Multi-dimensional data-driven personalized conditioning regimen optimization method and system

Abstract

The application discloses a multi-dimensional data driven personalized conditioning scheme optimization method and system, and belongs to the technical field of traditional Chinese medicine conditioning scheme optimization. The method comprises the following steps: obtaining time series data of a target user, preprocessing and feature extraction, and constructing a time series feature set; calling a benchmark feature set, using transfer learning to generate an adaptive feature set; setting an initial observation window, combining the adaptive feature set to perform center adjustment and extension of the extraction range, generating an observation window set, and constructing a conditioning evolution curve; combining the mapping of a low-dimensional latent variable space to construct a state sequence; taking each observation window as a node, and taking the health state evolution of the state sequence between adjacent observation windows as a directed edge to construct a time series logic network; obtaining a current conditioning scheme, generating multiple groups of to-be-verified schemes, combining the time series logic network to simulate and deduce each group of to-be-verified schemes, generating a simulated state sequence, and obtaining a conditioning optimization scheme for the target user through dynamic conflict evaluation and multi-dimensional comparison and screening.

Description

Technical Field

[0001] This invention belongs to the field of conditioning program optimization technology, and relates to a method and system for optimizing personalized conditioning programs driven by multi-dimensional data. Background Technology

[0002] With the rapid development of artificial intelligence technology and its deep integration with the treatment field, the intelligentization, standardization, and precision of treatment have become the core trends in the industry's development.

[0003] A Chinese patent with publication number CN121051235A discloses a method and system for optimizing traditional Chinese medicine (TCM) treatment plans. The method includes: scanning rule texts within each rule module of a TCM treatment plan database; extracting a first-function term from a first rule text; querying the corresponding opposite term based on a preset TCM function term relationship table; scanning all rule texts within the rule module to which the first-function term belongs; extracting a second-function term from a second rule text; determining whether the second-function term is the same as the opposite term of the first-function term; if they are the same, marking the first rule text and the second rule text as a logically opposite rule pair, writing the logically opposite rule pair into a rule conflict log table and providing a notification; if they are different, continuing to scan other rule texts to extract new second-function terms.

[0004] Although existing technologies have achieved the effect of avoiding contradictory suggestions output by the system due to rule conflicts, their data processing architecture is essentially still a static rule matching mechanism, lacking the ability to dynamically correlate and analyze time-series data. Specifically, the conditioning process is a dynamic intervention process based on the evolution of the body's state over time, indicating that the effectiveness of the rules in the conditioning plan and their interrelationships have a strong time dependence. The data processing architecture of existing technologies adopts a static rule matching method, which makes it difficult to construct and process complex time-series logic across time windows. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for optimizing personalized treatment plans driven by multi-dimensional data. By processing time-series data from multiple treatment cycles of users, combining transfer learning to generate adaptive features, and constructing a time-series logic network to characterize the temporal correlation of states, the method replaces static rule matching and solves the problem that it cannot handle complex time-series logic across time windows.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Multi-dimensional data-driven methods for optimizing personalized treatment plans include:

[0008] Acquire time-series data of the target user over multiple treatment cycles, preprocess and extract features from the time-series data, and construct a time-series feature set;

[0009] By utilizing transfer learning, a baseline feature set is invoked as the source domain data, and the time-series feature set is used as the target domain data to generate an adaptive feature set.

[0010] The initial observation window is set based on the conditioning cycle. According to the adaptation feature set, the center adjustment and extraction range extension are performed on each initial observation window to generate an observation window set.

[0011] Based on the features of the observation window, the evolution law of features with conditioning cycle is fitted and continuousized to generate conditioning evolution curve. Combined with the mapping of low-dimensional latent variable space, a state sequence is constructed. Using the state sequence of each observation window as nodes, a time-series logic network is constructed.

[0012] The system obtains the treatment plan for the current treatment cycle and adjusts the baseline plan based on preset optimization rules to generate multiple sets of plans to be verified. It then uses a time-series logic network to simulate and extrapolate each set of plans to be verified, generating a simulated state sequence. By performing dynamic conflict evaluation and multi-dimensional comparison and screening on the simulated state sequence, the system obtains the treatment optimization plan for the target user.

[0013] Specifically, the steps for constructing a time-series feature set include:

[0014] For the acquired time-series data, a unique cycle number is assigned to each cycle according to the order of the treatment cycle; the time-series data includes physiological data and text data;

[0015] By using two-dimensional convolution, pixel-level features are extracted from tongue and facial images in physiological data to obtain image features;

[0016] Pulse characteristics are obtained by frequency domain transformation of pulse waveforms in physiological data using fast Fourier transform and continuous wavelet transform.

[0017] Based on the entity annotation set of the conditioning domain, entity recognition is performed on text data to extract domain entities, and Word2Vec is used for vectorized embedding to generate text features;

[0018] Quantitative data in physiological data are integrated into structural features, and feature vectors are constructed based on image features, pulse features, text features, and structural features under the same treatment cycle.

[0019] The feature vectors are arranged chronologically according to the period number to construct the temporal feature set of the target user.

[0020] Specifically, the steps for generating the adaptation feature set include:

[0021] Collect general conditioning time-series data, and construct a general feature set through preprocessing and feature extraction;

[0022] Based on the temporal feature set, the conditioning tags of the target users are extracted and encoded into target tag vectors;

[0023] Match the corresponding conditioning label to each feature data in the general feature set and generate a validation label vector;

[0024] By using cosine similarity, the label similarity between the target label vector and the validation label vector is calculated, and feature data with label similarity below the similarity threshold are removed, thereby generating a benchmark feature set;

[0025] Based on the baseline feature set and the time series feature set, the distribution difference is calculated using the maximum mean difference. With the goal of minimizing the distribution difference, the feature space is numerically calibrated dimension by dimension to obtain the aligned source domain feature set and target domain feature set, which are then divided into three feature layers.

[0026] For a single treatment cycle, the feature saliency and feature similarity of each target domain are obtained, and the migration weight is configured for each feature layer by combining the preset personalized coefficient and domain adaptation coefficient.

[0027] Based on the transfer weights, the source domain feature set is fused into the target domain feature set periodically and layer by layer, and the adaptive feature set is obtained according to the temporal relationship of the conditioning cycle.

[0028] Specifically, the steps for generating the observation window set include:

[0029] Set the initial window length, use a sliding window with the adjustment period as the sliding step, and perform directional partitioning of the adaptation feature set to generate the initial observation window set;

[0030] For each initial observation window, the conditioning period at the middle position within the window is selected as the central period, and the boundary period is defined.

[0031] Based on the feature extraction range of the initial observation window, the extension period is obtained by extending the window both before and after it.

[0032] The feature changes between the central period and adjacent periods within the window are used as transition features within the window, and the feature changes between the boundary period and the extended period are used as transition features between windows.

[0033] Based on the difference between the feature vectors of the two conditioning cycles, the transition feature value is calculated, and the transition feature set of the initial observation window is obtained through normalization.

[0034] Based on each initial observation window, the feature vector of the central period and the transition feature set are obtained, and the vectors are concatenated to generate the window vector;

[0035] The window vectors are standardized and all windows are integrated to generate an observation window set.

[0036] Specifically, the steps for generating conditioning evolution curves include:

[0037] Each window vector is split into sub-vectors of three feature layers, while the preset fitting parameters are called.

