An electrocardiosignal classification method based on a dual-threshold neuromorphic reservoir

Abstract

The application discloses a kind of electrocardiosignal classification methods based on double threshold neural morphological reservoir, comprising the following steps: step 1, receive electrocardio time series signal, the adjacent time series nonlinear coding operation is carried out to the electrocardio time series signal, and the enhanced feature vector containing original signal and second-order interaction feature is generated;Step 2, the enhanced feature vector is input to double threshold neural morphological reservoir, and the double-channel pulse response sequence is generated using double threshold pulse neuron to distinguish signal intensity state;Step 3, according to the double-channel pulse response sequence, construct joint feature vector, and input classifier to obtain the classification result of electrocardiosignal.The application solves the problem that traditional liquid state machine is poor in linear separability and neuron is easy to saturate by the cooperation of feature enhancement and double threshold mechanism, while greatly reducing the number of neurons, improves the accuracy and energy efficiency ratio of arrhythmia classification.

Description

Technical Field

[0001] This invention relates to the fields of neuromorphic computing and artificial intelligence, and in particular to a method for classifying electrocardiogram signals based on a dual-threshold neuromorphic reservoir. Background Technology

[0002] Liquid state machines (LSMs), as a type of recurrent spiking neural network (SNN), have a natural advantage in processing time-series signals. However, existing LSM techniques face a challenge in balancing computational power and performance when applied to the analysis of bioelectrical signals (such as ECG). First, poor linear separability and insufficient feature mining: Traditional LSM directly injects the raw signal into the reservoir, ignoring the higher-order interactions within the input signal. The raw signal has poor linear separability in low-dimensional space, resulting in the need for a huge reservoir size to distinguish different types of signals through high-dimensional projection.

[0003] Secondly, information compression due to a single threshold: existing LIF neuron models typically employ only a single firing threshold. This design compresses input stimuli of varying intensities into a uniform impulse response, failing to distinguish between "weak stimuli" and "strong stimuli," resulting in underutilization of the reservoir's expressive capacity.

[0004] Finally, high resource consumption: In order to obtain high-precision classification results, traditional methods often rely on increasing the number of neurons (e.g., 512 or more), which leads to huge computational overhead (MAC operation) and energy consumption, making it difficult to deploy on resource-constrained wearable neuromorphic hardware. Summary of the Invention

[0005] The purpose of this invention is to provide a method for classifying electrocardiogram (ECG) signals based on a dual-threshold neuromorphic reservoir. This classification method offers advantages such as high computational efficiency, accurate classification, and low power consumption.

[0006] The technical solution of this invention: A method for classifying electrocardiogram signals based on a dual-threshold neuromorphic reservoir, comprising the following steps: Step 1: Receive the ECG time-series signal, perform adjacent time-series nonlinear encoding operation on the ECG time-series signal, and generate an enhanced feature vector containing the original signal and second-order interactive features; Step 2: Input the enhanced feature vector into the dual-threshold neuromorphic reservoir, wherein the dual-threshold neuromorphic reservoir is composed of multiple dual-threshold spiking neurons, used to generate a dual-channel impulse response sequence based on the enhanced feature vector; Step 3: Construct a joint feature vector based on the dual-channel impulse response sequence, and input the joint feature vector into a classifier to obtain the classification result of the electrocardiogram signal.

[0007] The aforementioned ECG signal classification method based on a dual-threshold neuromorphic reservoir, specifically the adjacent temporal nonlinear encoding operation on the ECG time-series signal, involves: encoding the length of the signal... input vector Calculate the product of elements at adjacent time steps to generate adjacent feature vectors. : ; The input vector With the adjacent feature vector The enhanced feature vector is formed by concatenating the features. : .

[0008] The aforementioned ECG signal classification method based on a dual-threshold neuromorphic reservoir, wherein the dual-threshold spiking neurons are configured with low-threshold... and high threshold ,in ; The dual-threshold spiking neurons at each time step Based on its membrane potential The state, respectively, generates a first spike output corresponding to the low threshold. (i.e., the channel for capturing weak signals) and the second spike output corresponding to the high threshold. (i.e., the channel for capturing strong signals): ; .

