A multi-stage excavation and multi-modal meta-fusion fetal acidosis classification system

Abstract

A multi-stage data mining and multi-modal meta-fusion fetal acidosis classification system is proposed, belonging to the field of medical artificial intelligence and signal processing technology. It addresses the severe class imbalance problem and lack of targeted processing for difficult samples in existing fetal acidosis analysis studies. A three-stage progressive training and meta-fusion framework is proposed, systematically decomposing the fetal acidosis classification task into three stages: basic learning, difficult sample reinforcement learning, and intelligent decision fusion. This provides a novel learning model for solving the challenges of class imbalance and difficult samples in medical signal analysis. An adaptive dynamic weighted loss function based on real-time recall is designed. This loss function dynamically adjusts the weight of each category in the loss calculation during model training, adaptively focusing on categories that are difficult to classify correctly, effectively alleviating the severe class imbalance problem in clinical data.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence and signal processing technology, specifically relating to a fetal acidosis classification system based on multi-stage mining and multimodal fusion. Background Technology

[0002] Intrapartum fetal acidosis is primarily caused by fetal hypoxia, which stems from disruptions in the oxygen exchange process between mother and fetus. This condition is significantly associated with serious adverse outcomes such as neonatal asphyxia and encephalopathy, constituting a significant public health burden. Clinically, fetal heart rate monitoring is a crucial non-invasive method for assessing fetal status and screening for acidosis; however, its interpretation has long relied on physician experience, resulting in high subjectivity and inconsistent results. Therefore, while international guidelines recommend standardized classification methods, cardiac blood gas (CTG) is ultimately an indirect assessment. Currently, umbilical artery blood gas analysis (especially pH) is widely recognized as an objective standard for diagnosing intrapartum acidosis, directly reflecting the fetal acid-base balance.

[0003] However, there are two main challenges when modeling this type of data using machine learning: First, the data suffers from severe class imbalance. In real clinical datasets, normal samples often constitute the vast majority, while acidosis samples (especially severe acidosis) are extremely rare. This distribution easily leads to model training bias towards the majority class, severely weakening the ability to identify the minority but clinically significant classes. Second, there is a lack of targeted handling for difficult samples. The data contains a large number of samples that are at the classification boundary, have ambiguous features, or are easily confused. General model structures struggle to effectively learn from these samples, limiting further improvements in overall performance.

[0004] To address these challenges, existing methods primarily focus on three levels: at the data level, oversampling or undersampling techniques such as SMOTE are used to adjust the class distribution, but this may introduce synthetic sample bias or lose valuable information; at the loss function level, Focal Loss, Class-Balanced Loss, or dynamically weighted loss are introduced to improve the model's attention to minority classes or difficult samples, but these methods may be sensitive to hyperparameters or may impair overall generalization ability due to over-focusing on minority classes; at the model level, strategies such as ensemble learning or Stacking are employed to improve the discriminative power for difficult samples and reduce the risk of overfitting by combining the advantages of multiple base learners, but traditional Stacking methods usually require multiple classifiers to be trained in parallel on the entire dataset, failing to explicitly optimize for difficult samples, and the fusion method is relatively simple, making it difficult to deeply explore the complementarity and uncertainty information in the prediction behavior of different models. Summary of the Invention

[0005] To address the serious class imbalance problem and lack of targeted processing for difficult samples in existing fetal acidosis analysis studies, this application provides a fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion.

[0006] The system includes: Data preprocessing and feature extraction unit: acquires raw fetal heart rate signal and synchronous clinical features; preprocesses the fetal heart rate signal, standardizes the clinical features, and aligns them with the preprocessed fetal heart rate signal to form multimodal sample input; Phase A Basic Classification Unit: Construct a basic classification network based on a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a multi-head self-attention mechanism. Perform end-to-end training on the preprocessed fetal heart rate signal and output the classification results. Introduce an adaptive dynamic weighted loss function based on recall to dynamically adjust the loss weights of each category and identify difficult samples, including misclassified samples and low-confidence samples, to provide training data for the next phase. Stage B Difficult Sample Multimodal Enhancement Training Unit: For the difficult samples identified in Stage A, a multimodal enhancement classifier is constructed for classification. The classifier combines the clinical features corresponding to the difficult samples for fusion processing and outputs the second classification result. The multimodal enhancement classifier uses depthwise separable convolution and SE attention mechanism to extract subtle features from the fetal heart rate signal. Stage C Meta-learning Fusion Decision Output Unit: Based on the predicted outputs of Stage A and Stage B on the validation set, construct meta-feature vectors, input them into the meta-feature fusion model for training, adaptively fuse the classification preferences and uncertainties of the two-stage models, and output the final classification result to achieve robust decision-making. Furthermore, the preprocessing of the fetal heart rate signal specifically includes: initializing the original fetal heart rate signal by replacing the 0 value of FHR with NaN; detecting missing values, if the maximum missing value is less than 15 seconds, linear interpolation is used for local repair, otherwise it is directly deleted; detecting spike artifacts, if the absolute value between two adjacent points is greater than 25 bpm, it is considered noise, and the region with a difference of less than 10 bpm between five consecutive heartbeats is defined as a stable segment, and linear interpolation is used to replace the value of the noise with the linear transition value between the previous stable segment and the subsequent new stable segment; detecting abnormal ranges, if the FHR value is greater than 200 bpm or less than 50 bpm, Hermite spline interpolation is used for filling.