[0038] For each feature layer's sub-vectors, a polynomial fitting method is used, with the window center period number as the horizontal axis and the feature layer sub-vector value as the vertical axis, to generate the initial conditioning curve for each feature layer.

[0039] Based on the coefficient of determination, the goodness of fit between the initial conditioning curve of each feature layer and the actual value of the window feature is calculated.

[0040] If the goodness of fit is lower than the fitting accuracy threshold, the curve fitting is deemed unsatisfactory, and the fitting parameters are readjusted to perform fitting until the curve fitting of all feature layers meets the standard.

[0041] If the goodness of fit is not lower than the fitting accuracy threshold, the curve fitting is deemed satisfactory.

[0042] In response to the curve fitting meeting the standard, fusion weights are assigned to each feature layer, and the initial conditioning curves of the three feature layers are fused to generate the conditioning evolution curve of the target user.

[0043] Specifically, the steps for constructing a state sequence include:

[0044] Based on each observation window, the fitted values ​​of the conditioning evolution curve and the fitted values ​​of the initial conditioning curves of each feature layer are called, and the conditioning state is quantized and calibrated to generate the state vector of the observation window.

[0045] The conditioning state vectors of all observation windows are integrated, arranged in order according to the window number, and an initial conditioning state sequence is generated, which is then filtered and smoothed.

[0046] Using cosine similarity, the conditioning similarity between any two adjacent state vectors is calculated, and temporal coherence is verified by combining it with a preset coherence threshold.

[0047] If the conditioning similarity is below the coherence threshold, the state vector is recalibrated based on the conditioning evolution curve.

[0048] The smoothed and verified state vectors are rearranged in time sequence to generate a state sequence.

[0049] Specifically, the steps for constructing a temporal logic network include:

[0050] Each state vector in the state sequence is used as a state node, and the state evolution relationship between two adjacent state nodes is used as a network edge.

[0051] For any network edge, edge similarity is calculated using the cosine similarity of the state vectors at the endpoints, and edge correlation is calculated using the transition feature values.

[0052] The edge weight is calculated by weighted summation of edge similarity and edge correlation.

[0053] Arrange all state nodes linearly according to window number, and combine them with edge weights to construct the initial network;

[0054] Iterate through all edges of the initial network and remove network edges whose edge weights are less than the edge association threshold;

[0055] Traverse all non-adjacent state node pairs in the initial network, calculate cosine similarity, filter out node pairs with cosine similarity greater than the potential evolution threshold, add indirect association edges, assign edge weights, and label them as potential evolutionary association edges;

[0056] By integrating state nodes, adjacent edges, and potential evolutionary association edges, a temporal logic network is finally generated.

[0057] Specifically, the simulation steps include:

[0058] The last state vector in the state sequence is used as the initial state for deduction;

[0059] The simulation parameters are standardized, including simulation step size, influence coefficient, maximum intervention intensity, and correction coefficient.

[0060] For each simulation step, the intervention intensity of the scheme to be verified is obtained, and the rate of change of state is calculated in combination with the simulation parameters;

[0061] Traverse the temporal logic network to obtain the edge weights that are consistent with the conditioning stage of the scheme to be verified. Calculate the predicted edge weights by the arithmetic mean and combine them with the initial state and the deduction parameters to calculate the predicted state vector for the corresponding deduction step.

[0062] After a complete deduction step by step, all predicted state vectors are arranged in time sequence to generate a simulated state sequence of the scheme to be verified.

[0063] Using cosine similarity, the trend fit between the simulated state sequence and the conditioning evolution curve is calculated;

[0064] Set an effective threshold for the extrapolation. When the trend fit is greater than the effective threshold, the extrapolation is deemed effective.

[0065] All schemes to be verified are simulated in turn, and corresponding simulated state sequences are generated and the validity of the simulation is determined.

[0066] Specifically, the steps of dynamic conflict assessment and multi-dimensional comparative screening include:

[0067] Using time sequence conflict, intervention conflict, and label conflict as evaluation indicators, the evaluation indicator values ​​of the schemes to be screened were calculated.

[0068] The preset conflict weights are used to perform a weighted summation of the evaluation index values ​​to obtain the overall conflict degree of the solutions to be screened.

[0069] If the overall conflict level is less than the conflict threshold, the solution is retained; otherwise, the solution is discarded.

[0070] Integrate all the retained candidate schemes, plot the predicted evolution curve for each scheme, and simultaneously call the conditioning evolution curve to calculate the evolutionary stability through cosine similarity.

[0071] If the evolutionary stability is greater than the preset longitudinal stability threshold, the prognosis is determined to be stable and the scheme is retained; otherwise, the scheme is eliminated.

[0072] Calculate the treatment conversion rate and average status value of each screening scheme, perform quantitative scoring according to the preset prognosis evaluation rules, and calculate the horizontal comparison score of each screening scheme through arithmetic average calculation.

[0073] All options to be screened are sorted in descending order based on the horizontal comparison score, and the option with the highest score is selected as the conditioning and optimization option.

[0074] A multi-dimensional data-driven personalized treatment plan optimization system includes: a data acquisition module, an evolution module, and an optimization simulation module;

[0075] The data acquisition module is used to acquire time-series data of the target user, construct a time-series feature set through preprocessing and feature extraction, call the benchmark feature set, optimize the time-series feature set using transfer learning, generate an adaptive feature set, set an initial observation window, and capture the transition features of the boundary region of the observation window through center adjustment and extraction range extension to generate an observation window set.

[0076] The evolution module is used to perform evolution fitting and continuous processing on the features in the observation window set, construct the conditioning evolution curve of a single feature, integrate the conditioning evolution curves of all feature dimensions using low-dimensional mapping, construct the state sequence, and construct the temporal logic graph with the observation window as the node and the healthy evolution state of the state sequence between adjacent observation windows as the directed edge.

[0077] The optimization simulation module constructs multiple sets of verification plans by calling the current treatment cycle's conditioning plan and preset optimization rules. It then uses a time-series logic network to simulate and deduce each set of verification plans, generating a simulated state sequence. By performing dynamic conflict evaluation and multi-dimensional comparison and screening on the simulated state sequence, the module obtains the conditioning optimization plan for the target user.

