Clinical research data analysis method based on machine learning

Abstract

The invention discloses a clinical research data analysis method based on machine learning, and the method comprises the steps: constructing a multi-modal variable structured causal map, and building a direction adjustable mechanism of a causal path; constructing a bidirectional nested structure attention mechanism, and capturing a cross-modal dependency and dynamic evolution relationship between variables; recording each layer of information propagation path and variable participation degree, and realizing reverse reconstruction of a model decision path in a reasoning stage; target-oriented attribution path regularization is introduced to carry out regularization constraint on an attribution path set of the key target variables; and constructing a nested attribution graph visualization system, and realizing interactive presentation of interpretation sub-graphs corresponding to prediction results so as to improve cognitive trust of model output. According to the method, from structure expression, path tracing and causal constraint to visual presentation, the core problems that a deep model is poor in interpretability, clinicians are not trusted, and existing interpretation tools are insufficient in applicability are solved in a full-link mode.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical research, and particularly relates to a method for analyzing clinical research data based on machine learning. Background Art

[0002] With the rapid accumulation of medical data and the rapid development of artificial intelligence technology, the method for analyzing clinical research data based on machine learning has shown unprecedented application potential in multiple fields such as disease prediction, diagnostic support, and treatment plan recommendation. Clinical research data usually has high heterogeneity, time dependence, and causal relevance. Its data forms include not only static electronic medical record data, laboratory test indicators, and biochemical parameters, but also multi-modal structures such as dynamically changing vital sign monitoring data, medical image features, and clinical pathway information. In this context, machine learning technology, especially deep learning models, has become the core support tool for clinical data modeling due to its powerful feature extraction and non-linear modeling capabilities.

[0003] However, although these models continue to break through in terms of prediction accuracy, their black-box internal mechanisms also pose severe challenges. Current mainstream deep learning models often struggle to present clear decision-making paths and causal chains, making it difficult for clinicians to accurately understand the basis for the diagnosis or prediction results given by the models, thus generating trust barriers. In addition, in actual clinical scenarios, there are often complex interaction relationships and temporal dependencies between different variables, and traditional machine learning models are difficult to effectively capture these structural features. Even when using existing model interpretation tools (such as SHAP, LIME, etc.) for posterior interpretation, problems such as poor result generalization ability, coarse attribution granularity, and unclear clinical semantics often occur.

[0004] Therefore, how to construct a machine learning analysis method with structural expression ability, strong causal interpretability, and capable of tracking the prediction decision-making path in multi-modal clinical data has become an important technical problem that urgently needs to be solved in the current field of medical artificial intelligence. Especially in clinical decision-making assistance systems for doctors, how to effectively explain the decision-making process of existing complex models not only concerns the credibility of the models, but also directly affects their implementation and application in the real medical environment. Summary of the Invention

[0005] To solve the above problems, the present invention provides a method for analyzing clinical research data based on machine learning.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: A method for analyzing clinical research data based on machine learning, comprising: Step 1: Based on medical ontology and medical domain knowledge, construct a structured causal graph of multi-modal variables using structure-preserving graph neural modeling, and establish a direction-adjustable mechanism for causal paths by introducing reversible relationship weight parameters; Step 2: Construct a bidirectional nested structure attention mechanism on the causal graph structure, and capture the cross-modal dependencies and dynamic evolution relationships between variables at both the graph structure level and the time series level, realizing the joint modeling of causal strength learning and reverse interpretation paths; Step 3: Record the information propagation paths and variable participation degrees of each layer during training through the information flow tensor inversion mechanism, and realize the reverse reconstruction of the model decision path during the inference stage, so as to ensure the causal consistency and traceability of the interpretation; Step 4: Introduce target-oriented attribution path regularization to regularize the set of attribution paths of key target variables, enhance the semantic stability and clinical rationality of the model decision basis, and prevent the drift of attribution paths; Step 5: Construct a nested attribution graph visualization system to realize the interactive presentation of the interpretation subgraph corresponding to the prediction result, so as to enhance the cognitive trust in the model output.

