Drug activity prediction method and system based on multi-modal medical big data

Abstract

The invention discloses a drug activity prediction method and system based on multi-modal medical big data, and relates to the technical field of artificial intelligence. The method comprises the steps of obtaining multi-modal medical big data, and performing multi-modal coverage collection based on the multi-modal medical big data to obtain a multi-modal data set; carrying out biological feature recognition based on the multi-modal data set and carrying out clustering analysis to determine a plurality of biological feature classes; performing drug activity influence analysis according to the biological characteristic classes, and constructing a characteristic-activity mapping list to screen target candidate drugs; and extracting biological characteristic data prediction activity based on the target candidate drugs, and generating drug activity report feedback optimization data. The technical problems of inaccurate prediction result and poor generalization ability caused by dependence on single-modal data and lack of correlation analysis on complex biological characteristics of a traditional drug activity prediction method are solved, and high-precision prediction of drug activity is realized by utilizing multi-modal fusion and biological characteristic clustering analysis, so that the method is suitable for large-scale popularization and application. And the medicine screening efficiency and the intelligent level are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically to a method and system for predicting drug activity based on multimodal medical big data. Background Technology

[0002] With the accelerated integration of precision medicine and drug development, multimodal big data, including genomics, proteomics, metabolomics, medical imaging, physiological signals, and electronic medical records, is continuously generated in the medical field. Current drug activity prediction methods largely rely on single-modality or weakly fused approaches, making it difficult to simultaneously characterize molecular mechanisms of action and individual biological differences, resulting in insufficient prediction accuracy and generalization ability. Furthermore, the heterogeneity of data sources, varying quality, noise, and missing data, coupled with dynamic changes in acquisition parameters and scenarios, limits end-to-end stability, making it difficult for traditional batch processing data pipelines and model training to provide timely feedback and optimization. On the other hand, activity assessment often needs to consider both population and individual levels, increasing the requirements for interpretability, traceability, and clinical usability. Simple threshold determinations or black-box models are insufficient to meet regulatory compliance and practical application needs, leading to the failure of many candidate drugs in subsequent clinical trials due to the inability to accurately predict their activity and toxicity in different populations or under specific biological characteristics, resulting in wasted resources. Summary of the Invention

[0003] This application provides a drug activity prediction method and system based on multimodal medical big data, which solves the technical problems of inaccurate prediction results and poor generalization ability caused by traditional drug activity prediction methods relying on single-modal data and lacking correlation analysis of complex biological characteristics.

[0004] The first aspect of this application provides a method for predicting drug activity based on multimodal medical big data, the method comprising: Multimodal medical big data is acquired through intelligent data acquisition devices. Multimodal coverage acquisition is then performed on this multimodal medical big data to obtain a multimodal dataset. Biometric identification is performed on this multimodal dataset to obtain a biometric set. The multimodal dataset is then clustered according to the biometric set to determine multiple biometric classes. The impact of these multiple biometric classes on drug activity is analyzed, and multiple activity impact levels are associated with these biometric classes to construct a feature-activity mapping list for screening candidate drug parameters and determining target candidate drugs. Based on the target candidate drugs, target biometric data is extracted for activity prediction to obtain prediction results. A drug activity report is generated according to drug activity standards and the prediction results to provide optimization feedback to the multimodal medical big data.

[0005] A second aspect of this application provides a drug activity prediction system based on multimodal medical big data, the system comprising: Data Acquisition Module: Acquires multimodal medical big data through intelligent data acquisition equipment, and performs multimodal coverage acquisition based on the multimodal medical big data to obtain a multimodal dataset; Data Analysis Module: Performs biometric identification based on the multimodal dataset to obtain a biometric set, and performs cluster analysis on the multimodal dataset according to the biometric set to determine multiple biometric classes; Drug Screening Module: Performs impact analysis on drug activity according to the multiple biometric classes, sets multiple activity impact levels and associates them with the multiple biometric classes, constructs a feature-activity mapping list to screen candidate drug parameters, and determines target candidate drugs; Activity Prediction Module: Extracts target biometric data based on the target candidate drugs to perform activity prediction, obtains prediction results, and generates a drug activity report based on drug activity standards and the prediction results to provide optimization feedback to the multimodal medical big data.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, multimodal medical big data is acquired and integrated using intelligent data acquisition equipment, forming a multimodal dataset by comprehensively collecting data from different modalities. Then, key biomarkers are identified, a biomarker set is constructed, and cluster analysis is used to divide the samples into multiple biomarker classes to reveal the characteristic differences between different individuals or samples. Based on this, the correlation between each biomarker class and drug activity is analyzed, an activity influence level system is established, and a mapping list between features and activities is constructed for screening parameters of candidate drugs and identifying target candidate drugs with potential efficacy. Finally, activity prediction is performed based on the corresponding biomarker data of the target drug, generating prediction results, and combining these with drug activity standards to form a drug activity report. Simultaneously, the results are fed back into the multimodal medical big data, enabling dynamic optimization of the model and data, thereby improving the accuracy and intelligence of drug activity prediction. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of the drug activity prediction method based on multimodal medical big data provided in the embodiments of this application.

[0009] Figure 2 This is a schematic diagram of the structure of a drug activity prediction system based on multimodal medical big data provided in an embodiment of this application.

[0010] Figure labeling: Data acquisition module 11, data analysis module 12, drug screening module 13, activity prediction module 14. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] Example 1, as Figure 1 As shown, this application provides a drug activity prediction method based on multimodal medical big data, the method including: Multimodal medical big data is acquired through intelligent data acquisition devices, and multimodal coverage acquisition is performed based on the multimodal medical big data to obtain a multimodal dataset.

[0013] In this embodiment, a data acquisition channel is established between multiple data sources, including hospital information systems, medical imaging equipment, biochemical detection devices, and wearable monitoring terminals. Built-in intelligent integrated circuits control the synchronous operation of multiple sensing modules. During acquisition, dynamic switching between modes is achieved using MOS transistor switching technology, enabling parallel acquisition of multimodal data such as genomic data, proteomic data, metabolomic data, imaging data, and clinical records, forming multimodal medical big data. Subsequently, intelligent data acquisition equipment performs multimodal coverage acquisition according to set parameters, ensuring accurate matching of the temporal, spatial, and content correspondences of each modality of data. After acquisition, the acquired data is formatted, cleaned, and structured for storage, forming a multimodal dataset that provides complete data support for subsequent biometric identification and drug activity prediction.

