AI-driven disease diagnosis and curative effect monitoring analysis system

Abstract

The invention discloses an AI-driven disease diagnosis and curative effect monitoring analysis system, which comprises a data processing module used for collecting multi-source diagnosis and treatment data and constructing a multi-modal sample set; the diagnosis prediction module is used for generating a diagnosis label, a diagnosis confidence value and a prediction curative effect trend through a multi-task learning model; the curative effect comparison module is used for collecting actual curative effect data, constructing a curative effect observation sequence and generating a curative effect deviation sequence based on the curative effect observation sequence and the predicted curative effect trend; the credibility evaluation module is used for executing credibility review based on the curative effect deviation sequence and the diagnosis confidence value; and the path evolution module is used for recording continuous multi-round diagnosis correction results, generating a jump type diagnosis path and updating a diagnosis and treatment data chain. The dynamic closed-loop verification mechanism between the multi-mode diagnosis and treatment data and the predicted curative effect trend improves the accuracy of disease diagnosis, the reliability of curative effect prediction and the adaptability of the diagnosis and treatment process.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent diagnosis, in particular to an AI-driven disease diagnosis and curative effect monitoring and analysis system. BACKGROUND

[0002] With the continuous development of artificial intelligence technology and clinical informatization construction, multi-source data fusion analysis for disease diagnosis and curative effect monitoring has gradually become an important direction of medical intelligence. The existing disease diagnosis and curative effect evaluation method mainly depends on single modal medical images, laboratory test results or medical record texts and other single-source information, and it is difficult to comprehensively represent the real pathological state of the patient. At the same time, there are significant differences in the collection frequency, time stamp accuracy and format structure of multi-modal data, which leads to high difficulty in time alignment and modal mapping between different data sources, and feature loss or time sequence mismatch is easily generated in the information fusion process. In addition, the traditional diagnosis model often processes the diagnosis and curative effect prediction processes separately, lacks a dynamic diagnosis correction mechanism based on curative effect feedback, and the diagnosis result is insufficiently responsive to long-term curative effect changes, and the confidence evaluation lacks time sequence continuity and credibility backtracking ability. In the aspect of diagnosis and treatment path optimization, the existing method mainly depends on manual review or simple statistical analysis, and cannot form a compressible and traceable structured diagnosis evolution graph, resulting in problems such as path redundancy, poor decision-making jump and insufficient explainability in the diagnosis process.

[0003] Therefore, how to provide an AI-driven disease diagnosis and curative effect monitoring and analysis system is a problem to be solved by those skilled in the art. SUMMARY

[0004] An object of the present application is to provide an AI-driven disease diagnosis and curative effect monitoring and analysis system. The present application makes full use of multi-source diagnosis and treatment data collection, multi-modal feature fusion, multi-task learning model, dynamic confidence back evaluation and diagnosis path compression and other technical means, and describes in detail the whole process of constructing a multi-modal sample set on the basis of unified patient identification and time alignment, jointly generating diagnosis labels and predicting curative effect trends, performing dynamic diagnosis correction based on curative effect deviation and constructing a jump diagnosis path. Through multi-modal feature fusion, the collaborative accuracy of diagnosis and prediction is improved, through dynamic confidence back evaluation, the matching degree of diagnosis result and real curative effect is enhanced, through evolution graph modeling and path compression, the structural compactness and traceability of the diagnosis and treatment chain are optimized, and the system has the advantages of high diagnosis accuracy, high curative effect monitoring sensitivity, high path optimization efficiency and strong system explainability.

[0005] According to the AI-driven disease diagnosis and curative effect monitoring and analysis system provided by the embodiment of the present application, the system comprises:

[0006] A data processing module is configured to collect multi-source diagnosis and treatment data and construct a multi-modal sample set.

