A bladder cancer diagnosis auxiliary system based on machine learning

Through multi-spectral detection and explanatory enhancement of graph convolutional long and short-term memory network, the problems of data quality and feature extraction accuracy in the bladder cancer diagnosis assistive system are solved, and the reliability and interpretability of bladder cancer diagnosis are improved.

CN118658610BActive Publication Date: 2024-11-15SECOND MEDICAL CENT OF CHINESE PLA GENERAL HOSPITAL +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411132051.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-15
Estimated Expiration
2044-08-19

AI Technical Summary

Technical Problem

The existing bladder cancer diagnosis assistive system is difficult to collect and analyze image data, lack of accuracy and accuracy of single peak analysis of biomarkers, the messy multimodal analysis data quality and complex multimodal data types, resulting in feasibility and accuracy of model construction. The rate is low, and the timing and local feature extraction accuracy of multi-spectral detection is high, but the model interpretability is poor.

Method used

Using a multispectral detection method combined with urine detection data, data preprocessing is performed through missing value optimization, extreme value normalization, data alignment and data fusion, and using an interpretable enhanced graph convolutional length short-term memory network to satisfy global, local and timing model of features and optimize interpretability through attention mechanism and SHAP value calculation.

Benefits of technology

It improves the feasibility of data source and model prediction performance of the bladder cancer diagnosis auxiliary system, improves the reliability and interpretability of the diagnosis, and solves the problems of data quality and feature extraction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118658610B_ABST
    Figure CN118658610B_ABST
Patent Text Reader

Abstract

The present invention discloses a bladder cancer diagnosis assistance system based on machine learning, including a label data collection module, a data preprocessing module, a multispectral detection module and a diagnosis assistance module. The present invention belongs to the technical field of bladder cancer biomarker detection, specifically a bladder cancer diagnosis assistance system based on machine learning. The present invention combines the multispectral detection method of urine test data to perform bladder cancer diagnosis assistance, optimizes the data source suggestion and overall feasibility of bladder cancer diagnosis assistance; adopts missing value optimization, extreme value normalization, data alignment optimization and data fusion methods for data preprocessing, improves the data quality of subsequent multispectral detection, and indirectly improves the prediction performance of the model; adopts an interpretability-enhanced graph convolutional long short-term memory network, and optimizes interpretability through attention mechanism and SHAP value calculation, thereby improving the feasibility and reliability of diagnosis assistance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of bladder cancer biomarker detection, and specifically is a bladder cancer diagnosis auxiliary system based on machine learning. Background Art

[0002] A machine learning-based bladder cancer diagnosis assistance system is a tool that uses advanced algorithms and data analysis techniques to automatically analyze patients' medical images, pathological data, clinical information, or biological characteristics of the disease, thereby helping doctors diagnose bladder cancer more accurately and quickly. Its main function is to improve the accuracy of diagnosis, reduce the misdiagnosis rate, optimize treatment plans, and improve medical efficiency.

[0003] However, in the existing bladder cancer diagnosis assistance system, there is a technical problem that the traditional bladder cancer diagnosis assistance system often focuses on the single-peak analysis of basic medical images or biomarkers, and the collection and analysis of image data is difficult, and it is difficult to ensure reliability and promotion feasibility, while the single-peak analysis of biomarkers lacks the support of accuracy and precision; in the existing bladder cancer diagnosis assistance system, there is a technical problem that the multi-peak analysis biomarkers in the traditional method have unavoidable messy data quality and complex multimodal data types, which in turn affects the feasibility of model construction and subsequent accuracy; in the existing bladder cancer diagnosis assistance system, there is a technical problem that when multi-spectral detection is implemented, the extraction accuracy of time series features and local protein features is high, and the fused multi-spectral data has multimodal characteristics, which requires consideration of both global features and local features and time series features, and an overly complex model design will reduce the interpretability of the model, which is not conducive to the specific application of bladder cancer diagnosis assistance. Therefore, the existing methods have technical problems that need to be improved in terms of global, time series, local features and interpretability. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a bladder cancer diagnosis assistance system based on machine learning. In view of the technical problem that in the existing bladder cancer diagnosis assistance system, traditional bladder cancer diagnosis assistance systems often focus on the single-peak analysis of basic medical images or biomarkers, and the collection and analysis of image data are difficult, and it is difficult to ensure reliability and promotion feasibility, and the single-peak analysis of biomarkers lacks the support of accuracy and precision, this solution creatively adopts a multi-spectral detection method combined with urine test data for bladder cancer diagnosis assistance, and optimizes the data source suggestion and overall feasibility of bladder cancer diagnosis assistance; in view of the technical problem that in the existing bladder cancer diagnosis assistance system, there is an inevitable messy data quality and complex multimodal data type in the traditional method of multi-peak analysis of biomarkers, which in turn affects the feasibility of model construction and the subsequent accuracy, this solution creatively adopts missing value optimization, extreme value normalization, data alignment optimization and data Data preprocessing is performed according to the fusion method, which improves the data quality of subsequent multispectral detection and indirectly improves the prediction performance of the model. In the existing bladder cancer diagnosis assistance system, there is a high requirement for the extraction accuracy of temporal features and local protein features during the specific implementation of multispectral detection, and the fused multispectral data has multimodal characteristics, which requires consideration of both global features and local features and temporal features. Overly complex model design will reduce the interpretability of the model, which is not conducive to the specific application of bladder cancer diagnosis assistance. Therefore, the existing methods have technical problems that need to be improved in terms of global, temporal, local features and interpretability. This scheme creatively adopts an interpretability-enhanced graph convolutional long short-term memory network, and constructs a basic subnetwork that satisfies the processing of global features, local features and temporal features by constructing stacked convolution, long short-term and graph networks. The interpretability is collaboratively optimized through the attention mechanism and SHAP value calculation, which improves the feasibility and reliability of diagnosis assistance.