[0078] The beneficial effects of this invention are:

[0079] This paper constructs a time-series feature set for target user time-series data and uses transfer learning to fuse general conditioning knowledge from the source domain with personalized features from the target domain to generate an adapted feature set. This addresses the problems of sparse time-series data for single users and the lack of cross-domain knowledge reuse capabilities in static architectures. Based on conditioning cycle division and optimized observation windows, discrete cycle features are transformed into window-level features with time-series correlations, building an analysis carrier across time windows and overcoming the shortcomings of existing technologies in constructing cross-cycle feature correlation carriers. Through conditioning evolution curve fitting, state sequence construction, and time-series logic network construction, scattered features are transformed into a quantifiable and correlated dynamic time-series logic model, comprehensively depicting the... The study examines the time-dependent relationships between the body's state evolution over time during the treatment process, overcoming the core shortcoming of static mechanisms that lack dynamic temporal correlation analysis capabilities. Based on a temporal logic network, it simulates and extrapolates multiple treatment plans, and combines dynamic conflict assessment and multi-dimensional screening to obtain optimized plans. This allows the entire plan optimization process to be based on the user's personalized temporal evolution patterns and dynamically optimized in combination with the user's temporal evolution characteristics. This not only enables the reuse of general treatment knowledge but also fully preserves the personalized characteristics of individual treatment, making the treatment plan highly compatible with the temporal changes in the body's state. This significantly improves the rationality and dynamic adaptability of the plan, making it more in line with the clinical reality of gradual and dynamic intervention in treatment. Attached Figure Description

[0080] Figure 1 A schematic diagram illustrating a method for optimizing personalized treatment plans driven by multidimensional data;

[0081] Figure 2 This is a flowchart of the process for generating the adaptation feature set in this invention;

[0082] Figure 3 This is a structural diagram of the window vector construction in this invention;

[0083] Figure 4 This is a flowchart of the simulation and deduction process in this invention. Detailed Implementation

[0084] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0085] Example 1

[0086] refer to Figures 1 to 4 As shown, this embodiment introduces a method for optimizing personalized treatment plans driven by multidimensional data, including the following steps:

[0087] Step S1: Obtain time-series data of the target user across multiple treatment cycles, and preprocess and extract features from the time-series data to achieve data characterization and standardization, thereby constructing a time-series feature set for the target user. This includes pixel-level feature extraction from images, frequency domain transformation of waveforms, entity recognition and vector embedding of text. The time-series data includes physiological data and text data. Physiological data includes pulse waveforms, tongue images, and facial images. Text data includes historical treatment plans and efficacy feedback data.

[0088] Step S2: Due to the insufficient sample size and sparse feature information of a single target user, and the analysis based solely on fixed local rules and limited data, lacking cross-domain features or knowledge reuse capabilities, the optimization of treatment plans for a single individual will lack sufficient feature support, easily leading to analytical bias. Transfer learning is utilized, calling the baseline feature set as the source domain data and the current target user's temporal feature set as the target domain data. Through feature space alignment, the difference in feature distribution between the source and target domains is eliminated, enhancing and completing the target user's temporal features. By dynamically allocating the transfer weights of features, the source domain features are prevented from masking the target user's personalized features, preserving the target user's personalized feature attributes. Ultimately, an adaptive feature set is generated for the target user, incorporating both general TCM treatment knowledge and perfectly matching the target user's individual characteristics. The general temporal feature set contains a large amount of publicly available TCM treatment annotation data and general treatment temporal features for people with the same symptoms / constitution.

[0089] Step S3: Using the treatment cycle as the basic time unit, set an initial observation window for the target user. According to the time sequence of the treatment cycle, divide the adaptive feature set into the corresponding initial observation window in turn. At the same time, considering that the health status of the target user changes gradually during the treatment process, and that the boundary transition features reflect the state evolution, perform central adjustment and extraction range extension on each initial observation window to capture the transition features of the boundary area of ​​the observation window, so that the features are closer to the real health status of the target user, thereby generating an observation window set.

[0090] Step S4: Using the features in the observation window set as the analysis object and the treatment cycle as the time basis, fit the trajectory function of the feature evolution with the treatment cycle. Since the feature points are discrete, use the kernel smoothing method to make the discrete feature points in the trajectory function continuous, eliminate the fluctuation error of discrete data, and generate the treatment evolution curve to characterize the dynamic change process of the target user's health status under treatment intervention. Extract the feature vector corresponding to each treatment cycle from the treatment evolution curve, map the feature vector in each observation window to the low-dimensional latent variable space, remove redundant information in the high-dimensional features, extract the feature dimension that characterizes the target user's health status, construct a state sequence for a single observation window of the target user, and use each observation window as a state sequence node to connect the scattered observation windows through the state evolution relationship, construct a temporal logic network for the target user, and realize the modeling and visualization of the complex temporal logic across time windows in the multi-cycle treatment process of the target user.

[0091] Step S5: Obtain the treatment plan for the current treatment cycle as the baseline plan. Adjust the baseline plan based on preset optimization rules to generate multiple sets of plans to be verified. Based on the causal relationship between the adjustment plan and health status in the time-series logic network, simulate and extrapolate each set of plans to be verified to extrapolate the health status evolution of the target user in the next observation window, thereby generating a simulated state sequence. By performing dynamic conflict evaluation and multi-dimensional comparison and screening on the simulated state sequence, obtain the optimized treatment plan for the target user. At the same time, after the end of each treatment cycle, collect new efficacy feedback data to update the time-series feature set and provide the latest data support for the plan optimization and simulation extrapolation of the next treatment cycle. The optimization rules are set by those skilled in the art, such as the optimization range of physiotherapy frequency, physiotherapy combination methods, and prescription compatibility adjustment criteria.

[0092] Furthermore, the steps for constructing a temporal feature set include:

[0093] For the acquired time-series data, a unique cycle number is assigned to each cycle according to the order of the conditioning cycles. ;in, , The total number of treatment cycles completed by the target user;

[0094] According to the characteristics of different types of data in time-series data, perform classification preprocessing, including: for structured physiology, use Grubbs' test to detect and remove outliers from quantified data, use the difference between adjacent cycles to complete missing values, and eliminate the difference in data dimensions through normalization. For unstructured physiology, use wavelet denoising to eliminate baseline drift and filter noise for pulse waveforms, and use Gaussian filtering to perform grayscale conversion, size normalization, and deblurring for images. For text data, combine the Traditional Chinese Medicine Dictionary to construct a domain-specific word segmentation dictionary for the Traditional Chinese Medicine field, use the Jieba word segmentation method to accurately segment the text, avoid the fragmentation of Traditional Chinese Medicine professional vocabulary caused by general word segmentation, construct a stop word list for the Traditional Chinese Medicine field that includes words without actual semantic meaning, such as "de", "jinxing", "jianyi", remove the stop words in the text after word segmentation, and retain the core semantic vocabulary;

[0095] Through two-dimensional convolution, perform pixel-level feature extraction on tongue image and facial image to obtain image features, including: segment the tongue body and tongue coating regions of the tongue image, and extract the texture features, color features, and morphological features of each region respectively, and extract the facial color, freckle, and luster features from the facial image;

[0096] Perform frequency domain transformation on the pulse waveform through fast Fourier transform and continuous wavelet transform to obtain pulse features, including: extract the frequency domain features in the pulse waveform through fast Fourier transform, such as the main frequency, frequency distribution, and energy spectrum, and extract the time-frequency joint features of the pulse waveform through continuous wavelet transform;