[0014] Furthermore, the method involves acquiring multimodal medical big data through intelligent data acquisition devices, and then performing multimodal coverage acquisition based on the multimodal medical big data to obtain a multimodal dataset. Multiple data acquisition channels are controlled by intelligent integrated circuits. Modulation switching of multimodal data sources is performed based on MOSFET switching technology, and initial acquisition parameters are set. When a multimodal data source is available, iterate through the multimodal data sources to identify data and divide the data acquisition area. The intelligent data acquisition device performs multimodal coverage acquisition in the data acquisition area according to the initial acquisition parameters to obtain multimodal medical big data. Based on the multimodal medical big data, the initial acquisition parameters of the intelligent integrated circuit are adjusted to set target acquisition parameters. The intelligent data acquisition device is then activated according to the target acquisition parameters to perform multimodal coverage acquisition to obtain the multimodal dataset.

[0015] Preferably, the system first initializes and configures multiple sensing channels, including genome sequencing, proteomics detection, metabolomics analysis, medical imaging acquisition, and physiological signal monitoring, using a built-in control chip. After configuration, the intelligent integrated circuit dynamically switches and selects the input channels for each modal data source based on MOS transistor switching technology, ensuring that different types of data do not cause signal interference or resource contention during multimodal parallel acquisition. Before acquisition begins, initial acquisition parameters are set according to the acquisition task requirements, including but not limited to sampling frequency, acquisition resolution, signal amplification factor, acquisition delay, and noise filtering threshold. The intelligent integrated circuit calibrates the operating status of each channel based on these parameters to ensure the consistency and stability of the acquisition. When a multimodal data source is detected to be available, the system initiates a traversal procedure to identify and mark each modal data source. During this process, the acquisition area is spatially divided based on sensor acquisition and feature mapping, generating a corresponding data acquisition area index table to facilitate subsequent data fusion and correlation analysis. Subsequently, the intelligent data acquisition device performs multimodal coverage acquisition on the pre-defined data acquisition area according to preset initial acquisition parameters. This involves synchronous signal acquisition through a multi-channel parallel acquisition mechanism, ensuring consistency of data across different modalities over time, thereby obtaining raw multimodal medical big data. After acquisition, the multimodal medical big data undergoes quality analysis and noise detection to identify data interference factors present during acquisition, such as signal drift, missing data, and abnormal noise. Based on the analysis results, the initial acquisition parameters are adaptively adjusted, dynamically optimizing parameters such as acquisition frequency, acquisition resolution, and signal amplification ratio to generate target acquisition parameters. Then, based on the optimized target acquisition parameters, the intelligent data acquisition device is reactivated to perform multimodal coverage acquisition again, obtaining a higher-quality and more structurally complete multimodal dataset, providing an accurate data foundation for subsequent biometric identification and drug activity prediction.

[0016] Furthermore, when the multimodal data source is available, the multimodal data source is traversed to identify the data and divide the data collection area. Methods include: When the multimodal data source is available, the first sensor and the second sensor are activated; the first sensor is used to collect genomic data to obtain a first sensor dataset; the second sensor is used to collect proteomic data to obtain a second sensor dataset; a feature space coordinate system is constructed based on the multimodal dataset, and the first sensor dataset and the second sensor dataset are mapped to the feature space coordinate system to obtain the feature distribution parameters of the multimodal dataset; the multimodal data source is identified based on the feature distribution parameters of the multimodal dataset, and the data collection area is divided.

[0017] Optionally, when a multimodal data source is detected to be available, the system will activate multiple sensors for different biological modalities. For example, the first sensor is mainly used for genomic data acquisition, and the second sensor is used for proteomic data acquisition. Subsequently, the first sensor, which can employ a high-sensitivity molecular detection chip or photoelectric conversion unit, is used to sense and acquire genomic information from the multimodal data source, enabling the reading and signal conversion of nucleic acid molecular sequences. During the acquisition process, intelligent integrated circuits are used to calibrate signal intensity, background noise, and temperature drift in real time to ensure the accuracy and stability of the genomic data. The acquired raw signals are converted from analog to digital to generate structured genomic feature data, forming the first sensing dataset. Simultaneously, the second sensor is activated to acquire proteomic data. This second sensor can be a multi-channel sensing device based on mass spectrometry, fluorescent labeling, or surface plasmon resonance technology, capable of capturing multi-dimensional feature information such as protein expression levels, spatial conformation, and post-translational modifications. During the acquisition process, signal standardization processing is performed based on sampling intensity and dynamic range to eliminate systematic errors between different batches and samples, ultimately obtaining proteomic-level feature information, forming the second sensing dataset. After acquiring the first and second sensor datasets, a multimodal feature space coordinate system is constructed. The first and second sensor datasets are then mapped to feature vectors, and feature alignment and weighted normalization are performed based on their respective biological attributes, such as gene loci, expression levels, and structural similarity. Principal component analysis, linear discriminant analysis, or autoencoder networks are then used to calculate the correlation between multimodal features, forming a unified feature space coordinate system. Within this feature space coordinate system, distribution analysis is performed on the multimodal features, calculating the density, distance, and similarity parameters of each feature point to obtain the feature distribution parameters of the multimodal dataset. These feature distribution parameters reflect the clustering and diversity of genomic and proteomic features in multidimensional space. Based on these feature distribution parameters, the system identifies and classifies the multimodal data sources, grouping regions with similar feature distribution trends into the same data acquisition area and assigning each area a unique identifier for subsequent multimodal data fusion, feature extraction, and activity analysis, thereby achieving precise positioning and efficient acquisition of multimodal medical data.

[0018] Furthermore, based on the multimodal medical big data, the initial acquisition parameters of the intelligent integrated circuit are adjusted, and target acquisition parameters are set, including: Data quality analysis is performed based on the multimodal medical big data to identify multiple data interference data, including noisy interference data and missing interference data. The intelligent data acquisition device is then adjusted according to the noise interference data and its characteristics to determine the acquisition frequency adjustment information. Similarly, the intelligent data acquisition device is adjusted according to the missing interference data and its characteristics to determine the acquisition intensity adjustment information. Based on the acquisition frequency adjustment information and the acquisition intensity adjustment information, the initial acquisition parameters are adjusted, and the target acquisition parameters are set.