[0005] The present invention provides a bladder cancer diagnosis auxiliary system based on machine learning, comprising a label data collection module, a data preprocessing module, a multi-spectral detection module and a diagnosis auxiliary module;

[0006] The marker data collection module is used to collect protein marker information, obtain bladder cancer protein marker raw data through marker data collection, and send the bladder cancer protein marker raw data to the data preprocessing module;

[0007] The data preprocessing module is used to optimize the collected data, obtain bladder cancer protein marker optimization data through data preprocessing, and send the bladder cancer protein marker optimization data to the multi-spectral detection module;

[0008] The multispectral detection module is used to design multi-peak analysis feature engineering of biomarkers and perform machine learning multispectral detection of bladder cancer protein marker prediction, obtain bladder cancer biomarker detection data through multispectral detection, and send the bladder cancer biomarker detection data to the diagnosis auxiliary module;

[0009] The diagnosis assistance module is used to assist in bladder cancer diagnosis using the model prediction results, and obtain bladder cancer diagnosis assistance reference information through diagnosis assistance.

[0010] Furthermore, the marker data collection is used to collect protein marker information, specifically, to obtain bladder cancer derived urine detection protein marker raw data by collecting clinical medical data and bladder cancer urine test data, and to extract bladder cancer protein marker raw data from the bladder cancer derived urine detection protein marker raw data;

[0011] The bladder cancer protein marker raw data specifically includes fluorescence spectrum detection raw data, coupled inductively coupled plasma mass spectrometry raw data and ultraviolet visible absorption spectrum detection raw data.

[0012] Furthermore, the data preprocessing is used to optimize the collected data, specifically to perform data cleaning, data normalization, data alignment and data fusion operations on the original data of bladder cancer protein markers to obtain bladder cancer protein marker optimized data, and specifically includes the following steps:

[0013] Missing value processing, specifically using multiple interpolation methods to process missing values, generating fill-in data by training a generative adversarial network or a random generation method with a limited data range, performing missing value optimization, and obtaining a missing fill-in data set;

[0014] Data normalization processing, specifically using extreme value normalization method to normalize data, optimize the influence of abnormal values ​​generated in the process of missing value processing and the original abnormal values ​​of the data, and obtain a normalized data set;

[0015] The data alignment process, specifically, performing overall data alignment process of the normalized data set through time alignment and spectrum alignment, comprises the following steps:

[0016] Time alignment processing, specifically, using a dynamic time warping method to perform time alignment processing to obtain dynamic time warping data;

[0017] Spectral alignment processing, specifically using a wavelet transform method to perform spectral alignment processing on the dynamic time warping data to obtain spectral alignment data;

[0018] Data fusion processing, specifically, using an embedded representation method to embed and fuse the fluorescence spectrum detection raw data, coupled inductively coupled plasma mass spectrometry raw data, and ultraviolet visible absorption spectrum detection raw data that have undergone missing value processing, data normalization processing, and data alignment processing to obtain embedded fused data;

[0019] The integrated preprocessing is specifically to obtain the embedded fusion data and then perform integrated preprocessing on the original data of the bladder cancer protein marker to obtain the optimized data of the bladder cancer protein marker.