[0007] Furthermore, the basic classification network includes a four-layer one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a multi-head attention mechanism.

[0008] Furthermore, the adaptive dynamic weighted loss function is set as follows: S41 Weight Initialization and Update Objective: At the start of training, the weights of all classes are initialized to 1. After each training cycle, the recall of each class is calculated based on the model's performance on the current training set. ,in Representing the category index, a target weight vector is generated. :

[0009] in, The total number of categories, As the amplification factor, Constants to prevent numerical instability; S42 Smooth Update Strategy: Use a smoothing method to adjust the current category weights Transition to target weights, training rounds during a predefined warm-up phase. Inside, the next round of sub-weighting Update using linear interpolation:

[0010] ;in, Indicates the current training round; After the warm-up phase, the next round of weighting The model will be updated smoothly using an exponentially weighted average method to avoid instability in model training caused by sudden changes in weights. ;in, For smoothing coefficients; Combining S43 with the loss function: The category weights are dynamically calculated using a smooth update strategy. When applied to the standard cross-entropy loss function, the resulting cross-entropy loss function is:

[0011] in, It is the batch size, when the sample belongs to the first... When class, ,otherwise , It is the predicted probability distribution vector. This represents the summation over all categories.

[0012] Furthermore, the identification of difficult samples specifically involves: using the trained Stage A basic classification network to classify the original training set, calculating the classification probability of each sample on its true label as a confidence index, and statistically analyzing the average confidence level of the entire training set. The identification of difficult samples can be determined by satisfying any of the following conditions: misclassification: the model's predicted category is inconsistent with the true label; low confidence: although the prediction is correct, the confidence level is lower than the average confidence level of the training set.

[0013] Furthermore, the multimodal augmentation classifier includes a two-channel data processing path. The input in the first channel is the difficult sample identified in stage A, and the input in the second channel is the clinical features corresponding to the difficult sample. In the first channel, the hard sample input feature extraction module performs feature extraction and attention weighting to obtain the feature sequence output. The feature sequence is input into the BilSTM module to capture long-term time-series information and obtain the time-series feature output. The time-series feature is input into the attention pooling layer to learn the attention weights of each time step, and the time-series features are weighted and summed to output the global attention aggregated features. The global attention aggregated features are then transformed nonlinearly through a fully connected layer to output the final fetal heart rate signal feature representation. In the second channel, the clinical features corresponding to the input difficult samples are first processed by the batch normalization module to obtain standardized and normalized clinical feature vectors. Then, the clinical feature vectors are nonlinearly transformed by the fully connected layer to obtain intermediate feature representations. The intermediate feature representations are then standardized again by the batch normalization module to output stable distributed multidimensional features. Finally, the multidimensional features are nonlinearly transformed by the fully connected layer to output the final clinical feature representation. The final fetal heart rate signal feature representation and the final clinical feature representation are concatenated by a fully connected layer to obtain fused features. The fused features are then mapped to the original classification probability output of the three categories by the output layer.

[0014] Furthermore, the meta-feature vector includes the following features: Basic probability characteristics: including the category probability distributions for stages A and B; Confidence and uncertainty characteristics: including classification confidence for stages A and B, the difference in confidence between the two models for stages A and B, and the entropy of classification for stages A and B; Prediction consistency features: including whether the predicted categories of the two models in stages A and B are consistent; Probability difference characteristics: including the maximum and average absolute difference between the probability vectors of the two models in stage A and stage B, and the probability difference between the two models in stage A and stage B by category.

[0015] Furthermore, the meta-feature fusion model contains three hidden layers and an output layer. In the hidden layer, the input meta-feature vector is linearly transformed, normalized, activated by the ReLU function, and regularized by Dropout. Finally, the data is mapped to three categories through the output layer.

[0016] The beneficial effects of the system described in this invention are as follows: This system, based on the CTU-CHB real-world dataset (containing raw CTG signals and synchronous clinical information), proposes a novel three-stage progressive training and meta-fusion framework. This framework not only effectively alleviates the class imbalance problem but also introduces multimodal information for specialized processing of difficult samples. Furthermore, it adaptively fuses the generalization ability of the base model with the fine-grained discriminative ability of the difficult sample model, ultimately outputting more robust and accurate classification decisions. Specifically: A three-stage progressive training and meta-fusion framework is proposed. This framework systematically decomposes the fetal acidosis classification task into three stages: basic learning, hard sample reinforcement learning, and intelligent decision fusion, providing a novel learning model for solving the problems of class imbalance and hard samples in medical signal analysis.

[0017] An adaptive dynamic weighted loss function based on real-time recall was designed. This loss function can dynamically adjust the weight of each category in the loss calculation during model training, thereby adaptively focusing on categories that are difficult to classify correctly and effectively alleviating the problem of severe class imbalance in clinical data.

[0018] A multimodal fusion classifier for difficult samples was constructed. For difficult samples identified in stage A, a multimodal training strategy was designed that integrates deep separable convolution, SE attention mechanism, and clinical table information, which significantly improved the model's ability to extract and distinguish subtle features of difficult samples with blurred boundaries and indistinct features.

