from the acceleration and angular velocity signals after wavelet packet decomposition. The epilepsy targeting features include an electromyographic activity index composed of the root mean square value and zero-crossing rate extracted from electromyographic signals, and an autonomic nerve activity index composed of the average skin conductivity level and skin conductivity change rate extracted from skin conductivity signals. The dual-disease auxiliary feature is the ratio of low-frequency to high-frequency heart rate variability (LF / HF) extracted from photoelectric pulse wave signals. The embedded neural network processor is loaded with a pre-trained convolutional long short-term memory network model. This convolutional long short-term memory network model is an end-to-end multi-classification model. It performs spatiotemporal fusion analysis by assigning preset basic differential contribution weights to Parkinson's disease targeting features, epilepsy targeting features, and auxiliary features of both diseases. When the tremor index is higher than a preset first feature threshold, the model increases the output allocation weight of the Parkinson's disease risk probability by a first preset ratio, while simultaneously decreasing the output allocation weights of the epilepsy risk probability and the normal state probability by a proportional ratio. When the electromyography activity index or the autonomic nerve activity index is higher than a preset second feature threshold, the model increases the output allocation weight of the epilepsy risk probability by a second preset ratio, while simultaneously decreasing the output allocation weights of the Parkinson's disease risk probability and the normal state probability by a proportional ratio. Finally, it outputs the risk probabilities of Parkinson's disease and epilepsy, and the normal state probability, respectively. The early warning feedback module, which is electrically connected to the disease feature analysis module, includes graded LED indicator lights and a vibration motor, and is used to output differentiated visual, auditory and tactile early warning prompts according to the respective risk probabilities of Parkinson's disease and epilepsy. The wireless communication module is electrically connected to the disease feature analysis module and the early warning feedback module, and supports Bluetooth Low Energy protocol and NB-IoT cellular Internet of Things communication protocol. It is used to selectively synchronize monitoring data to the user terminal or send emergency alarm information to preset contacts according to the trigger status of the early warning feedback module.

[0007] The principle of this invention lies in: A multi-modal physiological signal acquisition module accurately acquires multi-dimensional physiological signals related to Parkinson's disease and epilepsy, covering the specific physiological characteristics of both diseases and providing a suitable data source for subsequent pathological targeted feature extraction; a microprocessor control unit uniformly schedules the synchronous sampling timing of each sensor and completes analog-to-digital conversion, ensuring the temporal consistency of multi-source data and laying the foundation for subsequent feature fusion and analysis; a pathological targeted feature extraction algorithm unit extracts specific targeted features and dual-disease auxiliary features for the core pathological features of Parkinson's disease and epilepsy, achieving disease differentiation at the feature layer and avoiding cross-interference of different disease features; and a pre-trained convolutional long short-term memory network model loaded by the embedded neural network processor locally performs multi-dimensional... The system performs spatiotemporal fusion analysis on pathological target features with preset basic differentiated contribution weights. It does not rely on cloud computing power and can dynamically adjust the output allocation weight of disease risk probability according to the feature thresholds set by clinical pathology. The basic contribution weight ensures the priority of the core features of the disease, while the dynamic adjustment of the output allocation weight ensures that when the specific pathological features of a certain disease are significant, the output proportion of the risk probability of that disease is strengthened, accurately outputting the risk probability and normal state probability of Parkinson's disease and epilepsy respectively. The early warning feedback module outputs differentiated visual, auditory and tactile early warning prompts according to the risk probabilities of the two diseases. The wireless communication module flexibly switches the communication mode according to the early warning trigger status, realizing the closed loop of daily synchronization of monitoring data of the two diseases and emergency alarms, ensuring the real-time monitoring and the effectiveness of early warning.

[0008] The beneficial effects of this invention are as follows: This invention achieves simultaneous monitoring of Parkinson's disease and epilepsy through multimodal signal acquisition and multidimensional feature extraction targeting pathological conditions. It designs exclusive targeted features for each disease, ensuring that feature extraction highly matches the core pathological characteristics of the diseases, thereby improving the specificity and accuracy of monitoring from the source. The convolutional long short-term memory network model employs a spatiotemporal fusion analysis method that enhances the recognition accuracy of each disease by using preset basic differentiated contribution weights and dynamic output allocation weights triggered by pathological thresholds, thus reducing false positives and false negatives. The pre-trained lightweight model is deployed on an edge-embedded neural network... Through a network processor, local real-time analysis is achieved, avoiding the data transmission latency and network dependence risks brought by cloud computing. It can respond quickly to situations such as acute epileptic seizures and worsening tremors in Parkinson's disease. The differentiated graded early warning mechanism designed for dual diseases can output exclusive early warning prompts according to their respective risk probabilities, avoiding the ambiguity of information in general early warnings. The dual-mode communication protocol balances the low power consumption requirements of daily monitoring with the high reliability requirements of emergency alarms, meeting the dual needs of clinical comorbidity monitoring and daily health management. The overall device is a wearable design, with a compact structure, convenient operation, and is suitable for long-term, all-weather monitoring.