[0019] Optionally, various modal data are extracted from multimodal medical big data. Signal feature recognition algorithms, such as power spectral density analysis, variance detection, and wavelet transform, are used to identify random interference and systematic noise in the data. When abnormal high-frequency components or non-physiological fluctuations are detected in the signal frequency domain, such as data with a signal-to-noise ratio below a set threshold, this data is marked as noise interference data. This noise interference data typically originates from factors such as signal fluctuations in the acquisition channel, electromagnetic interference, or physiological signal drift. Simultaneously, time series interpolation analysis and sample distribution density calculation are used to determine if there are interruptions or abnormal intervals in the data. When the interval between consecutive sampled data exceeds the preset sampling period limit, or when abnormal null values ​​or zero-value segments exist in the data matrix, this data is marked as missing interference data. This missing interference data is mainly caused by partial data loss due to acquisition delay, transmission packet loss, or signal instability. By aggregating the marked noise interference data and missing interference data, multiple interference data sets are formed. For these noise interference data sets, feature detection algorithms, such as wavelet denoising and fast Fourier transform frequency domain analysis, are used to identify noise frequency bands and signal drift trends. Subsequently, the intelligent integrated circuit dynamically adjusts the operating parameters of the acquisition device based on the noise distribution characteristics, determining the acquisition frequency adjustment information. For example, when an increase in the proportion of high-frequency noise is detected, the acquisition frequency is automatically reduced or a bandpass filter mechanism is activated to weaken the interference. If the low-frequency drift signal is significant, the data acquisition effect is optimized by increasing the sampling rate and short-time averaging window. For these missing interference data, the distribution characteristics of the missing data are extracted, including the missing location, missing proportion, time interval, and modality type. Then, feature modeling is performed on the missing region of each modality to generate a missing feature description matrix, which is used to characterize the range and severity of the missing data. For example, for time series data (such as physiological signals), the matrix includes time interval deviation, sampling point loss rate, and waveform continuity; while for spatial data (such as medical images), it includes indicators such as pixel defect density, spatial coverage, and signal intensity distribution gradient. Next, acquisition intensity compensation parameters are calculated based on the missing feature description matrix. Typically, when the missing regions are concentrated and continuous, the acquisition power and signal excitation amplitude of the corresponding channel in that region are increased to enhance the penetration and captureability of the target signal. When the missing data is discretely distributed and the proportion is low, the local sampling repetition rate or the number of short-term samples is increased to improve data recovery accuracy. When the missing data mainly appears in low signal strength ranges, the acquisition time window is automatically extended to ensure that weak signals are fully recorded. After analyzing the noise interference data and the missing interference data, the initial acquisition parameters are adaptively corrected based on the acquisition frequency adjustment information and the acquisition intensity adjustment information to form optimized target acquisition parameters. These target acquisition parameters include updated sampling frequency, signal gain coefficient, sampling duration, filtering threshold, and channel weight coefficients.The system then activates the intelligent data acquisition device based on the target acquisition parameters to re-acquire multimodal coverage data, thereby obtaining a more stable, complete, and high-quality multimodal dataset, providing reliable data support for subsequent feature recognition and activity prediction.

[0020] Biometric identification is performed based on the multimodal dataset to obtain a biometric set. Then, cluster analysis is performed on the multimodal dataset according to the biometric set to determine multiple biometric classes.

[0021] In one embodiment, after obtaining the multimodal dataset, a multimodal feature extraction channel is used to identify biomarkers from the heterogeneous information in the dataset, including genomics, proteomics, metabolomics, imaging data, and clinical test indicators. For example, the genomic feature extraction channel extracts gene mutation sites, base variation types, and sequence fragment features; the proteomics feature extraction channel extracts protein expression levels, domain conformations, and molecular interaction parameters. After feature extraction, the feature data from different modalities are normalized and standardized to eliminate dimensional differences and data scale bias. These features are then processed by splicing and fusion to form a biomarker set. Next, cluster centers are initialized based on this biomarker set, and the biomarker set is clustered according to a predefined clustering algorithm. After multiple rounds of iterative calculations, multiple biomarker classes can be determined. Each biomarker class represents a group of samples with specific biomarker patterns, providing a clear data foundation for subsequent drug activity impact analysis and candidate drug screening.

[0022] Furthermore, based on the multimodal dataset, biometric identification is performed to obtain a biometric set. Then, the multimodal dataset is clustered according to the biometric set to determine multiple biometric classes. The method includes: A multimodal data feature extraction channel is constructed, and parallel feature extraction is performed on the multimodal dataset through the multimodal data feature extraction channel to obtain initial multimodal data features. The initial multimodal data features are normalized and combined to construct a multidimensional biometric vector. Based on the multidimensional biometric vector, weights are assigned to obtain a multidimensional weighted feature vector. The multidimensional weighted feature vector is then concatenated and fused to obtain a biometric set. Cluster centers are initialized based on the biometric set, and the multimodal dataset is clustered according to the cluster centers to determine the multiple biometric classes.

[0023] Preferably, independent multimodal data feature extraction channels are first constructed for different data sources of the multimodal dataset, such as a genomic feature extraction channel for the genome and a proteomics feature extraction channel for the proteome. Each channel is controlled by an intelligent integrated circuit and pre-trained with an adapted convolutional neural network. The training steps typically include forward propagation, loss calculation, backpropagation, and parameter optimization. For example, the genome can be adapted to a one-dimensional convolutional neural network, and the proteome can be adapted to a two-dimensional convolutional neural network. After receiving the multimodal dataset, the multimodal data feature extraction channel traverses the corresponding multimodal dataset in parallel to extract the corresponding initial data features, forming multimodal initial data features. Subsequently, due to possible differences in data scale, such as numerical range, units, and dimensions, the system uses Z-score normalization, min-max normalization, and other methods to map the multimodal data to a unified numerical space. Then, by using unique identifiers, such as patient IDs or test batch numbers, sample-level alignment is performed to form biometric vectors for different modalities. By storing these biometric vectors, a multidimensional biometric vector is formed. Subsequently, based on the importance of the features and in conjunction with expert decision-making, weights are assigned to the multi-dimensional biofeedback vectors. The assigned weight coefficients are then multiplied by the corresponding dimension of the biofeedback vector to form a multi-dimensional weighted feature vector. These multi-dimensional weighted feature vectors are then concatenated to form a unified feature space, creating a biofeedback set. This biofeedback set comprehensively reflects an individual's genetic characteristics, protein expression, metabolic levels, and physiological performance, serving as the core input for subsequent clustering analysis. Then, initial cluster centers are generated using random sampling or density-based initialization methods to avoid clustering bias. Euclidean distance and Mahalanobis distance are then used to measure the distance between different features, assigning each feature to the nearest cluster center. Finally, the cluster centers are updated to the average value of the features within each cluster. Through multiple iterations and the minimization of squared error criterion, the system gradually converges to a stable cluster structure, forming multiple biofeedback classes. Each biofeedback class represents a group with similar biological characteristics, metabolic patterns, or drug response characteristics, providing a scientific and scalable data foundation for drug activity prediction.

