Parkinson's disease early biomarker recognition system and device based on non-invasive electroencephalogram

Abstract

The invention relates to the technical field of intelligent diagnosis systems for neurodegenerative diseases, and discloses a system and equipment for identifying an early biomarker of Parkinson's disease based on non-invasive electroencephalogram. The Parkinson's disease early-stage biomarker identification system comprises a multi-mode signal acquisition module which is provided with a 64-lead dry electrode electroencephalogram cap, an integrated preamplifier and an ADC converter and is used for synchronously acquiring resting-state electroencephalogram and event-related potential, dynamically completing switching of the resting-state electroencephalogram sampling rate and transmitting acquired original data to a preprocessing module. According to the method, revolutionary improvement is realized in three dimensions of acquisition, analysis and calculation through system-level collaborative innovation, a nanocrystalline shielding dry electrode is combined with a dynamic impedance adjustment technology, so that the signal-to-noise ratio of a home environment is increased to 28dB, and the pathological correlation of key markers such as frontal lobe-basal node loop gamma entropy is verified by Granger causality and has a good application prospect. The cloud distributed model server supports 2000 paths of concurrent analysis, and the real-time bottleneck of home screening is thoroughly solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent diagnostic systems for neurodegenerative diseases, and in particular to a non-invasive electroencephalogram (EEG)-based early biomarker identification system and equipment for Parkinson's disease. Background Art

[0002] The current status of early Parkinson's screening is as follows: clinical-grade EEG equipment relies on wet electrodes and electromagnetically shielded rooms, failing to meet the needs of universal home screening. Traditional portable EEG devices, due to insufficient motion artifact suppression, have a detection rate of less than 40% for microvolt-level pathological oscillations (such as gamma-band abnormalities). Existing analysis software often uses single-dimensional features, such as frequency domain power alone, ignoring the fusion of time-frequency nonlinear features and lacking sensitivity to compensatory changes in prodromal neurological function. In addition, at the model level, the model's deep learning lacks decision-making explainability, and doctors are unable to verify the pathological correlation between biomarkers and substantia nigra degeneration, resulting in a clinical adoption rate of less than 30%. Moreover, the system's closed architecture cannot be continuously iterated, and new pathological discoveries require redevelopment of software, delaying the window for technology upgrades.

[0003] Regarding the defects of early Parkinson's screening: dry electrode impedance fluctuations lead to signal distortion, conventional filtering algorithms have difficulty in distinguishing real neural activity from contact noise, and motion artifact compensation requires millimeter-level displacement accuracy; secondly, biomarker stability verification requires cross-time data comparison, but the traditional system lacks a re-test reliability management module. When clinical data is integrated, the scale assessment and EEG acquisition time are asynchronous, resulting in the failure of cross-modal correlation. Finally, the transmission of raw EEG data requires a bandwidth of >2Mbps, which cannot be supported by the home network environment. Local encryption calculations are limited by the computing power of edge devices, making it difficult to meet real-time diagnosis needs. In addition, there is a natural conflict between the medical data privacy protection requirements and the model online update mechanism. Summary of the Invention

[0004] The purpose of the present invention is to provide an early biomarker identification system and equipment for Parkinson's disease based on non-invasive EEG. Through system-level collaborative innovation, revolutionary improvements are achieved in the three dimensions of acquisition, analysis and calculation. The nanocrystalline shielded dry electrode is combined with dynamic impedance adjustment technology to increase the signal-to-noise ratio in the home environment to 28dB. At the same time, the six-degree-of-freedom motion compensation module compresses the electrode displacement error to ±0.3mm, significantly improving the detection rate of weak beta oscillations in the prodromal period at 12-30Hz, aiming to solve the problems in the existing technology.