[0019] An adaptive model fusion mechanism based on meta-learning is proposed. By constructing a meta-feature vector containing the classification probabilities, confidence levels, and uncertainties of stages A and B as input to train the meta-learner of stage C, the fusion mechanism achieves an adaptive trade-off between the classification preferences and uncertainties of the two models, resulting in a more robust and accurate final classification decision mode than traditional fusion methods. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the fetal heart rate signal after preprocessing in an embodiment of the present invention; Figure 2 This is a structural diagram of the basic classification network in an embodiment of the present invention; Figure 3 This is a structural diagram of the multimodal augmentation classifier in an embodiment of the present invention; Figure 4 This is a structural diagram of the meta-feature fusion model in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0022] Example 1 This embodiment provides a fetal acidosis classification system that combines multi-stage mining and multimodal meta-fusion, the system comprising: Data preprocessing and feature extraction unit: acquires raw fetal heart rate signal and synchronous clinical features; preprocesses the fetal heart rate signal, standardizes the clinical features, and aligns them with the preprocessed fetal heart rate signal to form multimodal sample input; Phase A Basic Classification Unit: Construct a basic classification network based on a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a multi-head self-attention mechanism. Perform end-to-end training on the preprocessed fetal heart rate signal and output the classification results. Introduce an adaptive dynamic weighted loss function based on recall to dynamically adjust the loss weights of each category and identify difficult samples, including misclassified samples and low-confidence samples, to provide training data for the next phase. Stage B Difficult Sample Multimodal Enhancement Training Unit: For the difficult samples identified in Stage A, a multimodal enhancement classifier is constructed for classification. The classifier combines the clinical features corresponding to the difficult samples for fusion processing and outputs the second classification result. The multimodal enhancement classifier uses depthwise separable convolution and SE attention mechanism to extract subtle features from the fetal heart rate signal. Stage C Meta-learning Fusion Decision Output Unit: Based on the predicted outputs of Stage A and Stage B on the validation set, construct meta-feature vectors, input them into the meta-feature fusion model for training, adaptively fuse the classification preferences and uncertainties of the two-stage models, and output the final classification result to achieve robust decision-making.

[0023] Example 2 This embodiment further defines Embodiment 1. Currently, umbilical artery blood gas analysis (especially pH value) is recognized as an objective standard for diagnosing intrapartum acidosis, directly reflecting the fetal acid-base balance. This embodiment further classifies acidosis into two categories based on fetal electrocardiogram ST segment analysis results: moderate acidosis (pH > 7.05 but ≤ 7.15) and severe acidosis (pH ≤ 7.05), and uses this standard for data labeling. The dataset used contains 439 normal records, 69 cases of moderate acidosis, and 44 cases of severe acidosis, exhibiting a significant class imbalance.

[0024] The workflow of the data preprocessing and feature extraction unit is further explained below: (1) The CTU-CHB dataset is a real-world intrapartum fetal heart rate monitoring database constructed by the Czech Technical University in collaboration with Prague University Hospital. This dataset contains 552 complete intrapartum monitoring records with a signal sampling frequency of 4Hz and an average duration of approximately 90 minutes, covering key stages of the labor process. It provides raw fetal heart rate monitoring signals, which are closer to actual clinical application scenarios, but also correspondingly increases the complexity of data processing. A significant problem with this dataset is the large-scale absence of uterine contraction signals, making it difficult to determine whether fetal heart rate deceleration is related to uterine contractions.

[0025] A key advantage of this dataset is the provision of clinical textual information (see Table 1), offering an objective standard based on biochemical indicators for validating the clinical relevance of classification algorithms. Addressing the prevalent issue of missing uterine contraction signals, most current studies based on the CTU-CHB dataset have shifted to extracting features solely from fetal heart rate signals and defining or validating classification results based on the objective indicator of umbilical artery blood pH (typically using pH ≤ 7.15 as the acidosis threshold), thereby achieving fetal status classification or acidosis identification. This method has become a common approach in the field for evaluating the clinical relevance of models.

[0026] Table 1:

[0027] (2) In this embodiment, MATLAB is used for fetal heart rate signal preprocessing. The specific process is as follows: The original fetal heart rate signal is initialized, and the 0 value of FHR is replaced with NaN. Missing value detection: if the maximum missing value is less than 15s, linear interpolation is used for local repair; otherwise, it is directly deleted. Peak artifact detection: if the absolute value between two adjacent points is greater than 25bpm, it is considered noise. The region with a difference of less than 10bpm between five consecutive heartbeats is defined as a stable segment. Linear interpolation is used to replace the value of this noise with the linear transition value between the previous stable segment and the subsequent new stable segment. Abnormal range detection: if the FHR value is greater than 200bpm or less than 50bpm, Hermite spline interpolation is used for filling. The processing effect is as follows: Figure 1 As shown.

[0028] Example 3 This embodiment further defines Embodiment 1. The basic classification unit of Stage A is further explained as follows: Phase A Basic Classification Unit: Construct a basic classification network based on a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a multi-head self-attention mechanism. Perform end-to-end training on the preprocessed fetal heart rate signal and output the classification results. Introduce an adaptive dynamic weighted loss function based on recall to dynamically adjust the loss weights of each category and identify difficult samples, including misclassified samples and low-confidence samples, to provide training data for the next phase.

[0029] The structure diagram of the basic classification network is as follows: Figure 2 As shown.

[0030] (1) Convolutional Neural Network (CNN) Convolutional Neural Networks (CNNs) are deep learning models specifically designed for processing grid-like topological data. Their core idea is to extract shared weight features from local receptive fields by sliding convolutional kernels across the input data. In this embodiment, Stage A model employs a four-layer one-dimensional convolutional neural network (1D CNN) to process fetal electrocardiogram (ECG) time-series signals. Each of the first three layers includes batch normalization and dropout regularization after the convolutional operation. Batch normalization accelerates training convergence by standardizing intermediate layer activation values, while dropout randomly discards neurons with a probability of 0.4 to prevent overfitting. The core function of the convolutional module is to capture local morphological features in the ECG signal, identify key waveforms, and provide rich local feature representations for subsequent time-series modeling.