[0007] Further: The said Step 1 includes: Perform standardization processing on the original multi-modal clinical data; for numerical sequence variables, use normalization and sliding window mechanisms to construct a time series nested structure; for image variables, extract structure embeddings through lightweight CNN; for categorical labels, use one-hot encoding combined with medical ontology word embeddings for vectorization; Utilize medical ontology and clinically causal experience rules annotated by experts to construct an initial candidate edge set, and define each edge as a triple form including clinical semantic causal relationships; Assign two learnable weights, positive and negative, to each edge. The positive weight represents the direct influence of the independent variable, and the negative weight is used for subsequent reverse tracing of causal paths; Design a graph neural embedding mechanism, introduce a reversible weighting term for the direction of causal edges, ensure that the topological relationship of the graph structure is retained during the embedding process, and complete the construction of the structured causal graph.

[0008] Further: The said Step 2 includes: Based on the structured causal graph obtained in Step 1, combine the embedding vector of each node with time series information to form the input of the nested attention mechanism; Design a structure bidirectional attention module, use positive and negative weights, combine with learnable attention vectors, calculate the structure attention scores between nodes, and update the structure representation of nodes; For variables with a time dimension, design a time-aware attention module, calculate the importance of each time step for the current state of the variable, and obtain the weighted aggregated time series representation; Fuse the structural-level and temporal-level representations through a gating mechanism, retain the reversible causal path and dynamic temporal contribution, generate the output of the nested attention mechanism, and use it as the input for downstream prediction or causal attribution module; Introduce a path annotation function to record the direct or indirect influence paths between variables, establish a closed-loop path direction, and provide the basic data structure for the subsequent causal attribution module.

[0009] Further: The step 3 includes: During the forward propagation of the model, construct a multi-order information flow tensor to record the information conduction intensity from the input layer to the output layer, including the forward and backward weights in the structural-level attention and the amplitude weight of the flowing information; Introduce a global information flow conservation constraint to ensure that the information of each output node comes from the input variables through legal paths and the reverse flow is not weakened, and constrain the conservativeness of information flow by minimizing the net loss of information flow; During the inference stage, use the recorded information flow tensor to establish a reverse path reconstruction function, restore the variable paths and their intensities that affect the prediction result, and output an ordered path set for causal chain visualization; Based on the output of the reverse path reconstruction function, generate an explanation path map, display the key variables that affect the prediction variable and their actual information flow paths in the structural causal map, and present the logical path of the prediction through a visualization tool.

[0010] Further: The step 4 includes: According to the edge weights in the structured causal map, domain knowledge, and expert priors, set the attribution path set corresponding to each target clinical event variable to ensure the authority and interpretability of the path semantics; Define the total path dependence weight for each candidate path to measure the actual attribution intensity of each path by the model during training; Design an attribution path regularization loss function to encourage the model to rely more on key medical paths while suppressing the dependence on non-structurally reasonable paths, so that the structurally reasonable paths are actively strengthened during training; Combine the attribution path regularization term with the main task loss function and the information flow conservation constraint to construct the joint optimization objective of the model.

[0011] Further: The step 5 includes: Divide the map into a global level, a local level, and a time-dynamic level; Based on the information flow tensor, extract the causal path tensor activated by the sample, calculate the total attribution contribution degree of each variable, and determine the attribution direction in combination with the gradient information output by the model, and present it with node size, color gradient, and edge weight annotation in the map.

[0012] Compared with the prior art, the technical progress achieved by the present invention lies in: The present invention realizes the structural transformation from a black-box model to a causally transparent model. Through a structure-preserving graph neural modeling and nested attribution graph construction mechanism, the present invention introduces clear causal structure constraints at the model input stage, organizing the original multi-modal clinical variables (such as time series, physiological signals, imaging features, etc.) into a structured causal graph with clear node-edge relationships. This causal graph modeling method not only strengthens the model's structural expression ability but also naturally has path traceability, laying a solid foundation for subsequent interpretive analysis.