[0009] The membrane potential described in the aforementioned ECG signal classification method based on dual-threshold neuromorphic reservoirs State updates follow these rules:

[0010] in, It is the membrane potential decay factor. It is the total synaptic current, which includes the input current and the circulating current within the reservoir. This represents the membrane resistance of a neuron.

[0011] The aforementioned ECG signal classification method based on a dual-threshold neuromorphic reservoir generates a dual-channel impulse response sequence based on the enhanced feature vector, specifically including: Within a preset time window T, record all first peak output sequences and second peak output sequences generated by each neuron in the reservoir. All first peak output sequences and second peak output sequences recorded within the time window T are concatenated to form the joint feature vector.

[0012] The aforementioned ECG signal classification method based on dual-threshold neuromorphic reservoir uses a linear classifier whose weights are obtained through supervised learning. During training, an alternative gradient method is used for error backpropagation.

[0013] The aforementioned ECG signal classification method based on a dual-threshold neuromorphic reservoir further includes a data preprocessing step before inputting the enhanced feature vector into the dual-threshold neuromorphic reservoir: Extract signal segments centered on the R peak from the electrocardiogram signal database; The signal segment is divided into multiple sub-segments using a sliding window method to form the input time-series electrocardiogram signal.

[0014] A system for implementing the aforementioned electrocardiogram signal classification method includes: The nonlinear timing coding module is used to perform adjacent timing nonlinear coding operations on the input ECG timing signal to generate an enhanced feature vector containing second-order dynamic evolution information of the signal. A dual-threshold neuromorphic reservoir, composed of multiple dual-threshold spiking neurons, is used to receive the enhanced feature vector and generate a dual-channel pulse response sequence. Each neuron is configured with a low threshold and a high threshold, and low-energy and high-energy response channels are established in parallel through a threshold-level response mechanism. The low-energy response channel is configured to fire a pulse when the neuronal membrane potential exceeds the low threshold, enabling highly sensitive capture of weak morphological features in electrocardiogram signals and preventing the loss of weak signals such as P waves or T waves. The high-energy response channel is configured to fire a pulse when the neuronal membrane potential exceeds the high threshold, responding to high-intensity significant rhythmic features and preventing homogeneous saturation of neurons when processing high-amplitude QRS complexes. The full-time readout module is used to construct a joint feature vector based on the dual-channel impulse response sequence and output the classification result using a classifier.

[0015] In the above system, the nonlinear timing coding module and the dual-threshold neuromorphic reservoir are implemented based on digital logic circuits or neuromorphic computing devices. The digital logic circuit includes a field-programmable gate array, a digital signal processor, or an application-specific integrated circuit, which performs adjacent-sequential nonlinear coding operations and neuron state updates through a pipelined architecture. The neuromorphic computing device includes memristor arrays, phase-change memories, or two-dimensional material floating-gate transistors, and performs in-memory computing using the physical characteristics of the device.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. A balance between low computational complexity and enhanced nonlinear features. This invention employs an adjacent-time-series nonlinear coding strategy to extract second-order interactive features of adjacent time steps of the signal. This achieves a computational complexity comparable to the O(n) required for traditional full-matrix Hadamard product operations. 2 Compared to other methods, this invention reduces computational complexity to linear O(n), enriching the feature dimensions of the input signal without increasing the number of learnable parameters. This design not only effectively improves the problem of poor linear separability of the original signal in low-dimensional space, enabling clearer distinction between different types of ECG signals before entering the reservoir; at the same time, the nonlinearity introduced by adjacent temporal nonlinear encoding effectively alleviates the neuron "uniform saturation" phenomenon commonly found in traditional liquid state machines, allowing the reservoir to maintain a sparse and unsaturated efficient computational state across a wide input scale, preserving a larger dynamic capacity to process complex ECG signals.