[0097] This study employs BERT named entity recognition combined with a therapeutic domain entity annotation set to perform entity recognition on cleaned text data. Core entities in the TCM domain are extracted, such as prescription names, herbal names, treatment methods, symptom names, and constitution types. The identified text data is then vectorized and embedded using Word2Vec to convert discrete text entities into numerical forms, thereby generating text features. The therapeutic domain entity annotation set was constructed by those skilled in the art based on TCM disease and syndrome classifications and codes, and TCM clinical diagnostic terminology. Entity classifications include TCM syndrome entities, symptom and sign entities, Chinese herbal medicine / prescription entities, therapeutic intervention entities, and therapeutic cycle entities. This study identifies entities that are contraindicated in traditional Chinese medicine (TCM) and uses the BIO annotation system. It covers the entire process of TCM internal medicine terminology, TCM constitution identification, and TCM intervention protocols. Entity recognition is performed using the BERT named entity recognition model. The network structure includes a BERT-Chinese base model pre-trained based on TCM, employing a 12-layer Transformer encoder, 768 hidden layers, 12 self-attention heads, and 3072 fully connected layers. The output layer consists of a 6-class fully connected layer (corresponding to 6 major TCM entity classes) and a CRF layer. The training dataset is a publicly available standardized TCM dataset. Key hyperparameters include a maximum sequence length of 128 and a batch size of 16. The learning rate was 2e-5, dropout rate 0.1, optimizer AdamW, weight decay 1e-4, and pre-trained BERT weights were loaded. The parameters of the bottom 6 encoder layers were frozen. The model was fine-tuned using a labeled dataset. The first two epochs only trained the output layer, and subsequent epochs allowed training of the entire network. After each training round, the entity recognition F1 score was calculated on the validation set. The convergence criterion was set as follows: if the entity recognition F1 score on the validation set was ≥0.92 and the F1 score improvement was <0.001 for three consecutive epochs, training was stopped. Word2Vec used a Skip-gram model, which includes an input layer, a projection layer, and an output layer. Negative sampling is used to optimize training efficiency. The training dataset is consistent with the BERT model training dataset. User-generated text corpus is supplemented to ensure terminology coverage. Core hyperparameters include word vector dimension 128, context window size 5, negative sampling number 5, number of training iterations 10, initial learning rate 0.025, and minimum learning rate 1e-4. The TCM text corpus is preprocessed by word segmentation and stop word removal, a vocabulary is constructed, the word vector matrix is ​​initialized, and iterative training is performed according to the Skip-gram model. The learning rate is updated after each iteration, and the convergence criterion is set to stop training if the model loss function decreases by less than 0.001 and there is no significant decrease for two consecutive iterations.

[0098] Based on image features, pulse features, text features, and structural features, Z-score normalization is used to eliminate scale differences within features. Simultaneously, features from the same conditioning cycle are concatenated and fused to generate a conditioning cycle. eigenvectors ;in, , For image feature vectors, This is the pulse feature vector. For text feature vectors, For structural feature vectors;

[0099] Based on the cycle number, the feature vectors of the treatment cycle are arranged temporally to construct a temporal framework with the treatment cycle as the time axis and the single-cycle feature vector as the core element, thus building a temporal feature set for the target user. ;in, , A 3D matrix, is the dimension of the single-period feature vector.

[0100] Furthermore, the steps for generating the adaptation feature set include:

[0101] We collected publicly available TCM conditioning annotation data and general conditioning time-series data for people with the same symptoms / constitution. Based on the feature dimensions in the time-series feature set, we performed same-dimensional classification. At the same time, through preprocessing and feature extraction, we constructed a general feature set.

[0102] Based on the time-series feature set, the conditioning tags of the target users are extracted, including disease type, constitution type, core signs, and intervention history, and encoded into target tag vectors as the basis for source domain screening. For example, the target tag vector is chronic gastritis, phlegm-dampness constitution, swollen tongue with white coating, and moxibustion conditioning.

[0103] Match a corresponding conditioning label to each feature data in the general feature set, and bind a validation label vector to the feature data;

[0104] By using cosine similarity, the label similarity between the target label vector and the verification label vector is calculated. The general feature set is then targeted and filtered by a preset similarity threshold to remove irrelevant feature data with label similarity below the similarity threshold, thereby generating a benchmark feature set.

[0105] Based on the baseline feature set and the time series feature set, the distribution variability is calculated using the maximum mean difference. ;

[0106] With minimizing distributional dissimilarity as the optimization objective, the feature space of the baseline feature set and the temporal feature set is numerically calibrated dimension by dimension to obtain the aligned source domain feature set and target domain feature set, including: setting a convergence threshold. and with As a standard for spatial alignment, for each feature dimension, the mean deviation and variance deviation of the source and target domains are calculated separately. The source domain features in the corresponding dimension are then calibrated and their variance scaled based on the distribution dissimilarity. After a single dimensional calibration, the distribution dissimilarity is recalculated. If the updated distribution dissimilarity meets the spatial alignment standard, the calibration result is retained; otherwise, the calibration is withdrawn, the calibration amplitude is adjusted, and the operation is repeated until the distribution dissimilarity is less than the convergence threshold. The convergence threshold is set by those skilled in the art based on the actual fusion requirements of TCM conditioning features, such as... Mean calibration is to bring the source domain feature values ​​closer to the mean of that dimension of the target domain, eliminating the mean shift between domains. Variance scaling is to scale the source domain feature values ​​according to the variance ratio between the source domain and the target domain, matching the degree of dispersion of the numerical distribution of the two.

[0107] For example, the target domain feature set focuses on the characteristic dimensions of users with phlegm-dampness constitution, selecting two core features: tongue moisture value and pulse slipperiness value. Three treatment cycles have been completed. In cycle 1, the tongue moisture value was 0.8 and the pulse slipperiness value was 0.7. In cycle 2, the feature values ​​were 0.75 and 0.72, and in cycle 3, the feature values ​​were 0.82 and 0.68. Therefore, the mean of the tongue moisture value... ,variance The mean value of the pulse slip rate is variance is ;

[0108] The source domain contains three samples: sample 1 has feature values ​​of 0.6 and 0.5, sample 2 has feature values ​​of 0.65 and 0.55, and sample 3 has feature values ​​of 0.58 and 0.52. The corresponding mean values ​​of tongue image moisture are... ,variance The mean value of the slippery pulse rate ,variance At the same time, set a convergence threshold. ;

[0109] Distribution variability using a linear kernel function ;in, It is the Euclidean L2 norm, used to quantify the difference between the means of two populations distributed in a high-dimensional space;

[0110] For tongue moisture values, mean calibration is first performed, using the initial feature value plus the difference between the target domain mean and the source domain mean for the corresponding feature dimension. For example, the feature value of sample 1 after mean calibration is... Then, variance scaling is performed, with the scaling factor being the ratio of the standard deviation of the target domain to the standard deviation of the source domain in the corresponding dimension. However, after scaling sample 1 by the scaling factor, the feature value is Similarly, the feature value of sample 2 is 0.799 and the feature value of sample 3 is 0.733. The calibrated mean is... The deviation from the mean of the target domain is 0.029;

[0111] Similarly, after calibration, the characteristic values ​​of the pulse slip rate for each sample are 0.643, 0.69, and 0.662, respectively. The calibrated mean value at this point is... The deviation from the mean of the target domain is 0.035.

[0112] At this point, the degree of distribution difference The calibrated sample values ​​are retained.

[0113] The aligned source and target feature sets are divided into three feature layers: high personalization, medium personalization, and low personalization, denoted as , ... , , Among them, high personalization includes image features and pulse features, medium personalization includes structural features, and low personalization includes text features;

[0114] For a single conditioning cycle, the feature saliency of each target domain is obtained, the feature similarity between each feature layer and the target domain is calculated, and the feature saliency and feature similarity are weighted by a preset personalized coefficient and domain adaptation coefficient, thereby configuring a transfer weight for each feature layer, as shown in the following expression:

[0115]

[0116] In the formula, For the first The target user in the feature layer Migration weight for each conditioning cycle , For the first The domain adaptation coefficients of the feature layer are set by those skilled in the art based on the distributional differences between the source and target domains and the amount of data in the target domain, such as... , , , For the first The source domain and target domain in the feature layer The feature similarity of each treatment cycle is calculated using cosine similarity. For the first The individualization coefficient of the feature layer, summed with the domain adaptation coefficient, equals 1. This coefficient is used to balance source domain fusion and target domain preservation. For target users The feature significance of each treatment cycle was obtained by normalizing the Pearson correlation coefficient between the user's multi-cycle time series data and the TCM syndrome labels.

