Non-invasive intracellular action potential reconstruction method and system based on transfer learning

Abstract

The application relates to the technical field of biomedical signal processing, in particular to a non-invasive intracellular action potential reconstruction method and system based on transfer learning. The method constructs a paired data set by collecting extracellular and intracellular action potential signals, and carries out signal preprocessing, quality screening and feature standardization, and on the basis, a mapping relationship from extracellular action potential to intracellular action potential is established to form a basic reconstruction model. Further, a small amount of samples corresponding to a target size microelectrode are used for transfer learning optimization, so that the model is adapted to different microelectrode sizes and detection conditions, and finally, the extracellular action potential signal of a cell to be detected is used. The application realizes non-invasive prediction and reconstruction of intracellular action potential, and improves the accuracy and application range of cell electrophysiological detection.

Description

Technical Field

[0001] This application relates to the field of biomedical signal processing technology, and in particular to a non-invasive method and system for reconstructing intracellular action potentials based on transfer learning. Background Technology

[0002] In related technologies, with the development of cell electrophysiology research, drug evaluation, and disease mechanism analysis, intracellular action potential detection has become an important technical means for analyzing cellular functional status. Existing methods for acquiring intracellular action potentials mainly include patch-clamp techniques and electrophysiological detection techniques based on microelectrode arrays (MEAs). Patch-clamp techniques can directly measure transmembrane potential changes with high detection accuracy, but they require a high-resistance seal between the electrode and the cell membrane or membrane puncture, resulting in complex operation, low throughput, and difficulty in long-term stable recording. While microelectrode array techniques can achieve multi-channel, high-throughput electrical activity acquisition, traditional methods mainly acquire extracellular action potential signals, making it difficult to directly obtain key parameters reflecting the characteristics of intracellular electrical activity.

[0003] However, existing methods for detecting intracellular action potentials still have limitations in long-term monitoring and cross-platform applications. While intracellular recording methods based on electroporation or micro / nanostructured electrodes can acquire intracellular action potential signals for short periods, the self-repairing properties of the cell membrane make it difficult to maintain the intracellular coupling state between the electrode and the cell over a long period, resulting in limited signal duration. Furthermore, repeated electroporation may affect the normal physiological state of the cell. In recent years, AI-based action potential reconstruction methods have learned the mapping relationship between extracellular and intracellular action potential signals, enabling the prediction of intracellular action potential waveforms using stably acquired extracellular signals. However, existing methods typically train models for microelectrodes of fixed sizes. When the electrode size changes, the coupling characteristics between the cell and the electrode, the signal amplitude, and the noise distribution all change, leading to differences in extracellular action potential signals acquired by electrodes of different sizes. This makes it difficult to directly transfer and apply existing models.

[0004] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0005] The main objective of this application is to propose a non-invasive intracellular action potential reconstruction method and system based on transfer learning, which can improve the accuracy, adaptability and convenience of action potential reconstruction under different microelectrode conditions while reducing the invasiveness of detection.

[0006] To achieve the above objectives, one aspect of this application proposes a non-invasive intracellular action potential reconstruction method based on transfer learning, the method comprising the following steps: Acquire extracellular action potential signals and corresponding intracellular action potential signals, and construct an action potential pairing dataset; The paired action potential dataset is preprocessed and quality screened to obtain standardized action potential training data. Based on the standardized action potential training data, a mapping model from extracellular action potential to intracellular action potential is constructed to obtain the basic action potential reconstruction model. Based on a small amount of action potential pairing data corresponding to the target size microelectrode, the basic action potential reconstruction model is optimized by transfer learning to obtain an action potential reconstruction model adapted to the target size. Based on the extracellular action potential signal corresponding to the cell to be detected, and input into the action potential reconstruction model adapted to the target size, the signal is reconstructed to obtain the corresponding intracellular action potential reconstruction result.

[0007] In some embodiments, the step of performing signal preprocessing and quality screening on the action potential paired dataset to obtain standardized action potential training data includes: The paired action potential dataset is subjected to periodic segmentation to obtain a single-cycle action potential signal; Based on the feature points corresponding to the single-cycle action potential signal, time alignment and fixed-length truncation are performed to obtain a uniform-length action potential sequence. The amplitude of the action potential sequence is normalized to obtain normalized action potential data; The normalized action potential data is classified based on a signal quality classification model, and action potential data that meets preset quality conditions are selected to obtain standardized action potential training data.

[0008] In some embodiments, the step of classifying the normalized action potential data based on a signal quality classification model and filtering action potential data that meet preset quality conditions to obtain standardized action potential training data includes: A signal quality classification model is constructed based on a one-dimensional convolutional neural network, and convolutional feature extraction is performed on the normalized action potential data to obtain action potential quality features. Based on the action potential quality characteristics, the normalized action potential data is classified into quality levels to generate signal quality classification results. Based on preset quality screening conditions, target action potential pairing data are screened according to the signal quality classification results. The target action potential pairing data is converted into a training format to obtain standardized action potential training data.

[0009] In some embodiments, constructing a mapping model from extracellular action potentials to intracellular action potentials based on the standardized action potential training data to obtain a basic action potential reconstruction model includes: Based on a convolutional neural network, local feature extraction is performed on the extracellular action potential signal in the standardized action potential training data to obtain local waveform features of the action potential. Based on a long short-term memory network, the local waveform features of the action potential are modeled in a temporal correlation to extract the temporal change features corresponding to the extracellular action potential signal. Based on the aforementioned temporal variation characteristics, a mapping relationship is established with the corresponding intracellular action potential signal to generate intracellular action potential prediction results; Based on the error between the intracellular action potential prediction results and the intracellular action potential signals in the standardized action potential training data, the mapping model parameters are iteratively optimized to obtain the basic action potential reconstruction model.

[0010] In some embodiments, the step of extracting local features from the extracellular action potential signals in the standardized action potential training data based on a convolutional neural network to obtain local waveform features of the action potential includes: The extracellular action potential signals in the standardized action potential training data are serialized to generate action potential input sequences; Convolutional feature extraction is performed on the action potential input sequence based on a one-dimensional convolutional layer to obtain the local waveform response features of the action potential. The local waveform response features are enhanced based on a nonlinear activation function to extract waveform change features from the extracellular action potential signal. The waveform change features are reduced in dimensionality using a pooling layer to obtain key waveform features. The key waveform features are encoded based on multi-layer convolutional feature fusion to obtain local waveform features of action potential.

[0011] In some embodiments, the step of performing time-series correlation modeling on the local waveform features of the action potential based on a long short-term memory network and extracting the time-series change features corresponding to the extracellular action potential signal includes: The local waveform features of the action potential are temporally encoded based on a long short-term memory network to generate temporal hidden features of the action potential. The state of the action potential temporal latent features is updated based on the gating mechanism of the long short-term memory network, and the long-term dependency features corresponding to the periodic changes of the action potential are extracted. Based on the long-term dependent features, the dynamic evolution features corresponding to the extracellular action potential signal are extracted to obtain the action potential temporal change features; The temporal variation features of the action potential are fused to obtain the temporal variation features corresponding to the extracellular action potential signal.

[0012] In some embodiments, the step of performing transfer learning optimization on the basic action potential reconstruction model based on a small amount of action potential pairing data corresponding to the target-size microelectrode to obtain a target-size-adapted action potential reconstruction model includes: Based on the basic action potential reconstruction model trained with source domain action potential paired data, the source domain feature parameters in the basic action potential reconstruction model are transferred to the target domain to obtain the transfer initialization model. Based on a small amount of action potential pairing data corresponding to the target domain, the transfer initialization model is adapted to target domain features to obtain the target domain action potential feature mapping relationship. Based on the target domain action potential feature mapping relationship and the reconstruction error between the corresponding intracellular action potential signals, the migration initialization model is optimized to obtain the target domain adaptation model. Based on the target domain adaptation model, the action potential data corresponding to the target size microelectrode are reconstructed and verified to obtain the action potential reconstruction model adapted to the target size.