[0017] 2. Dual-threshold mechanism enhances dynamic range and coding accuracy. This invention employs dual-threshold spiking neurons, simulating the nonlinear response characteristics of biological nervous systems or neuromorphic hardware to stimuli of varying intensities through the synergistic configuration of low and high thresholds. This mechanism effectively distinguishes between "weak stimuli" (such as T waves) and "strong stimuli" (such as R waves) in the input signal, generating two sets of parallel pulse response sequences. This significantly improves the reservoir's ability to capture subtle temporal features of ECG signals with a low number of neurons, solving the problem of information loss or saturation that is common in traditional single-threshold models.

[0018] 3. Extremely high computational efficiency and resource compression ratio. Thanks to the optimized algorithm architecture, this invention significantly reduces the number of reservoir neurons (e.g., from the traditional 512 to 139), maintaining state-of-the-art classification accuracy while significantly reducing memory consumption. Experiments show that this invention requires only about 56 multiply-accumulate (MAC) operations per time step during the inference phase. Compared to traditional LSTM or SNN models, the computational overhead is reduced by several orders of magnitude, perfectly matching the microwatt-level power budget of implantable chips or passive wearable devices.

[0019] 4. Hardware-friendly and software-hardware synergy potential. The core operators of this invention (adjacent product, threshold comparison) are simple and efficient. They not only support low-cost implementation on general-purpose digital hardware such as FPGA / DSP through pipeline technology, but are also naturally adapted to in-memory computing (Processing-In-Memory) architectures based on memristors or two-dimensional materials, and have broad engineering application prospects. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2The average electrocardiogram waveforms of five types of heartbeats (N, SVEB, VEB, F, Q) in the MIT-BIH database; Figure 3 This is a schematic diagram of the original electrocardiogram signal characteristic distribution; Figure 4 A schematic diagram of the ECG signal characteristics after adjacent temporal nonlinear coding enhancement; Figure 5 It demonstrates the feature dimension (number of neurons) requirements for each strategy under different window sizes; Figure 6 The classification accuracy of different model configurations under various window sizes is presented; Figure 7 A two-dimensional graph of the "optimal decay path" for threshold and decay rate; Figure 8 for Figure 7 3D performance surface of classification accuracy corresponding to 2D graph; Figure 9 A comparison of neuronal activation heatmaps between traditional LSM and the HLSM (Hadamard-enhanced LSM) of this invention; Figure 10 This is the loss function curve during model training; Figure 11 The accuracy curve during model training; Figure 12 This is the confusion matrix of the model on the MIT-BIH test set. Detailed Implementation

[0021] The present invention will be further described below with reference to embodiments, but these embodiments are not intended to limit the scope of the invention.

[0022] Example: A method for classifying electrocardiogram signals based on a dual-threshold neuromorphic reservoir, such as... Figure 1 As shown, it includes the following steps: Step 1: Receive the ECG time-series signal, perform adjacent time-series nonlinear encoding operation on the ECG time-series signal, and generate an enhanced feature vector containing the original signal and second-order interactive features; In this step, the MIT-BIH arrhythmia database was selected as the data source for both training and testing. This database contains 48 half-hour ECG recordings with a sampling frequency of 360Hz, and only modified limb lead II (MLII) signals were used for analysis. Records containing pacemakers were excluded according to the AAMI criteria, and the ECG signal labels were mapped into five categories (as shown in Table 1: normal heartbeat N, supraventricular ectopic heartbeat SVEB, ventricular ectopic heartbeat VEB, fused heartbeat F, and unknown heartbeat Q). The typical morphological differences of each category of heartbeat can be identified through... Figure 2(Average ECG waveforms for each category) Visual observation: The QRS complex of a normal heartbeat is sharp and symmetrical, while the amplitude and width of an ectopic heartbeat waveform deviate significantly, providing morphological basis for subsequent classification.

[0023] Table 1

[0024] The differential thresholding method is used to identify R-peaks in electrocardiogram (ECG) signals (characteristic peaks corresponding to ventricular depolarization, which are key reference points for the cardiac cycle). Specifically, the original signal undergoes a first-order differential operation, a reasonable threshold is set to filter noise interference, and when the amplitude of the differential signal exceeds the threshold and meets the peak duration condition, it is determined to be an R-peak, and its timestamp and amplitude are recorded.