[0117] Based on the hierarchical transfer weights, the source domain feature set is fused into the target domain feature set periodically and layer by layer, and the adaptive feature set is obtained according to the temporal relationship of the conditioning cycle.

[0118] For example, The layer's domain fit coefficient is 0.2, and its personalization coefficient is 0.8. In the second conditioning cycle for the target user, the calibrated feature significance is 0.9, and the similarity to the target domain features is 0.85. Therefore, the corresponding transfer weight... ;

[0119] when Given the pulse slippage value in the layer, the eigenvalue of the target domain is 0.72, and the mean value of the source domain after calibration is 0.665, then the fused eigenvalue is... .

[0120] Furthermore, the steps for generating the observation window set include:

[0121] Set the initial window length A sliding window is used, with the adjustment cycle as the sliding step size. Starting from the first conditioning cycle, it is continuously... Using a rule of 1-minute window per period, the adaptive feature set is directionally divided to generate initial observation windows, which are then numbered and denoted as follows: The corresponding feature vector is The initial observation windows are integrated to generate an initial observation window set. ;in, Set by those skilled in the art, and is a positive integer, such as , , The initial total number of windows satisfies ;

[0122] For each initial observation window The treatment cycle covered is Select the treatment cycle in the middle of the window as the central cycle. As the core benchmark for window features, it avoids interference from boundary periodic errors and simultaneously defines the boundary period. , Based on the feature extraction range of the initial observation window, the adjustment period is extended by one conditioning cycle before and after the window, and is denoted as the extension cycle. , To capture transition features between adjacent windows, if the window is the first window, it extends only backward; if it is the last window, it extends only forward. Where, if If it is an odd number, then ,like If it is an odd number, then , for , for ;

[0123] Based on the central cycle Adjacent periods within the window ( , The characteristic changes between the boundary period and the extended period are used as the transition features within the window, and the characteristic changes between the boundary period and the extended period are used as the transition features between windows. In other words, the transition features within the window are... to , to The characteristic changes, the transition characteristics between windows are to , to The feature changes are such that the boundary window only extracts one-sided transition features;

[0124] Based on the difference in eigenvectors between two conditioning cycles, the transition eigenvalues ​​are calculated and min-max normalization is performed to obtain the initial observation window. transition feature set ;in, The total number of transition features; the expression is as follows:

[0125]

[0126] In the formula, For period To the cycle Transition eigenvalues For period eigenvectors, For period eigenvectors, For period With period The period difference is always 1, which is used to standardize the gradient value and avoid the period difference affecting the quantization of the change amplitude.

[0127] Based on each initial observation window, the feature vector and transition feature set of the corresponding central period are obtained. The vectors are concatenated to generate the window vector of the initial observation window. The scale difference between different window features is eliminated by the Z-score normalization method to ensure that the features of each window are comparable. The observation window set is generated by sorting the window numbers in order.

[0128] Furthermore, the steps for generating conditioning evolution curves include:

[0129] Using the window vectors in the observation window set as the basic data for fitting the conditioning evolution curve, each window vector is split into... , , The sub-vectors of the three feature layers simultaneously call preset fitting parameters, including fitting period, fitting accuracy threshold, and smoothing coefficient. Among them, the fitting coefficient is calibrated by those skilled in the art based on the characteristics of TCM conditioning cycle. For example, the fitting period is consistent with the conditioning cycle corresponding to the window, and the center period of the window is used as the fitting node. The fitting accuracy threshold is used to judge the degree of fit between the fitting curve and the actual feature value. For example, the fitting accuracy threshold is set to 0.85. The smoothing coefficient is used to balance the trend of the curve and the noise suppression effect. For example, the smoothing coefficient is set to 0.1.

[0130] For the sub-vectors of each feature layer, a polynomial fitting method is used, with the window center period number as the horizontal axis and the feature layer sub-vector value as the vertical axis, to generate the initial conditioning curve for each feature layer. This includes: because highly personalized feature fluctuations better reflect individual user conditioning changes, a cubic polynomial fitting method is used. The sub-vectors of the layer preserve the evolution trend of the target user's personalized features, reduce smoothing processing, and are fitted using a quadratic polynomial. The sub-vectors of the layer, balancing smoothness and trend, conform to the gradual change pattern of structured physiological characteristics, and are fitted using a first-order polynomial. The sub-vectors of the layer enhance smoothness and highlight the evolutionary trend of the general conditioning principles of traditional Chinese medicine;

[0131] Based on the coefficient of determination, the goodness of fit between the initial conditioning curve of each feature layer and the actual value of the window feature is calculated.

[0132] If the goodness of fit is lower than the fitting accuracy threshold, the curve fit is deemed substandard, and the fitting parameters are readjusted. Change the cubic polynomial of the layer to a quadratic polynomial, or adjust the smoothing coefficient to 0.15, and repeat the fitting until the curve fitting of all feature layers meets the standard. If the standard is still not met after 3 consecutive adjustments, switch the fitting algorithm to local weighted regression.

[0133] If the goodness of fit is not lower than the fitting accuracy threshold, the curve fitting is deemed satisfactory, and the current fitting parameters are retained.

[0134] In response to the curve fitting criteria being met, and based on the logic of prioritizing personalized features, fusion weights are assigned to each feature layer, such as... The layer weight is 0.4. The layer weight is 0.3. The weight of the layer is 0.3, which is used to fuse the initial conditioning curves of the three feature layers to generate a conditioning evolution curve exclusive to the target user.

[0135] Furthermore, the steps for constructing the state sequence include:

[0136] Based on the central period of each observation window, the fitted values ​​of the conditioning evolution curve and the fitted values ​​of the initial conditioning curves of each feature layer are retrieved, and the conditioning state is quantified and calibrated to generate the observation window. state vector ;in, for The state values ​​of the layer are determined by The initial conditioning curve of the layer is in the first stage. The fitted values ​​at the center period of each observation window were obtained after normalization. for The state values ​​of the layer are determined by The initial conditioning curve of the layer is in the first stage. The fitted values ​​at the center period of each observation window were obtained after normalization. for The state values ​​of the layer are determined by The initial conditioning curve of the layer is in the first stage. The fitted values ​​at the center period of each observation window were obtained after normalization. The overall state value is determined by... , , The summation is obtained by weighting the fusion weights of each feature layer, which represents the overall health conditioning status of the user within the corresponding observation window.

[0137] For example, the central period of the third observation window is the seventh conditioning period, after extraction and normalization. The normalized result of the layer fitting value is 0.75. The normalized result of the layer fitting value is 0.68. The normalized result of the layer fitting value is 0.72, combined with The layer weight is 0.4. The layer weight is 0.3. If the layer weight is 0.3, then the overall state value is Then the state vector of the third observation window ;

[0138] The conditioning state vectors of all observation windows are integrated, arranged in order according to the window number, and an initial conditioning state sequence is generated. The initial conditioning state sequence is then smoothed using the moving average filtering method to eliminate the noise of state fluctuations in a single cycle, ensuring the continuity and stability of the state sequence and conforming to the gradual change law of TCM conditioning.