[0024] Furthermore, the method for initializing cluster centers based on the biometric set, and then clustering the multimodal dataset according to the cluster centers to determine the multiple biometric classes includes: A distance matrix is ​​constructed by combining intelligent integrated circuits with the cluster centroids. A cluster centroid update strategy is formulated by dynamically adjusting the cluster centroids using MOS transistor switching technology and the distance matrix. Based on the update strategy, the similarity between the multimodal dataset and the cluster centroids is calculated to obtain a similarity coefficient. The density between samples within a cluster is calculated based on the similarity coefficient and the distance matrix, defining compactness. The separation between cluster centroids within a cluster is calculated based on the similarity coefficient and the distance matrix, defining separation. The compactness and separation are jointly optimized to determine a clustering scheme. The clustering scheme is then applied to the multimodal dataset to determine the multiple biometric classes.

[0025] Optionally, after the extraction and fusion of biometric sets, the intelligent integrated circuit performs parallel computation on the high-dimensional feature vectors of the multimodal data. During this process, using cluster centers as references, the feature distances between each modality and different centers are calculated to form an initial distance matrix. This initial distance matrix includes the Euclidean distances between each modality and between each modality and the centers. This is typically calculated in real-time by the matrix acceleration unit built into the integrated circuit, ensuring computational efficiency and accuracy for large-scale data. Subsequently, MOS transistor switching technology is used to dynamically switch and control the input channels of each cluster center. By monitoring the changes in the distance distribution between each cluster center and the modality data in real time, it is determined whether the cluster structure is converging. If the cluster centers are too close together or the intra-cluster sample distribution is divergent, the cluster centers are corrected according to a dynamic adjustment strategy. Typically, when the intra-cluster variance exceeds a threshold, the cluster center of that cluster is automatically moved to a higher-density region; when the inter-cluster distance is too close, a penalty factor is introduced to move the cluster centers further apart, preventing cluster overlap. Based on the updated centroid positions, cosine similarity is used to calculate the similarity coefficient between each modality and each cluster centroid. Higher similarity indicates a stronger match between the sample and the cluster. Then, the relationships between samples within a cluster are analyzed based on the similarity coefficient and distance matrix, defining two metrics: density and separation. Density is the reciprocal of the average distance between samples within a cluster, measuring the degree of aggregation within a cluster; a higher density value indicates similar features and concentrated distribution within the cluster. Separation is the minimum distance between different cluster centroids, assessing the degree of distinction between clusters; a higher separation value indicates clearer boundaries between different biometric classes. Density and separation are then jointly optimized to determine the clustering scheme. In this process, the silhouette coefficient is used as the core evaluation metric for the clustering scheme. The silhouette coefficient comprehensively reflects the matching degree of each modality in the current cluster, with a value range of [-1, 1], where a value closer to 1 indicates a better clustering effect. The system calculates the mean silhouette coefficient for each biometric class and re-divides classes with values ​​below a set threshold (e.g., 0.5). During the re-segmentation process, cluster boundaries are readjusted based on the similarity of modal data to optimize classification accuracy and stability. After obtaining the optimized clustering results, each biomarker class is validated against known biological classification systems, such as gene expression profiling, protein function grouping, or clinical phenotypic labels. Validation methods include chi-square tests and correlation coefficient matching. Through this process, a consistency confidence score is calculated for each biomarker class; a higher confidence score indicates a stronger fit between the cluster and the actual biomarker.Finally, after the clustering partitioning scheme is executed, multiple biomarker classes are generated, each containing a biomarker class, a corresponding sample index, cluster center parameters, density and separation indices, and confidence scores. These biomarker classes not only ensure the objectivity and reproducibility of the clustering partitioning, but also provide a highly reliable biomarker structural basis for subsequent drug activity impact analysis and candidate drug screening.

[0026] The influence of the multiple biomarker classes on drug activity is analyzed, multiple activity influence levels are set and associated with the multiple biomarker classes, a feature-activity mapping list is constructed to screen candidate drug parameters, and target candidate drugs are determined.

[0027] In one embodiment, after obtaining multiple biomarker classes, the contribution of each biomarker class to drug activity is calculated, generating an activity impact factor. This activity impact factor measures the weight of a particular biomarker in the drug response; a higher value indicates a more significant impact on drug activity. Subsequently, numerical boundary analysis and interval division are performed based on these activity impact factors, dividing continuous impact factor intervals into multiple activity impact level intervals, such as high impact level, medium impact level, and low impact level. A unique identifier is then generated for each biomarker class, and a mapping relationship is established between it and the corresponding activity impact level, forming a feature-activity mapping list to clarify the quantitative relationship between each biomarker class and the drug activity level. Afterward, the molecular structure parameters, biological action sites, and metabolic pathway characteristics of the candidate drug are matched with the aforementioned feature-activity mapping list. During the matching process, the system extracts target information, biomarker characteristics, and mechanisms of action corresponding to each biomarker class for the candidate drug and calculates the matching degree coefficient. If the candidate drug has a high matching degree in a high-impact biomarker class, it is assigned a higher potential activity score; if the matching degree is low, the activity score is correspondingly reduced. Then, a weighted scoring method is used to comprehensively calculate the matching degree and influence weight of multiple biomarker classes, forming a comprehensive activity score vector for each candidate drug. Based on a preset screening threshold, the candidate drugs are then ranked and prioritized, retaining drug molecules that perform exceptionally well in high-influence biomarker classes. Finally, based on the comprehensive activity score and mapping matching results, target candidate drugs are selected. These target candidate drugs exhibit significant advantages in their correlation with key biomarker classes, providing fundamental support for subsequent activity prediction, mechanism analysis, and drug optimization.

[0028] Furthermore, the method involves analyzing the impact of the multiple biomarker classes on drug activity, setting multiple activity impact levels and associating them with the multiple biomarker classes to construct a feature-activity mapping list. This method includes: The influence weights of drug activity are calculated based on the multiple biomarker classes to construct activity influence factors. Numerical boundary analysis is performed on the activity influence factors to obtain the influence value range. The activity influence factors are dynamically clustered according to the influence value range to determine the natural boundary points of the activity influence factors. Based on the natural boundary points, the activity influence factors are divided into multiple continuous intervals, and the multiple continuous intervals are labeled with influence levels to set multiple activity influence levels. The multiple biomarker classes are traversed to assign labels, generating biomarker class identifiers. The biomarker class identifiers are associated with the multiple activity influence levels to construct the feature-activity mapping list.