[0031] (2) Bidirectional Long Short-Term Memory Network (BiLSTM) Long Short-Term Memory (LSTM) networks are a variant of recurrent neural networks specifically designed for processing sequential data. Through gating mechanisms (input gate, forget gate, output gate), they effectively address the vanishing gradient problem of traditional RNNs and can model long-distance temporal dependencies. In this model, a bidirectional LSTM architecture is employed. The forward layer processes data from the beginning to the end of the sequence, while the backward layer processes data from the end to the beginning. Finally, the hidden states from both directions are concatenated to form a complete contextual representation. The input dimension is set to 128, matching the output of the previous transition layer; the number of hidden units is 64, and the bidirectional output dimension is 128. The core function of the BiLSTM module is to capture long-range temporal patterns in fetal electrocardiogram signals, such as heart rate variability and rhythm changes, overcoming the limitation of CNNs that only focus on local features. The input of the BiLSTM module is the output feature tensor of the entire convolutional neural network (corresponding to the output of the entire convolutional neural network), and the output is the bidirectional hidden state at each time step, specifically a 128-dimensional bidirectional encoded sequence (a concatenation of 64-dimensional forward and backward LSTMs).

[0032] The output of the entire convolutional neural network is input into the bidirectional long short-term memory network and also into the global average pooling layer for average pooling along the time dimension. That is, the average value of each channel is taken in the time dimension, and each sample is compressed into a 128-dimensional global feature vector. This retains the channel-level statistical information and discards the time dimension information to obtain the global statistical feature output of each channel.

[0033] The temporal feature sequence output from the BiLSTM is input into a multi-head attention mechanism for processing. Multiple parallel attention heads learn different attention patterns, enhancing the model's expressive power. This embodiment employs a multi-head self-attention mechanism with eight attention heads, and the embedding dimension is set to 128 to match the BiLSTM output. This module calculates the important information that each time step in the BiLSTM output sequence can focus on from all other time steps. A weighted average is used to fuse the dependencies within the sequence, ultimately obtaining a 128-dimensional attention aggregation feature. The core function of multi-head attention is to identify the differences in importance between different time regions in the ECG signal, achieving adaptive feature enhancement.

[0034] The fully connected layer (FC-Layer) concatenates the attention-aggregated features and the global average pooling features along the feature dimension to form a 128-dimensional fused feature vector, fully capturing the global statistical characteristics and temporal dependencies of the signal. A Dropout layer with a dropout rate of 0.5 is applied after the output of the fully connected layer, randomly dropping 50% of the neuron connections during training to effectively prevent overfitting and improve generalization performance. Finally, the output layer (128→3-dimensional) maps the Dropout layer representation to the logits output of the three classes, corresponding to the three-class classification case.

[0035] Example 4 This embodiment further defines Embodiment 1. It provides further explanation of the adaptive dynamic weighted loss function based on recall and the identification of difficult samples.

[0036] (1) Adaptive weighted loss function The dynamic weighted loss function used in this embodiment aims to alleviate the class imbalance problem in fetal heart rate monitoring data. Its core mechanism is to adaptively adjust the weights of each class in the loss function by monitoring the model's recall rate during training, thereby making the model focus more on samples that are difficult to classify correctly. Its workflow and weight update formula are as follows: 1.1 Weight Initialization and Update Objectives At the start of training, the weights of all classes are initialized to 1. At the end of each training epoch, the recall rate for each class is calculated based on the model's performance on the current training set. (in This indicates a category index, as described below. , , , In (Both represent this meaning), generating a target weight vector. Target weight. The calculation follows the principle that the lower the recall rate, the higher the weight, and the specific formula is as follows:

[0037] in, The total number of categories, This is the magnification factor (set to 1.5). For a very small constant (e.g.) To prevent numerical instability, the normalization operation of the denominator ensures that the average value of the target weights is 1, avoiding drastic changes in the loss scale.

[0038] 1.2 Smooth Update Strategy To avoid drastic fluctuations in weights, a smoothing method is used to adjust the current weights. To target weight The transition occurs within the predefined warmup training epochs (set to 15 epochs), where the weights for the next epoch are adjusted. Update using linear interpolation:

[0039] ;in, Indicates the current training round; After the warm-up phase, the next round of weighting The model will be updated smoothly using an exponentially weighted average method to avoid instability in model training caused by sudden changes in weights. ; in This is the smoothing coefficient (set to 0.3).

[0040] 1.3 Combination with loss function Cross-entropy loss is a widely used loss function for classification tasks. Its core idea is to measure the difference between the true class probability distribution and the model's predicted probability distribution, thus guiding the model's optimization direction. Its goal is to minimize the cross-entropy, making the model's classification probability distribution as close as possible to the true class distribution. This method originates from information entropy theory, and its core idea is to optimize the model's decision by maximizing information gain. The mathematical expression of the cross-entropy loss function is as follows:

[0041] in, It is the OneHot encoded vector of the true category. Indicates the batch size. The total number of categories, when the sample belongs to the first category. When class, ,otherwise . It is the probability distribution vector predicted by the model. This represents the summation over all categories.