[0139] Using cosine similarity, the conditioning similarity between any two adjacent state vectors is calculated, and temporal coherence is verified by combining it with a preset coherence threshold. If the conditioning similarity is lower than the coherence threshold, it indicates a sudden change in state, which does not conform to the gradual characteristics of TCM conditioning. In this case, the original features of the corresponding cycle are extracted first to correct the feature extraction error. If there is no error in the features, the state vectors are recalibrated using linear interpolation based on the state vectors of the two cycles before and after. After recalibration, the coherence is verified again. If it still does not meet the standard, it is marked as an abnormal cycle and smoothed using the moving average method. The coherence threshold is set by those skilled in the art based on the clinical laws of gradual changes in TCM conditioning and the statistical distribution of adjacent state similarity, such as 0.85.

[0140] The smoothed and verified state vectors are rearranged in time sequence to generate a state sequence. .

[0141] Furthermore, the steps for constructing a temporal logic network include:

[0142] Each state vector in the state sequence is taken as a state node, denoted as Each node is labeled with its corresponding conditioning cycle, window number, sub-feature layer state value and comprehensive state value, and is also associated with transitional feature values; among them, the conditioning cycle number of the state node is the center window number of the observation window;

[0143] The state evolution relationship between two adjacent state nodes is taken as a network edge, denoted as , indicating from node Pointing to node The edges are oriented in the same direction as the conditioning cycle, and the weight of the edges represents the correlation strength between two adjacent state nodes.

[0144] For any network edge, edge similarity is calculated using the cosine similarity of the state vectors at the endpoints, edge correlation is calculated using the transition feature values, and edge weight is calculated by weighted summation of edge similarity and edge correlation. The expression is as follows:

[0145]

[0146] In the formula, For network edge edge weights, The weighting balancing coefficients are set by those skilled in the art based on their clinical experience in traditional Chinese medicine treatment, such as... Used to balance the influence of state similarity and transition features. for The and the first Cosine similarity of state vectors For the first The corresponding period for each node is up to the [number]th node. The transition characteristic value of each node corresponds to a period;

[0147] All state nodes are linearly arranged according to the order of their conditioning cycles to form a chain-like node structure. Edge weights are assigned to each pair of adjacent state nodes to construct the initial network, denoted as […]. ;

[0148] Traverse all edges of the initial network and remove network edges whose edge weights are less than a preset edge association threshold; wherein, the edge association threshold is set by those skilled in the art in combination with the principles of traditional Chinese medicine conditioning, such as setting the edge association threshold to 0.5;

[0149] Traverse all non-adjacent state node pairs in the initial network, such as and Calculate the cosine similarity of non-adjacent state node pairs, filter out node pairs with a cosine similarity greater than a preset potential evolution threshold, and add indirect association edges to the selected node pairs, labeling them as potential evolution association edges. Configure the corresponding edge weights in accordance with the edge weight configuration formula of the network edges.

[0150] By integrating state nodes, adjacent edges, and potential evolutionary association edges, a temporal logic network is finally generated.

[0151] Furthermore, the simulation steps include:

[0152] Using the last state vector in the state sequence as the initial state for deduction, denoted as... It includes the last treatment cycle completed by the target user, and marks the initial comprehensive status value, sub-feature layer status value and corresponding transition feature value. At the same time, a set of solutions to be verified is randomly selected, and each set of solutions includes the intervention type, number of intervention cycles and intervention intensity.

[0153] Standardize the simulation parameters, including the simulation step size. Influence coefficient Maximum intervention intensity Correction coefficient The deduction parameters were all preset by those skilled in the art based on the principles of traditional Chinese medicine conditioning, and the deduction step size was one deduction step size for a single conditioning cycle. , The number of intervention cycles in the scheme to be verified; the influence coefficient is used to quantify the driving effect of intervention on state changes, such as setting it to 0.9; the maximum intervention intensity is used to avoid over-intervention, such as setting it to 0.8; and the correction coefficient is used to conform to the gradual conditioning principle of traditional Chinese medicine, such as setting it to 0.95.

[0154] For each simulation step To obtain the intervention intensity of the proposed solution. The rate of change of state is calculated by combining the inference parameters. This characterizes the magnitude of state change per unit conditioning cycle; the expression is shown below:

[0155]

[0156] Obtain the conditioning stages of the solution to be verified, including the initial, stable, and final stages. Simultaneously, traverse the conditioning stages of each edge in the temporal logic network to obtain the edge weights in the same conditioning stage as the solution to be verified, and calculate the predicted edge weights using the arithmetic mean. Combined with the initial state of the deduction Calculate the first... with the derivation parameters. The predicted state vector of each deduction step size The expression is as follows:

[0157]

[0158] After a complete deduction step by step, all predicted state vectors are arranged in time sequence to generate a simulated state sequence of the scheme to be verified.

[0159] Using cosine similarity, the trend fit between the simulated state sequence and the conditioning evolution curve is calculated, and an effective threshold for the extrapolation is set, such as 0.82. Only when the trend fit is greater than the effective threshold for the extrapolation is the extrapolation considered valid and recorded as a scheme to be screened; otherwise, the scheme to be verified is directly removed.

[0160] All schemes to be verified are simulated in sequence, and corresponding simulated state sequences are generated and the simulations are judged to be valid. Multiple sets of schemes to be screened and their corresponding simulated state sequences are generated.

[0161] Furthermore, the steps of dynamic conflict assessment and multi-dimensional comparative screening include:

[0162] The evaluation index values ​​of the selected solutions are calculated using time-series conflict, intervention conflict, and label conflict as evaluation indicators. Among them, time-series conflict is calibrated by trend fit, that is, the sum of time-series conflict and trend fit is 1. Intervention conflict is the absolute difference between the intervention intensity and the final state change rate of the selected solution. Label conflict is graded and quantified by those skilled in the art based on the adaptability of the intervention method and the user's adjustment label. For example, when it is fully adapted, the label conflict quantification value is 0.1, and the label conflict quantification value corresponding to mild conflict is 0.7.

[0163] The preset conflict weights are invoked to perform a weighted summation of the evaluation index values ​​to obtain the comprehensive conflict degree of the schemes to be screened; wherein, the conflict weights are set by those skilled in the art in combination with clinical experience in traditional Chinese medicine, such as being set to 0.4, 0.2, and 0.4 respectively.

[0164] Set a conflict threshold and conduct a conflict assessment; if the overall conflict level is less than the conflict threshold, it is determined that there is no obvious conflict and the solution to be screened is retained; if the overall conflict level is not less than the conflict threshold, it is determined that there is a serious conflict and the solution to be screened is removed.

[0165] Integrate all the retained schemes to be screened, calculate the state difference between adjacent inference steps based on the comprehensive state value, and calculate the conversion rate of each inference step with the difference between 1 and the comprehensive state value of the current inference step as the denominator and the state difference as the numerator. Extract the predicted comprehensive state value and conversion rate from each scheme, plot the predicted evolution curve, and call the conditioning evolution curve to calculate the evolution stability through cosine similarity.