[0020] Furthermore, the multispectral detection is used to design multi-peak analysis feature engineering of biomarkers and to predict bladder cancer protein markers by machine learning multispectral detection, specifically, based on the bladder cancer protein marker optimization data, an interpretability-enhanced graph convolutional long short-term memory network is used to perform multispectral detection to obtain bladder cancer biomarker detection data;

[0021] The graph convolutional long short-term memory network with enhanced interpretability specifically includes a multi-layer convolutional pooling subnet, a bidirectional stacked long short-term memory subnet, a graph neural subnet, an attention mechanism, a classification output layer, and an interpretability optimization operator;

[0022] The multi-layer convolutional pooling subnetwork is used to extract local features of multispectral detection from fused data;

[0023] The bidirectional stacked long short-term memory subnetwork is used to optimize the capture of temporal features in the fused data and enhance the prediction coherence;

[0024] The graph neural subnetwork is used to fuse local features of different spectral data and extract global relationship features;

[0025] The attention mechanism is used to extract important features while optimizing model performance and interpretability;

[0026] The explanatory optimization operator is used to assist in feature importance calculation in combination with the attention mechanism to enhance model interpretability;

[0027] The step of using an interpretability-enhanced graph convolutional long short-term memory network to perform multispectral detection and obtain bladder cancer biomarker detection data includes constructing a multi-layer convolutional pooling subnet, constructing a bidirectional stacked long short-term memory subnet, constructing a graph neural subnet, constructing an attention mechanism, constructing an interpretable optimization operator, model optimization training and multispectral detection, and specifically includes the following steps:

[0028] Construct a multi-layer convolutional pooling subnetwork, specifically, extract local features by constructing a convolutional layer and perform feature pooling through a maximum pooling layer to obtain a local feature set of spectral data;

[0029] Constructing a bidirectional stacked long short-term memory subnet, specifically, sequentially constructing a forward stacked long short-term memory network and a backward stacked long short-term memory network, performing bidirectional stacked long short-term memory subnet construction, and performing time series feature processing by constructing the bidirectional stacked long short-term memory subnet to obtain a spectral time series feature output;

[0030] Construct a graph neural subnetwork, specifically, construct a multi-layer graph convolutional network, and obtain aggregated feature data by stacking the outputs of multiple layers of graph convolutional networks;

[0031] Construct an attention mechanism to calculate the importance weight of each spectral feature and optimize interpretability, and perform attention weighting to obtain weighted features;

[0032] Construct an interpretable optimization operator, specifically by calculating the SHAP value to represent the importance of each spectral feature data, and use it as an indicator to measure the interpretability of the model. The calculation formula is:

[0033]

[0034] In the formula, Φ i is the SHAP value corresponding to the i-th feature, which is used to measure the importance of the feature and the interpretability of the model prediction results. N is the total set of features, i is the feature index, and S N is a feature subset, ! is a factorial operator, |·| is a modulo operator, f(·) is a model output representation function, and ∪ is a union operator;

[0035] Model optimization training, specifically, performing initial model training by constructing a multi-layer convolutional pooling subnet, constructing a bidirectional stacked long short-term memory subnet, constructing a graph neural subnet, and constructing an attention mechanism, calculating the SHAP value obtained by the explanatory optimization operator, and performing iterative training optimization by using a cross entropy loss function and an Adam optimization algorithm to obtain an optimized multispectral detection model Model MD ;

[0036] Multispectral detection, specifically using the optimized multispectral detection model Model MD , based on the bladder cancer protein marker optimization data, multispectral detection is performed to obtain bladder cancer biomarker detection data.

[0037] Furthermore, the diagnostic assistance is used to use the model prediction results to assist in the diagnosis of bladder cancer, specifically, based on the bladder cancer biomarker detection data, a biomarker analysis of bladder cancer urine samples is performed, and through extracellular vesicle characterization analysis, bladder cancer diagnostic assistance reference information is obtained.

[0038] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0039] (1) In view of the technical problems that the existing bladder cancer diagnosis assistance systems often focus on the single-peak analysis of basic medical images or biomarkers, the image data collection and analysis are difficult to ensure reliability and promotion feasibility, and the single-peak analysis of biomarkers lacks the support of accuracy and precision, this scheme creatively adopts a multi-spectral detection method combined with urine test data for bladder cancer diagnosis assistance, optimizing the data source suggestion and overall feasibility of bladder cancer diagnosis assistance;

[0040] (2) In the existing bladder cancer diagnosis auxiliary system, there are technical problems that the traditional methods of multi-modal analysis of biomarkers have unavoidable data quality problems and complex multi-modal data types, which in turn affect the feasibility of model construction and subsequent accuracy. This solution creatively uses missing value optimization, extreme value normalization, data alignment optimization and data fusion methods for data preprocessing, which improves the data quality of subsequent multi-spectral detection and indirectly improves the prediction performance of the model.