[0009] Furthermore, the microprocessor control unit is configured to synchronously trigger the accelerometer and gyroscope at a sampling frequency of 64Hz to 128Hz to acquire acceleration and angular velocity signals for assessing resting tremor and bradykinesia in Parkinson's disease.

[0010] This sampling frequency is highly compatible with the 4-12Hz core pathological characteristic frequency band of resting tremor in Parkinson's disease, and can accurately capture subtle changes in the acceleration and angular velocity of hand movements, ensuring that the tremor index can truly reflect the abnormal movement characteristics of Parkinson's disease.

[0011] Furthermore, the electromyography sensor adopts a dry electrode structure with an electrode spacing of 20 mm. The microprocessor control unit is configured to trigger the electromyography sensor at a sampling frequency of 500 Hz to 1000 Hz to collect electromyographic signals of the forearm muscle group used to identify myoclonic features during epileptic seizures.

[0012] The dry electrode structure eliminates the need for conductive gel, simplifying the wearing process and improving long-term comfort while avoiding skin irritation. The 20mm electrode spacing and 500-1000Hz sampling frequency precisely match the electromyographic signal characteristics of myoclonus during epileptic seizures, effectively capturing the firing frequency and amplitude changes of motor unit potentials in the forearm muscles. This provides reliable data support for epilepsy targeting characteristics, improving the accuracy of epileptic myoclonus seizure identification and the timeliness of early warning.

[0013] Furthermore, the skin conductance sensor consists of a pair of stainless steel electrodes, with an excitation current of 10... A; The photoelectric pulse wave sensor includes a green LED light source and a photodetector, and the microprocessor control unit is configured to trigger the photoelectric pulse wave sensor at a sampling frequency of 100Hz to acquire photoelectric pulse wave signals for extracting heart rate variability parameters.

[0014] Stainless steel electrodes possess good biocompatibility and conductivity, and can fix 10 An excitation current ensures the safety and stability of skin conductance signal acquisition; the green LED light source of the photoelectric pulse wave sensor is compatible with the 100Hz sampling frequency, enabling stable acquisition of pulse wave signals during daily activities, providing a high-quality data source for LF / HF ratio extraction, ensuring the accuracy of heart rate variability analysis, and thus improving the auxiliary judgment capability of dual-disease monitoring.

[0015] Furthermore, the feature extraction process of the pathological targeted feature extraction algorithm unit is as follows: after performing a 5Hz low-pass filter on the acceleration and angular velocity signals, the signals are decomposed into 3 layers of wavelet packets to extract the energy of the 4-12Hz frequency band and normalize it using the L2 norm to obtain the tremor index; the electromyography signal is divided into 200ms frames with 50% inter-frame overlap processing, and the root mean square value and zero-crossing rate of each frame are calculated to obtain the electromyography activity index; the skin conductance signal is low-pass filtered at 3Hz, and the average conductance and mean rate of change within 10 seconds are calculated to obtain the autonomic nerve activity index; after filtering and correcting the photoelectric pulse wave signal, the RR interval sequence is extracted, and the power spectral density ratio of the 0.04-0.15Hz low-frequency band to the 0.15-0.4Hz high-frequency band is calculated using fast Fourier transform to obtain the LF / HF.

[0016] The signal noise is effectively removed through filtering, decomposition, and normalization, ensuring the effectiveness and uniqueness of the features. At the same time, the extraction processes for Parkinson's disease targeted features, epilepsy targeted features, and dual-disease auxiliary features are independent of each other, avoiding cross-interference between features of different diseases.

[0017] Furthermore, the convolutional long short-term memory network model is trained and deployed in the following manner: Multimodal physiological signal samples are collected from clinically diagnosed Parkinson's disease patients, epilepsy patients, and healthy control groups, and each sample is labeled with its corresponding disease state label; the initial convolutional long short-term memory network model is trained offline on a cloud server using a time-series backpropagation algorithm, assigning preset basic differential contribution weights to Parkinson's disease targeting features, epilepsy targeting features, and dual-disease auxiliary features to obtain a trained network model; the trained network model is weighted and compressed to generate a lightweight model suitable for deployment on embedded devices; the lightweight model is burned into the non-volatile memory of the embedded neural network processor to achieve real-time dual-disease analysis on the edge.

[0018] Furthermore, the structure of the convolutional long short-term memory network model, from input to output, includes: a one-dimensional convolutional layer for extracting local temporal features of each pathological target feature sequence; a pooling layer using max pooling to reduce feature dimensionality while retaining key disease-targeting features; a long short-term memory layer containing 128 hidden units with a dropout rate of 0.3 for learning the temporal dynamic dependencies of multi-dimensional feature sequences; and a fully connected layer that outputs the probabilities of normal state, Parkinson's disease risk state, and epilepsy risk state through a Softmax activation function, with a total probability sum of 1.