[0007] a diagnosis prediction module configured to generate a diagnosis label, a diagnosis confidence value, and a predicted efficacy trend through a multi-task learning model;

[0008] an efficacy comparison module configured to collect actual efficacy data, construct an efficacy observation sequence, and generate an efficacy deviation sequence based on the efficacy observation sequence and the predicted efficacy trend;

[0009] a credibility evaluation module configured to perform credibility reevaluation based on the efficacy deviation sequence and the diagnosis confidence value;

[0010] a path evolution module configured to record the results of continuous rounds of diagnosis correction, generate a jump diagnosis path, and update the diagnosis and treatment data chain.

[0011] Optionally, the modules are implemented through the following steps:

[0012] Step one: collect multi-source diagnosis and treatment data of a patient, and construct a multi-modal sample set;

[0013] Step two: input the multi-modal sample set into a multi-task learning model to generate a diagnosis label, a diagnosis confidence value, and a predicted efficacy trend; the multi-task learning model comprises a diagnosis task sub-module and an efficacy prediction task sub-module;

[0014] Step three: construct a diagnosis and treatment data chain based on the diagnosis label, the diagnosis confidence value, and the predicted efficacy trend;

[0015] Step four: collect actual efficacy data of the patient within a preset time window to form an efficacy observation sequence; and perform deviation comparison based on the efficacy observation sequence and the predicted efficacy trend to generate an efficacy deviation sequence;

[0016] Step five: perform credibility reevaluation based on the efficacy deviation sequence and the diagnosis confidence value, and calculate a dynamic confidence score;

[0017] Step six: if the dynamic confidence score is less than a preset confidence threshold, perform diagnosis correction operation;

[0018] Step seven: record the results of continuous rounds of the diagnosis correction, and construct an evolution graph;

[0019] Step eight: perform path compression operation based on the evolution graph to generate a jump diagnosis path, and update the diagnosis and treatment data chain.

[0020] Optionally, the step one is specifically:

[0021] collect medical image data; the medical image data includes computed tomography images, magnetic resonance imaging images, and ultrasonic imaging images, and is stored in a standard image format;

[0022] Collecting clinical test data; the clinical test data includes blood biochemical indexes, immunological detection indexes and pathogenic detection indexes, and is stored in a structured test report data format;

[0023] Collecting physiological signal data; the physiological signal data includes electrocardiogram signals, respiratory signals and blood pressure signals, and a sampling frequency is associated with a time label and a continuous waveform sequence is recorded;

[0024] Collecting electronic medical record text data; the electronic medical record text data includes medical history description, complaint content and doctor observation record, and is preprocessed by a natural language processing method and converted into a structured text embedding;

[0025] The medical image data, the clinical test data, the physiological signal data and the electronic medical record text data are subjected to unified patient identification association, time alignment processing and modality mapping processing, and a unified format multi-modal sample set is constructed.

[0026] Optionally, the multi-task learning model includes a diagnosis task sub-module and a therapeutic effect prediction task sub-module, and specifically:

[0027] The diagnosis task sub-module is based on an improved Transformer structure, uses fusion feature representation to perform feature attention extraction and time sequence association modeling, and outputs disease diagnosis labels and diagnosis confidence values; wherein the diagnosis label represents a structured disease classification result, and the diagnosis confidence value represents the credibility of the model for the classification judgment;

[0028] The improved Transformer structure includes: a modality gating mechanism is arranged before each self-attention head, the modality gating mechanism is composed of a gating function generated based on multi-modal feature statistical indicators; the gating function is fitted by modality distribution features and significance weight parameters in the training sample; a task guide vector is embedded in the position encoding stage, the task guide vector is generated from the structural features of the diagnosis task label space, and the position encoding representation is constructed by embedding the input sequence together; a task-specific channel weight adjustment structure is arranged in the feedforward neural network module; the Transformer structure is constructed in a multi-layer stacking form, and a residual connection structure and layer normalization operation are introduced between layers;

[0029] The therapeutic effect trend prediction sub-module adopts a structure of fusion time sequence convolution network and causal attention mechanism, models the historical treatment response trajectory, extracts long-term dependence relationship and phase mutation signal; the sub-module generates a predicted therapeutic effect trend sequence under the condition control of the diagnosis task output result;

[0030] Perform parameter update through a joint loss function of the multi-task learning model; the joint loss function is composed of a cross-entropy loss term of the diagnosis task and a mean square error loss term of the efficacy trend prediction task; further introduce the diagnosis confidence value and the sample efficacy trend fluctuation index as weight factors to perform differentiated adjustment on the loss weight of each task.