[0013] Capture cross-modal - temporal dependencies of clinical variables through a bidirectional nested attention mechanism. This method innovatively designs a graph structure level + time series level nested attention mechanism. Through the bidirectional modeling of the forward causal strength path and the reverse interpretation path, it systematically depicts the structural dependencies and temporal evolution relationships between variables. This mechanism overcomes the deficiency of traditional interpretation methods (such as SHAP, LIME) in expressing ability when facing temporal data and complex dependencies, and is more in line with the true characteristics of non-independent and dynamic interactions between variables in clinical scenarios.

[0014] Introduce an information flow tensor inversion mechanism to achieve reverse interpretability in the inference stage. Compared with posterior interpretation tools based on gradients or sampling, the information flow tensor inversion mechanism proposed by the present invention continuously records the information propagation path and variable participation degree during the model training process, supporting the reverse tracing from the inside out in the inference stage. This mechanism ensures that the interpretation path is consistent with the actual prediction path, significantly improving the authenticity, accuracy, and consistency of interpretability.

[0015] Achieve clinical semantic stability through causal attribution path regularization. This method particularly introduces a target-oriented attribution path regularization to structurally constrain the attribution path, ensuring that each prediction of the model relies on a set of variables with stable semantics and clinical rationality as much as possible. This attribution optimization method effectively avoids the problem that the attribution path of the model jumps too much between different samples and lacks repeatability, enhancing the reliability of the results and the acceptance of physicians.

[0016] Construct a nested attribution graph visualization system to assist doctors in understanding and trusting. Finally, the present invention converges all the interpretation mechanisms into a multi-level interactive visualization system. Clinical doctors can, based on any prediction result, view the corresponding nested interpretation subgraph in real time, including the contribution degree of each variable to the result, the promotion or inhibition direction, and the causal interaction between variables. This system not only provides doctors with an intuitive interpretation view but also enhances the operability of the model to be trustworthy and usable in medical practice.

[0017] Generally speaking, from structural expression, path tracing, causal constraint to visual presentation, the present invention comprehensively solves the core problems such as poor interpretability of deep models, distrust of clinicians, and insufficient applicability of existing interpretation tools, and has the following comprehensive advantages: Strong structure: Modeled based on causal graphs, adapting to the structural requirements of clinical multi-modal variables; Dynamically interpretable: Can trace the causal chain and support reverse path restoration; Clinically semantically stable: The attribution path has medical rationality and prediction consistency; Strong interactivity: Through graph interaction visualization, enhancing doctors' understanding and trust; High innovation: Propose a number of native mechanisms to break through the technical bottlenecks of existing tools. Description of the Drawings

[0018] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention.

[0019] In the drawings: Figure 1 is the flowchart of the present invention. Detailed Embodiments

[0020] The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below with reference to the drawings.

[0021] As Figure 1 shown, the present invention discloses a method for analyzing clinical research data based on machine learning, including: Step 1: Based on medical ontology and medical domain knowledge, use Structure-preserving Graph Neural Encoding to construct a structured causal graph of multi-modal variables, and establish a direction-adjustable mechanism for causal paths by introducing reversible relationship weight parameters, providing a structural basis for subsequent traceable modeling.

[0022] Step1.1: Standardization and structured representation of multi-modal data First, perform unified standardization processing on the original clinical research data to ensure the comparability of features in different modalities (such as time series laboratory test data, biochemical index data, high-dimensional vectors of medical images, physiological sensor outputs, etc.) in the same semantic space. The following structured representation methods are used for different modality data: For numerical sequence variables : Normalize and use a sliding window mechanism to construct a time series nested structure; For image - type variables : Use lightweight CNN to extract structural embeddings ; For categorical labels (such as pathological diagnosis conclusions), use one - hot encoding + medical ontology word embeddings (such as UMLS vector embeddings) for unified vectorization; Finally, each variable is uniformly represented in the form of an embedding vector:

[0023] Step1.2: Construct the initial causal edge set by fusing medical ontologies Utilize the hyponym - hypernym structure and attribute - dependency graphs in domain ontologies (such as UMLS, SNOMED CT), and combine with clinically - annotated causal experience rules by experts to construct an initial candidate edge set: Define each edge in the form of a triple:

[0024] Among them, , represents the set of clinical semantic causal relationships, including causes, leadsto, inhibits, isriskfactorfor, etc.