[0041] (3) In the existing bladder cancer diagnosis assistance system, there is a problem that when multispectral detection is implemented, the extraction accuracy of temporal features and local protein features is high, and the fused multispectral data has multimodal characteristics. It is necessary to consider both global features and local features and temporal features. Overly complex model design will reduce the interpretability of the model, which is not conducive to the specific application of bladder cancer diagnosis assistance. Therefore, the existing methods have technical problems that need to be improved in terms of global, temporal, local features and interpretability. This scheme creatively adopts an interpretability-enhanced graph convolutional long short-term memory network. By constructing stacked convolution, long short-term and graph networks, a basic subnetwork that satisfies the processing of global features, local features and temporal features is constructed. The interpretability is synergistically optimized through the attention mechanism and SHAP value calculation, thereby improving the feasibility and reliability of diagnosis assistance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of the structure of a bladder cancer diagnosis auxiliary system based on machine learning provided by the present invention;

[0043] Figure 2 A flowchart of the steps performed by the data preprocessing module;

[0044] Figure 3 A schematic diagram of the process of the steps performed by the multi-spectral detection module;

[0045] Figure 4 Schematic diagram for visualizing the output of the multispectral detection module.

[0046] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] Example 1, see Figure 1 and Figure 2 , a bladder cancer diagnosis auxiliary system based on machine learning provided by the present invention includes a label data collection module, a data preprocessing module, a multispectral detection module and a diagnosis auxiliary module;

[0049] The marker data collection module is used to collect protein marker information, obtain bladder cancer protein marker raw data through marker data collection, and send the bladder cancer protein marker raw data to the data preprocessing module;

[0050] The data preprocessing module is used to optimize the collected data, obtain bladder cancer protein marker optimization data through data preprocessing, and send the bladder cancer protein marker optimization data to the multi-spectral detection module;

[0051] The multispectral detection module is used to design multi-peak analysis feature engineering of biomarkers and perform machine learning multispectral detection of bladder cancer protein marker prediction, obtain bladder cancer biomarker detection data through multispectral detection, and send the bladder cancer biomarker detection data to the diagnosis auxiliary module;

[0052] The diagnosis assistance module is used to assist in bladder cancer diagnosis using the model prediction results, and obtain bladder cancer diagnosis assistance reference information through diagnosis assistance.

[0053] By performing the above operations, in view of the technical problem that in the existing bladder cancer diagnosis assistance system, the traditional bladder cancer diagnosis assistance system often focuses on the single-peak analysis of basic medical images or biomarkers, and the image data collection and analysis are difficult, it is difficult to ensure reliability and promotion feasibility, and the single-peak analysis of biomarkers lacks the support of accuracy and precision, this solution creatively adopts a multi-spectral detection method combined with urine test data for bladder cancer diagnosis assistance, and optimizes the data source suggestion and overall feasibility of bladder cancer diagnosis assistance.

[0054] Embodiment 2, this embodiment is based on the above embodiment, see Figure 1 The marker data collection is used to collect protein marker information, specifically, to obtain bladder cancer derived urine detection protein marker raw data from clinical medical data and bladder cancer urine test data, and to extract bladder cancer protein marker raw data from the bladder cancer derived urine detection protein marker raw data;

[0055] The bladder cancer protein marker raw data specifically includes fluorescence spectrum detection raw data, coupled inductively coupled plasma mass spectrometry raw data and ultraviolet visible absorption spectrum detection raw data.

[0056] Embodiment 3, this embodiment is based on the above embodiment, see Figure 2 The data preprocessing is used to optimize the collected data, specifically to perform data cleaning, data normalization, data alignment and data fusion operations on the original data of bladder cancer protein markers to obtain bladder cancer protein marker optimized data, and specifically includes the following steps:

[0057] Missing value processing, specifically using multiple interpolation methods to process missing values, generating fill-in data by training a generative adversarial network or a random generation method with a limited data range, performing missing value optimization, and obtaining a missing fill-in data set;

[0058] Data normalization processing, specifically using extreme value normalization method to normalize data, optimize the influence of abnormal values ​​generated in the process of missing value processing and the original abnormal values ​​of the data, and obtain a normalized data set;

[0059] The calculation formula of the extreme value normalization method is:

[0060]