[0025] Centered on each R peak, a signal segment with a length of 180 sampling points is extracted to cover the complete cardiac cycle (including core features such as P wave, QRS complex, and T wave), ensuring that the segment can reflect the complete morphological characteristics of the heartbeat.

[0026] The extracted signal segments are divided a second time using a sliding window method. The window size (a hyperparameter, recommended range 64–128 sampling points) and step size (recommended step size 1 / 5–1 / 3 of the window size to ensure data continuity) are set. Each complete cardiac beat segment is divided into multiple overlapping sub-segments. Each sub-segment is normalized, mapping the signal amplitude to the [-1,1] interval to eliminate amplitude differences under different recording conditions and ensure the consistency of the model input. This forms the input time-series ECG signal vector. (n is the window length). To address the class imbalance problem in the database, random oversampling is used for a minority of classes and random undersampling is used for the majority of classes to construct a balanced training set, with 5000 instances in each class, for a total of 25000 training samples.

[0027] For the preprocessed input vector The adjacent time step element-wise multiplication strategy is used to perform adjacent temporal nonlinear coding operations to generate adjacent feature vectors. The calculation formula is: To visually verify the effectiveness of adjacent temporal nonlinear coding feature enhancement, this embodiment performs t-SNE two-dimensional visualization analysis on the distribution of reservoir input data. Figure 3 The distribution of the raw, unprocessed input data is shown, revealing significant overlap between data points of different categories. This indicates poor linear separability in the original feature space, making direct classification difficult. Figure 4The distribution of the data after adjacent temporal nonlinear encoding enhancement is shown. Compared with the original data, the enhanced data exhibits significant inter-class separability and higher intra-class compactness in the feature space. This result confirms that the Hadamard transform successfully enriches the nonlinear structure of the input signal, enabling signals of different arrhythmia categories to have better separability before entering the reservoir, thereby reducing the difficulty of subsequent classification tasks.

[0028] The original input vector and adjacent feature vectors Enhanced feature vectors are formed by concatenating dimensions. The splicing method is The concatenated vector has a length of 2n-1, which retains the original time-domain features and adds second-order interaction information, thus improving the linear separability of the signal. Figure 5 and Figure 6 The trade-offs between computational complexity and model performance for different Hadamard augmentation strategies were analyzed. Figure 5 The feature dimension (number of neurons) required for full Hadamard (HN_all) and partial Hadamard (Adj2 / 3 / 4) strategies were compared under different window sizes. The results show that the complexity of the full Hadamard strategy increases quadratically with increasing window size. The preferred adjacent-temporal nonlinear coding (Adj) strategy of this invention maintains linear growth. This greatly reduces the demand for computing resources. Figure 6 The classification accuracy of six different model configurations under different window sizes is shown. The six models are DT_HLSM_Adj2 (double threshold and Hadamard-enhanced liquid state machine, adjacent 2nd-order features), DT_HLSM_Adj3 (double threshold and Hadamard-enhanced liquid state machine, adjacent 3rd-order features), DT_HLSM_Adj4 (double threshold and Hadamard-enhanced liquid state machine, adjacent 4th-order features), HLSM_Adj2 (single threshold and Hadamard-enhanced liquid state machine, adjacent 2nd-order features), HLSM_Adj3 (single threshold and Hadamard-enhanced liquid state machine, adjacent 3rd-order features), and HLSM_Adj4 (single threshold and Hadamard-enhanced liquid state machine, adjacent 4th-order features). Comparing the dashed line (double threshold DT model) with the solid line (single threshold model), it can be seen that the accuracy of all models is improved after introducing the double threshold mechanism (DT). Figure 5 and Figure 6 As can be seen, the DT_HLSM_Adj architecture proposed in this invention (as shown by the dark blue dashed line) achieves classification performance that surpasses the high-computational-consumption full Hadamard model while maintaining linear low computational complexity, thus achieving the best balance between efficiency and accuracy.