[0025] This edge set constitutes the initial topological structure of the causal graph .

[0026] Step1.3: Introduce reversible relationship weight parameters (Reversible Edge Weighting) In order to enable the subsequent model to reverse - infer the causal path from any prediction result, this embodiment proposes a reversible relationship weight parameter mechanism. Each edge is respectively assigned two learnable weights, one for the forward direction and one for the reverse direction:

[0027] Among them: is the Sigmoid activation function, ensuring that the weight value range is in (0, 1); is the learnable parameter matrix; The forward weight indicates that the dependent variable is directly affected by ; The reverse weight is then used for the causal traceability mechanism in the subsequent path reverse - inference process.

[0028] Finally, each edge is represented as a reversible weighted edge including direction:

[0029] Step1.4: Structure-preserving Graph Neural Encoding After the initial causal graph is constructed, this embodiment designs a structure-preserving graph neural embedding mechanism to ensure that the topological relationship of the graph structure is not lost during the embedding process. The core idea is: introducing a reversible weighting term for the causal edge direction in the graph neural update, so that the directionality of the causal structure path is explicitly retained in each node state update.

[0030] For each node the embedding vector is updated as follows:

[0031] where: represents the node representation of the th layer; is the direction-discriminating linear transformation matrix of the graph neural network layer; is the set of adjacent nodes of node ; This update rule explicitly distinguishes the forward and backward structural information flows, preventing the causal structure from being blurred due to information averaging.

[0032] Step1.5: Visualization Verification and Structural Stability Evaluation To ensure the interpretability of this causal graph in real clinical scenarios, this embodiment introduces the following structural stability metric: Causal Chain Consistency:

[0033] where represents the set of target variable paths, and the higher the score, the more stable the causal path.

[0034] Reversibility Gap:

[0035] The closer it is to zero, the more symmetric the causal pair is, which is convenient for traceability interpretation.

[0036] The construction of this structure-preserving causal graph not only provides structural constraints and traceable information flow directions for subsequent nested attention mechanisms, but also provides a theoretical basis and technical channel for final interpretive modeling by introducing reversible relationship weight parameters. In addition, in this step, a closed-loop process from original variable embedding, structure composition, graph neural coding to visualization verification is maintained to ensure that all dependent information in subsequent steps originates from this rigorous graph structure modeling system.

[0037] Step 2: Construct a bidirectional nested structural attention mechanism on the causal graph structure, capturing cross-modal dependencies and dynamic evolution relationships between variables at both the graph structure level and the time series level, and realizing the joint modeling of causal strength learning and reverse interpretation paths.

[0038] Step2.1: Construct the input expression basis of the nested attention structure In Step 1, this embodiment has obtained a structure-preserving causal graph , and each node represents a structured clinical variable, and its embedding vector is . In addition, if the variable has a time series dimension (such as continuously recorded biochemical indicators), then the variable is represented as:

[0039] This embodiment will use as the input, and jointly design a bidirectional nested encoder with structural attention and time attention.

[0040] Step2.2: Structural-level Attention Modeling Structural-level attention is used to capture the semantic dependencies between clinical variables in the causal graph. Based on the forward weight and reverse weight defined in Step 1, this embodiment constructs a structural bidirectional attention module as follows:

[0041]

[0042] Among them: is a learnable attention vector; is a linear transformation matrix for query and key; ‖ represents vector concatenation; and They are the forward and reverse structural attention scores respectively, which incorporate the structural causal relationship weights.

[0043] The structural representation of the final node is updated to:

[0044] Step2.3: Temporal-level Attention Modeling For clinical variables with a time dimension , in this embodiment, a time-aware attention module is designed to capture the importance of each time step for the current state of the variable.