[0029] Preferably, after classifying biomarkers, each biomarker is treated as an independent influencing factor. Data corresponding to these influencing factors are extracted from historical data, and the influence weight of each influencing factor on drug activity is quantified using the Pearson correlation coefficient. This yields the activity influencing factor for each biomarker. The larger the activity influencing factor value, the more significant the role of that biomarker in drug activity changes. Subsequently, all activity influencing factors are sorted from largest to smallest, and their distribution characteristics, including mean, standard deviation, and distribution density, are calculated. Boundary analysis of these values ​​is then performed using the natural breakpoint method to identify statistically significant inflection points and boundary intervals in the distribution of influencing factors, forming a preliminary division of the influence value range. Afterward, based on the obtained influence value range, the activity influencing factors are dynamically clustered to determine their natural boundary points. The dynamic clustering process automatically groups the activity influencing factors through iterative optimization, ensuring minimal differences within each group and maximum differences between groups. During operation, the clustering algorithm dynamically adjusts the number of clusters and their boundary positions based on the distribution density of influence factors and changes in local gradients, ultimately determining the natural boundary points that best distinguish the levels of activity contribution. Then, the activity influence factors are divided into multiple continuous intervals according to these natural boundary points, such as high-influence, medium-influence, and low-influence intervals. These intervals are then assigned influence level labels, setting multiple activity influence levels. These level labels can be numerical or hierarchical, and each corresponds one-to-one with the actual magnitude of the activity contribution. After setting the influence levels, all biomarker classes are traversed, generating a unique identifier for each class. This identifier is then mapped to the corresponding activity influence level, forming a one-to-one feature-activity level relationship. Finally, these level relationships are integrated to construct a complete feature-activity mapping list. This list records the number of each biomarker class, its corresponding activity influence factor, natural boundary point interval, and activity influence level, supporting subsequent candidate drug parameter matching and screening operations, thus improving the accuracy and efficiency of drug screening and activity prediction.

[0030] Based on the target candidate drug, target biomarker data is extracted to predict activity, and prediction results are obtained. A drug activity report is generated according to the drug activity standard and combined with the prediction results to optimize and provide feedback to the multimodal medical big data.

[0031] In one embodiment, target biomarker data of the target candidate drug is first acquired, and then input into an activity prediction model for calculation. This activity prediction model can be built based on a deep neural network. During operation, the model calculates the interaction strength between the drug and the target biomarker through forward propagation, extracting deep implicit feature parameters to obtain accurate activity prediction results. Subsequently, the prediction results are compared and analyzed with a preset drug activity standard. This drug activity standard can be set based on international pharmacology databases, historical experimental data, or pharmacopoeia indicators to define judgment thresholds for different levels of activity, such as high, medium, and low activity. Through comparison, the activity level, potential indications, and risk factors of the target candidate drug under various biomarker conditions are identified, forming a structured drug activity report. This report includes the target drug's predicted activity value, activity level, key influencing features, potential target pathways, and risk warnings. Simultaneously, the report records the drug's activity distribution curves and high-frequency feature contributions under different biomarker categories to visually demonstrate the interaction patterns between the drug and the biological system. Finally, the generated drug activity report results are fed back to a multimodal medical big data platform to optimize the feedback mechanism. This mechanism adaptively adjusts data acquisition parameters, feature extraction algorithms, and model weight allocation based on prediction biases and feature weight changes in the report, thereby continuously improving data quality and prediction model performance. Through this closed-loop process, an intelligent feedback loop is achieved from drug activity prediction to model optimization and then to data improvement, ensuring more accurate and reliable drug activity prediction results and enhancing the intelligence level of drug development and clinical application.

[0032] Furthermore, based on the target candidate drug, target biomarker data is extracted for activity prediction to obtain prediction results. A drug activity report is then generated according to drug activity standards and the prediction results. The method includes: Target biomarker data of the target candidate drug is extracted from the multimodal dataset and filtered to determine the target biomarker dimension. Forward propagation calculation is performed on the target biomarker data according to the target biomarker dimension to obtain the calculation result. Deep analysis is performed based on the calculation result to extract deep feature parameters. Activity prediction is performed based on the calculation result and the deep feature parameters to obtain the prediction result. The prediction result is compared with drug activity standards to generate activity risk data. Continuous statistics are performed on the prediction result according to the activity risk data to determine high-frequency activity patterns. The high-frequency activity patterns are correlated with time-series data of drug molecular parameters to obtain multiple parameter combinations, which are then randomly mined to obtain an optimized parameter combination. The optimized parameter combination is added to the drug activity report.

[0033] Preferably, multi-source biomarker data related to the target candidate drug are first extracted from a multimodal dataset. This includes genomic mutation sites and expression variations, proteomic domain distribution and binding site activity, and metabolic pathway flux and metabolite concentration changes. After initial extraction, principal component analysis or mutual information maximization is used to calculate the correlation coefficient or mutual information between each feature and the target drug activity value. Highly correlated biomarker dimensions are selected as core analytical inputs to determine the final target biomarker dimensions, thereby reducing redundant information and improving computational efficiency and prediction accuracy. Subsequently, forward propagation calculations are performed on the selected target biomarker dimensions based on a pre-trained activity prediction model. The biomarker data corresponding to the input target biomarker dimensions are processed through weighted calculations and nonlinear activation functions in a multi-layer neural network, extracting complex relationships between features layer by layer to obtain the intermediate layer calculation results. Based on these results, further in-depth feature analysis was conducted to extract deep feature parameters, including molecular interaction energy gradients, target binding potential well distribution, time-series dynamic response parameters, and signal pathway coupling strength. These deep features can reveal the nonlinear relationship between the drug and the biological system, providing high-dimensional characterization support for subsequent activity prediction. Subsequently, the calculation results and deep feature parameters were fused and analyzed. The direct reaction characteristics reflected in the calculation results and the implicit structural characteristics embodied in the deep parameters were comprehensively calculated to obtain the activity prediction results of the target candidate drug. The prediction results were then compared with preset drug activity standards to determine the drug's activity level under specific conditions. Through comparison, corresponding activity risk data was generated to reflect the degree of deviation of the prediction results from the standards. Then, time series and statistical analysis was performed on the activity risk data to calculate the frequency and trend of activity responses under different feature dimensions, thereby identifying high-frequency activity patterns. These high-frequency activity patterns reflect the high-activity performance of the target drug under different biological feature dimensions and can reveal key feature combinations or typical reaction scenarios. After identifying high-frequency activity patterns, these patterns are correlated with time-series data of drug molecular parameters, such as molecular conformational change trajectories, interaction energy curves, and pharmacokinetic change sequences. This yields multiple parameter combinations corresponding to the high-frequency activity patterns. Then, random mining algorithms, such as genetic algorithms, Bayesian optimization, or Monte Carlo search, are used to find the optimal parameter combination within the parameter space, thus obtaining an optimized parameter combination that significantly improves drug activity. Finally, the obtained optimized parameter combinations and high-frequency activity pattern results are integrated into the drug activity report to guide data updates and model optimization, thereby achieving an efficient and intelligent process for drug screening, validation, and improvement.