[0005] The present invention is implemented as follows: a non-invasive EEG-based early biomarker identification system for Parkinson's disease comprises: The multimodal signal acquisition module is equipped with a 64-lead dry electrode EEG cap, an integrated preamplifier and an ADC converter, and is used to synchronously acquire resting-state EEG and event-related potentials. The resting-state EEG sampling rate is dynamically switched, and the collected raw data is transmitted to the preprocessing module. The preprocessing module is used to eliminate baseline drift of the collected raw data and separate the blink component. It implements spectrum notching and acceleration data-driven Kalman filter correction by constructing a 50 / 60 Hz harmonic dictionary, eliminates drift of the collected raw data signal, and outputs a 2-second EEG segment that has been re-referenced in sections. A multidimensional feature extraction module is used to calculate the coefficient of variation of latency, peak-to-peak fluctuation, and Hjorth mobility parameter in segmented, re-referenced 2-second EEG segments. The Welch method is used to generate the power spectral density matrix and inter-band coupling coefficients for the δ / θ / α / β / γ bands. Multiscale entropy analysis and fractal dimension are used to calculate the weighted phase lag index to construct a whole-brain functional connectivity network. The clustering coefficient of the frontal lobe-basal ganglia circuit is extracted from the whole-brain functional connectivity network. The biomarker optimization module is used to perform intraclass correlation coefficient testing on the extracted frontal lobe-basal ganglia circuit clustering coefficients. It triggers wavelet packet reconstruction and re-extraction of features that do not meet the criteria in the circuit clustering coefficients. Once all features meet the criteria, the maximum correlation minimum redundancy algorithm is used to retain features significantly associated with the UPDRS score. Regularized logistic regression is used to generate a 20-dimensional core biomarker encoding vector. An interpretable classification module is used to perform early Parkinson's disease identification based on the 20-dimensional core biomarker encoding vector generated by regression. After identification, a corresponding identification report is output, and when the classification confidence of the newly identified additional sample is less than 90%, the semi-supervised learning process is automatically triggered for deep learning. The clinical data fusion module is used to receive corresponding identification reports, analyze the evaluation results through the UPDRS-III scale, synchronize the clinical scale timestamp with the EEG event marker to ensure consistency in the spatiotemporal dimensions, and synchronize the baseline biomarkers to output the disease progression curve within a preset time in the future.

[0006] Furthermore, the multimodal signal acquisition module has a built-in three-axis accelerometer and an infrared eye tracker for real-time monitoring of head micro-movements and eyelid movements. The sampling rate is dynamically switched between 500-2000 Hz to adapt to different noise environments. After the adaptation is completed, the raw data is generated. The raw data is encrypted by Bluetooth 5.0 and then transmitted to the preprocessing module.

[0007] Furthermore, the multimodal signal acquisition module is further provided with an environmental adaptability unit, and the environmental adaptability unit includes: The electromagnetic shielding layer uses nanocrystalline alloy materials to suppress interference in the frequency band above 50Hz, with a shielding effectiveness of >60dB; Dynamic impedance adjustment circuit, which reduces the electrode-skin contact impedance in real time through micro-current injection, with the fluctuation range controlled within ±0.5kΩ; Artifact self-check unit, which performs a 3-second impedance scan and myoelectric noise baseline test before starting acquisition, and automatically switches unqualified leads to redundant electrodes; The dual-mode storage unit uses a local SD card to cache 72 hours of raw data and encrypted cloud storage for key event markers.

[0008] Furthermore, the preprocessing module includes: Adaptive bandpass filter unit, used to eliminate 0.5-70 Hz baseline drift in raw data; The CNN electrooculogram artifact recognition unit uses a three-layer convolution kernel structure to separate the blink component from the original data; Power frequency noise elimination unit, used to build a 50 / 60 Hz harmonic dictionary to achieve spectrum notching; The motion artifact compensation unit drives the Kalman filter through acceleration data to correct signal drift.

[0009] Furthermore, the motion artifact compensation unit includes: A six-degree-of-freedom motion modeler builds a head rotation / translation motion model based on accelerometer data and predicts the spatial displacement of electrodes; Kalman filter group, with independent configuration of filtering parameters for high-sensitivity forehead / temporal leads, and displacement compensation accuracy of ±0.3mm; Drift correction engine, using polynomial fitting to eliminate slow wave drift and retain >0.5Hz physiological signals; The quality assessment submodule calculates the variance ratio between segments within a threshold of <0.15 and automatically removes overly distorted data segments.

[0010] Furthermore, the multi-dimensional feature extraction module includes: Time domain engine for calculating latency coefficient of variation, peak-to-peak fluctuation, and Hjorth mobility parameter; Frequency domain engine, generates δ / θ / α / β / γ band power spectral density matrix and inter-band coupling coefficients through Welch method; Nonlinear engine, performing entropy analysis and fractal dimension calculations on scales 1-15; The connectomics engine constructs a whole-brain functional connectivity network based on the weighted phase lag index and extracts the clustering coefficient of the frontal lobe-basal ganglia circuit.

[0011] Furthermore, the connectomics engine includes: Frequency band connection matrix generator, used to independently calculate the whole brain weighted phase lag index matrix of δ / θ / α / β / γ frequency bands; A dynamic connection tracker, which captures transient changes in the frontal lobe-basal ganglia circuit connection strength with a 500ms sliding window and a step size of 100ms; A graph-theoretic parameter quantizer, used to convert brain networks into undirected graphs and calculate node centrality, global efficiency, and modularity index; An abnormal connectivity detector to identify pathological hyperconnections that are not present in healthy controls.