[0045] Define the time attention mechanism as follows:

[0046] Where: , .

[0047] are the key and query transformation matrices in the time attention; represents the variable at the time step's time attention weight; is the representation of the variable after structural update, used as the time attention query vector; Finally, the weighted aggregated temporal representation is obtained:

[0048] Step2.4: Structural-Temporal Bidirectional Nested Composition Update In this embodiment, the structural-level attention result and the time-level attention result are combined, and a gating mechanism is used to fuse these two representations to retain the reversible causal path and dynamic temporal contribution:

[0049] Where: is the Sigmoid function; represents the Hadamard element-wise product; is the variable representation output by the nested attention mechanism, used as the input for the downstream prediction or causal attribution module.

[0050] Step2.5: Upstream and Downstream Traceable Path Labeling and Closed-loop Constraint Based on bidirectional attention modeling, this embodiment introduces a path annotation function , which is used to record the direct or indirect influence paths of each variable on another variable and establish a closed-loop path direction:

[0051] This path annotation mechanism will serve as the basic data structure for causal chain backtracking and interpretable path graph generation in the subsequent causal attribution module.

[0052] Based on the directional edge weights and causal dependence structure defined in the continuous structure-preserving causal graph, in this step, a bidirectional nested structure attention mechanism is constructed to achieve, for the first time in clinical modeling tasks, the decoupled modeling and traceable combination of structure-dependent attention and temporal dynamic attention. This mechanism not only supports tracing the explanation link backward from the prediction but also has the ability to dynamically depict the causal interaction relationships between cross-modal variables, and is the core support component for the entire analysis system to achieve full-path transparency.

[0053] Step 3: Record the information propagation paths and variable participation degrees of each layer during training through the Information Flow Tensor Reversibility Mechanism, and reverse reconstruct the model decision path during the inference phase to ensure the causal consistency and traceability of the explanation.

[0054] Step 3.1: Construct an information flow record tensor During the forward propagation phase of the model, this embodiment constructs a multi-order information flow tensor for the update process of all variables in the nested structure attention module , which is used to record the information conduction intensity from the input layer to the output layer, where: represents the number of layers of the nested attention module in the model; represents the number of clinical variable nodes; represents the th layer, and represents the information flow intensity from variable to variable

[0055] Specifically defined as:

[0056] where: , are the forward and backward weights in the structure-level attention in the th layer respectively; Represents the amplitude weight of the flow information; This tensor is automatically accumulated in each round during the forward inference process of the model and supports the modeling of information dependencies between multimodal variables.

[0057] Step3.2: Define the information flow reversibility constraint term To ensure that the final output result of the model can be reversely restored along the path, this embodiment introduces a global information flow conservation constraint, that is, the information of each output node must come from the input variables propagated through a set of legal paths, and the reverse flow in the path should not be weakened. The specific constraint function is defined as follows:

[0058] Where: Represents the variable in the final prediction vector representation; Represents the variable in the input layer initial embedding; This loss function constrains the conservativeness of the information flow process by minimizing the net loss of information flow, thereby supporting the reverse reducibility of the prediction path.

[0059] Step3.3: Construct an information flow reversal function for path restoration explanation In the inference stage, this embodiment uses the recorded tensor to reconstruct the reverse path for any target prediction variable, that is, to establish a function for restoring the variable path and its strength that affect the prediction result:

[0060] Where: Is the variable to the variable total reverse path contribution to the prediction result; Is the path significance threshold; The output is an ordered path set, which can be directly used for causal chain visualization or highlighting significant variables.

[0061] Step3.4: Path graph generation and interpretable interface output Using the path information set constructed by the function This embodiment can further construct an explanatory path graph where: Are all the key variables that affect the prediction variable ; as its actual information flow path in the structural causal graph; A structural highlighting graph can be generated through a graph visualization tool (such as Graphviz or D3.js) to display the predicted logical path.