[0013] In some embodiments, the step of adapting the transfer initialization model to target domain features based on a small amount of action potential pairing data corresponding to the target domain, to obtain the target domain action potential feature mapping relationship, includes: Feature extraction is performed on source domain action potential paired data and target domain action potential paired data to obtain source domain action potential features and target domain action potential features. Based on the source domain action potential features and the target domain action potential features, feature distribution alignment is performed to determine the feature offset between the source domain and the target domain. Based on the feature offset, the feature extraction parameters in the migration initialization model are adjusted to match the source domain feature space with the target domain feature space. Based on the adjusted feature extraction parameters, feature mapping is performed on the target domain action potential data to obtain the target domain action potential feature mapping relationship.

[0014] In some embodiments, the step of reconstructing the signal based on the extracellular action potential signal corresponding to the cell to be detected and inputting the target size-adapted action potential reconstruction model to obtain the corresponding intracellular action potential reconstruction result includes: The extracellular action potential signal corresponding to the cell to be detected is standardized to obtain action potential data to be reconstructed. The action potential data to be reconstructed is input into the action potential reconstruction model adapted to the target size to obtain the intracellular action potential prediction signal. The intracellular action potential prediction signal is post-processed and reconstructed to obtain the intracellular action potential reconstruction waveform. Error analysis, correlation analysis, and stability analysis are performed based on the reconstructed intracellular action potential waveform, and the performance of the target-size-adapted action potential reconstruction model is evaluated to generate intracellular action potential reconstruction results.

[0015] To achieve the above objectives, another aspect of this application proposes a non-invasive intracellular action potential reconstruction system based on transfer learning, the system comprising: The data acquisition module is used to acquire extracellular action potential signals and corresponding intracellular action potential signals to construct a paired action potential dataset. The data processing module is used to perform signal preprocessing and quality screening on the action potential paired dataset to obtain standardized action potential training data. The model building module is used to construct a mapping model from extracellular action potentials to intracellular action potentials based on the standardized action potential training data, thereby obtaining a basic action potential reconstruction model. The transfer optimization module is used to perform transfer learning optimization on the basic action potential reconstruction model based on a small amount of action potential pairing data corresponding to the target size microelectrode, so as to obtain an action potential reconstruction model adapted to the target size. The signal reconstruction module is used to reconstruct the signal based on the extracellular action potential signal corresponding to the cell to be detected and input the action potential reconstruction model adapted to the target size, so as to obtain the corresponding intracellular action potential reconstruction result.

[0016] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0017] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0018] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0019] The embodiments of this application include at least the following beneficial effects: This application provides a non-invasive intracellular action potential reconstruction method and system based on transfer learning. This scheme constructs a paired dataset between extracellular action potential signals and intracellular action potential signals, and improves the reliability of training data by utilizing signal preprocessing and quality screening mechanisms, providing an effective data foundation for the subsequent construction of action potential mapping models. By establishing a mapping relationship between extracellular action potentials and intracellular action potentials, it achieves the reconstruction of intracellular action potentials without invading the cell interior, avoiding interference with cell state caused by traditional invasive detection methods such as patch clamps, and improving the safety and applicability of cell electrophysiological detection. By performing transfer learning optimization on the basic action potential reconstruction model based on a small amount of paired data corresponding to microelectrodes of the target size, the model can adapt to the signal feature differences generated by microelectrodes of different sizes, reducing the amount of data and time cost required to retrain the model for different detection conditions, and improving the model's generalization ability. By using the action potential reconstruction model adapted to the target size to reconstruct the extracellular action potential signal of the cell to be detected, the corresponding intracellular action potential information can be obtained, improving the convenience and detection efficiency of cell electrophysiological feature acquisition, and providing technical support for large-scale, long-term, non-invasive cell function analysis. In summary, this invention can improve the accuracy and adaptability of action potential reconstruction under different microelectrode conditions while reducing the invasiveness of detection. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of a non-invasive intracellular action potential reconstruction method based on transfer learning provided in an embodiment of this application; Figure 2 This is a comparison of the reconstruction errors of the basic action potential reconstruction model and the transfer learning optimization model under different microelectrode sizes provided in the embodiments of this application; Figure 3 This is a graph showing the comparison of the determination coefficients between the basic action potential reconstruction model and the transfer learning optimization model under different microelectrode sizes, provided in the embodiments of this application. Figure 4 These are signal reconstruction results of the non-invasive intracellular action potential reconstruction method based on transfer learning provided in this application under different microelectrode sizes. Figure 5 This is a schematic diagram of a module of a non-invasive intracellular action potential reconstruction system based on transfer learning provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0022] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0023] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0025] Before providing a detailed description of the embodiments of this application, some of the nouns and terms used in the embodiments of this application will be explained first. The nouns and terms used in the embodiments of this application shall be interpreted as follows: Action potential is an electrical signal formed by a rapid and transient change in the cell membrane potential after an excitable cell is stimulated. It is used to reflect the activity state of nerve cells, muscle cells, and other electrophysiological cells. In this application, action potential is used to characterize changes in intracellular and extracellular electrical activity and serves as the data basis for the reconstruction of intracellular action potential. Extracellular action potential (eAP) refers to the signal of local electric field change caused by cellular electrical activity, which is acquired by microelectrodes or microelectrode arrays located outside the cell. It is used to reflect the electrophysiological activity characteristics generated by cell populations or individual cells during excitation. In this application, the extracellular action potential signal is used as input data, and the corresponding intracellular action potential signal is predicted by the action potential reconstruction model optimized by transfer learning. Intracellular action potential (iAP) refers to action potential signals obtained by directly detecting changes in intracellular membrane potential, which can characterize the movement of ions across the cell membrane and the state of cellular electrophysiological activity. In this application, the intracellular action potential signal is used as the target reconstruction signal, and together with the corresponding extracellular action potential signal, they form an action potential pairing dataset, which is used to establish the mapping relationship between extracellular action potential signals and intracellular action potential signals. Non-invasive detection refers to a detection method that acquires cellular function information through external sensing devices without damaging the cell structure or directly entering the cell interior. In this application, intracellular action potentials are predicted by extracellular action potential signals, thereby reducing the invasiveness of the detection. A microelectrode array (MEA) is an electrophysiological detection device formed by arranging multiple microscale electrodes in a certain spatial structure, used to simultaneously acquire electrical activity signals from multiple cells or regions; in this application, it is used to acquire extracellular action potential signals. Transfer learning refers to the machine learning method of transferring learned model knowledge to new tasks, data distributions or application scenarios and adapting and optimizing it with a small amount of target domain data. In this application, a basic reconstruction model is trained using source domain action potential data and optimized based on a small amount of paired data corresponding to the target size microelectrode, so that the model is applicable to different microelectrode detection conditions. The source domain refers to a data space containing a large number of labeled training samples, used to train the basic model and learn general feature information; in this application, the source domain corresponds to the action potential pairing data used in the training phase of the basic action potential reconstruction model. The target domain refers to a new data space that differs from the source domain in data distribution but needs to be predicted or reconstructed using existing model knowledge; in this application, the target domain corresponds to action potential data acquired under different sized microelectrode acquisition conditions. A convolutional neural network (CNN) is a deep learning network structure that extracts features based on convolution operations. It can automatically learn local spatial features and complex pattern information from input data. In this application, a convolutional neural network is used to extract local waveform features from extracellular action potential signals. One-dimensional convolutional neural network (1D-CNN) refers to a convolutional neural network structure designed for one-dimensional sequence data, which extracts local variation features in the sequence through one-dimensional convolution operations; in this application, it is used for quality classification and feature extraction of action potential sequences. Long Short-Term Memory (LSTM) is an improved recurrent neural network that uses a gating structure to store and update long-term dependent information, effectively processing temporally correlated sequential data. In this application, it is used to perform temporal correlation modeling of local waveform features of action potentials and extract the dynamic change features of action potential signals. Signal reconstruction refers to the process of recovering the characteristics of a target signal based on existing observed signals through model calculations or algorithm processing; in this application, the corresponding intracellular action potential reconstruction waveform is predicted and generated based on the extracellular action potential signal.