[0012] Furthermore, the clinical data fusion module includes: Multi-source interface unit: used to analyze the UPDRS-III scale video assessment result parameters, such as joint range of motion, tremor frequency and Sniffin' Sticks olfactory test data; Feature alignment engine: used to synchronize clinical scale timestamps with EEG event markers; Cross-modal association miner: extracts EEG-clinical joint feature vectors using canonical correlation analysis; Longitudinal predictor: Based on the baseline biomarker data generated by the multi-source interface unit, feature alignment engine and cross-modal association miner, it inputs baseline biomarkers and outputs the disease progression curve for the next 36 months.

[0013] Furthermore, the interpretable classification module includes: The XGBoost classifier uses tree depth and learning rate features extracted from the 20-dimensional core biomarker encoding vector to complete early Parkinson's disease identification; The SHAP interpretability engine is used to visualize the spatial distribution of high-contribution biomarkers in the form of brain topography, plot the trigger sequence of key feature thresholds during the classification of specific samples, simulate the impact of changes in feature values ​​on classification results, quantify the causal contribution weights of Parkinson's pathological mechanisms, generate feature contribution rankings and brain topography decision path maps, and preview changes in classification results in real time to assist in clinical decision verification; Built-in dynamic update interface, automatically triggering the semi-supervised learning process when the classification confidence of new samples is <90%.

[0014] Compared with the existing technology, the non-invasive EEG-based early Parkinson's disease biomarker identification system and equipment provided by the present invention have the following beneficial effects: 1. Through system-level collaborative innovation, revolutionary improvements have been achieved in the three dimensions of acquisition, analysis, and computing. At the signal acquisition layer, nanocrystalline shielded dry electrodes combined with dynamic impedance adjustment technology increase the signal-to-noise ratio in home environments to 28dB. At the same time, the six-degree-of-freedom motion compensation module compresses the electrode displacement error to ±0.3mm, significantly improving the detection rate of weak beta oscillations in the prodromal phase at 12-30Hz. At the feature processing layer, a four-dimensional parallel extraction engine simultaneously generates 48 types of biomarkers. The intra-group correlation coefficient of features is improved through a three-level optimization pipeline. The pathological relevance of key markers such as the frontal lobe-basal ganglia circuit gamma entropy has been verified by Granger causality. The cloud-based distributed model server supports 2000 concurrent analysis channels, completely solving the real-time bottleneck of home screening. 2. The home-based closed-loop verification mechanism suppresses the false positive rate and can identify high-risk groups a median of 5.3 years in advance. Combined with the three-year individualized disease course curve output by the GRU longitudinal prediction model, it guides precise medication plans and significantly reduces patients' average annual medical expenses. At the universal level, the cost of a single test is reduced to 1 / 25 of PET. When the classification confidence of a newly identified additional sample is <90%, the semi-supervised learning process is automatically triggered for deep learning. The interpretable classification module can accumulate pathology in 50 samples for local deployment in county hospitals, reducing the misdiagnosis rate at the grassroots level. The feature contribution heat map generated by the interpretable AI engine reveals abnormal θ-γ cross-frequency coupling as a new treatment target. The blockchain-based biomarker database of 100,000 cases provides a real-world R&D platform for pharmaceutical companies. Related dry electrodes and edge computing technologies have been derived into Alzheimer's disease screening systems, which have huge commercial value for the late identification of Parkinson's disease.

[0015] An early biomarker identification device for Parkinson's disease based on non-invasive EEG is characterized by comprising a storage device and a processor, wherein the storage device is used to store a computer program, and the processor runs the computer program to enable the early biomarker identification device for Parkinson's disease to execute the early biomarker identification system for Parkinson's disease described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of the structure of the non-invasive EEG-based early biomarker identification system for Parkinson's disease proposed by the present invention; Figure 2 This is a schematic diagram of the structure of the multimodal signal acquisition module in the non-invasive EEG-based early Parkinson's disease biomarker identification system proposed by the present invention; Figure 3 This is a schematic diagram of the logical operation of the environmental adaptability unit in the multimodal signal acquisition module of the non-invasive EEG-based early biomarker identification system for Parkinson's disease proposed in the present invention; Figure 4This is a schematic diagram of the structure of the non-invasive EEG-based early Parkinson's disease biomarker identification device proposed by the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] The implementation of the present invention is described in detail below with reference to specific embodiments.