[0062] By performing a tensor-level cascade statistics on the directional modeling in the structural causal graph (step 1) and the dependent dynamic capture in the nested attention mechanism (step 2), this step innovatively designs a reversible information flow tracking framework. This framework not only introduces information conservation constraints during the model training phase to ensure semantic consistency, but also supports the reduction of causal chains based on path tensors during the model inference phase, realizing a truly traceable and interpretable deep clinical decision support system. This mechanism provides a structured, quantitative, and visual solution to the interpretability defects of existing black-box models.

[0063] Step 4: Its core objective is to make the model more inclined to select medically reasonable variable paths during the learning and inference processes through path regularization control at the structural level, thereby improving the stability, consistency, and clinical acceptability of the prediction process.

[0064] Introduce Target-oriented Attribution Path Regularization to regularize the set of attribution paths for key target variables, enhance the semantic stability and clinical rationality of the model's decision-making basis, and prevent attribution path drift.

[0065] Step 4.1: Define the set of key causal paths for target variables In this embodiment, first, according to the edge weights in the structured causal graph, as well as domain knowledge and expert priors, the set of attribution paths corresponding to each target clinical event variable (for example: sepsis outbreak, decreased survival rate, intracerebral hemorrhage progression, etc.) is set, denoted as:

[0066] Each path among them is a possible attribution path starting from multiple key variable nodes and reaching the target variable node along the directed edges in the graph and can be formally expressed as:

[0067] This set can be derived from standard medical guidelines (such as the Sepsis-3 diagnostic criteria), clinical practice data statistics, or be manually formulated by medical experts to ensure the authority and interpretability of the path semantics.

[0068] Step 4.2: Introduce the attribution path selection weight tensor In the defined information flow tensor Based on this, this embodiment provides each candidate path Define the total dependency weight of its path , calculated as follows:

[0069] The path dependency weight measures the actual attribution strength of the model during the training process. The higher the circulation strength on each side of the path, the greater the overall attribution weight.

[0070] Step 4.3: Construct a goal-oriented attribution regularization term In order to encourage the model to rely more on key medical pathways while suppressing reliance on non-structurally reasonable pathways, this embodiment designs an attribution path regularization loss function:

[0071] in: represents the set of all possible attribution paths (obtained by graph traversal); This regularization term makes the loss smaller when the target path weight accounts for a higher proportion of all path weights; This is equivalent to performing a set of normalized maximization processing on the attribution path distribution, so that the structurally reasonable paths are actively strengthened during the training process.

[0072] Step 4.4: Joint optimization design of the total loss function Finally, this embodiment combines the attribution path regularization term with the existing main task loss function (such as classification cross entropy or regression MSE) and the information flow conservation constraint (step 3) to construct the joint optimization objective of the model:

[0073] in: represents the main task prediction loss; , are weight hyperparameters, which respectively regulate the importance of the reversibility constraint and the attribution path regularization in the overall optimization objective; Through joint optimization, we can ensure prediction accuracy, path reproducibility and causal explanation rationality at the same time.

[0074] In the previous three steps, this embodiment has established a machine learning model framework with clear structure, explicit dependencies, and traceable paths. In this step, further optimization constraints are implemented on the process of how the model selects paths, enabling the model to not only restore the past prediction process but also learn to prefer reasonable explanation paths during training. This strategy not only overcomes the opacity problem of traditional machine learning models regarding variable selection paths but also avoids the model relying on statistical pseudo-correlation paths in the data during training, thereby significantly enhancing its clinical credibility and generalization robustness.

[0075] Step 5: Based on the structured causal graph, bidirectional nested structure attention mechanism, information flow tensor inversion mechanism, and causal attribution path regularization mechanism constructed in the previous four steps, this step further designs an interactive explanation interface for the final clinical users, aiming to achieve the semantic conversion from the model black box to a transparent glass box, so that clinicians can understand the causal mechanism and variable collaborative contribution behind each prediction result in a structured, visual, and operable manner.