[0042] Finally, the class weights are dynamically calculated through a smooth update strategy. It is applied to the standard cross-entropy loss function. The weighted cross-entropy loss is calculated as follows:

[0043] In this way, the model applies a larger gradient penalty to misclassifications of low-recall categories during backpropagation, thereby guiding the model to optimize its classification performance.

[0044] (2) Difficult sample identification is a key step in the three-stage training framework. Its goal is to filter out samples with poor model learning performance from the training results of stage A, providing a data foundation for the specialized training of stage B. The identification process adopts a dual-criteria strategy, comprehensively considering two dimensions: misclassification and classification confidence. Specifically, the trained stage A model is first used to classify the original training set, and the classification probability of each sample on its true label is calculated as a confidence index. The average confidence level of the entire training set is then calculated. Difficult samples are judged if they meet either of the following conditions: misclassification: the model's predicted class is inconsistent with the true label; low confidence: although the prediction is correct, the confidence level is lower than the average confidence level of the training set. This dual screening mechanism can simultaneously capture samples with obvious misclassification and boundary samples with ambiguous model judgment, ensuring that stage B can specifically learn more challenging sample features, thereby improving the model's discrimination ability in complex situations.

[0045] Example 5 This embodiment further defines Embodiment 1. It provides a further explanation of the multimodal augmentation training unit for stage B difficult samples: The structure of a multimodal augmentation classifier is as follows: Figure 3 As shown.

[0046] The multimodal augmentation classifier includes a two-channel data processing path. The input in the first channel is the difficult sample identified in stage A, and the input in the second channel is the clinical features corresponding to the difficult sample.

[0047] (1) The processing in the first channel is as follows: The hard sample input feature extraction module performs feature extraction and attention weighting to obtain a 128-dimensional feature sequence output. The feature extraction module includes a depthwise separable convolution module and a channel attention mechanism (SE Attention Mechanism).

[0048] For difficult input samples, a depthwise separable convolution module is used for feature extraction. Depthwise separable convolution is an efficient alternative to traditional convolution, decomposing standard convolution into two independent operations: depthwise convolution and pointwise convolution. Depthwise convolution performs spatial convolution independently on each input channel, with C parameters. in ×K Cin × K (Traditional convolution is C) in ×C out ×K C in × C out × K Pointwise convolution achieves cross-channel information fusion through 1×1 convolution. This module serves as a fundamental building block in the stage B signal encoder. Its core advantage lies in significantly reducing the number of parameters while maintaining feature extraction capabilities. It is particularly suitable for processing relatively small-scale, difficult sample datasets, effectively preventing overfitting and improving training efficiency.

[0049] The extracted features are dynamically enhanced by learning the importance weights of each feature channel using the SE-Net (Squeeze-and-Excitation Network) channel attention mechanism. The SE-Net module contains two core operations: squeezing and activation. The squeezing operation compresses the global information of each channel into a scalar descriptor through global average pooling; the activation operation learns the non-linear dependencies between channels through a fully connected layer with a bottleneck structure, outputting channel weights from 0 to 1. In the model of stage B, the SE-Net module is embedded after a deep separable convolution, enabling the model to adaptively emphasize the frequency bands and temporal patterns most relevant to the pathological features of fetal acidosis, enhancing the ability to identify subtle pathological changes in difficult samples.

[0050] The input to the BiLSTM is a 128-dimensional feature sequence output from the feature extraction module. By capturing long-range temporal patterns in fetal electrocardiogram signals, it outputs 256-dimensional features. The temporal features output by the BiLSTM are then fed into an attention pooling layer. By learning the attention weights for each time step, the 256-dimensional temporal features are weighted and summed to output a 256-dimensional global attention pooled feature. This 256-dimensional attention pooled feature is then non-linearly transformed through a fully connected layer to output the final 128-dimensional feature.

[0051] (2) The processing in the second channel is as follows: For the clinical features corresponding to the input difficult samples, the BatchNorm1d layer is first used to obtain standardized and normalized clinical feature vectors. Then, the standardized clinical feature vectors are nonlinearly transformed by the FC-Layer to obtain a 128-dimensional intermediate feature representation. The BatchNorm1d layer further standardizes the 128-dimensional features output from the first layer, calculates the mean and variance of the current batch data, eliminates internal covariate bias, and outputs a stable 128-dimensional feature distribution. Finally, the FC-Layer is used for nonlinear transformation to obtain a 64-dimensional clinical feature representation.

[0052] (3) The FC-Layer abstracts and represents the 128-dimensional fetal heart rate signal features output from channel 1 and the 64-dimensional clinical features output from channel 2, performs feature concatenation, obtains a 192-dimensional fused feature vector, and outputs 128-dimensional fused features. The OutputLayer maps the 128-dimensional fused features to the original classification probability output of 3 categories.

[0053] Example 6 This embodiment further defines Embodiment 1. The stage C-ary learning fusion decision output unit is further explained as follows: Stage C, as the decision fusion and optimization module of the entire classification framework, plays a core role as an intelligent "decision coordinator." It does not directly process the raw signal data, but rather learns how to optimally integrate classification information from Stage A and Stage B to achieve a final classification that surpasses any single model. Existing model fusion modes are mostly weighted fusion, adjusting the weighted fusion coefficients through corresponding computational mechanisms. However, the fusion mode in this embodiment differs. The module's working mechanism is as follows: First, the system extracts and constructs a 29-dimensional meta-feature vector from the outputs of the two base models, as shown in Table 2. This vector comprehensively encodes various information such as the basic probability distribution, classification confidence, uncertainty measure, and classification consistency. This vector is used as input to train the meta-feature fusion model. Training employs a five-fold cross-validation strategy to ensure its generalization ability. The meta-feature fusion model learns the complex nonlinear relationship between these meta-features and the true labels, adaptively weighing the contributions of the two base models, and ultimately outputting a fused, more accurate classification decision. That is, how to fuse the classification information from Stages A and B to achieve the best final result. The fundamental significance of this design lies in its ability to improve the robustness and accuracy of the system from the perspective of ensemble learning. It can effectively reconcile the advantages of different models on different subsets of data and correct potential individual model biases through the learned fusion strategy, thereby providing a more reliable and stable automated grading result for the clinical assessment of fetal status.