[0019] The same or similar numbers in the drawings of this embodiment correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "up", "down", "left", "right", etc. indicate directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. This is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0020] Reference Figure 1-3 As shown in the figure, the non-invasive EEG-based early biomarker identification system for Parkinson's disease includes: a multimodal signal acquisition module, equipped with a 64-lead dry electrode EEG cap, an integrated preamplifier and an ADC converter, for synchronously acquiring resting-state EEG and event-related potentials, dynamically switching the resting-state EEG sampling rate, and transmitting the acquired raw data to a preprocessing module. Specifically, the multimodal signal acquisition module has a built-in three-axis accelerometer and an infrared eye tracker for real-time monitoring of head micro-movements and eyelid movements, and dynamically switches the sampling rate between 500-2000 Hz to adapt to different noise environments. After the adaptation is completed, raw data is generated, encrypted by Bluetooth 5.0, and then transmitted to the preprocessing module; The preprocessing module is used to eliminate baseline drift in the collected raw data and separate blink components. It implements spectrum notching and acceleration data-driven Kalman filter correction by constructing a 50 / 60 Hz harmonic dictionary, eliminates drift in the collected raw data signal, and outputs a 2-second EEG segment that has been re-referenced in sections. Specifically, the preprocessing module includes: Adaptive bandpass filter unit, used to eliminate 0.5-70 Hz baseline drift in raw data; The CNN electrooculogram artifact recognition unit uses a three-layer convolution kernel structure to separate the blink component from the original data; Power frequency noise elimination unit, used to build a 50 / 60 Hz harmonic dictionary to achieve spectrum notching; Motion artifact compensation unit, which drives the Kalman filter through acceleration data to correct signal drift; The multidimensional feature extraction module is used to calculate the coefficient of variation of the latency, peak-to-peak fluctuation, and Hjorth mobility parameter in segmented, re-referenced 2-second EEG segments. The Welch method is used to generate the power spectral density matrix and inter-band coupling coefficient of the δ / θ / α / β / γ bands. Multiscale entropy analysis and fractal dimension are used to calculate the weighted phase lag index to complete the construction of the whole-brain functional connectivity network and extract the frontal lobe-basal ganglia circuit clustering coefficient from the whole-brain functional connectivity network. The biomarker optimization module is used to perform intraclass correlation coefficient testing on the extracted frontal lobe-basal ganglia circuit clustering coefficient, trigger wavelet packet reconstruction and re-extraction for features in the circuit clustering coefficient that do not meet the standards, and use the maximum correlation minimum redundancy algorithm to retain features significantly associated with the UPDRS score after all meet the standards. Regularized logistic regression is used to generate a 20-dimensional core biomarker encoding vector. The home-based closed-loop verification mechanism suppresses the false positive rate and can identify high-risk groups a median of 5.3 years in advance. Combined with the three-year individualized disease course curve output by the GRU longitudinal prediction model, it guides precise medication plans and significantly reduces patients' average annual medical expenses.

[0021] Specifically, it also includes an interpretable classification module, which is used to perform early Parkinson's disease identification on the 20-dimensional core biomarker encoding vector generated by regression, and output the corresponding identification report after the identification is completed. When the classification confidence of the newly identified incremental sample is <90%, the semi-supervised learning process is automatically triggered for deep learning; the clinical data fusion module is used to receive the corresponding identification report, and parse the evaluation results through the UPDRS-III scale, and synchronize the clinical scale timestamp with the EEG event marker to ensure the consistency of the spatiotemporal dimensions, synchronize the baseline biomarkers to output the disease progression curve within the preset time in the future, and realize the three dimensions of collection, analysis and calculation through system-level collaborative innovation. It has achieved revolutionary improvements. At the signal acquisition layer, the nanocrystalline shielded dry electrode combined with dynamic impedance adjustment technology has increased the signal-to-noise ratio in the home environment to 28dB. At the same time, the six-degree-of-freedom motion compensation module compresses the electrode displacement error to ±0.3mm, significantly improving the detection rate of weak beta oscillations in the prodromal period at 12-30Hz. At the feature processing layer, the four-dimensional parallel extraction engine simultaneously generates 48 types of biomarkers, and the intra-group correlation coefficient of the feature is improved through a three-level optimization pipeline. The pathological correlation of key markers such as the gamma entropy of the frontal lobe-basal ganglia loop has been verified by Granger causality. The cloud-based distributed model server supports 2000 concurrent analyses, which completely solves the real-time bottleneck of home screening.

