A traditional Chinese medicine tumor synergistic treatment clinical data analysis method

Abstract

The application discloses a traditional Chinese medicine tumor synergistic treatment clinical data analysis method, and relates to the field of medical and health information science.The content comprises the following steps: acquiring and preprocessing multi-modal clinical data, further performing preliminary feature extraction, and obtaining preliminary feature data;standardizing the preliminary feature data, and obtaining standardized preliminary feature data; performing mapping processing on the standardized preliminary feature data by using a time difference self-adaptive power logarithmic mapping algorithm with semantic correlation and noise suppression, and obtaining mapped preliminary feature data; performing fusion processing on the mapped preliminary feature data by using an adaptive perception attention fusion algorithm, and obtaining comprehensive feature data; and analyzing the comprehensive feature data by using a multi-task prediction model, and obtaining a comprehensive analysis result.The problems that multi-modal clinical data cannot be uniformly processed and analyzed and processed inaccurately due to differences in acquisition frequency, data structure, semantic expression and quality reliability are solved.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics, and in particular to a clinical data analysis method for synergistic treatment of tumors using traditional Chinese medicine. Background Technology

[0002] In the clinical treatment of cancer, the synergistic treatment of traditional Chinese medicine (TCM) and Western medicine is gradually becoming an important comprehensive treatment model. Western medicine treatment mainly uses methods such as surgery, radiotherapy, chemotherapy, targeted therapy, and immunotherapy, which can play a direct role in tumor control. TCM treatment, on the other hand, improves patients' quality of life and tolerance by differentiating syndromes and treating accordingly, regulating the body's overall state and immune function, and alleviating treatment side effects. However, in actual clinical practice, due to the diverse data sources of TCM and Western medicine, including structured laboratory test values, semi-structured imaging parameters, unstructured medical record texts, physiological signals from wearable devices, and patient self-report questionnaires, different modalities differ significantly in terms of collection frequency, temporal distribution, dimensional scale, semantic expression, and data quality, leading to significant challenges in the unified processing and analysis of multimodal data. Existing multimodal medical data processing methods generally suffer from problems such as feature fusion relying on manually set weights and the inability to dynamically perceive changes in data quality when dealing with asynchronous sampling, data noise, and semantic differences between modalities. This makes it difficult to fully utilize the temporal correlation and clinical value of information from each modality. Therefore, there is an urgent need to provide a synergistic treatment clinical data analysis method to solve the above problems. Summary of the Invention

[0003] This invention provides a clinical data analysis method for collaborative treatment of tumors using traditional Chinese medicine, in order to solve the problems of difficulty in unified processing and analysis and inaccurate processing caused by differences in the acquisition frequency, data structure, semantic expression and quality reliability of multimodal clinical data.

[0004] The present invention provides a clinical data analysis method for synergistic treatment of tumors using traditional Chinese medicine, comprising the following steps:

[0005] S1. Acquire and preprocess multimodal clinical data, perform preliminary feature extraction on the preprocessed multimodal clinical data to obtain preliminary feature data; standardize the preliminary feature data to obtain standardized preliminary feature data; map the standardized preliminary feature data using a temporal difference adaptive power-logarithmic mapping algorithm with semantic relevance and noise suppression to obtain mapped preliminary feature data.

[0006] S2. The preliminary feature data after mapping is fused using an adaptive perceptual attention fusion algorithm to obtain comprehensive feature data; the comprehensive feature data is then analyzed using a multi-task prediction model to obtain comprehensive analysis results.

[0007] Preferably, S1 specifically includes:

[0008] In the implementation of the time-difference adaptive power-logarithmic mapping algorithm for semantic relevance and noise suppression, semantic relevance and time offset quantization values ​​are introduced, and combined with the standardized preliminary feature data, the mapped preliminary feature data is generated.

[0009] Preferably, S2 specifically includes:

[0010] The adaptive perception attention fusion algorithm performs weighted fusion of the mapped preliminary feature data by calculating adaptive perception attention fusion weights.

[0011] Preferably, S2 specifically includes:

[0012] Based on the initial feature data after mapping, adaptive perceptual attention fusion weights are generated by combining semantic relevance gain term, data quality gain term, and time offset penalty term.

[0013] Preferably, S2 specifically includes:

[0014] The multi-task prediction model consists of a shared encoder and task-specific output heads.

[0015] Preferably, S2 specifically includes:

[0016] The acquired preprocessed historical multimodal clinical data is processed by a semantically correlated and noise-suppressed time-difference adaptive power-logarithmic mapping algorithm and an adaptive perceptual attention fusion algorithm to generate historical comprehensive feature data, which is then used as the input to the shared encoder.