[0029] Step 2: Input the enhanced feature vector into the dual-threshold neuromorphic reservoir, wherein the dual-threshold neuromorphic reservoir is composed of multiple dual-threshold spiking neurons, used to generate a dual-channel impulse response sequence based on the enhanced feature vector; In this step, a reservoir consisting of N dual threshold leak integral firing (DT-LIF) neurons is constructed. Random sparse connections are used between neurons, the connection probability (con) is set to 0.2, the connection weights are initialized to random values ​​in the interval [−0.5, 0.5], and the excitation-inhibition neuron ratio (EI ratio) is 0.2.

[0030] Each DT-LIF neuron (denoted as neuron i) is configured with two firing thresholds, a low threshold and a high threshold. With high threshold ,and The default values ​​are 0.3 and 0.5 respectively; the membrane potential decay factor (decay) is set to 0.5, and the membrane resistance... It is a fixed constant used to adjust the dynamic response characteristics of neurons. Figure 7 and Figure 8 This study revealed a synergistic relationship between two key hyperparameters of LIF neurons—the firing threshold and the membrane potential decay rate. Figure 7 The experiment demonstrates the "optimal attenuation path" in a two-dimensional plane. Experiments revealed that there is no globally optimal fixed attenuation rate; the optimal attenuation rate dynamically changes with the threshold. For example, a low threshold (approximately 0.2) requires a low attenuation rate, while a high threshold (>0.4) requires a high attenuation rate (0.8-0.95). Figure 8 The corresponding 3D performance surface is displayed, with the red path marking the "ridge" region with the highest accuracy. From Figure 7 The "optimal decay path" indicates that a low threshold (0.3) requires a decay rate of 0.5. Figure 8 The 3D performance surface accuracy reached its peak, verifying the rationality of the parameter settings.

[0031] Each DT-LIF neuron at time step Total synaptic current received Including input current and reservoir internal circulating current, the membrane potential update follows a discrete-time dynamic equation:

[0032] in, It is the membrane potential decay factor.

[0033] The dynamic process of membrane potential renewal can be advantageously achieved through... Figure 9(Comparison of neuronal activation heatmaps, where 0.1-1.0 represents the amplitude scaling factor) demonstrates that traditional liquid state machines (LSMs) exhibit severe neuronal saturation (yellow areas) at different input scales, leading to dynamic range compression. In contrast, the Hadamard-enhanced LSM (HLSM) in this embodiment effectively suppresses non-specific saturation firing over a wide input range, maintaining rich dynamic response characteristics.

[0034] At each time step Based on its membrane potential The state, respectively, generates a first spike output corresponding to the low threshold. (i.e., the channel for capturing weak signals) and the second spike output corresponding to the high threshold. (i.e., the channel for capturing strong signals): ; .

[0035] Set a preset time window T (consistent with the input vector time step length n), and record the time of each neuron within the window. Sequence and The sequence forms a dual-channel impulse response sequence.

[0036] In this implementation, the proposed liquid state machine model was implemented using the PyTorch 2.2.0 framework. The optimal hyperparameters were determined through an initial optimization process, which involved systematically changing each parameter one by one while keeping other parameters constant. The final key settings are summarized in Table 2: Table 2

[0037] All computational experiments were conducted on workstations equipped with an Intel Core i5-13400 CPU, 32 GB of RAM, and an NVIDIA GeForce RTX 4070 GPU.

[0038] Step 3: Construct a joint feature vector based on the dual-channel impulse response sequence, and input the joint feature vector into a classifier to obtain the classification result of the electrocardiogram signal.

[0039] In this step, the low-threshold spike sequences and high-threshold spike sequences of all DT-LIF neurons recorded within the time window T are concatenated sequentially to form a joint feature vector. : ; The joint eigenvector It fully integrates temporal evolution information under different excitation levels.

[0040] Finally, a trainable classifier is used for classification: .

[0041] The classifier used in this implementation is a linear classifier, and the output layer weight moments are... The dataset is trained through supervised learning. To address the non-differentiability of spike signals, an alternative gradient method is used for error backpropagation, with cross-entropy loss as the loss function. During training, the dataset is divided into training and test sets in a 7:3 ratio, and the number of training epochs is controlled by convergence criteria (accuracy fluctuation is verified to be less than 0.1% after 5 consecutive epochs). For prediction, the joint feature vector is input into the trained classifier, which outputs the probability of each class. The class with the highest probability is taken as the final classification result.