[0022] The operation logic flow of this technical solution is: Multimodal EEG data acquisition: Using a 64-lead international 10-20 system standard electrode cap, bimodal EEG signals (resting-state and event-related potential (ERP)) were collected in an electromagnetically shielded room. Resting-state acquisition included eyes-closed rest (5 minutes) and eyes-open rest (5 minutes). The ERP task employed the auditory oddball paradigm, with 80% standard stimulation and 20% deviant stimulation. The sampling rate was set to 1000 Hz, and the electrode impedance was controlled below 5 kΩ. ECG and EOG signals were recorded simultaneously for subsequent artifact removal. The acquisition process required that the subjects had not taken neuroactive medication within 48 hours and were in a state of wakefulness and relaxation. EEG signal preprocessing and enhancement: The raw EEG signals were bandpass filtered at 0.5-70 Hz, and independent component analysis (ICA) combined with wavelet threshold denoising was used to eliminate electrooculogram (EOG), electromyographic (EMG), and power frequency interference. An automatic artifact rejection algorithm with an amplitude threshold of ±100 μV and a gradient threshold of 50 μV / ms was used to remove transient noise. The signals were re-referenced using the whole-brain mean potential, segmented into 2-second time windows, and baseline corrected to generate a high-quality EEG segment library. Multi-dimensional feature joint extraction: Four types of features are extracted in parallel from the preprocessed signal: (a) Time domain characteristics: including peak amplitude, latency, zero-crossing rate, and Hjorth activity parameter; (b) Frequency domain features: Calculate the power spectral density and inter-band power ratio of the δ (1-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-30 Hz), and γ (30-50 Hz) bands; (c) Time-frequency characteristics: Morlet wavelet transform is used to extract the energy entropy and synchronization index of specific frequency bands; (d) Nonlinear dynamic characteristics: calculation of multiscale entropy, fractal dimension and Lyapunov exponent; Feature stability optimization and selection: A three-step screening strategy was used: first, the intraclass correlation coefficient (ICC>0.8) was used to assess the stability of features in repeated experiments; second, the maximum correlation minimum redundancy algorithm was used to select a subset of features significantly associated with Parkinson's disease course (p<0.01); finally, L1 regularized sparse logistic regression was used to compress feature dimensions, retaining the top 20 core biomarkers with the highest absolute weight coefficients; Dynamic classification model construction and validation: Selected features are input into the integrated learning framework to build an XGBoost-based Parkinson's early classification model. The model is trained using a 5-fold cross-validation strategy, and SMOTE oversampling is used to address class imbalance. Model performance is evaluated using AUC, sensitivity, and specificity. The model is ultimately deployed as a dynamic diagnostic system that can be updated online.

[0023] Specifically, the dual paradigms of resting state and ERP induce complementary neural responses, and synchronous behavioral monitoring ensures data validity, overcoming the lack of sensitivity of a single paradigm; time domain / frequency domain / nonlinearity / connectivity features are jointly extracted, and multi-scale entropy and functional connectivity matrix are integrated for the first time to comprehensively capture neurodegeneration information; the three-level mechanism of ICC stability screening + mRMR correlation filtering + L1 sparse optimization ensures that biomarkers are clinically reproducible; the SHAP framework is combined with the visualization of brain topography decision pathways to break through traditional limitations and provide evidence of association with pathological mechanisms; at the same time, the small sample migration solution solves the problem of scarcity of clinical data, and the PET verification feedback loop realizes model self-evolution, significantly improving the specificity of early warning.

[0024] In this embodiment, the multimodal signal acquisition module is further provided with an environmental adaptability unit, which includes: The electromagnetic shielding layer uses nanocrystalline alloy materials to suppress interference in the frequency band above 50Hz, with a shielding effectiveness of >60dB; Dynamic impedance adjustment circuit, which reduces the electrode-skin contact impedance in real time through micro-current injection, with the fluctuation range controlled within ±0.5kΩ; Artifact self-check unit, which performs a 3-second impedance scan and myoelectric noise baseline test before starting acquisition, and automatically switches unqualified leads to redundant electrodes; The dual-mode storage unit uses a local SD card to cache 72 hours of raw data and encrypted cloud storage for key event markers.

[0025] In this embodiment, the motion artifact compensation unit includes: A six-degree-of-freedom motion modeler builds a head rotation / translation motion model based on accelerometer data and predicts the spatial displacement of electrodes; Kalman filter group, with independent configuration of filtering parameters for high-sensitivity forehead / temporal leads, and displacement compensation accuracy of ±0.3mm; Drift correction engine, using polynomial fitting to eliminate slow wave drift and retain >0.5Hz physiological signals; The quality assessment submodule calculates the variance ratio between segments within a threshold of <0.15 and automatically removes overly distorted data segments.