[0017] Preferably, S2 specifically includes:

[0018] In a shared encoder, historical composite feature data undergoes positional encoding to obtain a positionally encoded feature sequence.

[0019] Preferably, S2 specifically includes:

[0020] The position-encoded feature sequence is fed into the short-term pattern extraction layer to obtain the short-term pattern extraction result; the short-term pattern extraction result is fed into the long-term dependency modeling layer to obtain the global aggregate representation.

[0021] Preferably, S2 specifically includes:

[0022] The global aggregated representation is passed to three task-specific output heads, including the tumor progression rate prediction head, the quality of life change trend prediction head, and the potential side effect risk prediction head, to obtain the comprehensive analysis results.

[0023] The beneficial effects of the technical solution of the present invention are:

[0024] 1. By introducing a time-difference adaptive power-logarithmic mapping algorithm that combines semantic relevance and noise suppression, a unified mapping processing of multimodal clinical data under asynchronous sampling, different dimensions, and different noise levels is achieved. This effectively suppresses the interference of high-noise features on the overall analysis and amplifies feature information that is highly relevant to the diagnostic and treatment goals while ensuring the cross-modal comparability of features, thereby improving the stability and accuracy of subsequent analysis.

[0025] 2. By using an adaptive perceptual attention fusion algorithm, the preliminary feature data after mapping is weighted by semantic relevance, data quality and time offset, and the contribution ratio of each modality to the comprehensive feature is dynamically allocated. This allows the fusion result to be adaptively adjusted according to the actual situation at different time points, overcoming the problem that traditional fixed weight fusion cannot adapt to data quality fluctuations. It achieves dual optimization of cross-modal features in terms of temporal consistency and information value.

[0026] 3. The multi-task prediction model adopts a combined architecture of shared encoder and task-specific output head. Through the synergistic effect of positional encoding, short-term pattern extraction and long-term dependency modeling, it simultaneously captures the local dynamic changes of clinical data and the global dependencies across time steps, thereby improving the model's generalization ability among different patients. Attached Figure Description

[0027] Figure 1 This is a flowchart of a clinical data analysis method for collaborative treatment of tumors using traditional Chinese medicine, as described in this invention. Detailed Implementation

[0028] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0030] The following description, in conjunction with the accompanying drawings, details a specific scheme for a clinical data analysis method for collaborative treatment of tumors using traditional Chinese medicine, provided by this invention.

[0031] See attached document Figure 1 The diagram illustrates a flowchart of a clinical data analysis method for collaborative treatment of tumors using traditional Chinese medicine, provided by an embodiment of the present invention. The method includes the following steps:

[0032] S1. Acquire and preprocess multimodal clinical data, perform preliminary feature extraction on the preprocessed multimodal clinical data to obtain preliminary feature data; standardize the preliminary feature data to obtain standardized preliminary feature data; map the standardized preliminary feature data using a semantically relevant and noise-suppressive time-difference adaptive power-logarithmic mapping algorithm to obtain mapped preliminary feature data.

[0033] Multimodal clinical data is collected from various sources, including traditional Chinese medicine (TCM) medical record systems, Western medicine HIS systems, imaging workstations, and wearable devices for patients, utilizing existing multimodal medical data acquisition technologies. This data includes TCM components, Western medicine components, and patient reports. The TCM components include electronic medical record texts (diagnosis and treatment records), high-resolution tongue images, and pulse sensor waveform signals. The Western medicine components include laboratory test values ​​(blood indicators, tumor markers, etc.), imaging data (tumor volume, morphological quantification parameters), and gene testing results. Patient reports include pain scores, sleep quality index, and changes in appetite. The multimodal clinical data is preprocessed using techniques familiar to those skilled in the art, such as Z-score standardization to remove outliers and Min-Max normalization to unify dimensions, resulting in preprocessed multimodal clinical data.

[0034] Pre-processed multimodal clinical data underwent preliminary feature extraction using existing feature engineering techniques to obtain preliminary feature data. Feature engineering techniques included: obtaining fixed-length symptom-syndrome semantic vectors from text using a medical domain language encoder; extracting features such as color histogram moments, texture consistency quantification, and tongue coating coverage from tongue images using convolutional networks; extracting time-frequency features such as dominant frequency, peak energy spectral density, and periodic variation coefficient from pulse waveforms; describing trends in structured numerical values ​​using differences and rolling mean from previous visits; and collecting quantitative indicators such as lesion volume, longest diameter, and density distribution quantiles from imaging data (details omitted here). Simultaneously, the preliminary feature data was standardized using Z-scores to obtain standardized preliminary feature data.