[0026] This application provides a non-invasive intracellular action potential reconstruction method and system based on transfer learning. This approach constructs a paired dataset between extracellular and intracellular action potential signals and utilizes signal preprocessing and quality screening mechanisms to improve the reliability of training data, providing an effective data foundation for the subsequent construction of an action potential mapping model. By establishing a mapping relationship between extracellular and intracellular action potentials, intracellular action potentials can be reconstructed without invading the cell, avoiding interference with cell state caused by traditional invasive detection methods such as patch clamps, thus improving the safety and applicability of cell electrophysiological detection. By optimizing the basic action potential reconstruction model through transfer learning based on a small amount of paired data corresponding to microelectrodes of the target size, the model can adapt to the signal feature differences generated by microelectrodes of different sizes, reducing the amount of data and time required to retrain the model for different detection conditions and improving the model's generalization ability. By using the action potential reconstruction model adapted to the target size to reconstruct the extracellular action potential signal of the cell to be detected, the corresponding intracellular action potential information can be obtained, improving the convenience and detection efficiency of cell electrophysiological feature acquisition and providing technical support for large-scale, long-term, non-invasive cell function analysis. In summary, this invention can improve the accuracy and adaptability of action potential reconstruction under different microelectrode conditions while reducing the invasiveness of detection.

[0027] This application provides a non-invasive intracellular action potential reconstruction method based on transfer learning, relating to the field of biomedical signal processing technology. This method can be applied to terminals, servers, or software running on either. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the non-invasive intracellular action potential reconstruction method based on transfer learning, but is not limited to the above forms.

[0028] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0029] Figure 1 This is an optional flowchart of a non-invasive intracellular action potential reconstruction method based on transfer learning provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S1 to S5: S1: Obtain extracellular action potential signals and corresponding intracellular action potential signals, and construct a paired action potential dataset; In this embodiment, a multi-size microelectrode array chip is first fabricated to simultaneously acquire cell electrophysiological signals under different electrode sizes. This microelectrode array chip integrates five recording electrodes of 20μm, 50μm, 100μm, 200μm, and 400μm sizes on the same glass substrate. Each electrode size is configured with multiple recording channels and is fabricated using a unified photolithography and metal lift-off process to reduce the impact of material properties and processing differences between electrodes of different sizes on the acquired signals. A cell culture region is set on the chip surface, and a reference electrode is configured thereon, thus forming a complete cell culture and electrophysiological detection platform.

[0030] Furthermore, the cultured cardiomyocytes were seeded onto the surface of a microelectrode array, and action potentials were acquired after the cells formed a stable, synchronous pulsating network. Specifically, on the third day of culture, extracellular action potential signals were continuously recorded using a multi-channel electrophysiological acquisition platform, while a short-duration electroporation pulse was applied to induce transient nanopore structures in the cell membrane, thereby acquiring corresponding intracellular action potential signals within the same cells and time window. The electroporation pulse parameters were set to an amplitude of 3V, a frequency of 20Hz, a pulse width of 200μs, and a duration of 1s, ensuring simultaneous acquisition of extracellular and intracellular signals while maintaining cell survival and rhythm stability.

[0031] Subsequently, the synchronously acquired extracellular action potential (APT) signals were matched with intracellular APT signals according to time series to establish a one-to-one correspondence of APT pairing samples. Data acquired from 20 μm electrodes was used as the source domain data to construct the basic training dataset; for electrodes of different target sizes (50–400 μm), a small amount of extracellular-intracellular pairing data was collected for each size to form the target domain adaptation dataset. Through this method, an APT pairing database encompassing different electrode sizes, culture states, and electrophysiological modulation conditions was constructed, providing a data foundation for subsequent model training and transfer learning.

[0032] S2: Perform signal preprocessing and quality screening on the action potential paired dataset to obtain standardized action potential training data; The process involves signal preprocessing and quality screening of the paired action potential dataset to obtain standardized action potential training data, including: S21: Perform periodic segmentation on the paired action potential dataset to obtain a single-cycle action potential signal; S22: Based on the feature points corresponding to the single-cycle action potential signal, perform time alignment and fixed-length truncation to obtain a uniform-length action potential sequence. S23: Perform amplitude normalization on the action potential sequence to obtain normalized action potential data; S24: Based on the signal quality classification model, the normalized action potential data is classified according to quality, and action potential data that meets the preset quality conditions are selected to obtain standardized action potential training data.

[0033] In this embodiment, the continuous extracellular action potential (APD) signals and their corresponding intracellular APD signals in the action potential pairing dataset are periodically segmented. The start and end positions of each APD cycle are determined based on the periodic variation characteristics of the APD waveform, dividing the continuous electrophysiological recording signal into multiple single-cycle APD signals. Since the cell pulsation rhythm remains stable during electroporation, the extracellular and intracellular APD signals have a temporal correspondence. Therefore, the two types of signals are synchronously segmented to establish a one-to-one correspondence between single-cycle extracellular and intracellular APD signals.

[0034] Time alignment is performed based on feature points in single-cycle action potential signals. Specifically, the trough positions of the cycle in extracellular action potential signals are extracted as reference feature points, and the corresponding peak positions in intracellular action potential signals are extracted as reference feature points. Time offsets from different signal acquisition processes are eliminated through feature point matching. Subsequently, the action potential cycle is truncated to a fixed length based on the reference feature points, ensuring that each action potential signal contains the complete depolarization, peak change, and repolarization processes, and is converted into a uniform-length action potential sequence. Simultaneously, amplitude normalization is performed on the uniform-length action potential sequence to reduce amplitude differences caused by different cell states, electrode sizes, and acquisition batches, thereby improving the consistency between data from different sources.

[0035] The normalized action potential data is input into a signal quality classification model for automated quality assessment. In this embodiment, a signal quality classification model based on a one-dimensional convolutional neural network is used to comprehensively analyze the signal-to-noise level, waveform integrity, and abnormal distortion in the action potential waveform, and classify the signals according to preset quality levels. Based on the classification results, action potential data that meets the quality conditions are selected, and only high-quality extracellular and intracellular action potential signals are retained as effective training samples. Data with significant noise interference, abnormal waveforms, or incomplete records are removed, ultimately obtaining standardized action potential training data, providing a reliable data foundation for subsequent training of the extracellular action potential to intracellular action potential mapping model.

[0036] Specifically, the normalized action potential data is classified according to the signal quality classification model, and action potential data that meets the preset quality conditions are selected to obtain standardized action potential training data, including: A signal quality classification model is constructed based on a one-dimensional convolutional neural network. Convolutional feature extraction is performed on normalized action potential data to obtain action potential quality features. Based on the action potential quality characteristics, the normalized action potential data is classified into quality levels to generate signal quality classification results. Based on preset quality screening conditions, target action potential pairing data are screened according to signal quality classification results; The target action potential paired data is converted into a training format to obtain standardized action potential training data.

[0037] In this embodiment, a signal quality classification model based on a one-dimensional convolutional neural network is first constructed to extract quality features from the normalized action potential data. Specifically, the extracellular action potential signal and the corresponding intracellular action potential signal are input into the signal quality classification model. The local morphological features and temporal variation features in the action potential waveform are extracted through convolutional layers, including features such as peak variation, depolarization process, repolarization process, and signal stability, to generate action potential quality features that characterize the signal quality state.

[0038] Based on the action potential quality characteristics, normalized action potential data is classified into quality levels to generate signal quality classification results. Specifically, the quality level of each group of action potential data is determined according to the waveform integrity, noise level, and feature stability of the action potential signal. Low-quality, abnormal waveform, and incomplete records are filtered out according to preset quality screening conditions, and only target action potential paired data that meet the training requirements are retained. In this embodiment, a four-level quality classification method (A to D) is used, retaining extracellular action potential A / B level data and intracellular action potential A level data as subsequent model training samples.