[0019] One-dimensional convolutional layers can accurately extract local temporal correlations of target features for Parkinson's disease and epilepsy, as well as auxiliary feature sequences for both diseases, adapting to the temporal characteristics of physiological signals; max pooling layers can effectively reduce feature dimensionality and end-side computation while accurately preserving key information of target features for each disease, avoiding feature loss; LSTM layers, with 128 hidden units and a dropout rate of 0.3, can efficiently learn the long-term dynamic dependencies of multi-dimensional pathological target feature sequences and capture the trend of physiological signal changes before and after disease onset, while effectively avoiding model overfitting; fully connected layers achieve accurate classification and probability output of three states through the Softmax activation function.

[0020] Furthermore, the early warning feedback module is configured with tiered early warning logic and presets a first risk threshold of 70% and a second risk threshold of 85%: when the probability of Parkinson's disease or epilepsy is greater than the first risk threshold but less than the second risk threshold, it is determined to be low risk for the corresponding disease, triggering a first-level early warning, the tiered LED indicator light is constantly yellow, and the corresponding disease is marked with a low-risk label; when the probability of Parkinson's disease or epilepsy is greater than or equal to the second risk threshold, it is determined to be high risk for the corresponding disease, triggering a second-level early warning, the tiered LED indicator light flashes red, the vibration motor vibrates continuously at a frequency of 2Hz for 3 seconds, and the corresponding disease is marked with a high-risk label.

[0021] This tiered early warning mechanism sets up risk assessment and early warning output logics for Parkinson's disease and epilepsy respectively, and clearly distinguishes between the two diseases through exclusive disease identifiers. The solid yellow low-risk warning conveys the potential risk of the corresponding disease to users while avoiding excessive interference with their daily activities. The flashing red high-risk warning, accompanied by vibration, uses multi-sensory stimulation to ensure that users or those around them can quickly perceive the warning information and identify the disease. The 70% and 85% thresholds were determined through statistical analysis of clinical samples, balancing the sensitivity and specificity of the warning. The standardization of the warning logic improves the intuitiveness of risk transmission, making it easier for different groups to quickly understand and take intervention measures for the corresponding diseases, thus taking into account both the effectiveness of the warning and the user experience.

[0022] Furthermore, the wireless communication module is configured with dual-mode communication logic: in daily monitoring mode, at 10-minute intervals, it synchronizes the pathological target feature data and risk probability of the two diseases to the user terminal device through the Bluetooth Low Energy protocol; when the second-level warning is triggered, it immediately sends an emergency alarm message containing timestamps, geographical coordinates and warning diseases to at least one preset emergency contact through the NB-IoT cellular Internet of Things communication protocol.

[0023] The dual-mode communication logic employs differentiated data transmission strategies for daily data synchronization and emergency alarms in dual-disease monitoring. The Bluetooth Low Energy protocol's 10-minute interval synchronization mode effectively reduces device power consumption and extends battery life while achieving complete synchronization of pathological target feature data and risk probabilities for Parkinson's disease and epilepsy. The NB-IoT protocol boasts advantages such as wide coverage, low power consumption, and strong anti-interference capabilities, ensuring that high-risk alarm information can be delivered quickly and stably even in complex environments. Furthermore, the alarm information includes the warning disease, timestamp, and geographical coordinates, providing crucial decision-making basis for medical staff and families to conduct emergency interventions for the corresponding diseases, thereby improving the timeliness and accuracy of clinical rescue.

[0024] Furthermore, the electronic monitoring system also includes a power management module, which is electrically connected to the microprocessor control unit. It uses a 200mAh rechargeable lithium polymer battery to power each module of the electronic monitoring system, supports magnetic wireless charging, and has a battery life of no less than 72 hours in continuous monitoring mode. The power management module is equipped with a low battery detection unit. When the battery level is lower than a preset threshold of 15%, the user is prompted to charge the battery by flashing blue LED indicators.

[0025] The combination of a 200mAh lithium polymer battery and magnetic wireless charging ensures more than 72 hours of continuous monitoring while simplifying charging operations and improving ease of use. The low battery detection and blue flashing indicator mechanism prevent monitoring interruptions due to battery depletion, ensuring the continuous operation of the device around the clock, reducing the probability of missed detections due to power outages, and improving the reliability and stability of the device. Attached Figure Description

[0026] Figure 1 This is a block diagram of the electronic monitoring system of an embodiment of the smart bracelet used to monitor neurological diseases according to the present invention. Detailed Implementation