[0035] Furthermore, the standardized preliminary feature data is mapped using a time-difference adaptive power-logarithmic mapping algorithm that combines semantic relevance and noise suppression to obtain the mapped preliminary feature data. The specific implementation formula is as follows:

[0036] ;

[0037] in, At any moment Next Preliminary feature data after modality mapping; Represents clinical data modalities; Indicates the current moment; and This ensures the numerical stability of features across orders of magnitude; At any moment Next The noise estimate of the standardized preliminary feature data for each modality, i.e., the variance of the standardized preliminary feature data, reflects the short-term volatility of the standardized preliminary feature data and is used to suppress unstable features. It is calculated over a central time window, the size of which is determined based on the specific application scenario and will not be elaborated here; It is semantic relevance, used to measure time. Next The correlation strength between the standardized preliminary feature data of each modality and the current diagnostic and treatment goals (such as syndrome ontology or tumor efficacy goals) is used to amplify the effective features. It is calculated using cosine similarity, with a reference value range of [0,1]. The absolute value ensures symmetrical scaling. At any moment Next Preliminary standardized feature data for each modality; It is the time offset quantization value, calculated by the first... The difference between the standardized preliminary feature data of each modality at the current time and the time determined according to the actual application scenario is obtained, which is used to characterize the time difference caused by asynchronous sampling; It is the time penalty coefficient, used to control the amplification of the denominator by the time offset. It is determined according to expert experience and the reference range is [0.01, 0.5]. It is the power-law benchmark coefficient, used to adjust the overall nonlinear enhancement intensity. It is determined according to expert experience and has a reference range of [0.5, 2.5]. It is the time difference penalty factor, used to map the time offset caused by asynchronous sampling into a continuously differentiable penalty; By compressing features across magnitudes into a comparable range, extreme values ​​are avoided from dominating.

[0038] S2. The preliminary feature data after mapping is fused using an adaptive perceptual attention fusion algorithm to obtain comprehensive feature data; the comprehensive feature data is then analyzed using a multi-task prediction model to obtain comprehensive analysis results.

[0039] Based on the mapped preliminary feature data, an adaptive perceptual attention fusion algorithm is introduced for fusion processing to obtain comprehensive feature data. The adaptive perceptual attention fusion algorithm fuses the mapped preliminary feature data by calculating adaptive perceptual attention fusion weights. The specific implementation formula is as follows:

[0040] ;

[0041] in, At any moment No. The adaptive perceptual attention fusion weights for each modality represent the relative contribution of that modality to the comprehensive feature data; At any moment Available modal set; It is an attention content matching matrix, which is obtained by jointly learning backpropagation and gradient descent based on historical multimodal clinical data extracted from existing databases; It is the attention projection vector, used to... The mapping is to scalar content scores, determined based on expert experience. Indicates transpose; It is the semantic relevance coefficient, used to control... The gain strength of the adaptive perceptual attention fusion weights is determined based on expert experience, with a reference range of values. ; It is a quality score coefficient, used to control... The gain strength of the adaptive perceptual attention fusion weights is determined based on expert experience, with a reference range of values. ; It is a data quality score, representing the data quality at time [time]. The mapped first The reliability of the preliminary characteristic data for each modality was determined based on expert experience, with a reference range of [value missing]. ; This is the time offset influence coefficient, determined based on expert experience, with a reference range of values. ; Indicates at time No. The information relevance of each modality in the current context; These represent the semantic relevance gain term, the data quality gain term, and the time offset penalty term, respectively.

[0042] Furthermore, the preliminary feature data after mapping is weighted and fused based on the adaptive perceptual attention fusion weight to obtain the comprehensive feature data at each time step.

[0043] Finally, the comprehensive feature data is fed into a multi-task prediction model for processing to obtain comprehensive analysis results, including tumor progression rate, trends in patients' quality of life, and potential side effect risks. The multi-task prediction model is built upon preprocessed historical multimodal clinical data extracted from the database (covering traditional Chinese medicine texts and images, Western medicine structured numerical data and images, follow-up questionnaires, and medication records, etc.), with the following specific architecture:

[0044] The multi-task prediction model employs a shared encoder and task-specific output heads. Preprocessed historical multimodal clinical data is pushed forward through a time-sliding window determined by the specific application scenario. This data is processed by a semantically relevant and noise-suppressed time-difference adaptive power-logarithmic mapping algorithm and an adaptive perceptual attention fusion algorithm to obtain historical comprehensive feature data, which is then used as input. In the shared encoder, the input is first processed by positional encoding, embedding time-step information into the historical comprehensive feature data. This ensures the multi-task prediction model can perceive the temporal relationships of the input, resulting in a positionally encoded feature sequence. The positionally encoded feature sequence then enters the short-term pattern extraction layer. Utilizing a gated one-dimensional convolutional structure, while maintaining sensitivity to missing data masks, it captures local temporal change trends and rapid fluctuation patterns within the window, yielding the short-term pattern extraction result. This result is then fed into the long-term dependency modeling layer, which consists of alternating stacks of lightweight multi-head attention structures and feedforward networks. This layer is used to mine global dependencies across time steps based on local patterns, resulting in a global aggregate representation. Multi-head attention structures can focus on the potential connections between different time slices in a sequence, while feedforward networks perform nonlinear transformations and combinations within the feature dimension to enhance the expressive power of features.

[0045] The output of the shared encoder, i.e., the global aggregate representation, is passed to three task-specific output heads. The tumor progression rate prediction head extracts key time-slice features through time-series pooling and combines them with the global aggregate representation to predict the rate and range of tumor volume change at future follow-ups.

[0046] The life quality change trend prediction head adopts a smooth residual structure to focus on capturing long-term trend changes, obtain modality coverage and data quality statistics, and weight the modality coverage and data quality statistics at the input end to reduce the interference of low-quality data on trend judgment.

[0047] The potential side effect risk prediction head uses a feature screening and discrimination structure to prioritize feature channels related to drug dosage fluctuations, laboratory anomalies, and symptom mutations. It outputs the probability of adverse events occurring within a future medical cycle and provides key modalities and time slices that trigger risk determination.

[0048] In summary, a clinical data analysis method for synergistic treatment of tumors with traditional Chinese medicine has been developed.

[0049] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0050] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for analyzing clinical data on synergistic treatment of tumors using traditional Chinese medicine, characterized in that, Includes the following steps: S1. Acquire and preprocess multimodal clinical data. Perform preliminary feature extraction and standardization on the preprocessed multimodal clinical data. Map the standardized preliminary feature data using a time-difference adaptive power-logarithmic mapping algorithm that combines semantic relevance and noise suppression to obtain the mapped preliminary feature data. The formula is as follows: ； in, At any moment Next Preliminary feature data after modality mapping; Represents clinical data modalities; Indicates the current time; At any moment Next Noise estimates of the standardized preliminary feature data for each modality; It is the semantic relevance, calculated using cosine similarity, with a value range of [0,1]. At any moment Next Preliminary standardized feature data for each modality; It is the time offset quantization value, calculated by the first... The difference between the standardized preliminary feature data of each modality at the current time and the time determined according to the actual application scenario is obtained; This is the time penalty coefficient, determined based on expert experience, with a value range of [0.01, 0.5]. It is the power exponent benchmark coefficient, determined by expert experience, with a value range of [0.5, 2.5]. S2. The preliminary feature data after mapping is fused using an adaptive perceptual attention fusion algorithm to obtain comprehensive feature data. The adaptive perceptual attention fusion algorithm fuses the preliminary feature data after mapping by calculating adaptive perceptual attention fusion weights, as shown in the following formula: ； in, At any moment No. Adaptive perceptual attention fusion weights for each modality; At any moment Available modal set; It is an attention content matching matrix, which is obtained by jointly learning backpropagation and gradient descent based on historical multimodal clinical data extracted from existing databases; It is the attention projection vector, determined based on expert experience. Indicates transpose; It is the semantic relevance coefficient; This is the quality score coefficient, determined based on expert experience, and its value range is [range missing]. ; It is a data quality score, representing the data quality at time [time]. The mapped first The reliability of the preliminary feature data for each modality was determined based on expert experience, with a value range of [value range missing]. ; This is the time offset influence coefficient, determined based on expert experience, with a value range of [value missing]. ; A multi-task prediction model consisting of a shared encoder and task-specific output heads is used to analyze comprehensive feature data. Preprocessed historical multimodal clinical data is processed using a temporal difference adaptive power-logarithmic mapping algorithm with semantic relevance and noise suppression, and an adaptive perceptual attention fusion algorithm to generate historical comprehensive feature data. This historical comprehensive feature data is then used as input to the shared encoder. Within the shared encoder, the historical comprehensive feature data undergoes positional encoding to obtain a positionally encoded feature sequence. This positionally encoded feature sequence is then fed into a short-term pattern extraction layer to obtain short-term pattern extraction results. These results are then fed into a long-term dependency modeling layer to obtain a global aggregate representation. Finally, this global aggregate representation is passed to three task-specific output heads—one predicting tumor progression rate, one predicting quality of life trends, and one predicting potential side effect risks—to obtain the comprehensive analysis results.

Images