[0054] Table 2:

[0055] like Figure 4As shown, the meta-feature fusion model consists of three hidden layers and one output layer. It aggregates all meta-feature vectors and extracts the validation set for stage C based on the index of the validation set at each fold. Other data is used for training to ensure consistency of the validation sets across the three stages, thus demonstrating the model's improvement. The input meta-feature vectors undergo three layers of linear transformation, normalization, ReLU activation, and Dropout regularization. Finally, the output layer maps the data to three categories for output.

[0056] Example 7 This embodiment further illustrates Embodiments 1-6. Embodiments 1-6 propose a three-stage learning framework based on signal priority and conditional enhancement to systematically address the common class imbalance and difficult classification problems in fetal heart rate monitoring signals. The overall framework consists of three interconnected stages with different functional focuses. Stage A constructs a basic classification model based on time-series signals, extracts features using a CNN-BiLSTM-attention architecture, and introduces a dynamic class-weighted loss function to alleviate class imbalance while identifying difficult samples. Stage B enhances the training of difficult time-series signal samples identified in Stage A, combining depthwise separable convolution and a Squeeze-and-Excitation attention mechanism to strengthen the modeling of signal features, and integrates the clinical text information corresponding to the difficult time-series signal samples to optimize the difficult samples. Stage C fuses the classification outputs of the first two stages through a meta-feature fusion framework, constructing a fusion network containing 29-dimensional meta-features to achieve adaptive decision-making based on model classification preferences and uncertainties. The model evaluation adopts a five-fold cross-validation strategy. Throughout the entire experimental process, phases A, B, and C always use the same validation set partition, forming a complete modeling process from basic learning to fine optimization and then to adaptive fusion.

[0057] Experimental environment and parameter configuration: This embodiment was completed on a computing platform equipped with an NVIDIA RTX 3090 GPU and an Intel Xeon Platinum 8358P CPU (15 vCPU @ 2.60 GHz). All models were implemented based on the PyTorch deep learning framework, using CUDA 11.1 and cuDNN 8.0 for GPU acceleration. The experimental code ran in a Python 3.8 environment.

[0058] The network structure, hyperparameters, and training strategies for each stage in this embodiment are detailed in Table 3. All experiments were trained using 500 epochs with an early stopping mechanism and a patience value of 100. Five-fold cross-validation was used for evaluation, and the validation set data used in stages A, B, and C remained consistent throughout the process. This rigorously ensured that differences in data partitioning would not introduce information leakage risks during model evaluation, guaranteeing the rigor of the experiments and the comparability of the results. The parameters used in different stages are shown in Table 3.

[0059] Table 3:

[0060] Ablation experiment: Table 4 shows the results of the ablation experiment. As can be seen from Table 4, after introducing the proposed dynamic class weighting strategy in stage A, the F1 score improved from 70.88% ± 0.60% to 76.10% ± 1.71% (an absolute improvement of approximately 5.2 percentage points). Simultaneously, recall and precision tended to balance (recall decreased from 79.53% to 75.98%, while precision increased from 64.38% to 78.81%). This indicates that dynamic weighting effectively alleviates the class imbalance problem, enhances attention to the minority class, and thus significantly improves classification performance without changing the model structure or data distribution.

[0061] Comparative Experiment 6 (using only stacking to replace stage C, F1 = 72.07±0.78%) shows that simple stacking cannot fully learn the complementary information between stage A and stage B models in terms of prediction uncertainty and classification preference, resulting in lower F1 and precision rates compared to the complete framework using meta-learning fusion. The method of this invention further improves the F1 score by approximately 6.8 percentage points compared to the stacking alternative, with a particularly significant improvement in precision (approximately 14.1%), indicating that the meta-learning fusion approach can more effectively suppress false positives and improve decision stability.

[0062] To further investigate the impact of each component in Stage B on classification performance, ablation experiments were conducted on internal modules. Results showed that removing the depthwise separable convolutional module reduced the F1 score to 69.97±1.70%, while removing the SE-Net attention mechanism reduced it to 72.58±2.29%. Furthermore, training using only single-modality fetal heart rate signals (i.e., without integrating clinical text information) also resulted in a decrease in the F1 score to 74.62±2.11%. These results indicate that relying solely on signal data or lacking in-depth analysis of signal structure is insufficient to improve model performance and may even lead to performance degradation. Therefore, deep fusion of clinical text information and fetal heart rate signals is essential to effectively enhance the model's ability to discriminate difficult samples, thereby achieving better classification results. The above results support the following conclusions: First, the dynamic weighting strategy can effectively alleviate class imbalance and improve the ability to identify minority classes. Second, stage B enhances the representation ability of difficult samples through structured enhancement (depth-separable convolution, SE attention, and multimodal fusion), while stage C adaptively fuses the classification preferences and uncertainties of the two models through meta-learning. The synergistic effect of these two approaches significantly reduces false positives / false negatives, thereby achieving a comprehensive and stable improvement in F1, precision, and recall. These ablation results verify the effectiveness and clinical applicability of the framework proposed in this invention for medical signal classification tasks with few samples and class imbalance.