[0039] The selected target action potential pairing data undergoes training format conversion processing. Specifically, extracellular action potential signals are used as model input, and corresponding intracellular action potential signals are used as training labels. The data length, arrangement, and labeling information of the samples are standardized to form standardized action potential training data. This processing improves the consistency and reliability of the training data, providing a data foundation for the subsequent construction of a mapping model from extracellular to intracellular action potentials.

[0040] Specifically, this embodiment constructs a signal classification model based on a one-dimensional convolutional neural network (1D-CNN) to automate the quality screening of long-sequence electrophysiological signals, and performs quality grading (A–D) on extracellular and intracellular action potentials. This model aims to extract robust representations from high-dimensional time-series signals and distinguish between low signal-to-noise ratio, waveform distortion, and incomplete recordings. Before inputting the signal into the model, the original signal undergoes uniform preprocessing. The extracellular action potential signal (eAP) length is normalized to 2000 sampling points, and the intracellular signal is normalized to 1000 sampling points. Subsequently, amplitude normalization is performed on each signal segment: ; in, This represents the normalized action potential signal value; and These represent the minimum and maximum amplitude values ​​of each signal segment, respectively. This operation effectively eliminates amplitude drift caused by cross-cell and cross-electrode recordings, allowing the model to focus on waveform morphology features. The classification model employs a two-stage convolutional feature extraction architecture. The first stage consists of shallow convolutional modules (10 kernels, size=5) to capture local waveform structure features of action potentials; the second stage expands to 200 convolutional kernels (kernel size=7) to learn higher-order temporal dependencies and overall discharge patterns. All convolutional layers use zero-padding to maintain temporal consistency and ReLU to introduce nonlinear expressive power. Each convolutional layer is followed by max pooling (stride=2) for progressive downsampling, and its output is defined as the local window maximum response. ; In the formula, This represents the result obtained after max pooling. Each output feature value is used to represent the maximum response feature of the local region of the action potential within the corresponding time window; This represents the input feature map data; m and n represent the local index parameters within the max pooling window, used to limit the positions of the feature values ​​to be compared in the current pooling region.

[0041] After two pooling operations, the feature map significantly reduces computational complexity while preserving key peak information. To avoid information loss caused by local pooling, a global average pooling (GAP) layer is introduced after convolutional feature extraction to statistically compress the entire time dimension. ; Where z represents the global feature representation vector obtained after global average pooling; L is the feature map length. Let t be the time step and t be the corresponding feature vector. This step maps the dynamic temporal signal into a global representation vector, thereby enhancing the model's robustness to the overall action potential morphology. The final features are mapped to four output classes (A–D) through a fully connected layer, and a probability distribution is generated using the Softmax function.

[0042] In the formula, Indicates input action potential signal Classified as the first The predicted probability of a signal quality class is used to represent the confidence level of the current signal belonging to a certain quality class, corresponding to the probability results of A, B, C, and D levels output by the signal quality classification model. This indicates the signal quality classification result; Indicates the corresponding number The classification weight vector of the signal quality category is used to measure the matching relationship between the feature vector z and the category; Indicates the first The classification weight parameters corresponding to the quality category.

[0043] Model training employed a stochastic gradient descent (SGD) optimizer for parameter updates, with a learning rate of 0.1 and cross-entropy as the loss function. During training, the model exhibited stable convergence on the validation set and demonstrated good discriminative ability for signals with low SNR or waveform distortion. After training, only extracellular signals classified as A / B by the CNN and intracellular signals classified as A were retained for subsequent eAP–iAP pairing modeling. This automated selection mechanism reduced reliance on manual annotation while improving the consistency and physiological reliability of the training data, thus providing a high-purity input foundation for cross-size transfer learning models.

[0044] S3: Construct a mapping model from extracellular action potentials to intracellular action potentials based on standardized action potential training data to obtain a basic action potential reconstruction model; Among them, a mapping model from extracellular action potentials to intracellular action potentials is constructed based on standardized action potential training data to obtain a basic action potential reconstruction model, including: S31: Based on a convolutional neural network, local feature extraction is performed on the extracellular action potential signal in the standardized action potential training data to obtain the local waveform features of the action potential. S32: Based on the long short-term memory network, the temporal correlation model of the local waveform features of action potential is performed to extract the temporal change features corresponding to the extracellular action potential signal; S33: Based on the temporal variation characteristics, establish the mapping relationship between the corresponding intracellular action potential signal and generate intracellular action potential prediction results; S34: Based on the error between the intracellular action potential prediction results and the intracellular action potential signals in the standardized action potential training data, the mapping model parameters are iteratively optimized to obtain the basic action potential reconstruction model.

[0045] In this embodiment, a mapping model from extracellular action potentials to intracellular action potentials is constructed based on standardized action potential training data. Specifically, the extracellular action potential signal from the standardized training data is used as the model input. A convolutional neural network is used to extract local features from the action potential waveform, obtaining local waveform features of the action potential that include information such as depolarization rising edge, peak change, plateau phase characteristics, and repolarization trend. This enables the model to learn key morphological features in the extracellular action potential signal. A CNN-LSTM structure is used, where the convolutional neural network is used to extract local waveform features of the action potential.

[0046] Local waveform features of action potentials are input into a long short-term memory network for temporal correlation modeling. By learning the continuous variation pattern of action potential signals over time, the temporal variation features corresponding to extracellular action potential signals are extracted. Based on these temporal variation features, a mapping relationship is established between them and intracellular action potential signals. The corresponding intracellular action potential waveform is then predicted by the model, enabling it to learn the nonlinear correspondence between extracellular and intracellular action potentials and achieve the reconstruction of extracellular signals into intracellular signals.

[0047] The predicted intracellular action potentials (IPPs) were compared with the actual IPP signals from the standardized IPP training data, and the model parameters were iteratively optimized based on the prediction error. During training, the parameters of the convolutional network, temporal network, and mapping layer were continuously adjusted to gradually approximate the actual signals in terms of overall morphology, depolarization process, repolarization process, and IPP duration, ultimately yielding a basic IPP reconstruction model. This basic model was trained on 20μm electrode data to learn the mapping relationship between extracellular and intracellular IPPs and to provide an initial model for subsequent transfer learning optimization.

[0048] Specifically, to achieve high-fidelity reconstruction from extracellular signals to iAP, this embodiment constructs a basic action potential reconstruction model based on convolutional neural networks and long short-term memory networks (CNN-LSTM), and establishes a baseline model (CrossMEA) using 20μm electrode data as the source domain training set. 20 This model is used as a performance baseline for subsequent cross-size transfer learning.

[0049] During the data input phase, all eAP–iAP paired signals undergo uniform preprocessing. The original 10,000-point waveform is resampled to 6,000 points to reduce computational complexity while preserving key action potential dynamics information.

[0050] The model employs a hierarchical temporal feature extraction structure, in which the CNN module is used to extract local waveform morphological features (such as depolarization rising edge, plateau details, and repolarization slope changes), while the LSTM module is used to model the long-range temporal dependence and periodic dynamic evolution of action potentials.

[0051] Specifically, based on a convolutional neural network, local feature extraction is performed on the extracellular action potential signal in the standardized action potential training data to obtain local waveform features of the action potential, including: The extracellular action potential signals in the standardized action potential training data are serialized to generate action potential input sequences. Convolutional feature extraction is performed on the action potential input sequence based on a one-dimensional convolutional layer to obtain the local waveform response features of the action potential. Feature enhancement of local waveform response features is performed based on nonlinear activation functions to extract waveform change features from extracellular action potential signals; The key waveform features are obtained by dimensionality reduction of waveform change features based on the pooling layer. Feature encoding of key waveform features is performed based on multi-layer convolution feature fusion to obtain local waveform features of action potential.