[0166] If the evolutionary stability is greater than the preset longitudinal stability threshold, the prognosis is determined to be stable and the scheme is retained; if the evolutionary stability is not greater than the longitudinal stability threshold, the prognosis is determined to be abnormal and the corresponding scheme is eliminated.

[0167] The conditioning conversion rate of each candidate scheme is calculated by averaging the conversion rates of all the extrapolation steps, and the average state value of each candidate scheme is obtained by averaging all the comprehensive state values.

[0168] Based on the preset prognostic evaluation rules, the average status value and treatment conversion rate are quantitatively scored, and the horizontal comparison score of each candidate solution is calculated by arithmetic average calculation. For example, the prognostic evaluation rules include: average status value ≥ 0.85 gets 10 points, 0.8-0.85 gets 8 points, 0.75-0.8 gets 6 points, and < 0.75 gets 4 points; treatment conversion rate ≥ 80% gets 10 points, 70%-80% gets 8 points, 60%-70% gets 6 points, and < 60% gets 4 points.

[0169] All options to be screened are sorted in descending order based on the horizontal comparison score, and the option with the highest score is selected as the conditioning and optimization option.

[0170] Example 2

[0171] Another embodiment of the present invention: a multi-dimensional data-driven personalized treatment plan optimization system, comprising: a data acquisition module, an evolution module, and an optimization simulation module;

[0172] The data acquisition module is used to acquire time-series data of the target user in multiple treatment cycles and update the time-series data in real time after optimization simulation. Through preprocessing and feature extraction, a time-series feature set is constructed for the target user. At the same time, the baseline feature set is called and the personalized features of the target user are supplemented by transfer learning to optimize the time-series feature set, thereby generating an adaptive feature set for the target user. This realizes the reuse of general knowledge and the preservation of personalized features. At the same time, the adaptive feature set is divided through a preset initial observation window. Based on each initial observation window after division, the transition features of the observation window boundary area are captured through central adjustment and extraction range extension to generate an observation window set.

[0173] The evolution module is used to perform evolution fitting and continuous processing on the features in the observation window set, thereby constructing a conditioning evolution curve for a single feature. Using low-dimensional mapping, a state sequence is constructed for a single observation window of the target user through the conditioning evolution curves of all feature dimensions. At the same time, the health state evolution is obtained through the state sequences of adjacent observation windows. With the observation window as the node and the health evolution state as the directed edge, a temporal logic graph is constructed for the target user.

[0174] The optimization simulation module constructs multiple sets of verification plans by calling the current treatment cycle's conditioning plan and preset optimization rules. Through the causal relationship between the adjustment plan and health status in the time-series logic network, it simulates and extrapolates the health status evolution of the target user in the next observation window, thereby generating a simulated state sequence. By performing dynamic conflict evaluation and multi-dimensional comparison and screening on the simulated state sequence, the optimization plan for the target user is obtained.

[0175] Working principle and effects:

[0176] This study acquires time-series data from multiple treatment cycles of target users. After preprocessing and feature extraction, a time-series feature set is constructed. Transfer learning is used to fuse source domain baseline features with target domain personalized features to generate an adaptive feature set, effectively compensating for the sparsity of single-user time-series data. Initial observation windows are divided and optimized based on treatment cycles to form a set of observation windows with temporal correlations, thus building a feature analysis platform across time windows. Treatment evolution curves are fitted based on window features and processed continuously. A state sequence is constructed through low-dimensional latent variable mapping, and a time-series logic network is built using the state sequence as nodes, overcoming the limitation of static rule matching in handling complex time-series logic across time windows. Simulations are conducted on multiple sets of proposed solutions based on the time-series logic network. Through dynamic conflict assessment and multi-dimensional comparison and screening, an optimized treatment solution suitable for the target user is obtained. This dynamic time-series analysis architecture replaces the traditional static matching mechanism, ensuring that the treatment solution conforms to the temporal evolution of the body's state, taking into account both general treatment knowledge and individual personalized characteristics. This significantly improves the dynamic adaptability and clinical rationality of the solution, effectively avoiding intervention bias caused by static rules.

[0177] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing personalized treatment plans driven by multidimensional data, characterized in that, include: Acquire time-series data of the target user over multiple treatment cycles, preprocess and extract features from the time-series data, and construct a time-series feature set; By utilizing transfer learning, a baseline feature set is invoked as the source domain data, and the time-series feature set is used as the target domain data to generate an adaptive feature set. The initial observation window is set based on the conditioning cycle. According to the adaptation feature set, the center adjustment and extraction range extension are performed on each initial observation window to generate an observation window set. Based on the features of the observation window, the evolution law of features with conditioning cycle is fitted and continuousized to generate conditioning evolution curve. Combined with the mapping of low-dimensional latent variable space, a state sequence is constructed. Using the state sequence of each observation window as nodes, a time-series logic network is constructed. The system obtains the treatment plan for the current treatment cycle and adjusts the baseline plan based on preset optimization rules to generate multiple sets of plans to be verified. It then uses a time-series logic network to simulate and extrapolate each set of plans to be verified, generating a simulated state sequence. By performing dynamic conflict evaluation and multi-dimensional comparison and screening on the simulated state sequence, the system obtains the treatment optimization plan for the target user.

2. The method for optimizing personalized treatment plans driven by multidimensional data according to claim 1, characterized in that, The steps for constructing a time series feature set include: For the acquired time-series data, a unique cycle number is assigned to each cycle according to the order of the treatment cycle; the time-series data includes physiological data and text data; By using two-dimensional convolution, pixel-level features are extracted from tongue and facial images in physiological data to obtain image features; Pulse characteristics are obtained by frequency domain transformation of pulse waveforms in physiological data using fast Fourier transform and continuous wavelet transform. Based on the entity annotation set of the conditioning domain, entity recognition is performed on text data to extract domain entities, and Word2Vec is used for vectorized embedding to generate text features; Quantitative data in physiological data are integrated into structural features, and feature vectors are constructed based on image features, pulse features, text features, and structural features under the same treatment cycle. The feature vectors are arranged chronologically according to the period number to construct the temporal feature set of the target user.

3. The method for optimizing personalized treatment plans driven by multidimensional data according to claim 2, characterized in that, The steps for generating the adaptation feature set include: Collect general conditioning time-series data, and construct a general feature set through preprocessing and feature extraction; Based on the temporal feature set, the conditioning tags of the target users are extracted and encoded into target tag vectors; Match the corresponding conditioning label to each feature data in the general feature set and generate a validation label vector; By using cosine similarity, the label similarity between the target label vector and the validation label vector is calculated, and feature data with label similarity below the similarity threshold are removed to generate a benchmark feature set. Based on the baseline feature set and the time series feature set, the distribution difference is calculated using the maximum mean difference. With the goal of minimizing the distribution difference, the feature space is numerically calibrated dimension by dimension to obtain the aligned source domain feature set and target domain feature set, which are then divided into three feature layers. For a single treatment cycle, the feature saliency and feature similarity of each target domain are obtained, and the migration weight is configured for each feature layer by combining the preset personalized coefficient and domain adaptation coefficient. Based on the transfer weights, the source domain feature set is fused into the target domain feature set periodically and layer by layer, and the adaptive feature set is obtained according to the temporal relationship of the conditioning cycle.