[0026] In this embodiment, the multi-dimensional feature extraction module includes: Time domain engine for calculating latency coefficient of variation, peak-to-peak fluctuation, and Hjorth mobility parameter; Frequency domain engine, generates δ / θ / α / β / γ band power spectral density matrix and inter-band coupling coefficients through Welch method; Nonlinear engine, performing entropy analysis and fractal dimension calculations on scales 1-15; The connectomics engine constructs a whole-brain functional connectivity network based on the weighted phase lag index and extracts the clustering coefficient of the frontal lobe-basal ganglia circuit.

[0027] In this embodiment, the connectomics engine includes: a frequency-band connection matrix generator, which is used to independently calculate the whole-brain weighted phase lag index matrix of the δ / θ / α / β / γ frequency bands; a dynamic connection tracker, which captures transient changes in the connection strength of the frontal lobe-basal ganglia loop with a 500ms sliding window and a step size of 100ms; a graph theory parameter quantizer, which is used to convert the brain network into an undirected graph and calculate node centrality, global efficiency and modularity index; and an abnormal connection detector, which is used to identify pathological hyperconnectivity that does not exist in the healthy control group.

[0028] In this embodiment, the clinical data fusion module includes: Multi-source interface unit: used to analyze the UPDRS-III scale video assessment result parameters, including joint range of motion, tremor frequency, and Sniffin' Sticks olfactory test data; Feature alignment engine: used to synchronize clinical scale timestamps with EEG event markers; Cross-modal association miner: extracts EEG-clinical joint feature vectors using canonical correlation analysis; Longitudinal predictor: Based on the baseline biomarker data generated by the multi-source interface unit, feature alignment engine and cross-modal association miner, it inputs baseline biomarkers and outputs the disease progression curve for the next 36 months.

[0029] In this embodiment, the interpretable classification module includes: The XGBoost classifier uses tree depth and learning rate features extracted from the 20-dimensional core biomarker encoding vector to complete early Parkinson's disease identification; The SHAP interpretability engine is used to visualize the spatial distribution of high-contribution biomarkers in the form of brain topography, plot the threshold trigger sequence of key features in the classification process of specific samples, simulate the impact of changes in feature values ​​on classification results, quantify the causal contribution weights of Parkinson's disease pathology, generate feature contribution rankings and brain topography decision path maps, and preview changes in classification results in real time to assist in clinical decision verification. A three-level optimization pipeline improves the intra-group correlation coefficient of features, and the pathological relevance of key markers such as the frontal lobe-basal ganglia circuit gamma entropy is verified by Granger causality. The cloud-based distributed model server supports 2,000 concurrent analyses, completely solving the real-time bottleneck of home screening. The built-in dynamic update interface automatically triggers the semi-supervised learning process when the classification confidence of a new sample is <90%. It can also interpret the feature contribution heat map generated by the AI ​​engine to reveal θ-γ cross-frequency coupling abnormalities as new therapeutic targets. The blockchain-based database of 100,000 biomarkers provides a real-world R&D platform for pharmaceutical companies. Related dry electrodes and edge computing technologies have been derived into Alzheimer's disease screening systems, which have huge commercial value for the late identification of Parkinson's disease.

[0030] Through system-level collaborative innovation, this technical solution has achieved revolutionary improvements in the three dimensions of acquisition, analysis, and computing. At the signal acquisition layer, nanocrystalline shielded dry electrodes combined with dynamic impedance adjustment technology have increased the signal-to-noise ratio in the home environment to 28dB. Moreover, when the classification confidence of a newly identified sample is <90%, the semi-supervised learning process is automatically triggered for deep learning. The interpretable classification module can complete the local deployment of county hospitals with accumulated pathology in 50 samples, reducing the misdiagnosis rate at the grassroots level. The feature contribution heat map generated by the interpretable AI engine reveals θ-γ cross-frequency coupling abnormalities as a new treatment target, and the blockchain-based biomarker database of 100,000 cases provides a real-world R&D platform for pharmaceutical companies. Related dry electrodes and edge computing technologies have been derived into Alzheimer's disease screening systems, which have huge commercial value for the late identification of Parkinson's disease.