[0052] In this embodiment, the extracellular action potential (ET) signals in the standardized action potential training data are serialized, converting the preprocessed ET waveforms into a continuous ET input sequence according to the time sampling order. Specifically, each group of ET signals is arranged with a uniform length, and waveform change information within the action potential cycle is preserved, enabling the input sequence to fully characterize local dynamic features such as depolarization initiation, peak changes, and repolarization processes. Subsequently, the ET input sequence is input into a one-dimensional convolutional network, and feature extraction of local waveform changes in the time series is performed through a multi-layer convolutional structure to obtain the local waveform response features of the ET under different time windows. A CNN-LSTM structure is adopted, in which the convolutional neural network is used to extract local waveform features in the ET, including information such as the depolarization rising edge, plateau details, and repolarization slope changes.

[0053] Nonlinear feature enhancement is applied to the local waveform response features to improve the model's ability to represent complex action potential morphological changes. Specifically, nonlinear activation processing enhances the convolutional network's ability to identify different amplitude variations, waveform transitions, and subtle structural differences, enabling the convolutional output to more accurately reflect key change regions in the extracellular action potential signal. Subsequently, pooling layers are used to reduce the dimensionality of the enhanced waveform features, reducing the size of the feature data while preserving the main response information in the action potential waveform, reducing the computational complexity of subsequent models, and improving the model's robustness to local signal changes.

[0054] Based on a multi-layer convolutional feature fusion approach, the key waveform features after pooling are further encoded. Specifically, shallow local waveform information and deep semantic features are gradually fused through different levels of convolutional structures, enabling the model to simultaneously acquire short-term temporal scale variation features and overall waveform structure features in the action potential signal. After multi-layer convolutional feature extraction and fusion, local waveform features of action potentials that can characterize the local morphological changes of extracellular action potentials are generated. These local waveform features are then input into a subsequent long short-term memory network for temporal correlation modeling to further learn the mapping relationship between extracellular and intracellular action potentials.

[0055] Specifically, the Convolutional Neural Network (CNN) consists of multiple layers of one-dimensional convolution and pooling operations. The convolution operation is defined as: ; in, This represents the output of the convolutional layer. Indicates the input signal. This corresponds to the nonlinear activation function (ReLU). represents the convolution kernel weights, and b is the bias. The pooling layer uses max pooling to achieve temporal compression; this structure reduces sequence redundancy while preserving key peak information. .

[0056] Specifically, based on long short-term memory networks, temporal correlation modeling is performed on the local waveform features of action potentials to extract the temporal variation features corresponding to extracellular action potential signals, including: Based on the Long Short-Term Memory Network, the local waveform features of the action potential are temporally encoded to generate the temporal hidden features of the action potential. The state of the action potential temporal hidden features is updated based on the gating mechanism of the long short-term memory network, and the long-term dependency features corresponding to the periodic changes of the action potential are extracted. Based on long-term dependent features, dynamic evolution features corresponding to extracellular action potential signals are extracted to obtain action potential temporal change features; Feature fusion was performed on the temporal variation characteristics of action potentials to obtain the temporal variation characteristics corresponding to extracellular action potential signals.

[0057] In this embodiment, the local waveform features of action potentials are temporally encoded based on a Long Short-Term Memory (LSTM) network. Specifically, the local waveform features of action potentials extracted by a convolutional neural network are input into the LSTM network in chronological order of the action potential cycle. A recurrent structure encodes the feature information in continuous time segments, enabling the network to learn the correlation between different time stages and generate temporally implicit features containing dynamic changes in action potentials. This process transforms the original local waveform features into a high-dimensional feature representation with temporal dependencies, providing a foundation for subsequent extraction of long-term variation patterns. A CNN-LSTM structure is adopted, where the LSTM network is used to perform temporal modeling of the action potential features extracted by convolution, learning the dynamic characteristics of the action potential signal changing over time.

[0058] A gating mechanism based on Long Short-Term Memory (LSTM) networks is used to update the state of the temporal latent features of action potentials (APs) and extract long-term dependent features corresponding to the periodic changes in action potentials. Specifically, through input control, information retention, and output control processes within the LTM network, action potential features at different time stages are filtered and transmitted, enabling the network to retain important information related to the periodic changes in action potentials while reducing the impact of short-term noise and local abnormal fluctuations on the extraction of temporal features. By performing correlation analysis on feature changes within multiple action potential cycles, dependent features reflecting the long-term changes in extracellular action potentials are extracted, thereby enhancing the model's ability to express complex electrophysiological dynamic processes.

[0059] This study extracts dynamic evolution features corresponding to extracellular action potential (APD) signals based on long-term dependency characteristics and fuses these temporal variation features. Specifically, by analyzing the changes in long-term dependency characteristics at different time stages, dynamic evolution information corresponding to depolarization, peak changes, and repolarization processes in the extracellular APD signals is extracted. Temporal features at different time scales are then fused and encoded to form temporal variation features that comprehensively reflect the periodic variation patterns of APDs. The resulting extracellular APD temporal variation features are used to establish a mapping relationship with intracellular APD signals, enabling the prediction and reconstruction of intracellular APDs based on extracellular signals.

[0060] Specifically, in this embodiment, the feature sequences extracted by the CNN are input into the LSTM module for time-dependent modeling. The LSTM controls the information flow through gating mechanisms, including forget gates, input gates, and output gates, thereby achieving selective memorization and updating of long-term electrophysiological dynamics. Its core state update is as follows: ; ; In the formula, This indicates the LSTM cell state at the current time step t; This indicates the LSTM cell state corresponding to the previous time step t; The output of the forget gate is used to control the cell state at the previous time step. The proportion of information retained in the middle; The output of the input gate controls the degree to which the current input features affect the cell state update. It represents the candidate cell state generated at the current time step, and is used to represent the candidate memory information calculated based on the current input action potential characteristics; This represents the hidden state of the LSTM output at the current time step, used to represent the implicit features of the action potential timing after gating and updating. This represents the output of the output gate, used to control the cell state. Which information can be output as the current hidden state? This represents the result of a nonlinear mapping of the current cell state.

[0061] This invention employs a 128-hidden-unit LSTM layer to capture the cross-cycle dependence and morphological evolution characteristics of action potentials. To improve the physiological consistency of iAP reconstruction, an electrophysiological constraint term is introduced on top of the standard mean squared error (MSE) to construct a composite loss function. The MSE is defined as: ; in, This represents the measured intracellular potential. This represents the model's predicted value, where N is the total number of sampling points. This term ensures that the model minimizes amplitude and morphological differences across the entire waveform scale, thereby improving overall fitting accuracy.

[0062] To further enhance the constraints on electrophysiological significance, three indicators related to the duration of electrophysiological activity were introduced into the loss function. The first is the 50% repolarization duration of the action potential (APD). 50 This parameter reflects a key characteristic of complex polarization dynamics, and its calculation method is as follows: ; in, The time point when the membrane potential drops to half of its maximum amplitude. This indicates the depolarization start time point. During training, the model predicts and compares the measured APD using the following methods. 50 Differences: ; In the formula, The duration of 50% repolarization of the action potential (APD) 50 The error loss is used to measure the deviation between the model-predicted action potential and the actual action potential in terms of the repolarization dynamics characteristics. The APD corresponding to the actual intracellular action potential signal 50 The time value, i.e. the repolarization duration corresponding to the measured action potential decreasing from the depolarization peak to half of the maximum amplitude, is used as a real electrophysiological reference value during model training. The APD predicted by the model is represented. 50 The time value, i.e., the predicted repolarization duration calculated based on the reconstructed intracellular action potential waveform, is used to compare with the actual APD. 50 The models were compared to evaluate their ability to recover from action potential dynamics.

[0063] In addition, to maintain the temporal consistency of each stage of the action potential, time error constraints are introduced for the depolarization and repolarization stages, defined as follows: ; ; in, and This represents the time error loss during the depolarization stage and the time error loss during the repolarization stage. and These represent the start times of depolarization and repolarization in the measured signal, respectively. and The model predicts the corresponding time points. By simultaneously optimizing these key temporal features, the model can better reconstruct the dynamic characteristics of action potentials while maintaining overall waveform accuracy. The final total loss function... Defined as: ; in, , , and These are weighting coefficients used to balance the importance of numerical errors and physiological constraints. This multi-level loss structure not only ensures the waveform accuracy of the reconstructed signal but also guarantees the consistency of the prediction results across physiological time scales.