[0042] In this embodiment, the stability of the linear classifier training process can be achieved through... Figure 10 loss curve and Figure 11 Accuracy curve validation: The training loss continuously decreased to below 0.1 with each epoch, while the validation accuracy remained stable above 98%, and the difference between training and validation accuracy was small, proving that the model did not overfit and had strong generalization ability. Classification performance passed... Figure 12 (Confusion matrix results) show that: the diagonal elements all exceed 97%, and the off-diagonal elements are less than 2.5%. In particular, the classification accuracy of ventricular ectopic heartbeats (VEB) is close to 100%. The recognition accuracy of key clinical categories is excellent, meeting the needs of electrocardiogram diagnosis.

[0043] The validation results of this embodiment on the MIT-BIH test set are as follows: Classification accuracy: The model achieves a classification accuracy of 98.38%, which is close to the performance of the traditional 512-neuron LSM model (98.79%), but the number of neurons is reduced by 73%, demonstrating the advantage of lightweight design.

[0044] Computational efficiency: Only 56 MAC operations are required per step (more than 99% less than traditional methods), and the model maintains high accuracy with low complexity, making it suitable for resource-constrained devices.

[0045] Robustness: Enhanced feature and parameter co-optimization enables the model to work stably under different input scales and noise environments, meeting the needs of practical clinical applications.

[0046] Example 2: This example provides a system for implementing the method in Example 1, comprising: A nonlinear time-series coding module receives continuous electrocardiogram (ECG) time-series signals and performs nonlinear interactive operations between adjacent time steps to generate an enhanced feature vector containing second-order dynamic evolution information of the signal. Specifically, the nonlinear interactive operation performed by the nonlinear time-series coding module involves performing a sliding window-based point-by-point multiplication operation on the input signal vector, i.e., calculating the product of the current time step sampling point and the previous time step sampling point, and using this product as the second-order interactive feature. This operational strategy eliminates the O(n) operation required by traditional full-matrix second-order interactive methods. 2 The computational complexity is reduced to linear O(n) by extracting only the temporal correlation information between adjacent time points.

[0047] A dual-threshold neuromorphic reservoir, composed of multiple dual-threshold spiking neurons, is used to receive the enhanced feature vector and generate a dual-channel pulse response sequence. Each neuron is configured with a low threshold and a high threshold, and low-energy and high-energy response channels are established in parallel through a threshold-level response mechanism. The low-energy response channel is configured to fire a pulse when the neuron's membrane potential exceeds the low threshold, for highly sensitive capture of weak morphological features in the electrocardiogram signal, preventing the loss of weak signals such as P waves or T waves. The high-energy response channel is configured to fire a pulse when the neuron's membrane potential exceeds the high threshold, for responding to high-intensity significant rhythmic features and preventing uniform saturation of neurons when processing high-amplitude QRS complexes. The membrane potential update of the dual-threshold spiking neurons follows a discrete-time dynamic equation, and the low and high thresholds are nonlinearly co-configured based on the charge capture probability distribution of the physical device. The low threshold is used to simulate the rapid capture and release of charge carriers by shallow-level traps, corresponding to the rapid decay characteristics of the signal; the high threshold is used to simulate the strong capture and retention of charge carriers by deep-level traps, corresponding to the slow decay characteristics of the signal.

[0048] The full-time readout module is used to construct a joint feature vector based on the dual-channel impulse response sequence and output the classification result using a classifier. The full-time readout module employs a dual-path parallel readout strategy, recording the weak signal pulse sequence generated by the low-energy response channel and the strong signal pulse sequence generated by the high-energy response channel, respectively. This dual-path parallel readout strategy fully preserves the temporal evolution information of the ECG signal from subtle morphological changes to significant high-amplitude peaks, avoiding feature loss or signal distortion caused by threshold trade-offs in traditional single-threshold models.