[0027] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: A smart bracelet for monitoring neurological diseases includes a wristband body and an electronic monitoring system embedded in the wristband body, the electronic monitoring system comprising: A multimodal physiological signal acquisition module is used to acquire multimodal physiological signals, including acceleration signals acquired by an accelerometer, angular velocity signals acquired by a gyroscope, electromyographic signals acquired by an electromyographic sensor, skin conductance signals acquired by a skin conductance response sensor, and photoelectric pulse wave signals acquired by a photoelectric pulse wave sensor. The microprocessor control unit is electrically connected to the multimodal physiological signal acquisition module and is used to control the synchronous sampling timing and analog-to-digital conversion of each sensor; The disease feature analysis module is electrically connected to the microprocessor control unit and includes a pathological targeted feature extraction algorithm unit and an embedded neural network processor. The pathological targeting feature extraction algorithm unit is used to extract multi-dimensional features from the multimodal physiological signals. These multi-dimensional features include Parkinson's disease targeting features, epilepsy targeting features, and dual-disease auxiliary features. The Parkinson's disease targeting features include a tremor index composed of 4-12Hz frequency band energy extracted from the acceleration and angular velocity signals after wavelet packet decomposition. The epilepsy targeting features include an electromyographic activity index composed of the root mean square value and zero-crossing rate extracted from electromyographic signals, and an autonomic nerve activity index composed of the average skin conductivity level and skin conductivity change rate extracted from skin conductivity signals. The dual-disease auxiliary feature is the ratio of low-frequency to high-frequency heart rate variability (LF / HF) extracted from photoelectric pulse wave signals. The embedded neural network processor is loaded with a pre-trained convolutional long short-term memory network model. This convolutional long short-term memory network model is an end-to-end multi-classification model. It performs spatiotemporal fusion analysis by assigning preset basic differential contribution weights to Parkinson's disease targeting features, epilepsy targeting features, and auxiliary features of both diseases. When the tremor index is higher than a preset first feature threshold, the model increases the output allocation weight of the Parkinson's disease risk probability by a first preset ratio, while simultaneously decreasing the output allocation weights of the epilepsy risk probability and the normal state probability by a proportional ratio. When the electromyography activity index or the autonomic nerve activity index is higher than a preset second feature threshold, the model increases the output allocation weight of the epilepsy risk probability by a second preset ratio, while simultaneously decreasing the output allocation weights of the Parkinson's disease risk probability and the normal state probability by a proportional ratio. Finally, it outputs the risk probabilities of Parkinson's disease and epilepsy, and the normal state probability, respectively. The early warning feedback module, which is electrically connected to the disease feature analysis module, includes graded LED indicator lights and a vibration motor, and is used to output differentiated visual, auditory and tactile early warning prompts according to the respective risk probabilities of Parkinson's disease and epilepsy. The wireless communication module is electrically connected to the disease feature analysis module and the early warning feedback module, and supports Bluetooth Low Energy protocol and NB-IoT cellular Internet of Things communication protocol. It is used to selectively synchronize monitoring data to the user terminal or send emergency alarm information to preset contacts according to the trigger status of the early warning feedback module.

[0028] The electronic monitoring system also includes a power management module, which is electrically connected to the microprocessor control unit. It uses a 200mAh rechargeable lithium polymer battery to power each module of the electronic monitoring system, supports magnetic wireless charging, and has a battery life of no less than 72 hours in continuous monitoring mode. The power management module is equipped with a low battery detection unit. When the battery level is lower than a preset threshold of 15%, the user is prompted to charge the battery by flashing blue LED indicators.

[0029] In this embodiment, the wristband body is made of medical-grade soft silicone material, and its length can be adjusted within the range of 15cm-22cm to fit users of different ages and wrist sizes. The core functional modules of the electronic monitoring system are integrated into the dial area of ​​the wristband body.

[0030] In the multimodal physiological signal acquisition module, each sensor establishes a stable connection with the microprocessor control unit via an I2C communication bus, ensuring the synchronization and reliability of data transmission: The accelerometer uses a high-precision triaxial MEMS accelerometer, model ADXL34, to acquire acceleration signals and accurately capture minute hand tremors; the gyroscope uses a six-axis integrated gyroscope, model MPU6050, to acquire angular velocity signals; the electromyography (EMG) sensor adopts a medical-grade dry electrode structure, possessing good biocompatibility and conductivity, with a fixed electrode spacing of 20mm, eliminating the need for additional conductive gel application, and is used to acquire EMG signals from the forearm muscles, adapting to the requirements of epileptic myoclonus feature recognition; the skin conductance sensor consists of a pair of symmetrically arranged stainless steel electrodes, with a constant excitation current of 10... The operating voltage range is 3.3V-5V. It reflects the excitation level of the autonomic nerve by measuring the change in electrical conductance between two points on the skin. The photoelectric pulse wave sensor integrates a green LED light source (wavelength 525nm) and a high-sensitivity photodetector. The model is MAX30102. It has anti-motion interference capability and can stably collect pulse wave signals under the user's daily activity state.

[0031] The microprocessor control unit uses a low-power ARM Cortex-M4 core microcontroller, model STM32L476RG. This processor integrates a high-speed ADC module (12-bit resolution, sampling rate up to 1Msps), multi-channel I2C interface, and UART interface to meet the needs of multi-sensor data acquisition and multi-module communication. Sampling timing control: A precise trigger signal is generated through the processor's built-in timer to control the sampling frequency of the accelerometer and gyroscope, which is adjustable within the range of 64Hz-128Hz (default configuration is 100Hz), and the sampling frequency of the electromyography sensor, which is adjustable within the range of 500Hz-1000Hz (default configuration is 100Hz). The sampling frequency of the photoelectric pulse wave sensor is fixed at 100Hz (assuming a Hz configuration of 800Hz), and the sampling frequency of the skin conductance response sensor is fixed at 10Hz. The sampling trigger signals of all sensors are generated synchronously to ensure the consistency of timestamps for multi-source data. Analog-to-digital conversion and data encapsulation: The analog signals output by each sensor are received and converted into 16-bit digital signals through the built-in ADC module. The signals are encapsulated into data frames in the format of "sensor type-timestamp-data value". The data frames are checked using CRC32 to ensure the integrity of data transmission. The data frames are then transmitted to the disease feature analysis module in real time.