[0034] Furthermore, based on the calculation results and the deep feature parameters, activity prediction is performed to obtain the prediction result, the method includes: The system determines whether the target biomarker data is in the feature-activity mapping list. If the target biomarker data is in the feature-activity mapping list, a high-activity prediction instruction is generated, and the target candidate drug is marked as a high-activity drug using the high-activity prediction instruction. Based on the calculation result, the high-activity drug is matched with the deep feature parameters to generate a first prediction result. If the target biomarker data is not in the feature-activity mapping list, a low-activity prediction instruction is generated, and the target candidate drug is marked as a low-activity drug using the low-activity prediction instruction. Based on the calculation result, the low-activity drug is matched with the deep feature parameters to generate a second prediction result.

[0035] Preferably, the system first calls the pre-constructed feature-activity mapping list and uses Euclidean distance or cosine similarity to determine whether the target biomarker data is in the feature-activity mapping list. When the target feature and the biomarker class in the mapping list reach a preset threshold in terms of Euclidean distance or cosine similarity (e.g., similarity ≥ 0.85), the target feature is determined to belong to a known feature pattern in the mapping list. At this time, the system generates a high-activity prediction instruction and marks the target candidate drug as a high-activity drug. This marking is used not only for classification within the model but also for subsequent report generation and database updates. Subsequently, the activity prediction model is used to perform multi-dimensional comparative analysis of the calculation results and deep feature parameter set to generate the first prediction result for the drug, reflecting the high-activity response pattern of the drug and the biological system and its credibility score. Conversely, if the target biomarker data is determined not to be in the feature-activity mapping list, it indicates that the drug feature has a low matching degree with the high-activity biomarker class and belongs to the unknown or low-association feature interval. At this time, a low-activity prediction instruction is generated, and the target candidate drug is marked as a low-activity drug. After labeling, the calculation results are compared and analyzed with deep feature parameters, but at this stage, the focus is on the drug's performance in low-response pathways, generating a second prediction result to characterize the drug's potential insufficient activity and risk of action. Finally, the first and second prediction results are stored separately and then aggregated during the drug activity report generation stage. The prediction results for highly active drugs are used for recommendations and subsequent optimization, while the results for low-activity drugs serve as reverse samples for model learning, continuously optimizing the model and achieving self-learning and closed-loop improvement in drug activity prediction.

[0036] In summary, the embodiments of this application have at least the following technical effects: First, multimodal medical big data is acquired through intelligent data acquisition equipment. Multimodal coverage collection is then performed on this big data to obtain a multimodal dataset. Next, biometric identification is performed on the multimodal dataset to obtain a biometric set. This biometric set is then clustered to identify multiple biometric classes. Then, the impact of these biometric classes on drug activity is analyzed. Multiple activity impact levels are associated with these biometric classes, and a feature-activity mapping list is constructed to screen candidate drug parameters and determine target candidate drugs. Finally, target biometric data is extracted from the target candidate drugs for activity prediction. Prediction results are obtained, and a drug activity report is generated based on the prediction results and drug activity standards to optimize and provide feedback to the multimodal medical big data. This approach solves the technical problems of inaccurate prediction results and poor generalization ability caused by traditional drug activity prediction methods that rely on single-modal data and lack correlation analysis of complex biometrics. It achieves high-precision drug activity prediction using multimodal fusion and biometric clustering analysis, improving drug screening efficiency and intelligence.

[0037] Example 2, based on the same inventive concept as the drug activity prediction method based on multimodal medical big data in the previous examples, such as... Figure 2 As shown, this application provides a drug activity prediction system based on multimodal medical big data, wherein the system includes: Data acquisition module 11: Acquires multimodal medical big data through intelligent data acquisition equipment, and performs multimodal coverage acquisition based on the multimodal medical big data to obtain a multimodal dataset; Data analysis module 12: Performs biometric identification based on the multimodal dataset to obtain a biometric set, and performs cluster analysis on the multimodal dataset according to the biometric set to determine multiple biometric classes; Drug screening module 13: Performs impact analysis on drug activity according to the multiple biometric classes, sets multiple activity impact levels and associates them with the multiple biometric classes, constructs a feature-activity mapping list to screen candidate drug parameters, and determines target candidate drugs; Activity prediction module 14: Extracts target biometric data based on the target candidate drugs to perform activity prediction, obtains prediction results, and generates a drug activity report based on the drug activity standards and the prediction results to provide optimization feedback to the multimodal medical big data.

[0038] Furthermore, the data acquisition module 11 is used to perform the following methods: Multiple data acquisition channels are controlled by intelligent integrated circuits. Modulation switching of multimodal data sources is performed based on MOSFET switching technology, and initial acquisition parameters are set. When a multimodal data source is available, iterate through the multimodal data sources to identify data and divide the data acquisition area. The intelligent data acquisition device performs multimodal coverage acquisition in the data acquisition area according to the initial acquisition parameters to obtain multimodal medical big data. Based on the multimodal medical big data, the initial acquisition parameters of the intelligent integrated circuit are adjusted to set target acquisition parameters. The intelligent data acquisition device is then activated according to the target acquisition parameters to perform multimodal coverage acquisition to obtain the multimodal dataset.

[0039] Furthermore, the data acquisition module 11 is used to perform the following methods: Data quality analysis is performed based on the multimodal medical big data to identify multiple data interference data, including noisy interference data and missing interference data. The intelligent data acquisition device is then adjusted according to the noise interference data and its characteristics to determine the acquisition frequency adjustment information. Similarly, the intelligent data acquisition device is adjusted according to the missing interference data and its characteristics to determine the acquisition intensity adjustment information. Based on the acquisition frequency adjustment information and the acquisition intensity adjustment information, the initial acquisition parameters are adjusted, and the target acquisition parameters are set.