[0049] The system provided in this embodiment can be implemented using various hardware architectures to adapt to different application scenarios and power consumption requirements: In a preferred embodiment, the core module of the system is built upon neuromorphic computing devices to achieve extreme energy efficiency. Specifically, the nonlinear timing coding module and the reservoir computing module are implemented using memristor crossbar arrays, phase-change memories (PCMs), or floating-gate transistors based on two-dimensional materials (such as MoS2). The physical characteristics of these devices (such as tunable conductance and charge trapping / releasing dynamics) are utilized to simulate the synaptic weights and membrane potential decay of neurons, thereby performing processing-in-memory computation in the analog domain and minimizing power consumption. For example, the dual-threshold neuronal reservoir utilizes the nonlinear current-voltage characteristics or charge relaxation characteristics of the devices, in conjunction with dual voltage comparators, to generate impulse responses for high and low thresholds respectively, naturally aligning with the multi-timescale processing mechanisms of biological nervous systems.

[0050] As an alternative implementation, this system can also be implemented based on digital logic circuits, offering advantages such as rapid deployment and logic reconfiguration. For example, Field-Programmable Gate Arrays (FPGAs), Digital Signal Processors (DSPs), or Application-Specific Integrated Circuits (ASICs) can be used. In this type of digital implementation, the nonlinear timing encoding module employs a digital multiplier pipeline design to implement the multiplication and concatenation of adjacent elements, with the computational complexity strictly controlled to linear O(n). The dual-threshold neuron reservoir uses parallel comparator logic circuits to record dual-path responses for high and low thresholds. The full-time readout module uses a data bus and register array to concatenate and buffer joint feature vectors. This digital implementation method has a short development cycle and can dynamically adjust the window size and threshold parameters through software configuration to adapt to ECG signals with different sampling rates.

[0051] In the preferred neuromorphic implementation, analog or mixed-signal domain designs are used to implement the element-wise multiplication and vector concatenation operations of adjacent time steps: the input vector is calculated using the conductance modulation characteristics of analog multipliers or memristors. The product of adjacent elements generates adjacent feature vectors. Then, the original vector is processed through a data selector. and Concatenate sequentially to output the enhanced feature vector. The computational complexity is controlled to be linear O(n), ensuring real-time processing (processing latency ≤ 10ms / frame). The parameters of the nonlinear timing coding module are configured to support dynamic adjustment of the input vector length n (corresponding to the sliding window size) (range 64-128 sampling points), adapting to ECG signals with different sampling frequencies (such as 360Hz, 250Hz), without requiring additional hardware circuit modifications, and can be switched only through software configuration.

[0052] The hardware implementation of the dual-threshold neuromorphic reservoir consists of N dual-threshold leak-integrated firing (DT-LIF) neuronal circuits (N can be configured to 100-200, with a default of 139). Random sparse connections are used between neurons (connection probability 20%-30%). Synaptic current calculation and membrane potential updates are preferably achieved through a memristor crossbar array or the synaptic nucleus of a neuromorphic chip. Each neuronal circuit includes an analog integrator or digital accumulator, as well as a dual-channel voltage comparator. The comparator compares the membrane potential with preset low and high thresholds, respectively, using analog or digital methods, and outputs two sets of spike signals in parallel.

[0053] The hardware implementation of the full-time readout module is divided into a feature concatenation unit and a linear classifier unit. The feature concatenation unit concatenates all peak sequences within the time window T in the order of "low threshold sequence and high threshold sequence" via the data bus to generate a joint feature vector. The linear classifier unit uses digital multipliers and accumulators to perform the multiplication and accumulation operation (MAC operation) between the weight matrix (obtained through supervised learning training, optimized by the alternative gradient method during the training phase) and the joint feature vector, outputs the probability values ​​of each category, and then determines the final classification result through a maximum value comparator.

[0054] In summary, this invention constructs an ECG signal classification scheme that combines high accuracy, low power consumption, and minimal hardware resource consumption through the synergistic innovation of adjacent temporal nonlinear coding and dual-threshold neuronal dynamics. It is suitable for deployment in wearable or implantable medical devices to achieve real-time and efficient cardiac health monitoring.