[0031] Reference Figure 4 As shown, a non-invasive EEG-based early biomarker identification device for Parkinson's disease is characterized by including a storage device and a processor, the storage device being used to store a computer program, and the processor running the computer program to enable the early biomarker identification device for Parkinson's disease to execute the above-mentioned early biomarker identification system for Parkinson's disease. Through system-level collaborative innovation, revolutionary improvements are achieved in the three dimensions of acquisition, analysis, and calculation. At the signal acquisition layer, nanocrystalline shielded dry electrodes combined with dynamic impedance adjustment technology increase the signal-to-noise ratio in the home environment to 28dB. At the same time, the six-degree-of-freedom motion compensation module compresses the electrode displacement error to ±0.3mm, significantly improving the detection rate of weak beta oscillations in the prodromal period at 12-30Hz. At the feature processing layer, a four-dimensional parallel extraction engine synchronously generates 48 types of biomarkers. The intra-group correlation coefficient of the features is improved through a three-level optimization pipeline. The pathological correlation of key markers such as the gamma entropy of the frontal lobe-basal ganglia loop is verified by Granger causality. The cloud-based distributed model server supports 2000 concurrent analyses, completely solving the real-time bottleneck of home screening. This technical solution combines the three-year individualized disease course curve output by the GRU longitudinal prediction model to guide precise medication plans and significantly reduce patients' average annual medical expenses. At the universal level, the cost of a single test is reduced to 1 / 25 of PET. When the classification confidence of a newly identified sample is <90%, the semi-supervised learning process is automatically triggered for deep learning. The interpretable classification module can accumulate pathology samples in 50 cases to complete local deployment in county hospitals, reducing the misdiagnosis rate at the grassroots level. In addition, the feature contribution heat map generated by the interpretable AI engine reveals θ-γ cross-frequency coupling abnormalities as new treatment targets.

[0032] In this embodiment, the entire operation process can be controlled by a computer to provide signal feedback to implement the steps in sequence. These are all conventional knowledge of current automated control and will not be described in detail in this embodiment.

[0033] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A non-invasive EEG-based early biomarker identification system for Parkinson's disease, characterized by: include: The multimodal signal acquisition module is equipped with a 64-lead dry electrode EEG cap, an integrated preamplifier and an ADC converter, and is used to synchronously acquire resting-state EEG and event-related potentials. The resting-state EEG sampling rate is dynamically switched, and the collected raw data is transmitted to the preprocessing module. The preprocessing module is used to eliminate baseline drift of the collected raw data and separate the blink component. It implements spectrum notching and acceleration data-driven Kalman filter correction by constructing a 50 / 60 Hz harmonic dictionary, eliminates drift of the collected raw data signal, and outputs a 2-second EEG segment that has been re-referenced in sections. A multidimensional feature extraction module is used to calculate the coefficient of variation of latency, peak-to-peak fluctuation, and Hjorth mobility parameter in segmented, re-referenced 2-second EEG segments. The Welch method is used to generate the power spectral density matrix and inter-band coupling coefficients for the δ / θ / α / β / γ bands. Multiscale entropy analysis and fractal dimension are used to calculate the weighted phase lag index to construct a whole-brain functional connectivity network. The clustering coefficient of the frontal lobe-basal ganglia circuit is extracted from the whole-brain functional connectivity network. The biomarker optimization module is used to perform intraclass correlation coefficient testing on the extracted frontal lobe-basal ganglia circuit clustering coefficients. It triggers wavelet packet reconstruction and re-extraction of features that do not meet the criteria in the circuit clustering coefficients. Once all features meet the criteria, the maximum correlation minimum redundancy algorithm is used to retain features significantly associated with the UPDRS score. Regularized logistic regression is used to generate a 20-dimensional core biomarker encoding vector. An interpretable classification module is used to perform early Parkinson's disease identification based on the 20-dimensional core biomarker encoding vector generated by regression. After identification, a corresponding identification report is output, and when the classification confidence of the newly identified additional sample is less than 90%, the semi-supervised learning process is automatically triggered for deep learning. The clinical data fusion module is used to receive corresponding identification reports, analyze the evaluation results through the UPDRS-III scale, synchronize the clinical scale timestamp with the EEG event marker to ensure consistency in the spatiotemporal dimensions, and synchronize the baseline biomarkers to output the disease progression curve within a preset time in the future.

2. The non-invasive EEG-based early biomarker identification system for Parkinson's disease according to claim 1, characterized in that: The multimodal signal acquisition module has a built-in three-axis accelerometer and an infrared eye tracker for real-time monitoring of head micro-movements and eyelid movements. The sampling rate is dynamically switched between 500-2000 Hz to adapt to different noise environments. After the adaptation is completed, the raw data is generated. The raw data is encrypted by Bluetooth 5.0 and then transmitted to the preprocessing module.

3. The non-invasive EEG-based early biomarker identification system for Parkinson's disease according to claim 2, characterized in that: The multimodal signal acquisition module is further provided with an environmental adaptability unit, and the environmental adaptability unit includes: The electromagnetic shielding layer uses nanocrystalline alloy materials to suppress interference in the frequency band above 50Hz, with a shielding effectiveness of >60dB; Dynamic impedance adjustment circuit, which reduces the electrode-skin contact impedance in real time through micro-current injection, with the fluctuation range controlled within ±0.5kΩ; Artifact self-check unit, which performs a 3-second impedance scan and myoelectric noise baseline test before starting acquisition, and automatically switches unqualified leads to redundant electrodes; The dual-mode storage unit uses a local SD card to cache 72 hours of raw data and encrypted cloud storage for key event markers.