[0040] Furthermore, the data acquisition module 11 is used to perform the following methods: When the multimodal data source is available, the first sensor and the second sensor are activated; the first sensor is used to collect genomic data to obtain a first sensor dataset; the second sensor is used to collect proteomic data to obtain a second sensor dataset; a feature space coordinate system is constructed based on the multimodal dataset, and the first sensor dataset and the second sensor dataset are mapped to the feature space coordinate system to obtain the feature distribution parameters of the multimodal dataset; the multimodal data source is identified based on the feature distribution parameters of the multimodal dataset, and the data collection area is divided.

[0041] Furthermore, the data analysis module 12 is used to perform the following methods: A multimodal data feature extraction channel is constructed, and parallel feature extraction is performed on the multimodal dataset through the multimodal data feature extraction channel to obtain initial multimodal data features. The initial multimodal data features are normalized and combined to construct a multidimensional biometric vector. Based on the multidimensional biometric vector, weights are assigned to obtain a multidimensional weighted feature vector. The multidimensional weighted feature vector is then concatenated and fused to obtain a biometric set. Cluster centers are initialized based on the biometric set, and the multimodal dataset is clustered according to the cluster centers to determine the multiple biometric classes.

[0042] Furthermore, the data analysis module 12 is used to perform the following methods: A distance matrix is ​​constructed by combining intelligent integrated circuits with the cluster centroids. A cluster centroid update strategy is formulated by dynamically adjusting the cluster centroids using MOS transistor switching technology and the distance matrix. Based on the update strategy, the similarity between the multimodal dataset and the cluster centroids is calculated to obtain a similarity coefficient. The density between samples within a cluster is calculated based on the similarity coefficient and the distance matrix, defining compactness. The separation between cluster centroids within a cluster is calculated based on the similarity coefficient and the distance matrix, defining separation. The compactness and separation are jointly optimized to determine a clustering scheme. The clustering scheme is then applied to the multimodal dataset to determine the multiple biometric classes.

[0043] Furthermore, the drug screening module 13 is used to perform the following methods: The influence weights of drug activity are calculated based on the multiple biomarker classes to construct activity influence factors. Numerical boundary analysis is performed on the activity influence factors to obtain the influence value range. The activity influence factors are dynamically clustered according to the influence value range to determine the natural boundary points of the activity influence factors. Based on the natural boundary points, the activity influence factors are divided into multiple continuous intervals, and the multiple continuous intervals are labeled with influence levels to set multiple activity influence levels. The multiple biomarker classes are traversed to assign labels, generating biomarker class identifiers. The biomarker class identifiers are associated with the multiple activity influence levels to construct the feature-activity mapping list.

[0044] Furthermore, the activity prediction module 14 is used to perform the following method: Target biomarker data of the target candidate drug is extracted from the multimodal dataset and filtered to determine the target biomarker dimension. Forward propagation calculation is performed on the target biomarker data according to the target biomarker dimension to obtain the calculation result. Deep analysis is performed based on the calculation result to extract deep feature parameters. Activity prediction is performed based on the calculation result and the deep feature parameters to obtain the prediction result. The prediction result is compared with drug activity standards to generate activity risk data. Continuous statistics are performed on the prediction result according to the activity risk data to determine high-frequency activity patterns. The high-frequency activity patterns are correlated with time-series data of drug molecular parameters to obtain multiple parameter combinations, which are then randomly mined to obtain an optimized parameter combination. The optimized parameter combination is added to the drug activity report.