[0032] The disease feature analysis module includes a pathological targeted feature extraction algorithm unit and an embedded neural network processor. The two achieve high-speed data interaction through an internal bus. After the pathological targeted feature extraction algorithm unit completes feature extraction, it transmits the pathological targeted feature sequence composed of tremor index, electromyographic activity index, autonomic nerve activity index, and LF / HF ratio to the embedded neural network processor in real time, where it performs real-time edge-side computation by a convolutional long short-term memory network model. The configuration parameters in this embodiment are as follows: Pathological feature thresholds: The preset first feature threshold is 0.6, and the preset second feature thresholds are 0.7 for electromyography activity index and 0.65 for autonomic nerve activity index; Basic differential contribution weights: Parkinson's disease targeted features account for 0.45%, epilepsy targeted features account for 0.4%, and dual-disease auxiliary features account for 0.15%. Dynamic adjustment of output allocation weights: When the tremor index is higher than 0.6, the model adjusts the output allocation weights for the Parkinson's disease risk probability at the output layer according to a preset ratio of "increasing the baseline value by 20%". The adjustment logic is that the other two items are "proportionally reduced". For example, if the initial output probabilities calculated by the model based on the base weights are: Parkinson's disease 0.45, epilepsy 0.1, normal state 0.45, the adjusted Parkinson's disease risk probability = 0.45 × (1 + 20%) = 0.54, and the value to be reduced is 0.54 - 0.45 = 0.09. Therefore, the adjusted epilepsy risk probability = 0.1 - 0.09 × (0.1 / (0.1 + 0.4)). 5))≈0.08, normal state probability=0.45-0.09×(0.45 / (0.1+0.45))≈0.38, the sum of the three is 1; when the electromyography activity index is higher than 0.7 or the autonomic nerve activity index is higher than 0.65, the model will adjust the weight of the epilepsy risk probability allocation in the output layer according to the preset ratio of "increase the baseline value by 37.5%"; if the pathological characteristic threshold conditions of the two types of diseases are met at the same time, the model will prioritize retaining the dynamic weight adjustment of the more significant triggering disease based on the significance of the trigger threshold (such as the tremor index exceeding the standard by a greater margin than the electromyography activity index exceeding the standard), to ensure the clinical rationality of the risk assessment.