Claims

1. A method for classifying electrocardiogram signals based on a dual-threshold neuromorphic reservoir, characterized in that: Includes the following steps: Step 1: Receive the ECG time-series signal, perform adjacent time-series nonlinear encoding operation on the ECG time-series signal, and generate an enhanced feature vector containing the original signal and second-order interactive features; Step 2: Input the enhanced feature vector into the dual-threshold neuromorphic reservoir, wherein the dual-threshold neuromorphic reservoir is composed of multiple single-membrane potential dual-threshold spiking neurons, which are used to generate a dual-channel impulse response sequence based on the enhanced feature vector; Generating a dual-channel impulse response sequence based on the enhanced feature vector specifically includes: Within a preset time window T, record all first peak output sequences and second peak output sequences generated by each neuron in the reservoir. All first peak output sequences and second peak output sequences recorded within the time window T are concatenated to form a joint feature vector; Step 3: Construct a joint feature vector based on the dual-channel impulse response sequence, and input the joint feature vector into a classifier to obtain the classification result of the electrocardiogram signal; The dual-threshold spiking neuron is configured with a low threshold and high threshold ,in ; The dual-threshold spiking neurons at each time step Based on its membrane potential The state, respectively, generates a first spike output corresponding to the low threshold. and the second spike output corresponding to the high threshold : ； ； The membrane potential State updates follow these rules: ； in, It is the membrane potential decay factor. It is the total synaptic current, which includes the input current and the circulating current within the reservoir. This represents the membrane resistance of a neuron.

2. The ECG signal classification method based on a dual-threshold neuromorphic reservoir according to claim 1, characterized in that: The process of performing adjacent-sequence nonlinear encoding on the ECG timing signal specifically involves: encoding a length of... input vector Calculate the product of elements at adjacent time steps to generate adjacent feature vectors. : ； The input vector With the adjacent feature vector The enhanced feature vector is formed by concatenating the features. : 。 3. The ECG signal classification method based on a dual-threshold neuromorphic reservoir according to claim 1, characterized in that: The classifier is a linear classifier, and its weights are obtained through supervised learning. During training, an alternative gradient method is used for error backpropagation.

4. The ECG signal classification method based on a dual-threshold neuromorphic reservoir according to claim 1, characterized in that: Before inputting the enhanced feature vector into the dual-threshold neuromorphic reservoir, the method further includes a data preprocessing step: Extract signal segments centered on the R peak from the electrocardiogram signal database; The signal segment is divided into multiple sub-segments using a sliding window method to form the input time-series electrocardiogram signal.

5. An electrocardiogram (ECG) signal classification system, employing the ECG signal classification method based on a dual-threshold neuromorphic reservoir as described in any one of claims 1 to 4, characterized in that: The system includes: The nonlinear timing coding module is used to perform adjacent timing nonlinear coding operations on the input ECG timing signal to generate an enhanced feature vector containing second-order dynamic evolution information of the signal. A dual-threshold neuromorphic reservoir, composed of multiple dual-threshold spiking neurons, is used to receive the enhanced feature vector and generate a dual-channel pulse response sequence. Each neuron is configured with a low threshold and a high threshold, and low-energy and high-energy response channels are established in parallel through a threshold-level response mechanism. The low-energy response channel is configured to fire a pulse when the neuronal membrane potential exceeds the low threshold, enabling highly sensitive capture of weak morphological features in electrocardiogram signals and preventing the loss of weak signals such as P waves or T waves. The high-energy response channel is configured to fire a pulse when the neuronal membrane potential exceeds the high threshold, responding to high-intensity significant rhythmic features and preventing homogeneous saturation of neurons when processing high-amplitude QRS complexes. The full-time readout module is used to construct a joint feature vector based on the dual-channel impulse response sequence and output the classification result using a classifier.

6. The system according to claim 5, characterized in that: The nonlinear timing coding module and the dual-threshold neuromorphic reservoir are implemented based on digital logic circuits or neuromorphic computing devices. The digital logic circuit includes a field-programmable gate array, a digital signal processor, or an application-specific integrated circuit, which performs adjacent-sequential nonlinear coding operations and neuron state updates through a pipelined architecture. The neuromorphic computing device includes memristor arrays, phase-change memories, or two-dimensional material floating-gate transistors, and performs in-memory computing using the physical characteristics of the device.

Images