[0076] Construct a nested attribution graph visualization system based on the output of the foregoing modules to realize the interactive presentation of the explanation subgraph corresponding to the prediction result, including variable contribution degree, influence direction, and causal chain structure, and enhance the doctor's cognitive trust in the model output.

[0077] Step5.1: Define the hierarchical structure of the nested attribution graph Combined with the structured causal graph constructed in the previous steps and the activated explanation subgraph in the specific prediction , this embodiment divides the entire graph into three visible levels: Global Causal View: Displays the structural causal relationships among all variables, that is, the complete , providing background information for the causal graph; Local Attribution View: Displays the activated causal subgraph for a specific prediction result , and the edge weights correspond to the path contribution weights of the current sample; Temporal Variable Trajectory: Embeds the time series dimension to display the change trends and dynamic attribution weights of key variables within the historical window.

[0078] Step5.2: Embed variable attribution weight and direction information Based on the information flow tensor , for the sample in this embodiment Extract the causal path tensor activated by this sample:

[0079] Among them, the weights are from the normalized output of the bidirectional nested structure attention mechanism.

[0080] Subsequently, define the total attribution contribution degree of each variable node as:

[0081] And combine the gradient information output by the model to determine the attribution direction (promote vs inhibit) of each node:

[0082] If : It means that the variable is a promoting factor; If : It means that the variable is an inhibitory factor.

[0083] Finally, in the graph visualization, map the contribution degree to the node size, map the directionality to the color gradient (e.g., green for inhibition, red for promotion), and label the flow intensity and reversibility weight on the edge.

[0084] Step5.3: Develop interactive visualization interface components This embodiment designs a front-end and back-end separated interactive interface system based on the Web technology stack (such as D3.js, Plotly, WebGL+React), including the following key modules: Variable overview panel: Display the contribution degree of each variable with a sorted bar chart , supporting filtering by category (such as biochemical indicators, vital signs, etc.); Attribution path navigation graph: Build a subgraph based on , which can dynamically expand nodes and display the propagation relationship between variables on the causal chain; Time series tracking pane: After selecting any variable, display its dynamic change curve and attention change trajectory at the past time points; Decision contribution summary card: Automatically generate a summary in natural language to display the main contributing variables, prediction credibility, and attribution path explanation of this prediction result.

[0085] Step5.4: Nested interactive refresh mechanism This system supports dynamic jumping and linkage from any level to any level. For example: When users click on a high-contribution variable node, they can automatically locate its path, its time series trend, and its downstream impact; If the user modifies the target prediction task (e.g., switching from survival rate prediction to postoperative infection prediction), the system can re-render based on the attribution path weight. ; By dragging the time slider, users can view the attention weight and predicted impact of a variable at different time points in real time.

[0086] Through this step, this embodiment builds an interpretable graph system based on the multiple constraints of structural attribution, flow reversibility, and path regularization, featuring a nested hierarchical structure, visible variable behavior, clear path direction, and responsive interactions. This makes the model not only interpretable but also capable of proactively presenting explanations. The ultimate goal is to maximize clinicians' understanding and trust in the model, thereby promoting the evolution of clinical research data analysis methods towards deployable, auditable, and collaborative approaches.

[0087] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for analyzing clinical research data based on machine learning, characterized in that, Including: Step 1: Based on medical ontology and medical domain knowledge, construct a structured causal graph of multimodal variables using structure-preserving graph neural modeling, and establish a direction-adjustable mechanism for causal paths by introducing reversible relationship weight parameters; Step 2: Construct a bidirectional nested structure attention mechanism on the causal graph structure, and capture cross-modal dependencies and dynamic evolution relationships between variables at both the graph structure level and the time series level to realize the joint modeling of causal strength learning and reverse interpretation paths; Step 3: Record the information propagation paths and variable participation degrees of each layer during training through the information flow tensor inversion mechanism, and realize the reverse reconstruction of the model decision path during the inference stage, so as to ensure the causal consistency and traceability of the interpretation; Step 4: Introduce a target-oriented attribution path regularization to regularize the set of attribution paths of key target variables, enhance the semantic stability and clinical rationality of the model decision basis, and prevent the attribution path from drifting; Step 5: Construct a nested attribution graph visualization system to realize the interactive presentation of the interpretation subgraph corresponding to the prediction result, so as to enhance the cognitive trust in the model output.