[0061] Where x′ is the normalized data set, x is the missing fill data set, MID(·) is the median calculation function, and IQR(·) is the interquartile range calculation function;

[0062] The data alignment process, specifically, performing overall data alignment process of the normalized data set through time alignment and spectrum alignment, comprises the following steps:

[0063] The time alignment process is specifically to use the dynamic time warping method to perform time alignment process to obtain dynamic time warping data. The calculation formula is:

[0064]

[0065] Where DTW(·) is the dynamic time warping function, A is the first time series data, which is used to represent any sequence data to be time aligned, B is the second time series data, K is the total number of data points, k is the data point index, d(·) is the Euclidean distance calculation function, and a ik is the kth data point corresponding to the i-th position in the first time series data, b jk is the kth data point corresponding to the jth position in the second time series data;

[0066] Spectral alignment processing, specifically using a wavelet transform method to perform spectral alignment processing on the dynamic time warping data to obtain spectral alignment data;

[0067] Data fusion processing, specifically, using an embedded representation method to embed and fuse the fluorescence spectrum detection raw data, coupled inductively coupled plasma mass spectrometry raw data, and ultraviolet visible absorption spectrum detection raw data that have undergone missing value processing, data normalization processing, and data alignment processing to obtain embedded fused data;

[0068] The calculation formula of the embedded fusion representation is:

[0069] E fused =Concat(E FL , E IM , E Abs );

[0070] In the formula, E fused is the embedded fusion data, Concat(·) is the vector concatenation operation, E FL is the fluorescence spectrum detection alignment data, E IM is the coupled inductively coupled plasma mass spectrometry alignment data, and EAbs is the UV-visible absorption spectroscopy detection alignment data;

[0071] The integrated preprocessing is specifically to obtain the embedded fusion data and then perform integrated preprocessing on the original data of the bladder cancer protein marker to obtain the optimized data of the bladder cancer protein marker.

[0072] By performing the above operations, in order to solve the technical problems in the existing bladder cancer diagnosis auxiliary system, there are inevitable data quality chaos and complex multimodal data types in traditional methods of multi-peak analysis biomarkers, which in turn affects the feasibility of model construction and subsequent accuracy. This solution creatively uses missing value optimization, extreme value normalization, data alignment optimization and data fusion methods for data preprocessing, which improves the data quality of subsequent multi-spectral detection and indirectly improves the prediction performance of the model.

[0073] Embodiment 4, this embodiment is based on the above embodiment, see Figure 3 and Figure 4The multispectral detection is used to design multi-peak analysis feature engineering of biomarkers and perform machine learning multispectral detection of bladder cancer protein marker prediction, specifically, based on the bladder cancer protein marker optimization data, using an interpretability-enhanced graph convolutional long short-term memory network to perform multispectral detection to obtain bladder cancer biomarker detection data;

[0074] The graph convolutional long short-term memory network with enhanced interpretability specifically includes a multi-layer convolutional pooling subnet, a bidirectional stacked long short-term memory subnet, a graph neural subnet, an attention mechanism, a classification output layer, and an interpretability optimization operator;

[0075] The multi-layer convolutional pooling subnetwork is used to extract local features of multispectral detection from fused data;

[0076] The bidirectional stacked long short-term memory subnetwork is used to optimize the capture of temporal features in the fused data and enhance the prediction coherence;

[0077] The graph neural subnetwork is used to fuse local features of different spectral data and extract global relationship features;

[0078] The attention mechanism is used to extract important features while optimizing model performance and interpretability;

[0079] The explanatory optimization operator is used to assist in feature importance calculation in combination with the attention mechanism to enhance model interpretability;

[0080] The step of using an interpretability-enhanced graph convolutional long short-term memory network to perform multispectral detection and obtain bladder cancer biomarker detection data includes constructing a multi-layer convolutional pooling subnet, constructing a bidirectional stacked long short-term memory subnet, constructing a graph neural subnet, constructing an attention mechanism, constructing an interpretable optimization operator, model optimization training and multispectral detection, and specifically includes the following steps:

[0081] Construct a multi-layer convolutional pooling subnet. Specifically, extract local features by constructing a convolutional layer and perform feature pooling through a maximum pooling layer to obtain a local feature set of spectral data. The calculation formula is:

[0082]

[0083] In the formula, H l is the convolution feature of spectral data, P l is the local feature set of spectral data after pooling, ReLU(·) is a nonlinear activation function, Conv1D(·) is a one-dimensional convolutional layer function, X is the original input data, which is used to represent the bladder cancer protein marker optimization data, W l is the convolutional layer weight, b lis the convolutional layer bias term, where l is the convolutional layer index, k is the convolution kernel size parameter, s is the step size parameter, MaxPooling1D(·) is the maximum pooling function corresponding to the one-dimensional convolutional layer, p is the pooling window size parameter, and s p is the pooling step size parameter;