[0063] Table 4:

[0064] Generalization experiments demonstrate the effectiveness of dynamic weighting: To verify the generalization performance of the proposed method under different data distributions and class imbalance scenarios, this embodiment further conducted generalization experiments on the MIT-BIH arrhythmia dataset. This dataset is an authoritative public benchmark in the field of electrocardiogram (ECG) signal analysis, containing 48 dual-channel dynamic ECG records (each approximately half an hour long, with a sampling frequency of 360Hz), and the heartbeat types were labeled by experts. The distribution of various heartbeat types in the dataset is highly imbalanced: normal beats (N) account for the vast majority, while some types (such as fusion heartbeats) have very few samples. This distribution characteristic is similar to the imbalance problem of fetal acidosis categories in the CTU-CHB dataset, and therefore is often used to evaluate the robustness and generalization ability of algorithms on real clinical imbalanced data. In this experiment, continuous ECG signals were divided into segments with a 10-second window, and a five-class classification task was performed. Since the segmented category F (fusion heartbeat) only contains 2 samples, it cannot constitute an effective training and testing segment in cross-validation, and it is also difficult to support the model to fully learn and reliably classify this category. Therefore, to ensure the validity of the experiment and the interpretability of the results, this study excluded class F in subsequent analyses and used only the remaining four classes to evaluate generalization performance. SMOTE (Synthetic Minority Over-sampling Technique) is a classic data-level class balancing algorithm. Its core idea is to alleviate class imbalance by interpolating minority class samples within the feature space to synthesize new samples. Specifically, for each minority class sample, SMOTE randomly selects one of its k (k=5) nearest neighbors and then randomly selects a point on the line connecting the two points to generate a new sample. This method effectively increases the diversity of minority class samples and avoids the risk of overfitting caused by simply copying samples. However, SMOTE has a limitation: it may generate samples without practical clinical significance, especially in high-dimensional or sparse feature spaces. Synthetic samples may fall into the majority class region, introducing noise and thus reducing the classifier's generalization performance on the real test set.

[0065] Focal Loss is a strategy that addresses class imbalance at the loss function level. By dynamically adjusting the loss weights for each sample, it makes the model more focused on samples that are difficult to classify. Its core improvement is the addition of a modulation factor to the standard cross-entropy loss. ,in It is the model's predicted probability of the true class. ( =2) is an adjustable focusing parameter. This factor reduces the number of easily classified samples ( The loss contribution of high (high) samples, while relatively increasing the number of hard-to-classify samples ( Focal Loss uses low (low) weights. It is particularly suitable for tasks with extreme foreground-background imbalance, such as object detection. However, its performance is highly dependent on hyperparameters. It is quite sensitive, and when dealing with severely imbalanced multi-category medical data, it may still be necessary to combine other strategies to stabilize the optimization direction.

[0066] Class-Balanced Loss (CB Loss) is based on the concept of "effective sample size" and theoretically derives how to set class weights. This loss function assumes that as the number of samples increases, the marginal benefit of adding new samples diminishes. Therefore, class weights should be inversely proportional to the effective sample size of that class, calculated as follows: , in For category The number of samples, ( =0.99) is a hyperparameter.

[0067] CB Loss assigns higher loss weights to classes with smaller sample sizes, thus giving them more attention in model optimization. Compared to simple inverse frequency weighting, CB Loss considers information redundancy caused by data overlap, providing a more reasonable weight estimate.

[0068] As shown in Table 5, the proposed dynamic class weighting method demonstrates significant advantages on multiple datasets. On the heart disease dataset, this method achieved the highest F1 score (97.37% ± 0.12%), with precision and recall both remaining above 97%, exhibiting excellent class imbalance discrimination ability. In the more challenging CTU-CHB fetal acidosis classification task, where class imbalance is more pronounced, the dynamic weighting method improved the F1 score to 76.10% ± 1.71%, significantly outperforming comparative methods such as inverse frequency weighting, SMOTE, Focal Loss, and CB Loss. Especially in minority class identification, dynamic weighting significantly improves precision, indicating that this method has significant advantages in mitigating class imbalance and improving classification accuracy. More importantly, additional validation on the MIT-BIH arrhythmia dataset also confirms that this method can stably improve model performance on physiological signal data with varying degrees of class imbalance. These results consistently demonstrate that the proposed mechanism for dynamically adjusting weights based on real-time recall can adaptively capture the learning states of various categories during training, effectively alleviating the limitations of static weights or resampling strategies, and thus exhibiting excellent generalization ability and clinical applicability potential under different data distributions.