2. The method for analyzing clinical research data based on machine learning according to claim 1, wherein The said Step 1 includes: Perform standardization processing on the original multimodal clinical data; use normalization and sliding window mechanisms for numerical sequence variables to construct a time series nested structure; extract structural embeddings for image variables through lightweight CNN; vectorize categorical labels using one-hot encoding combined with medical ontology word embeddings; Utilize medical ontology and clinically causal experience rules annotated by experts to construct an initial candidate edge set, and define each edge as a triple form including clinical semantic causal relationships; Assign two learnable weights, positive and negative, to each edge. The positive weight represents the direct influence of the independent variable, and the negative weight is used for subsequent reverse tracing of causal paths; Design a graph neural embedding mechanism, introduce a reversible weighting term for the direction of causal edges, ensure that the topological relationship of the graph structure is retained during the embedding process, and complete the construction of the structured causal graph.

3. The method for analyzing clinical research data based on machine learning according to claim 2, wherein The said Step 2 includes: Based on the structured causal graph obtained in Step 1, combine the embedding vector of each node with time series information to form the input of the nested attention mechanism; Design a structure bidirectional attention module, use positive and negative weights, combined with learnable attention vectors, to calculate the structural attention scores between nodes and update the structural representation of the nodes; For variables with a time dimension, design a time-aware attention module to calculate the importance of each time step for the current state of the variable, and obtain the weighted aggregated time series representation; Fuse the structural-level and time-level representations through a gating mechanism, retain the reversible causal path and dynamic time series contributions, generate the output of the nested attention mechanism, and use it as the input for the downstream prediction or causal attribution module; Introduce a path annotation function to record the direct or indirect influence paths between variables, establish a path direction closed loop, and provide the basic data structure for the subsequent causal attribution module.

4. A method for analyzing clinical research data based on machine learning according to claim 3, characterized in that The said Step 3 includes: During the forward propagation of the model, construct a multi-order information flow tensor to record the information conduction intensity from the input layer to the output layer, including the positive and negative weights in the structural-level attention and the amplitude weights of the flowing information; Introduce the global information flow conservation constraint to ensure that the information of each output node comes from the input variables propagated through legal paths and the reverse flow is not weakened, and constrain the conservativeness of information flow by minimizing the net loss of information flow; In the inference stage, use the recorded information flow tensor to establish a reverse path reconstruction function, restore the variable paths and their intensities that affect the prediction result, and output an ordered path set for causal chain visualization; Based on the output of the reverse path reconstruction function, generate an explanation path graph, display the key variables that affect the prediction variable and their actual information flow paths in the structural causal graph, and present the prediction logic path through a visualization tool.

5. The method for analyzing clinical research data based on machine learning according to claim 4, wherein Step 4 includes: According to the edge weights in the structured causal graph, domain knowledge and expert priors, set the attribution path set corresponding to each target clinical event variable to ensure the authority and interpretability of the path semantics; Define the total path dependence weight for each candidate path to measure the actual attribution intensity of each path by the model during training; Design an attribution path regularization loss function to encourage the model to rely more on key medical paths while suppressing the dependence on non-structurally reasonable paths, so that the structurally reasonable paths are actively strengthened during training; Combine the attribution path regularization term with the main task loss function and the information flow conservation constraint to construct the joint optimization objective of the model.

6. The method for analyzing clinical research data based on machine learning according to claim 5, wherein Step 5 includes: Divide the graph into a global level, a local level and a time dynamic level; Based on the information flow tensor, extract the causal path tensor activated by the sample, calculate the total attribution contribution degree of each variable, and determine the attribution direction in combination with the gradient information output by the model, and present it in the graph with node size, color gradient and edge weight annotation.

Images