[0033] The feature extraction algorithm unit extracts features as follows: Tremor Index Extraction: Acceleration and angular velocity signals are preprocessed (5Hz low-pass filtering to remove high-frequency noise), and 3-level wavelet packet decomposition is used (db4 is selected as the wavelet basis function). After decomposition, wavelet coefficients of the 4-12Hz frequency band of the two types of signals are extracted, and the energy values ​​of the two types of signals in this frequency band are calculated. The energy values ​​are normalized using the L2 norm (the normalized value range is 0-1). The frequency band energy of the acceleration signal and the frequency band energy of the angular velocity signal are used as the two dimensions of the tremor index, which together constitute the tremor index. Electromyography (EMG) activity index extraction: The EMG signal is processed in frames with a frame length of 200ms and an inter-frame overlap rate of 50%. The root mean square (RMS) value and zero-crossing rate (ZCR) are calculated for each frame. The calculation results are then standardized (the standardized values ​​range from 0 to 1). The standardized RMS value and ZCR are used as the two dimensions of the EMG activity index, which together constitute the EMG activity index. Extraction of autonomic nervous activity indicators: The skin conductance signal was subjected to a 3Hz low-pass filter, and the average skin conductance level within a unit time (10 seconds) and the average rate of change of conductance (absolute value) between adjacent sampling points were calculated. Both were standardized (the standardized values ​​range from 0 to 1) and used as two dimensions of the autonomic nervous activity indicators, which together constitute the autonomic nervous activity indicators. LF / HF ratio extraction: The photoelectric pulse wave signal is subjected to a 5Hz low-pass filter and baseline drift correction. The time interval (RR interval) sequence between two adjacent pulse wave peaks is extracted. The RR interval sequence is subjected to Fast Fourier Transform (FFT) to calculate the power spectral density in the low-frequency band (0.04-0.15Hz) and the high-frequency band (0.15-0.4Hz). The ratio of the two is LF / HF. The embedded neural network processor uses a high-performance edge computing chip, the NXP i.MXRT1170. This processor has a built-in neural network acceleration unit, supports 8-bit / 16-bit weight quantization operations, and has a processing speed of up to 1 TOPS, meeting the real-time analysis requirements of the edge. The training and deployment process of the loaded convolutional long short-term memory network model is as follows: Sample Collection and Labeling: 800 clinical multimodal physiological signal samples were collected, including 250 clinically diagnosed Parkinson's disease patients (aged 45-75 years), 250 epilepsy patients (aged 20-60 years), and 300 healthy controls (aged 20-75 years). Each sample contained 30 seconds of continuous synchronous multimodal physiological signals. Two senior neurologists jointly labeled the disease status (normal state / Parkinson's disease risk state / epilepsy risk state) and pathological feature threshold labels (tremor index 0.6, electromyographic activity index 0.7, and autonomic nerve activity index 0.65, determined based on clinicopathological features). Offline model training: The training environment was set up on a cloud server (configured with NVIDIA A100 GPU and 128GB of memory). The initial convolutional long short-term memory network model was built using the PyTorch framework. The training process used the temporal backpropagation algorithm, with a learning rate of 0.001, 100 training epochs, and a batch size of 32. The cross-entropy loss function was used to optimize the model parameters. The ratio of training set to validation set was 8:2. During training, a preset basic differential contribution weight of 0.45:0.4:0.15 was assigned to Parkinson's disease targeting features, epilepsy targeting features, and dual-disease auxiliary features. Lightweight model optimization: The trained model is quantized with 8-bit weights, and redundant convolutional kernels and fully connected layer neurons are removed by a structured pruning algorithm with a pruning rate of 40%. This compresses the model size while effectively reducing the loss of model accuracy, generating a lightweight model suitable for deployment on embedded devices. Model burning and deployment: Convert the lightweight model into a binary format supported by the processor and burn it into the non-volatile memory (model W25Q64, storage capacity 64MB) of the embedded neural network processor via the JTAG interface; The specific structure and operational logic of the convolutional long short-term memory network model: The hierarchical structure of the model from input to output is as follows: One-dimensional convolutional layer: convolutional kernel size is 3. 1. The convolutional kernels consist of 32 kernels, with ReLU activation function, used to extract local temporal correlation features of multi-dimensional feature sequences. The output feature map dimension is [batch_size, sequence_length, 32]. Pooling layer: Max pooling is used with a pooling window size of 2 and a stride of 2 to reduce feature dimension and computational load while retaining key feature information. The output feature map dimension is [batch_size, sequence_length / 2, 32]. Long Short-Term Memory (LSTM) layer: Contains 128 hidden units with a dropout rate of 0.3. It is used to learn the long-term temporal dependencies of feature sequences and capture the physiological signal change trends before and after disease onset. The output dimension is [batch_size, 128]. Fully connected layer: Contains 3 output nodes, with Softmax activation function. It maps the output of the LSTM layer to the initial probabilities of three disease states. After passing through the "pathological threshold-triggered output weight dynamic adjustment logic", the final normalized probabilities (normal state probability, Parkinson's disease risk probability, epilepsy risk probability) are output.

[0034] The early warning feedback module includes a three-color graded LED indicator (red, yellow, and blue) and a miniature vibration motor. It is connected to the disease feature analysis module through a GPIO interface to realize the visualization and tactile prompts of the risk level. The graded early warning logic parameter settings are as follows: the first preset threshold is 70%, and the second preset threshold is 85%. These thresholds are determined through statistical analysis of clinical samples to ensure a balance between the sensitivity and specificity of the early warning. Low-risk state triggering conditions: The probability of Parkinson's disease risk or the probability of an epileptic seizure warning is greater than 70% and less than 85%. At this time, the yellow LED indicator light is constantly on, and the dial displays the corresponding disease label "Low-risk Parkinson's" or "Low-risk Epilepsy", indicating that the user currently has a potential risk of the corresponding disease and needs to pay close attention to their physical condition and avoid strenuous activity. High-risk state triggering conditions: The probability of Parkinson's disease risk or the probability of an epileptic seizure warning is greater than or equal to 85%. At this time, the red LED indicator light flashes at a frequency of 1Hz, and the vibration motor vibrates continuously at a frequency of 2Hz for 3 seconds. The dial displays the corresponding disease label "High-risk Parkinson's" or "High-risk Epilepsy", ensuring that the user or those around them can quickly perceive the warning information and identify the warning disease. Low battery warning: When the power management module detects that the battery level is lower than 15%, the blue LED indicator light flashes at a frequency of 2Hz, prompting the user to charge in time to avoid monitoring interruption due to battery depletion.

[0035] The wireless communication module integrates a Bluetooth Low Energy chip and an NB-IoT cellular IoT chip. It connects to the disease feature analysis module and the early warning feedback module via a UART interface to achieve data synchronization and emergency alarm functions. The Bluetooth Low Energy chip is model CC2541, which supports the BLE4.2 protocol, has a communication distance of up to 10 meters, and a transmission rate of 1Mbps. In daily monitoring mode, it synchronizes monitoring data to the user terminal device (mobile APP) every 10 minutes. The synchronized data includes: mean tremor index, mean electromyography activity index, mean autonomic nerve activity index, LF / HF ratio, and risk probability output by the model. The NB-IoT cellular IoT chip is the Quectel BC95, which supports the NB-IoT networks of the three major telecom operators: China Telecom, China Unicom, and China Mobile. It features wide communication coverage and low power consumption. When a high-risk warning is triggered, NB-IoT communication is immediately activated to send an emergency alarm SMS to 1-3 preset emergency contacts (configurable via a mobile app, such as medical staff or family members). The SMS content format is: "[Disease Warning] Warning Type: Parkinson's Disease Risk / Epilepsy Risk Warning; Trigger Time: YYYY-MM-DD, HH:MM:SS; Current Location: XXX; Please take timely intervention measures." The geographical location information is provided by the built-in GPS / BeiDou module of the wristband. Communication priority setting: Emergency alarm communication has a higher priority than daily data synchronization. When the two conflict, emergency alarm information will be sent first to ensure timely information transmission in high-risk situations.