4. The non-invasive EEG-based early biomarker identification system for Parkinson's disease according to claim 3, characterized in that: The pre-processing module comprises: Adaptive bandpass filter unit, used to eliminate 0.5-70 Hz baseline drift in raw data; The CNN electrooculogram artifact recognition unit uses a three-layer convolution kernel structure to separate the blink component from the original data; Power frequency noise elimination unit, used to build a 50 / 60 Hz harmonic dictionary to achieve spectrum notching; The motion artifact compensation unit drives the Kalman filter through acceleration data to correct signal drift.

5. The non-invasive EEG-based early biomarker identification system for Parkinson's disease according to claim 4, characterized in that: The motion artifact compensation unit includes: A six-degree-of-freedom motion modeler builds a head rotation / translation motion model based on accelerometer data and predicts the spatial displacement of electrodes; Kalman filter group, with independent configuration of filtering parameters for high-sensitivity forehead / temporal leads, and displacement compensation accuracy of ±0.3mm; Drift correction engine, using polynomial fitting to eliminate slow wave drift and retain >0.5Hz physiological signals; The quality assessment submodule calculates the variance ratio between segments within a threshold of <0.15 and automatically removes overly distorted data segments.

6. The non-invasive EEG-based early biomarker identification system for Parkinson's disease according to claim 5, characterized in that: The multi-dimensional feature extraction module includes: Time domain engine for calculating latency coefficient of variation, peak-to-peak fluctuation, and Hjorth mobility parameter; Frequency domain engine, generates δ / θ / α / β / γ band power spectral density matrix and inter-band coupling coefficients through Welch method; Nonlinear engine, performing entropy analysis and fractal dimension calculations on scales 1-15; The connectomics engine constructs a whole-brain functional connectivity network based on the weighted phase lag index and extracts the clustering coefficient of the frontal lobe-basal ganglia circuit.

7. The non-invasive EEG-based early biomarker identification system for Parkinson's disease according to claim 6, characterized in that: The connectomics engine includes: Frequency band connection matrix generator, used to independently calculate the whole brain weighted phase lag index matrix of δ / θ / α / β / γ frequency bands; A dynamic connection tracker, which captures transient changes in the frontal lobe-basal ganglia circuit connection strength with a 500ms sliding window and a step size of 100ms; A graph-theoretic parameter quantizer, used to convert brain networks into undirected graphs and calculate node centrality, global efficiency, and modularity index; An abnormal connectivity detector to identify pathological hyperconnections that are not present in healthy controls.

8. The non-invasive EEG-based early biomarker identification system for Parkinson's disease according to claim 7, characterized in that: The clinical data fusion module includes: Multi-source interface unit: used to analyze the UPDRS-III scale video assessment result parameters, such as joint range of motion, tremor frequency and Sniffin' Sticks olfactory test data; Feature alignment engine: used to synchronize clinical scale timestamps with EEG event markers; Cross-modal association miner: extracts EEG-clinical joint feature vectors using canonical correlation analysis; Longitudinal predictor: Based on the baseline biomarker data generated by the multi-source interface unit, feature alignment engine and cross-modal association miner, it inputs baseline biomarkers and outputs the disease progression curve for the next 36 months.

9. The non-invasive EEG-based early biomarker identification system for Parkinson's disease according to claim 8, characterized in that: The explainable classification module includes: The XGBoost classifier uses tree depth and learning rate features extracted from the 20-dimensional core biomarker encoding vector to complete early Parkinson's disease identification; The SHAP interpretability engine is used to visualize the spatial distribution of high-contribution biomarkers in the form of brain topography, plot the trigger sequence of key feature thresholds during the classification of specific samples, simulate the impact of changes in feature values ​​on classification results, quantify the causal contribution weights of Parkinson's pathological mechanisms, generate feature contribution rankings and brain topography decision path maps, and preview changes in classification results in real time to assist in clinical decision verification; Built-in dynamic update interface, automatically triggering the semi-supervised learning process when the classification confidence of new samples is <90%.

10. A non-invasive EEG-based early biomarker identification device for Parkinson's disease, characterized by: The system comprises a storage device and a processor, wherein the storage device is used to store a computer program, and the processor runs the computer program to enable the device for identifying early biomarkers of Parkinson's disease to execute the system for identifying early biomarkers of Parkinson's disease according to any one of claims 1 to 9.

Images