4. The method for optimizing personalized treatment plans driven by multidimensional data according to claim 3, characterized in that, The steps for generating an observation window set include: Set the initial window length, use a sliding window with the adjustment period as the sliding step, and perform directional partitioning of the adaptation feature set to generate the initial observation window set; For each initial observation window, the conditioning period at the middle position within the window is selected as the central period, and the boundary period is defined. Based on the feature extraction range of the initial observation window, the extension period is obtained by extending the window both before and after it. The feature changes between the central period and adjacent periods within the window are used as transition features within the window, and the feature changes between the boundary period and the extended period are used as transition features between windows. Based on the difference between the feature vectors of the two conditioning cycles, the transition feature value is calculated, and the transition feature set of the initial observation window is obtained through normalization. Based on each initial observation window, the feature vector of the central period and the transition feature set are obtained, and the vectors are concatenated to generate the window vector; The window vectors are standardized and all windows are integrated to generate an observation window set.

5. The method for optimizing personalized treatment plans driven by multidimensional data according to claim 4, characterized in that, The steps for generating conditioning evolution curves include: Each window vector is split into sub-vectors of three feature layers, while the preset fitting parameters are called. For each feature layer's sub-vectors, a polynomial fitting method is used, with the window center period number as the horizontal axis and the feature layer sub-vector value as the vertical axis, to generate the initial conditioning curve for each feature layer. Based on the coefficient of determination, the goodness of fit between the initial conditioning curve of each feature layer and the actual value of the window feature is calculated. If the goodness of fit is lower than the fitting accuracy threshold, the curve fitting is deemed unsatisfactory, and the fitting parameters are readjusted to perform fitting until the curve fitting of all feature layers meets the standard. If the goodness of fit is not lower than the fitting accuracy threshold, the curve fitting is deemed satisfactory. In response to the curve fitting meeting the standard, fusion weights are assigned to each feature layer, and the initial conditioning curves of the three feature layers are fused to generate the conditioning evolution curve of the target user.

6. The method for optimizing personalized treatment plans driven by multidimensional data according to claim 5, characterized in that, The steps to construct a state sequence include: Based on each observation window, the fitted values ​​of the conditioning evolution curve and the fitted values ​​of the initial conditioning curves of each feature layer are called, and the conditioning state is quantized and calibrated to generate the state vector of the observation window. The conditioning state vectors of all observation windows are integrated, arranged in order according to the window number, and an initial conditioning state sequence is generated, which is then filtered and smoothed. Using cosine similarity, the conditioning similarity between any two adjacent state vectors is calculated, and temporal coherence is verified by combining it with a preset coherence threshold. If the conditioning similarity is below the coherence threshold, the state vector is recalibrated based on the conditioning evolution curve. The smoothed and verified state vectors are rearranged in time sequence to generate a state sequence.

7. The method for optimizing personalized treatment plans driven by multidimensional data according to claim 6, characterized in that, The steps for constructing a time-series logic network include: Each state vector in the state sequence is used as a state node, and the state evolution relationship between two adjacent state nodes is used as a network edge. For any network edge, edge similarity is calculated using the cosine similarity of the state vectors at the endpoints, and edge correlation is calculated using the transition feature values. The edge weight is calculated by weighted summation of edge similarity and edge correlation. Arrange all state nodes linearly according to window number, and combine them with edge weights to construct the initial network; Iterate through all edges of the initial network and remove network edges whose edge weights are less than the edge association threshold; Traverse all non-adjacent state node pairs in the initial network, calculate cosine similarity, filter out node pairs with cosine similarity greater than the potential evolution threshold, add indirect association edges, assign edge weights, and label them as potential evolutionary association edges; By integrating state nodes, adjacent edges, and potential evolutionary association edges, a temporal logic network is finally generated.

8. The method for optimizing personalized treatment plans driven by multidimensional data according to claim 7, characterized in that, The simulation steps include: The last state vector in the state sequence is used as the initial state for deduction; The simulation parameters are standardized, including simulation step size, influence coefficient, maximum intervention intensity, and correction coefficient. For each simulation step, the intervention intensity of the scheme to be verified is obtained, and the rate of change of state is calculated in combination with the simulation parameters; Traverse the temporal logic network to obtain the edge weights that are consistent with the conditioning stage of the scheme to be verified. Calculate the predicted edge weights by the arithmetic mean and combine them with the initial state and the deduction parameters to calculate the predicted state vector for the corresponding deduction step. After a complete deduction step by step, all predicted state vectors are arranged in time sequence to generate a simulated state sequence of the scheme to be verified. Using cosine similarity, the trend fit between the simulated state sequence and the conditioning evolution curve is calculated; Set an effective threshold for the extrapolation. When the trend fit is greater than the effective threshold, the extrapolation is deemed effective. All schemes to be verified are simulated in turn, and corresponding simulated state sequences are generated and the validity of the simulation is determined.

9. The method for optimizing personalized treatment plans driven by multidimensional data according to claim 8, characterized in that, The steps of dynamic conflict assessment and multi-dimensional comparative screening include: Using time sequence conflict, intervention conflict, and label conflict as evaluation indicators, the evaluation indicator values ​​of the schemes to be screened were calculated. The preset conflict weights are used to perform a weighted summation of the evaluation index values ​​to obtain the overall conflict degree of the solutions to be screened. If the overall conflict level is less than the conflict threshold, the solution is retained; otherwise, the solution is discarded. Integrate all the retained candidate schemes, plot the predicted evolution curve for each scheme, and simultaneously call the conditioning evolution curve to calculate the evolutionary stability through cosine similarity. If the evolutionary stability is greater than the preset longitudinal stability threshold, the prognosis is determined to be stable and the scheme is retained; otherwise, the scheme is eliminated. Calculate the treatment conversion rate and average status value of each screening scheme, perform quantitative scoring according to the preset prognosis evaluation rules, and calculate the horizontal comparison score of each screening scheme through arithmetic average calculation. All options to be screened are sorted in descending order based on the horizontal comparison score, and the option with the highest score is selected as the conditioning and optimization option.

10. A multi-dimensional data-driven personalized treatment plan optimization system, used to implement the multi-dimensional data-driven personalized treatment plan optimization method as described in any one of claims 1-9, characterized in that, include: Data acquisition module, evolution module, and optimization simulation module; The data acquisition module is used to acquire time-series data of the target user, construct a time-series feature set through preprocessing and feature extraction, call the benchmark feature set, optimize the time-series feature set using transfer learning, generate an adaptive feature set, set an initial observation window, and capture the transition features of the boundary region of the observation window through center adjustment and extraction range extension to generate an observation window set. The evolution module is used to perform evolution fitting and continuous processing on the features in the observation window set, construct the conditioning evolution curve of a single feature, integrate the conditioning evolution curves of all feature dimensions using low-dimensional mapping, construct the state sequence, and construct the temporal logic graph with the observation window as the node and the healthy evolution state of the state sequence between adjacent observation windows as the directed edge. The optimization simulation module constructs multiple sets of verification plans by calling the current treatment cycle's conditioning plan and preset optimization rules. It then uses a time-series logic network to simulate and deduce each set of verification plans, generating a simulated state sequence. By performing dynamic conflict evaluation and multi-dimensional comparison and screening on the simulated state sequence, the module obtains the conditioning optimization plan for the target user.

Images