[0064] The CNN-LSTM model was trained on 20μm electrode data and served solely as the source domain baseline model for CrossMEA transfer learning. After training, the model was able to stably reconstruct the complete action potential waveform, exhibiting high consistency in the depolarization rising edge, plateau phase, and repolarization phase. This indicates that the architecture can effectively learn the nonlinear mapping relationship between eAP and iAP and provides a stable initialization foundation for cross-size generalization models.

[0065] S4: Based on a small amount of action potential pairing data corresponding to the target size microelectrode, the basic action potential reconstruction model is optimized by transfer learning to obtain the target size-adapted action potential reconstruction model. Among them, based on a small amount of action potential pairing data corresponding to the target-size microelectrode, the basic action potential reconstruction model is optimized through transfer learning to obtain an action potential reconstruction model adapted to the target size, including: S41: Based on the basic action potential reconstruction model trained with source domain action potential paired data, the source domain feature parameters in the basic action potential reconstruction model are transferred to the target domain to obtain the transfer initialization model. S42: Based on a small amount of action potential pairing data corresponding to the target domain, the transfer initialization model is adapted to target domain features to obtain the target domain action potential feature mapping relationship. S43: Based on the target domain action potential feature mapping relationship and the reconstruction error between the corresponding intracellular action potential signals, the parameters of the migration initialization model are optimized to obtain the target domain adaptation model. S44: Based on the target domain adaptation model, the action potential data corresponding to the target size microelectrode are reconstructed and verified to obtain the action potential reconstruction model adapted to the target size.

[0066] In this embodiment, a basic action potential reconstruction model is first trained based on source domain action potential paired data, and then the source domain feature parameters in the basic model are transferred to the target domain. Specifically, extracellular action potential signals and corresponding intracellular action potential signals collected by a 20μm microelectrode are used as source domain data. The mapping relationship between extracellular and intracellular action potentials is learned through a convolutional neural network and a long short-term memory network to obtain the basic action potential reconstruction model. Subsequently, the action potential local waveform feature extraction parameters, temporal correlation parameters, and signal mapping parameters learned in the basic model are transferred to the data environment corresponding to the target-size microelectrode to form a transfer initialization model, providing initial parameters for subsequent target domain adaptation.

[0067] Based on a limited amount of action potential pairing data corresponding to the target-size microelectrode, the transfer initialization model is adapted to target domain features. Specifically, the extracellular action potential signals and corresponding intracellular action potential signals collected by the target-size microelectrode are input into the transfer initialization model. By adjusting the feature parameters related to the target size, the model adapts to changes in signal amplitude, waveform morphology, and electrode coupling state under different electrode sizes. Simultaneously, the universal action potential feature extraction capability of the basic model is retained, and the deep mapping parameters are optimized to establish the feature mapping relationship between extracellular and intracellular action potentials under the target-size microelectrode condition.

[0068] Based on the target domain action potential feature mapping relationship and the reconstruction error between the corresponding intracellular action potential signals, the migration initialization model is iteratively optimized, and the reconstruction is verified using microelectrode data of the target size. Specifically, the extracellular action potential signal of the target domain is input into the optimized model to obtain the predicted intracellular action potential result, which is then compared with the synchronously acquired real intracellular action potential signal. The model parameters are adjusted according to the waveform differences to gradually approximate the real signal. When the model output meets the preset accuracy requirements, it is determined as the action potential reconstruction model adapted to the target size, realizing non-invasive reconstruction of intracellular action potentials under microelectrode conditions of different sizes.

[0069] Specifically, based on a small amount of action potential pairing data corresponding to the target domain, the transfer initialization model is adapted to target domain features to obtain the target domain action potential feature mapping relationship, including: Feature extraction is performed on source domain action potential paired data and target domain action potential paired data to obtain source domain action potential features and target domain action potential features. The feature distribution is aligned based on the action potential features of the source domain and the action potential features of the target domain to determine the feature offset between the source domain and the target domain. The feature extraction parameters in the migration initialization model are adjusted based on the feature offset to match the feature space of the source domain with the feature space of the target domain. Based on the adjusted feature extraction parameters, feature mapping is performed on the target domain action potential data to obtain the target domain action potential feature mapping relationship.

[0070] In this embodiment, feature extraction is performed based on source domain action potential pairing data and target domain action potential pairing data to obtain corresponding source domain action potential features and target domain action potential features. Specifically, the source domain data uses extracellular action potential signals and corresponding intracellular action potential signals collected by a 20μm microelectrode, while the target domain data uses a small amount of extracellular-intracellular action potential pairing data collected by a microelectrode of the target size. Both types of data are input into a transfer initialization model, and local waveform features of the action potentials are extracted through a convolutional feature extraction layer. Periodic variation features are then extracted using a temporal network, resulting in source domain action potential features and target domain action potential features that characterize the morphology and dynamic changes of action potentials under different electrode sizes.

[0071] Feature distribution alignment is performed based on the action potential characteristics of the source and target domains to determine the feature offset between them. Specifically, due to differences in electrode area, cell coupling state, and signal acquisition range among microelectrodes of different sizes, action potential signals generated by the same cell exhibit offsets in amplitude, waveform structure, and feature distribution. Therefore, matching analysis is performed on the feature spaces of the source and target domains to determine the feature differences between the two types of data. Based on the feature offset, the feature extraction parameters in the transfer learning initialization model are adjusted so that the action potential features extracted by the model gradually adapt to the data distribution of the target-sized microelectrode, while retaining the learned general action potential waveform features in the source domain. Transfer learning reduces the data distribution differences caused by different electrode sizes, and target domain adaptation is achieved by freezing shallow parameters and adjusting deep parameters.

[0072] Based on the adjusted feature extraction parameters, feature mapping is performed on the target domain action potential data to obtain the target domain action potential feature mapping relationship. Specifically, the parameter-adjusted transfer initialization model is input into the action potential data corresponding to the target-size microelectrode, enabling the model to extract the adapted local waveform features and temporal variation features of the action potential according to the target domain feature distribution, and establish a mapping relationship with the corresponding intracellular action potential signal. The model is continuously optimized through a small amount of target domain paired data, so that the output features can accurately reflect the correspondence between extracellular and intracellular action potentials under the target-size microelectrode condition, and finally obtain the target domain action potential feature mapping relationship, providing a foundation for the subsequent reconstruction and optimization of the target size adaptation model.

[0073] S5: Based on the extracellular action potential signal corresponding to the cell to be detected, and inputting the action potential reconstruction model adapted to the target size, the signal is reconstructed to obtain the corresponding intracellular action potential reconstruction result.

[0074] Specifically, based on the extracellular action potential signal corresponding to the cell to be detected, and input into an action potential reconstruction model adapted to the target size, signal reconstruction is performed to obtain the corresponding intracellular action potential reconstruction results, including: S51: Standardize the extracellular action potential signal corresponding to the cell to be detected to obtain the action potential data to be reconstructed; S52: Input the action potential data to be reconstructed into the action potential reconstruction model adapted to the target size to obtain the intracellular action potential prediction signal; S53: Post-processing and signal reconstruction are performed on the intracellular action potential prediction signal to obtain the intracellular action potential reconstruction waveform; S54: Based on the intracellular action potential reconstruction waveform, perform error analysis, correlation analysis and stability analysis, and evaluate the performance of the action potential reconstruction model adapted to the target size to generate intracellular action potential reconstruction results.

[0075] In this embodiment, the extracellular action potential signal corresponding to the cell to be detected is standardized to obtain action potential data to be reconstructed. Specifically, the extracellular action potential signal generated by the cell to be detected is acquired through a microelectrode array of the target size, and the signal is processed according to the data processing flow of the training phase, including period segmentation, time alignment, fixed-length truncation, and amplitude normalization, to ensure that the action potential signal to be detected is consistent with the input data format of the target size adaptation model. Period segmentation is used to extract the complete action potential period, time alignment is used to eliminate time offset during acquisition, fixed-length truncation is used to ensure that different samples have a uniform length, and amplitude normalization is used to reduce the impact of different cell states and electrode acquisition differences on the reconstruction results, thereby obtaining action potential data to be reconstructed that meets the model input requirements.