[0045] Furthermore, the activity prediction module 14 is used to perform the following method: The system determines whether the target biomarker data is in the feature-activity mapping list. If the target biomarker data is in the feature-activity mapping list, a high-activity prediction instruction is generated, and the target candidate drug is marked as a high-activity drug using the high-activity prediction instruction. Based on the calculation result, the high-activity drug is matched with the deep feature parameters to generate a first prediction result. If the target biomarker data is not in the feature-activity mapping list, a low-activity prediction instruction is generated, and the target candidate drug is marked as a low-activity drug using the low-activity prediction instruction. Based on the calculation result, the low-activity drug is matched with the deep feature parameters to generate a second prediction result.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for predicting drug activity based on multi-modal medical big data, characterized in that, The method comprises: acquiring multi-modal medical big data through an intelligent data acquisition device, and performing multi-modal coverage collection based on the multi-modal medical big data to obtain a multi-modal data set; performing biological feature recognition based on the multi-modal data set to obtain a biological feature set, performing cluster analysis on the multi-modal data set according to the biological feature set, and determining a plurality of biological feature classes; performing influence analysis on drug activity according to the plurality of biological feature classes, associating a plurality of activity influence levels with the plurality of biological feature classes, constructing a feature-activity mapping list to screen candidate drug parameters, and determining a target candidate drug; based on the target candidate drug, extracting target biological feature data to perform activity prediction, obtaining a prediction result, and generating a drug activity report according to a drug activity standard combined with the prediction result to optimize and feed back the multi-modal medical big data. 2.The method of claim 1, wherein, acquiring multi-modal medical big data through an intelligent data acquisition device, and performing multi-modal coverage collection based on the multi-modal medical big data to obtain a multi-modal data set, the method comprising: controlling a plurality of data acquisition channels using an intelligent integrated circuit, switching modes of multi-modal data sources based on MOS tube switching technology, and setting initial collection parameters; when the multi-modal data source is available, traversing the multi-modal data source to identify data, and dividing data acquisition areas; performing multi-modal coverage collection on the data acquisition areas according to the initial collection parameters through the intelligent data acquisition device to obtain multi-modal medical big data; adjusting the initial collection parameters of the intelligent integrated circuit according to the multi-modal medical big data, setting target collection parameters, and activating the intelligent data acquisition device to perform multi-modal coverage collection according to the target collection parameters to obtain the multi-modal data set. 3.The method of claim 2, wherein, adjusting the initial collection parameters of the intelligent integrated circuit according to the multi-modal medical big data, setting target collection parameters, the method comprising: performing data quality analysis based on the multi-modal medical big data to determine a plurality of data interference data, the plurality of data interference data including noise interference data and missing interference data; performing collection adjustment on the intelligent data acquisition device according to the noise interference data and data characteristics to determine collection frequency adjustment information; performing collection adjustment on the intelligent data acquisition device according to the missing interference data and data characteristics to determine collection intensity adjustment information; adjusting the initial collection parameters based on the collection frequency adjustment information and the collection intensity adjustment information to set the target collection parameters. 4.The method of claim 2, wherein, when the multi-modal data source is available, traversing the multi-modal data source to identify data, and dividing data acquisition areas, the method comprising: when the multi-modal data source is available, starting a first sensor and a second sensor; performing sensing collection on genomic data through the first sensor to obtain a first sensing data set; performing sensing collection on proteomic data through the second sensor to obtain a second sensing data set; based on the multi-modal data set, constructing a feature space coordinate system, mapping the first sensing data set and the second sensing data set to the feature space coordinate system to obtain feature distribution parameters of the multi-modal data set; Data identification is performed on the multi-modal data source based on the feature distribution parameters of the multi-modal data set, and the data acquisition area is divided. 5.The method of claim 2, wherein, Biological feature recognition is performed based on the multi-modal data set to obtain a biological feature set, and the multi-modal data set is subjected to cluster analysis according to the biological feature set to determine a plurality of biological feature classes, and the method comprises: A multi-modal data feature extraction channel is constructed, and parallel feature extraction is performed on the multi-modal data set through the multi-modal data feature extraction channel to obtain multi-modal initial data features; The multi-modal initial data features are normalized and combined to construct a multi-dimensional biological feature vector; Weight distribution is performed based on the multi-dimensional biological feature vector to obtain a multi-dimensional weighted feature vector, and the multi-dimensional weighted feature vector is subjected to feature splicing and fusion to obtain a biological feature set; The cluster center points are initialized based on the biological feature set, and the multi-modal data set is clustered and divided according to the cluster center points to determine the plurality of biological feature classes. 6.The method of claim 5, wherein, The cluster center points are initialized based on the biological feature set, and the multi-modal data set is clustered and divided according to the cluster center points to determine the plurality of biological feature classes, and the method comprises: A distance matrix is constructed by calculating the cluster center points using an intelligent integrated circuit; The cluster center points are dynamically adjusted by combining the distance matrix through MOS tube switching technology, and a cluster center point update strategy is formulated; Based on the cluster center point update strategy, the similarity between the multi-modal data set and the cluster center points is calculated to obtain a similarity coefficient; Based on the similarity coefficient and the distance matrix, the calculation is performed between the intra-cluster samples to define the compactness; Based on the similarity coefficient and the distance matrix, the calculation is performed between the intra-cluster cluster center points to define the separation degree; The compactness and the separation degree are jointly optimized to determine a cluster division scheme; The multi-modal data set is subjected to the cluster division scheme to determine the plurality of biological feature classes. 7.The drug activity prediction method based on multi-modal medical big data according to claim 1, wherein, The influence of drug activity is analyzed according to the plurality of biological feature classes, a plurality of activity influence levels are associated with the plurality of biological feature classes, and a feature-activity mapping list is constructed, and the method comprises: An influence weight calculation of drug activity is performed based on the plurality of biological feature classes to construct an activity influence factor; Numerical boundary analysis is performed according to the activity influence factor to obtain an influence numerical range; The activity influence factor is dynamically clustered according to the influence numerical range to determine a natural division point of the activity influence factor; Based on the natural division point, the activity influence factor is divided into a plurality of continuous intervals, the plurality of continuous intervals are subjected to influence level identification, and a plurality of activity influence levels are set; The plurality of biological feature classes are iterated for identification and distribution to generate a biological feature class identifier, the biological feature class identifier is associated with the plurality of activity influence levels, and the feature-activity mapping list is constructed. 8.The drug activity prediction method based on multi-modal medical big data according to claim 1, wherein, Based on the target candidate drug, target biological feature data is extracted for activity prediction to obtain a prediction result, and a drug activity report is generated according to a drug activity standard combined with the prediction result, and the method comprises: According to the multi-modal data set, target biological feature data of the target candidate drug is extracted for screening, and a target biological feature dimension is determined; According to the target biological feature dimension, forward propagation calculation is performed on the target biological feature data to obtain a calculation result, deep analysis is performed according to the calculation result, and deep feature parameters are extracted; Based on the calculation result and the deep feature parameters, activity prediction is performed to obtain the prediction result; The prediction result is compared with the drug activity standard for activity comparison, and activity risk data is generated; According to the activity risk data, the prediction result is continuously counted to determine a high-frequency activity mode; The high-frequency activity mode is associated with the drug molecular parameter time series data to obtain a plurality of parameter combinations for random mining to obtain a parameter optimization combination; The parameter optimization combination is added to the drug activity report. 9.The drug activity prediction method based on multi-modal medical big data according to claim 8, wherein, Based on the calculation result and the deep feature parameters, activity prediction is performed to obtain the prediction result, the method comprising: Determine whether the target biological feature data is in the feature-activity mapping list; If the target biological feature data is in the feature-activity mapping list, a high-activity prediction instruction is generated, and the target candidate drug is marked as a high-activity drug through the high-activity prediction instruction; Based on the calculation result, the high-activity drug is matched with the deep feature parameters to generate a first prediction result; If the target biological feature data is not in the feature-activity mapping list, a low-activity prediction instruction is generated, and the target candidate drug is marked as a low-activity drug through the low-activity prediction instruction; Based on the calculation result, the low-activity drug is matched with the deep feature parameters to generate a second prediction result.

10. A drug activity prediction system based on multi-modal medical big data, characterized by, The system for implementing the drug activity prediction method based on multi-modal medical big data of any one of claims 1-9, the system comprising: A data acquisition module: multi-modal medical big data is acquired through intelligent data acquisition equipment, and multi-modal data sets are obtained through multi-modal coverage collection based on the multi-modal medical big data; A data analysis module: biological feature recognition is performed based on the multi-modal data sets to obtain a biological feature set, and the multi-modal data sets are clustered and analyzed according to the biological feature set to determine a plurality of biological feature classes; A drug screening module: influence analysis of drug activity is performed according to the plurality of biological feature classes, a plurality of activity influence levels are associated with the plurality of biological feature classes, a feature-activity mapping list is constructed to screen candidate drug parameters, and a target candidate drug is determined; An activity prediction module: target biological feature data is extracted based on the target candidate drug for activity prediction to obtain a prediction result, and a drug activity report is generated according to the prediction result in combination with the drug activity standard to optimize feedback on the multi-modal medical big data.

Images