[0084] The maximum value of the convolution layer index l is 7, which is used to indicate the total number of convolution layers;

[0085] Constructing a bidirectional stacked long short-term memory subnet, specifically, sequentially constructing a forward stacked long short-term memory network and a backward stacked long short-term memory network, performing bidirectional stacked long short-term memory subnet construction, and performing time series feature processing by constructing the bidirectional stacked long short-term memory subnet to obtain a spectral time series feature output;

[0086] The calculation formula for constructing the bidirectional stacked long short-term memory subnet is:

[0087]

[0088] In the formula, is the final hidden state output of the forward stacked LSTM network, where l′ is the level index of the stacked LSTM network, LSTM(·) is the basic LSTM neural network identification function, P is the local feature set of spectral data, is the weight value of the forward stacked LSTM network, b l′ is the bias term of the stacked LSTM network, is the final hidden state output of the backward stacked LSTM network, is the weight value of the backward stacked long short-term memory network, O l′ is the spectral time series feature output, [·,·] is the output aggregation operator;

[0089] The maximum value of the layer index l′ of the stacked long short-term memory network is set to 3, which is used to represent the number of layers of the stacked long short-term memory network;

[0090] Construct a graph neural subnet, specifically construct a multi-layer graph convolutional network, and obtain aggregated feature data by stacking the outputs of multiple layers of graph convolutional networks. The calculation formula is:

[0091]

[0092] In the formula, is the output feature of the Lth layer graph convolutional network, L is the graph convolutional network level index, Sig(·) is the S-type activation function, A is the graph structure adjacency matrix, W L is the weight value of the Lth layer graph convolutional network, H FIt is the aggregate feature data used to represent the total output of the graph convolutional network stacked at multiple levels. is the output feature of the graph convolutional network with a total of L′ layers, where L′ is the total number of multi-layer graph convolutional networks;

[0093] The specific value of the total number L′ of the multi-layer graph convolutional network is 5;

[0094] An attention mechanism is constructed to calculate the importance weight of each spectral feature and optimize the interpretability, and to perform attention weighting to obtain weighted features. The calculation formula is:

[0095]

[0096] In the formula, e i is the attention score data of the i-th feature, tanh(·) is the hyperbolic tangent function, W e is the attention score weight, O i is the original input feature, used to represent the aggregated feature data, b e is the attention score bias term, α i is the normalized attention weight, exp(·) is the natural base function, J is the total number of normalized summed indexes of attention scores, j is the normalized summed index of attention scores, and e j is the normalized summed attention score data of the jth feature, H att is the weighted feature, I is the total number of features, and i is the feature index;

[0097] Construct an interpretable optimization operator, specifically by calculating the SHAP value to represent the importance of each spectral feature data, and use it as an indicator to measure the interpretability of the model. The calculation formula is:

[0098]

[0099] In the formula, Φ i is the SHAP value corresponding to the i-th feature, which is used to measure the importance of the feature and the interpretability of the model prediction results. N is the total set of features, i is the feature index, and S N is a feature subset, ! is a factorial operator, |·| is a modulo operator, f(·) is a model output representation function, and ∪ is a union operator;

[0100] Model optimization training, specifically, performing initial model training by constructing a multi-layer convolutional pooling subnet, constructing a bidirectional stacked long short-term memory subnet, constructing a graph neural subnet, and constructing an attention mechanism, calculating the SHAP value obtained by the explanatory optimization operator, and performing iterative training optimization by using a cross entropy loss function and an Adam optimization algorithm to obtain an optimized multispectral detection model Model MD ;

[0101] Multispectral detection, specifically using the optimized multispectral detection model Model MD , based on the bladder cancer protein marker optimization data, multispectral detection is performed to obtain bladder cancer biomarker detection data.

[0102] See also Figure 4 As shown in the figure, the horizontal direction is the data points of multi-spectral data detection, and the vertical direction is the prediction results of the optimized multi-spectral detection model. Among them, the dark black waveform FL is used to represent the fluorescence spectrum detection prediction data, the gray waveform IM is used to represent the coupled inductive plasma mass spectrometry prediction data, and the light gray waveform Abs is used to represent the ultraviolet-visible absorption spectrum detection prediction data.