[0076] The action potential data to be reconstructed is input into a target-size-adapted action potential reconstruction model to obtain the corresponding intracellular action potential prediction signal. Specifically, the target-size-adapted model extracts features and performs temporal analysis on the input extracellular action potential signal based on the target domain action potential feature mapping relationship established during the transfer learning stage. It learns the correspondence between extracellular and intracellular action potentials under target-size microelectrode conditions and outputs the predicted intracellular action potential signal. Subsequently, the predicted signal undergoes post-processing and signal reconstruction, including waveform smoothing, amplitude calibration, time calibration, and continuous periodic splicing, to make the predicted signal closer to the real intracellular action potential waveform in terms of waveform morphology, duration, and dynamic change trend, thus obtaining the reconstructed intracellular action potential waveform. After model training is completed, only extracellular action potential signals need to be collected to reconstruct the corresponding intracellular action potential through the trained model, and the quality of the reconstructed signal is improved through post-processing.

[0077] Error analysis, correlation analysis, and stability analysis are performed on the reconstructed intracellular action potential waveform to evaluate the performance of the action potential reconstruction model adapted to the target size, generating the final intracellular action potential reconstruction result. Specifically, the reconstructed intracellular action potential waveform is compared and analyzed with the corresponding real detection data to evaluate the model performance in terms of waveform error, signal change trend, and long-term reconstruction stability. Error analysis is used to determine the deviation between the reconstructed waveform and the real intracellular action potential; correlation analysis is used to evaluate the consistency between the two in terms of action potential morphological changes; and stability analysis is used to evaluate the reliability of the model's output during continuous monitoring. When the model meets the preset performance requirements, the corresponding reconstructed waveform is used as the final intracellular action potential reconstruction result, realizing non-invasive long-term intracellular action potential monitoring without continuous electroporation.

[0078] refer to Figure 2 As shown, Figure 2 This tool is used to compare the reconstruction errors of the basic action potential reconstruction model and the target size adaptation model optimized by transfer learning under different microelectrode sizes. The horizontal axis represents the microelectrode size (20μm, 50μm, 100μm, 200μm, and 400μm), and the vertical axis represents the mean squared error (MSE). The MSE and MSE models represent the basic model trained on 20μm source domain data (CrossMEA). 20 ) and the model optimized through transfer learning (CrossMEA) The error distribution of ( ). Figure 2 It can be seen that the basic model has a low error under the condition of 20μm electrode, but as the electrode size changes, the model reconstruction error gradually increases due to the influence of electrode acquisition range, signal amplitude and cell coupling differences.

[0079] After optimization through transfer learning, CrossMEA The model exhibits lower error distribution under different microelectrode sizes (50μm, 100μm, 200μm, and 400μm), indicating that using a small amount of action potential pairing data corresponding to the target-size microelectrode for feature adaptation of the basic model can effectively reduce data distribution shifts between different electrode sizes and improve the model's cross-size reconstruction capability. This result validates that the non-invasive intracellular action potential reconstruction method based on transfer learning proposed in this application can maintain high reconstruction accuracy and stability under different microelectrode sizes.

[0080] like Figure 3 As shown, Figure 3 CrossMEA is used to demonstrate the basic action potential reconstruction model under different microelectrode sizes. 20 CrossMEA model optimized with transfer learning The results of the determination coefficient comparison are shown. The horizontal axis represents the microelectrode size, including 20μm, 50μm, 100μm, 200μm and 400μm; the vertical axis represents the determination coefficient R², the closer the value is to 1, the higher the consistency between the intracellular action potential waveform obtained by the model reconstruction and the measured intracellular action potential signal.

[0081] Depend on Figure 3 It can be seen that CrossMEA 20 It exhibits a high R² under 20μm source domain electrode conditions, but when directly applied to target size electrodes of 50–400μm, its R² significantly decreases and becomes more dispersed due to the shift in signal amplitude, waveform morphology, and characteristic distribution caused by changes in electrode size. After optimization through migration of a small amount of action potential paired data in the target domain, CrossMEA… The R² values ​​for each target electrode size are close to 1 and are more concentrated, indicating that transfer learning can effectively adapt to the action potential characteristics corresponding to microelectrodes of different sizes, and improve the correlation, accuracy and cross-size stability of intracellular action potential reconstruction waveforms.

[0082] like Figure 4 As shown, Figure 4 This method is used to demonstrate the signal reconstruction performance of a non-invasive intracellular action potential reconstruction method based on transfer learning under different microelectrode sizes. Figure 4 (i) Extracellular action potential signals acquired by microelectrodes of different sizes (Extra.AP). Figure 4 (ii) is the real intracellular action potential signal obtained by synchronous recording (Intra.AP). Figure 4(iii) CrossMEA optimized through transfer learning The model reconstructs intracellular action potential signals based on extracellular action potential signals. Figure 4 The changes in action potential waveforms under 20μm, 50μm, 100μm, 200μm and 400μm microelectrodes are shown to verify the model’s reconstruction capability under different electrode platforms.

[0083] Depend on Figure 4 It is evident that as the size of the microelectrode increases, the waveform amplitude, signal morphology, and noise level of the extracellular action potential signal change significantly, indicating that different sized microelectrodes lead to certain distribution differences in the acquired signals. However, after transfer learning optimization based on a small amount of action potential paired data of the target size, CrossMEA... The model can accurately reconstruct the corresponding intracellular action potential waveform based on extracellular action potential signals acquired by microelectrodes of different sizes. Figure 4 (iii) It can be seen that the reconstruction results are consistent with the real intracellular action potential signal in terms of peak position, action potential period, depolarization process, and repolarization trend. Figure 4 (ii) Maintaining high consistency indicates that the proposed method can effectively learn the mapping relationship between extracellular action potential and intracellular action potential under different microelectrode sizes, realize intracellular action potential reconstruction without continuous invasive detection, and improve the adaptability and reliability of non-invasive electrophysiological monitoring.

[0084] Please see Figure 5 This application also provides a non-invasive intracellular action potential reconstruction system based on transfer learning, the system comprising: The data acquisition module is used to acquire extracellular action potential signals and corresponding intracellular action potential signals to construct a paired action potential dataset. The data processing module is used to perform signal preprocessing and quality screening on the action potential paired dataset to obtain standardized action potential training data. The model building module is used to construct a mapping model from extracellular action potentials to intracellular action potentials based on standardized action potential training data, thereby obtaining a basic action potential reconstruction model. The transfer optimization module is used to perform transfer learning optimization on the basic action potential reconstruction model based on a small amount of action potential pairing data corresponding to the target size microelectrode, so as to obtain an action potential reconstruction model adapted to the target size. The signal reconstruction module is used to reconstruct the signal based on the extracellular action potential signal of the cell to be detected and input the action potential reconstruction model adapted to the target size, so as to obtain the corresponding intracellular action potential reconstruction result.

[0085] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0086] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0087] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0088] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0089] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0090] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0091] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0092] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0093] This application provides a non-invasive intracellular action potential reconstruction method and system based on transfer learning. This approach constructs a paired dataset between extracellular and intracellular action potential signals and utilizes signal preprocessing and quality screening mechanisms to improve the reliability of training data, providing an effective data foundation for the subsequent construction of an action potential mapping model. By establishing a mapping relationship between extracellular and intracellular action potentials, intracellular action potentials can be reconstructed without invading the cell, avoiding interference with cell state caused by traditional invasive detection methods such as patch clamps, thus improving the safety and applicability of cell electrophysiological detection. By optimizing the basic action potential reconstruction model through transfer learning based on a small amount of paired data corresponding to microelectrodes of the target size, the model can adapt to the signal feature differences generated by microelectrodes of different sizes, reducing the amount of data and time required to retrain the model for different detection conditions and improving the model's generalization ability. By using the action potential reconstruction model adapted to the target size to reconstruct the extracellular action potential signal of the cell to be detected, the corresponding intracellular action potential information can be obtained, improving the convenience and detection efficiency of cell electrophysiological feature acquisition and providing technical support for large-scale, long-term, non-invasive cell function analysis. In summary, this invention can improve the accuracy and adaptability of action potential reconstruction under different microelectrode conditions while reducing the invasiveness of detection.