[0069] Table 5:

Claims

1. A fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion, characterized in that, The system includes: Data preprocessing and feature extraction unit: acquires raw fetal heart rate signal and synchronous clinical features; preprocesses the fetal heart rate signal, standardizes the clinical features, and aligns them with the preprocessed fetal heart rate signal to form multimodal sample input; Phase A Basic Classification Unit: Construct a basic classification network based on a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a multi-head self-attention mechanism. Perform end-to-end training on the preprocessed fetal heart rate signal and output the classification results. Introduce an adaptive dynamic weighted loss function based on recall to dynamically adjust the loss weights of each category and identify difficult samples, including misclassified samples and low-confidence samples, to provide training data for the next phase. Stage B Difficult Sample Multimodal Enhancement Training Unit: For the difficult samples identified in Stage A, a multimodal enhancement classifier is constructed for classification. The classifier combines the clinical features corresponding to the difficult samples for fusion processing and outputs the second classification result. The multimodal enhancement classifier uses depthwise separable convolution and SE attention mechanism to extract subtle features from the fetal heart rate signal. Stage C Meta-learning Fusion Decision Output Unit: Based on the predicted outputs of Stage A and Stage B on the validation set, construct meta-feature vectors, input them into the meta-feature fusion model for training, adaptively fuse the classification preferences and uncertainties of the two-stage models, and output the final classification result to achieve robust decision-making.

2. The fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion according to claim 1, characterized in that, The preprocessing of the fetal heart rate signal specifically includes: initializing the original fetal heart rate signal by replacing the 0 value of FHR with NaN; detecting missing values, if the maximum missing value is less than 15 seconds, linear interpolation is used for local repair, otherwise the missing value is directly deleted; detecting spike artifacts, if the absolute value between two adjacent points is greater than 25 bpm, it is considered noise, and the region with a difference of less than 10 bpm between five consecutive heartbeats is defined as a stable segment, and linear interpolation is used to replace the value of the noise with the linear transition value between the previous stable segment and the subsequent new stable segment; detecting abnormal ranges, if the FHR value is greater than 200 bpm or less than 50 bpm, Hermite spline interpolation is used for filling.

3. The fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion according to claim 2, characterized in that, The basic classification network consists of a four-layer one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a multi-head attention mechanism.

4. The fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion according to claim 3, characterized in that, The adaptive dynamic weighted loss function is set as follows: S41 Weight Initialization and Update Objective: At the start of training, the weights of all classes are initialized to 1. After each training cycle, the recall of each class is calculated based on the model's performance on the current training set. ,in Representing the category index, a target weight vector is generated. : in, The total number of categories, As the amplification factor, Constants to prevent numerical instability; S42 Smooth Update Strategy: Use a smoothing method to adjust the current category weights Transition to target weights, training rounds during a predefined warm-up phase. Inside, the next round of sub-weighting Update using linear interpolation: ;in, Indicates the current training round; After the warm-up phase, the next round of weighting The model will be updated smoothly using an exponentially weighted average method to avoid instability in model training caused by sudden changes in weights. ;in, For smoothing coefficients; Combining S43 with the loss function: The category weights are dynamically calculated using a smooth update strategy. When applied to the standard cross-entropy loss function, the resulting cross-entropy loss function is: in, It is the batch size, when the sample belongs to the first... When class, ,otherwise , It is the predicted probability distribution vector. This represents the summation over all categories.

5. The fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion according to claim 4, characterized in that, The identification of difficult samples is as follows: the trained Stage A basic classification network is used to classify the original training set, the classification probability of each sample on its true label is calculated as a confidence index, and the average confidence level of the entire training set is calculated. Difficult samples can be identified if they meet any of the following conditions: misclassification: the model predicts a class that is inconsistent with the true label; low confidence: although the prediction is correct, the confidence level is lower than the average confidence level of the training set.

6. The fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion according to claim 5, characterized in that, The multimodal augmentation classifier includes a two-channel data processing path. The input in the first channel is the difficult sample identified in stage A, and the input in the second channel is the clinical features corresponding to the difficult sample. In the first channel, the hard sample input feature extraction module performs feature extraction and attention weighting to obtain the feature sequence output. The feature sequence is input into the BilSTM module to capture long-term time-series information and obtain the time-series feature output. The time-series feature is input into the attention pooling layer to learn the attention weights of each time step, and the time-series features are weighted and summed to output the global attention aggregated features. The global attention aggregated features are then transformed nonlinearly through a fully connected layer to output the final fetal heart rate signal feature representation. In the second channel, the clinical features corresponding to the input difficult samples are first processed by the batch normalization module to obtain standardized and normalized clinical feature vectors. Then, the clinical feature vectors are nonlinearly transformed by the fully connected layer to obtain intermediate feature representations. The intermediate feature representations are then standardized again by the batch normalization module to output stable distributed multidimensional features. Finally, the multidimensional features are nonlinearly transformed by the fully connected layer to output the final clinical feature representation. The final fetal heart rate signal feature representation and the final clinical feature representation are concatenated by a fully connected layer to obtain fused features. The fused features are then mapped to the original classification probability output of the three categories by the output layer.

7. A fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion according to claim 6, characterized in that, The meta-feature vector includes the following features: Basic probability characteristics: including the category probability distributions for stages A and B; Confidence and uncertainty characteristics: including classification confidence for stages A and B, the difference in confidence between the two models for stages A and B, and the entropy of classification for stages A and B; Prediction consistency features: including whether the predicted categories of the two models in stages A and B are consistent; Probability difference characteristics: including the maximum and average absolute difference between the probability vectors of the two models in stage A and stage B, and the probability difference between the two models in stage A and stage B by category.

8. A fetal acidosis classification system based on multi-stage mining and multimodal meta-fusion according to claim 7, characterized in that, The meta-feature fusion model consists of three hidden layers and an output layer. In the hidden layers, the input meta-feature vectors are linearly transformed, normalized, activated by the ReLU function, and regularized by Dropout. Finally, the output layer maps the data to three categories.

Images