[0036] In this embodiment, the power management module uses a high-performance power management chip (model TPS61200); it employs a 200mAh high-energy-density lithium polymer battery with a nominal voltage of 3.7V, supports magnetic wireless charging, has a charging interface compliant with the Qi standard, and a charging power of 5W; in continuous monitoring mode, the battery life after a single full charge is no less than 72 hours, meeting all-weather monitoring requirements; low battery detection and protection: the battery voltage is monitored in real time through a voltage sampling circuit. When the voltage is less than or equal to 3.6V (corresponding to 15% battery capacity), a low battery warning is triggered; when the voltage is less than or equal to 3.3V (corresponding to 5% battery capacity), it automatically enters a low-power sleep mode, retaining only the low battery detection and charging wake-up functions to prevent battery over-discharge damage.

[0037] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A smart bracelet for monitoring neurological diseases, comprising a wristband body and an electronic monitoring system embedded in the wristband body, characterized in that, The electronic monitoring system includes: A multimodal physiological signal acquisition module is used to acquire multimodal physiological signals, including acceleration signals acquired by an accelerometer, angular velocity signals acquired by a gyroscope, electromyographic signals acquired by an electromyographic sensor, skin conductance signals acquired by a skin conductance response sensor, and photoelectric pulse wave signals acquired by a photoelectric pulse wave sensor. The microprocessor control unit is electrically connected to the multimodal physiological signal acquisition module and is used to control the synchronous sampling timing and analog-to-digital conversion of each sensor; The disease feature analysis module is electrically connected to the microprocessor control unit and includes a pathological targeted feature extraction algorithm unit and an embedded neural network processor. The pathological targeting feature extraction algorithm unit is used to extract multi-dimensional features from the multimodal physiological signals. These multi-dimensional features include Parkinson's disease targeting features, epilepsy targeting features, and dual-disease auxiliary features. The Parkinson's disease targeting features include a tremor index composed of 4-12Hz frequency band energy extracted from the acceleration and angular velocity signals after wavelet packet decomposition. The epilepsy targeting features include an electromyographic activity index composed of the root mean square value and zero-crossing rate extracted from electromyographic signals, and an autonomic nerve activity index composed of the average skin conductivity level and skin conductivity change rate extracted from skin conductivity signals. The dual-disease auxiliary feature is the ratio of low-frequency to high-frequency heart rate variability (LF / HF) extracted from photoelectric pulse wave signals. The embedded neural network processor is loaded with a pre-trained convolutional long short-term memory network model. This convolutional long short-term memory network model is an end-to-end multi-classification model. It performs spatiotemporal fusion analysis by assigning preset basic differential contribution weights to Parkinson's disease targeting features, epilepsy targeting features, and auxiliary features of both diseases. When the tremor index is higher than a preset first feature threshold, the model increases the output allocation weight of the Parkinson's disease risk probability by a first preset ratio, while simultaneously decreasing the output allocation weights of the epilepsy risk probability and the normal state probability by a proportional ratio. When the electromyography activity index or the autonomic nerve activity index is higher than a preset second feature threshold, the model increases the output allocation weight of the epilepsy risk probability by a second preset ratio, while simultaneously decreasing the output allocation weights of the Parkinson's disease risk probability and the normal state probability by a proportional ratio. Finally, it outputs the risk probabilities of Parkinson's disease and epilepsy, and the normal state probability, respectively. The early warning feedback module, which is electrically connected to the disease feature analysis module, includes graded LED indicator lights and a vibration motor, and is used to output differentiated visual, auditory and tactile early warning prompts according to the respective risk probabilities of Parkinson's disease and epilepsy. The wireless communication module is electrically connected to the disease feature analysis module and the early warning feedback module, and supports Bluetooth Low Energy protocol and NB-IoT cellular Internet of Things communication protocol. It is used to selectively synchronize monitoring data to the user terminal or send emergency alarm information to preset contacts according to the trigger status of the early warning feedback module.

2. The smart bracelet for monitoring neurological diseases according to claim 1, characterized in that, The microprocessor control unit is configured to synchronously trigger the accelerometer and gyroscope at a sampling frequency of 64Hz to 128Hz to acquire acceleration and angular velocity signals for assessing resting tremor and bradykinesia in Parkinson's disease.