[0094] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0095] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0096] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0097] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0098] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A non-invasive method for reconstructing intracellular action potentials based on transfer learning, characterized in that, The method includes the following steps: Acquire extracellular action potential signals and corresponding intracellular action potential signals, and construct an action potential pairing dataset; The paired action potential dataset is preprocessed and quality screened to obtain standardized action potential training data. Based on the standardized action potential training data, a mapping model from extracellular action potential to intracellular action potential is constructed to obtain the basic action potential reconstruction model. Based on a small amount of action potential pairing data corresponding to the target size microelectrode, the basic action potential reconstruction model is optimized by transfer learning to obtain an action potential reconstruction model adapted to the target size. Based on the extracellular action potential signal corresponding to the cell to be detected, and input into the action potential reconstruction model adapted to the target size, the signal is reconstructed to obtain the corresponding intracellular action potential reconstruction result.

2. The method according to claim 1, characterized in that, The step of preprocessing and quality screening the paired action potential dataset to obtain standardized action potential training data includes: The paired action potential dataset is subjected to periodic segmentation to obtain a single-cycle action potential signal; Based on the feature points corresponding to the single-cycle action potential signal, time alignment and fixed-length truncation are performed to obtain a uniform-length action potential sequence. The amplitude of the action potential sequence is normalized to obtain normalized action potential data; The normalized action potential data is classified based on a signal quality classification model, and action potential data that meets preset quality conditions are selected to obtain standardized action potential training data.

3. The method according to claim 2, characterized in that, The process of classifying the normalized action potential data based on a signal quality classification model and selecting action potential data that meet preset quality conditions to obtain standardized action potential training data includes: A signal quality classification model is constructed based on a one-dimensional convolutional neural network, and convolutional feature extraction is performed on the normalized action potential data to obtain action potential quality features. Based on the action potential quality characteristics, the normalized action potential data is classified into quality levels to generate signal quality classification results. Based on preset quality screening conditions, target action potential pairing data are screened according to the signal quality classification results. The target action potential pairing data is converted into a training format to obtain standardized action potential training data.

4. The method according to claim 1, characterized in that, The process of constructing a mapping model from extracellular action potentials to intracellular action potentials based on the standardized action potential training data, resulting in a basic action potential reconstruction model, includes: Based on a convolutional neural network, local feature extraction is performed on the extracellular action potential signal in the standardized action potential training data to obtain local waveform features of the action potential. Based on a long short-term memory network, the local waveform features of the action potential are modeled in a temporal correlation to extract the temporal change features corresponding to the extracellular action potential signal. Based on the aforementioned temporal variation characteristics, a mapping relationship is established with the corresponding intracellular action potential signal to generate intracellular action potential prediction results; Based on the error between the intracellular action potential prediction results and the intracellular action potential signals in the standardized action potential training data, the mapping model parameters are iteratively optimized to obtain the basic action potential reconstruction model.

5. The method according to claim 4, characterized in that, The method based on a convolutional neural network extracts local features from the extracellular action potential signals in the standardized action potential training data to obtain local waveform features of the action potential, including: The extracellular action potential signals in the standardized action potential training data are serialized to generate action potential input sequences; Convolutional feature extraction is performed on the action potential input sequence based on a one-dimensional convolutional layer to obtain the local waveform response features of the action potential. The local waveform response features are enhanced based on a nonlinear activation function to extract waveform change features from the extracellular action potential signal. The waveform change features are reduced in dimensionality using a pooling layer to obtain key waveform features. The key waveform features are encoded based on multi-layer convolutional feature fusion to obtain local waveform features of action potential.

6. The method according to claim 4, characterized in that, The method based on a long short-term memory network performs temporal correlation modeling on the local waveform features of the action potential, extracting the temporal variation features corresponding to the extracellular action potential signal, including: The local waveform features of the action potential are temporally encoded based on a long short-term memory network to generate temporal hidden features of the action potential. The state of the action potential temporal latent features is updated based on the gating mechanism of the long short-term memory network, and the long-term dependency features corresponding to the periodic changes of the action potential are extracted. Based on the long-term dependent features, the dynamic evolution features corresponding to the extracellular action potential signal are extracted to obtain the action potential temporal change features; The temporal variation features of the action potential are fused to obtain the temporal variation features corresponding to the extracellular action potential signal.

7. The method according to claim 1, characterized in that, The basic action potential reconstruction model is optimized through transfer learning based on a small amount of action potential pairing data corresponding to the target-size microelectrode, resulting in an action potential reconstruction model adapted to the target size, including: Based on the basic action potential reconstruction model trained with source domain action potential paired data, the source domain feature parameters in the basic action potential reconstruction model are transferred to the target domain to obtain the transfer initialization model. Based on a small amount of action potential pairing data corresponding to the target domain, the transfer initialization model is adapted to target domain features to obtain the target domain action potential feature mapping relationship. Based on the target domain action potential feature mapping relationship and the reconstruction error between the corresponding intracellular action potential signals, the migration initialization model is optimized to obtain the target domain adaptation model. Based on the target domain adaptation model, the action potential data corresponding to the target size microelectrode are reconstructed and verified to obtain the action potential reconstruction model adapted to the target size.

8. The method according to claim 7, characterized in that, The step of adapting the transfer initialization model to target domain features based on a small amount of action potential pairing data corresponding to the target domain, to obtain the target domain action potential feature mapping relationship, includes: Feature extraction is performed on source domain action potential paired data and target domain action potential paired data to obtain source domain action potential features and target domain action potential features. Based on the source domain action potential features and the target domain action potential features, feature distribution alignment is performed to determine the feature offset between the source domain and the target domain. Based on the feature offset, the feature extraction parameters in the migration initialization model are adjusted to match the source domain feature space with the target domain feature space. Based on the adjusted feature extraction parameters, feature mapping is performed on the target domain action potential data to obtain the target domain action potential feature mapping relationship.

9. The method according to claim 1, characterized in that, The process involves reconstructing the intracellular action potential based on the extracellular action potential signal corresponding to the cell to be detected, and inputting it into the target size-adapted action potential reconstruction model to obtain the corresponding intracellular action potential reconstruction result, including: The extracellular action potential signal corresponding to the cell to be detected is standardized to obtain action potential data to be reconstructed. The action potential data to be reconstructed is input into the action potential reconstruction model adapted to the target size to obtain the intracellular action potential prediction signal. The intracellular action potential prediction signal is post-processed and reconstructed to obtain the intracellular action potential reconstruction waveform. Error analysis, correlation analysis, and stability analysis are performed based on the reconstructed intracellular action potential waveform, and the performance of the target-size-adapted action potential reconstruction model is evaluated to generate intracellular action potential reconstruction results.

10. A non-invasive intracellular action potential reconstruction system based on transfer learning, characterized in that, The system includes: The data acquisition module is used to acquire extracellular action potential signals and corresponding intracellular action potential signals to construct a paired action potential dataset. The data processing module is used to perform signal preprocessing and quality screening on the action potential paired dataset to obtain standardized action potential training data. The model building module is used to construct a mapping model from extracellular action potentials to intracellular action potentials based on the standardized action potential training data, thereby obtaining a basic action potential reconstruction model. The transfer optimization module is used to perform transfer learning optimization on the basic action potential reconstruction model based on a small amount of action potential pairing data corresponding to the target size microelectrode, so as to obtain an action potential reconstruction model adapted to the target size. The signal reconstruction module is used to reconstruct the signal based on the extracellular action potential signal corresponding to the cell to be detected and input the action potential reconstruction model adapted to the target size, so as to obtain the corresponding intracellular action potential reconstruction result.

Images