[0103] By performing the above operations, in the existing bladder cancer diagnosis assistance system, there is a problem that in the specific implementation of multispectral detection, the extraction accuracy of temporal features and local protein features is high, and the fused multispectral data has multimodal characteristics. It is necessary to consider both global features and local features and temporal features. Overly complex model design will reduce the interpretability of the model, which is not conducive to the specific application of bladder cancer diagnosis assistance. Therefore, the existing methods have technical problems that need to be improved in terms of global, temporal, local features and interpretability. This scheme creatively adopts an interpretability-enhanced graph convolutional long short-term memory network, and constructs a basic subnetwork that satisfies the processing of global features, local features and temporal features by constructing stacked convolution, long short-term and graph networks. The interpretability is collaboratively optimized through the attention mechanism and SHAP value calculation, thereby improving the feasibility and reliability of diagnosis assistance.

[0104] Embodiment 5, this embodiment is based on the above embodiment, see Figure 1 The diagnostic aid is used to use the model prediction results to assist in the diagnosis of bladder cancer, specifically to perform biomarker analysis of bladder cancer urine samples based on the bladder cancer biomarker detection data, and to obtain bladder cancer diagnostic aid reference information through extracellular vesicle characterization analysis.

[0105] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process and method including a series of elements include not only those elements, but also other elements not explicitly listed, or also include elements inherent to such process and method.

[0106] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and alterations may be made to the embodiments without departing from the principles and spirit of the invention.

[0107] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A bladder cancer diagnosis assistance system based on machine learning, characterized by: It includes a label data collection module, a data preprocessing module, a multispectral detection module and a diagnosis auxiliary module; The marker data collection module is used to collect protein marker information, obtain bladder cancer protein marker raw data through marker data collection, and send the bladder cancer protein marker raw data to the data preprocessing module; The data preprocessing module is used to optimize the collected data, obtain optimized bladder cancer protein marker data through data preprocessing, and send the optimized bladder cancer protein marker data to the multispectral detection module; The data preprocessing is used to optimize the collected data, specifically performing data cleaning, data normalization, data alignment and data fusion operations on the raw data of bladder cancer protein markers to obtain optimized data of bladder cancer protein markers, and specifically includes the following steps: Missing value processing, specifically using multiple interpolation methods to handle missing values, generating imputed data by training a generative adversarial network or a random generation method with a limited data range, performing missing value optimization, and obtaining a missing imputed data set; Data normalization processing, specifically using the extreme value normalization method to perform data normalization processing, optimize the impact of abnormal values ​​generated during the missing value processing and the original abnormal values ​​of the data, and obtain a normalized data set; Data alignment processing, specifically performing overall data alignment processing of the normalized data set through time alignment and spectral alignment; Data fusion processing, specifically, using an embedding representation method to embed and fuse the fluorescence spectrum detection raw data, coupled inductively coupled plasma mass spectrometry raw data, and ultraviolet-visible absorption spectrum detection raw data that have undergone missing value processing, data normalization processing, and data alignment processing to obtain embedded fused data; Integrated preprocessing, specifically, performing integrated preprocessing on the original data of the bladder cancer protein marker by obtaining the embedded fusion data to obtain optimized data of the bladder cancer protein marker; The multispectral detection module is used to design multimodal analysis feature engineering for biomarkers and perform machine learning multispectral detection prediction of bladder cancer protein markers. Specifically, based on the bladder cancer protein marker optimization data, an interpretability-enhanced graph convolutional long short-term memory network is used to perform multispectral detection to obtain bladder cancer biomarker detection data, and the bladder cancer biomarker detection data is sent to the diagnosis assistance module. The steps of performing multispectral detection using the interpretability-enhanced graph convolutional long short-term memory network to obtain bladder cancer biomarker detection data include constructing a multi-layer convolutional pooling subnet, constructing a bidirectional stacked long short-term memory subnet, constructing a graph neural subnet, constructing an attention mechanism, constructing an interpretability optimization operator, model optimization training, and performing multispectral detection. The diagnosis assistance module is used to assist in bladder cancer diagnosis using the model prediction results, and obtain bladder cancer diagnosis assistance reference information through diagnosis assistance.

2. The machine learning-based bladder cancer diagnosis assistance system according to claim 1, characterized in that: The marker data collection is used to collect protein marker information, specifically by collecting clinical medical data and bladder cancer urine test data to obtain bladder cancer derived urine detection protein marker raw data, and extracting bladder cancer protein marker raw data from the bladder cancer derived urine detection protein marker raw data; The bladder cancer protein marker raw data specifically includes fluorescence spectrum detection raw data, coupled inductively coupled plasma mass spectrometry raw data and ultraviolet-visible absorption spectrum detection raw data.