3. The smart bracelet for monitoring neurological diseases according to claim 1, characterized in that, The electromyography (EMG) sensor adopts a dry electrode structure with an electrode spacing of 20 mm. The microprocessor control unit is configured to trigger the EMG sensor at a sampling frequency of 500 Hz to 1000 Hz to collect EMG signals of the forearm muscle groups used to identify myoclonic features during epileptic seizures.

4. The smart bracelet for monitoring neurological diseases according to claim 1, characterized in that, The skin conductance sensor consists of a pair of stainless steel electrodes, with an excitation current of 10. A; The photoelectric pulse wave sensor includes a green LED light source and a photodetector, and the microprocessor control unit is configured to trigger the photoelectric pulse wave sensor at a sampling frequency of 100Hz to acquire photoelectric pulse wave signals for extracting heart rate variability parameters.

5. The smart bracelet for monitoring neurological diseases according to claim 1, characterized in that, The feature extraction process of the pathological targeted feature extraction algorithm unit is as follows: after performing a 5Hz low-pass filter on the acceleration and angular velocity signals, the signals are decomposed into 3 layers of wavelet packets to extract the energy of the 4-12Hz frequency band and normalize it with the L2 norm to obtain the tremor index; the electromyography signal is divided into 200ms frames with 50% inter-frame overlap, and the root mean square value and zero-crossing rate of each frame are calculated to obtain the electromyography activity index; the skin conductance signal is low-pass filtered at 3Hz, and the average conductance and mean rate of change of the conductance within 10 seconds are calculated to obtain the autonomic nerve activity index; after filtering and correcting the photoelectric pulse wave signal, the RR interval sequence is extracted, and the power spectral density ratio of the 0.04-0.15Hz low-frequency band to the 0.15-0.4Hz high-frequency band is calculated by fast Fourier transform to obtain the LF / HF.

6. The smart bracelet for monitoring neurological diseases according to claim 1, characterized in that, The convolutional long short-term memory network model is trained and deployed as follows: Multimodal physiological signal samples are collected from clinically diagnosed Parkinson's disease patients, epilepsy patients, and healthy controls, and each sample is labeled with its corresponding disease state. Using a time-series backpropagation algorithm, the initial convolutional long short-term memory network model is trained offline on a cloud server. Pre-defined basic differential contribution weights are assigned to Parkinson's disease targeting features, epilepsy targeting features, and dual-disease auxiliary features to obtain the trained network model. The trained network model is then weighted and compressed to generate a lightweight model suitable for deployment on embedded devices. The lightweight model is burned into the non-volatile memory of the embedded neural network processor to achieve real-time dual-disease analysis on the edge.

7. The smart bracelet for monitoring neurological diseases according to claim 6, characterized in that, The structure of the convolutional long short-term memory network model, from input to output, includes: a one-dimensional convolutional layer for extracting local temporal features of each pathological target feature sequence; a pooling layer using max pooling to reduce feature dimensionality while retaining key disease-targeting features; a long short-term memory layer containing 128 hidden units with a dropout rate of 0.3, used to learn the temporal dynamic dependencies of multi-dimensional feature sequences; and a fully connected layer that outputs the probabilities of normal state, Parkinson's disease risk state, and epilepsy risk state through a Softmax activation function, with a total probability sum of 1.

8. The smart bracelet for monitoring neurological diseases according to claim 1, characterized in that, The early warning feedback module is configured with a hierarchical early warning logic and presets a first risk threshold of 70% and a second risk threshold of 85%. When the probability of Parkinson's disease or epilepsy is greater than the first risk threshold and less than the second risk threshold, it is determined to be low risk for the corresponding disease, triggering the first-level early warning. The hierarchical LED indicator light is constantly yellow and marked with the low-risk label for the corresponding disease. When the probability of Parkinson's disease or epilepsy is greater than or equal to the second risk threshold, it is judged as a high risk for the corresponding disease, triggering a second-level warning. The graded LED indicator flashes red, the vibration motor vibrates continuously at a frequency of 2Hz for 3 seconds, and the high-risk mark for the corresponding disease is marked.

9. The smart bracelet for monitoring neurological diseases according to claim 8, characterized in that, The wireless communication module is configured with dual-mode communication logic: in daily monitoring mode, it synchronizes the pathological target feature data and risk probability of the two diseases to the user terminal device through the Bluetooth Low Energy protocol at 10-minute intervals; when the second-level warning is triggered, it immediately sends an emergency alarm message containing timestamps, geographical coordinates and warning diseases to at least one preset emergency contact through the NB-IoT cellular Internet of Things communication protocol.

10. The smart bracelet for monitoring neurological diseases according to any one of claims 1 to 9, characterized in that, The electronic monitoring system also includes a power management module, which is electrically connected to the microprocessor control unit. It uses a 200mAh rechargeable lithium polymer battery to power each module of the electronic monitoring system, supports magnetic wireless charging, and has a battery life of no less than 72 hours in continuous monitoring mode. The power management module is equipped with a low battery detection unit. When the battery level is lower than a preset threshold of 15%, the user is prompted to charge the battery by flashing blue LED indicators.

Images