3. The machine learning-based bladder cancer diagnosis assistance system according to claim 2, characterized in that: The multispectral detection is specifically performed based on the bladder cancer protein marker optimization data, using an interpretability-enhanced graph convolutional long short-term memory network to perform multispectral detection to obtain bladder cancer biomarker detection data; The graph convolutional long short-term memory network with enhanced interpretability specifically includes a multi-layer convolutional pooling subnet, a bidirectional stacked long short-term memory subnet, a graph neural subnet, an attention mechanism, a classification output layer, and an interpretable optimization operator; The multi-layer convolutional pooling subnetwork is used to extract local features of multispectral detection from the fused data; The bidirectional stacked long short-term memory subnetwork is used to optimize the capture of temporal features in the fused data and enhance prediction coherence; The graph neural subnetwork is used to fuse local features of different spectral data and extract global relationship features; The attention mechanism is used to extract important features while optimizing model performance and interpretability; The explanatory optimization operator is used to combine with the attention mechanism to assist in feature importance calculation and enhance model interpretability.

4. The machine learning-based bladder cancer diagnosis assistance system according to claim 3, characterized in that: The step of using an interpretability-enhanced graph convolutional long short-term memory network to perform multispectral detection and obtain bladder cancer biomarker detection data specifically includes: Construct a multi-layer convolutional pooling subnetwork, specifically by constructing a convolutional layer to extract local features and performing feature pooling through a maximum pooling layer to obtain a local feature set of spectral data; Constructing a bidirectional stacked long short-term memory subnetwork, specifically, sequentially constructing a forward stacked long short-term memory network and a backward stacked long short-term memory network, performing bidirectional stacked long short-term memory subnetwork construction, and performing time series feature processing by constructing the bidirectional stacked long short-term memory subnetwork to obtain a spectral time series feature output; Construct a graph neural subnetwork, specifically by building a multi-layer graph convolutional network and obtaining aggregated feature data by stacking the outputs of multiple layers of graph convolutional networks; Construct an attention mechanism to calculate the importance weight of each spectral feature and optimize interpretability, and perform attention weighting to obtain weighted features; Construct an interpretable optimization operator, specifically by calculating the SHAP value to represent the importance of each spectral feature data, and use it as an indicator to measure the interpretability of the model; Model optimization training, specifically, performing initial model training by constructing a multi-layer convolutional pooling subnet, constructing a bidirectional stacked long short-term memory subnet, constructing a graph neural subnet, and constructing an attention mechanism, and obtaining an optimized multispectral detection model by calculating the SHAP value obtained by the explanatory optimization operator and performing iterative training optimization using the cross entropy loss function and the Adam optimization algorithm. MD ; Multispectral detection, specifically using the optimized multispectral detection model Model MD , based on the bladder cancer protein marker optimization data, multispectral detection is performed to obtain bladder cancer biomarker detection data.

5. The machine learning-based bladder cancer diagnosis assistance system according to claim 4, characterized in that: The data alignment process specifically includes the following steps: Time alignment processing, specifically using the dynamic time warping method to perform time alignment processing to obtain dynamic time warping data, the calculation formula is: Where DTW(·) is the dynamic time warping function, A is the first time series data, which is used to represent any sequence data to be time aligned, B is the second time series data, K is the total number of data points, k is the data point index, d(·) is the Euclidean distance calculation function, and a ik is the kth data point corresponding to the i-th position in the first time series data, b jk is the kth data point corresponding to the jth position in the second time series data; The spectral alignment processing specifically adopts the wavelet transform method to perform spectral alignment processing on the dynamic time warping data to obtain spectral alignment data.

6. The machine learning-based bladder cancer diagnosis assistance system according to claim 5, characterized in that: The calculation formula for constructing the explanatory optimization operator is: Where, Φ i is the SHAP value corresponding to the i-th feature, which is used to measure the importance of the feature and the interpretability of the model prediction results. N is the total set of features, i is the feature index, and S N is a feature subset, ! is a factorial operator, |·| is a modulo operator, f(·) is the model output representation function, and ∪ is a union operator.

7. The machine learning-based bladder cancer diagnosis assistance system according to claim 6, characterized in that: The diagnostic aid is used to assist in bladder cancer diagnosis using model prediction results, specifically by performing biomarker analysis of bladder cancer urine samples based on the bladder cancer biomarker detection data, and obtaining bladder cancer diagnostic aid reference information through extracellular vesicle characterization analysis.