[0031] Optionally, the step three is specifically:

[0032] Encode the diagnosis label as an initial diagnosis node, and add a diagnosis confidence value as a weight attribute of the initial diagnosis node;

[0033] According to the predicted efficacy trend corresponding to the diagnosis label, construct an efficacy trend node, and establish a directional connection relationship, and mark the prediction time interval and the efficacy change direction;

[0034] Map the patient identity and the treatment time information to a diagnosis and treatment primary key index, and form a unified data structure with the diagnosis node and the efficacy trend node;

[0035] Introduce auxiliary attributes between the diagnosis node and the efficacy trend node, including diagnosis source, model version number and sample modal combination identifier, and add coding as diagnosis and treatment path generation conditions;

[0036] Perform time sorting on the diagnosis node, the efficacy trend node and the connection relationship to form a multi-round diagnosis and treatment trajectory structure;

[0037] Store the node information, connection attribute and auxiliary annotation in each trajectory structure in a graph structure to form a diagnosis and treatment data chain.

[0038] Optionally, the step four is specifically:

[0039] Collect multi-dimensional efficacy data covering image indicators, test values, physiological signals and subjective scores, and integrate them according to the patient identity and the time label to construct an efficacy observation sequence containing multiple efficacy observation nodes;

[0040] Perform scale unification processing on the efficacy observation sequence, and complete position alignment according to the time interval and change direction marked in the predicted efficacy trend;

[0041] Calculate the difference between each efficacy prediction node and the corresponding efficacy observation node, establish a residual deviation matrix, and extract the residual amplitude and direction marker;

[0042] Construct an efficacy deviation sequence according to the residual deviation matrix, and mark the residual strength level and deviation direction of each efficacy observation node;

[0043] Synchronously store the residual information of each efficacy observation node in the efficacy deviation sequence and the diagnosis confidence value, and set an optional evaluation marker field;

[0044] The efficacy bias sequence is embedded into the diagnosis and treatment data chain structure, and is consistent with the index and time topology sequence of the prediction efficacy trend node.

[0045] Optionally, the step five is specifically:

[0046] The residual strength level and deviation direction corresponding to each efficacy observation node in the efficacy bias sequence are extracted to construct a bias feature vector set;

[0047] The bias feature vector set and the diagnosis confidence value are modality fused to form a joint review feature representation;

[0048] The bias amplitude, deviation direction weight and diagnosis confidence factor in the joint review feature representation are extracted and input into the credibility evaluation module to perform a confidence score calculation process;

[0049] The confidence score calculation result is calibrated in combination with a preset confidence score standard and a dynamic adjustment factor, and a dynamic confidence score is output;

[0050] The dynamic confidence score is written into the diagnosis node attribute field in the diagnosis and treatment data chain structure, and the evaluation mark field is updated synchronously;

[0051] The multiple rounds of dynamic confidence scores are time series sorted to construct a credibility change trajectory, and the trend attribute and fluctuation amplitude are labeled.

[0052] Optionally, the diagnosis correction operation is specifically:

[0053] The fusion feature representation, diagnosis label and diagnosis confidence value associated with the diagnosis node are extracted to construct a diagnosis correction input tensor;

[0054] The diagnosis correction input tensor and the bias feature vector set in the efficacy bias sequence are spliced into a diagnosis correction feature set;

[0055] The diagnosis correction feature set is input into a diagnosis correction module, and the diagnosis correction module includes a residual sensitive attention submodule and a confidence regulation submodule;

[0056] The coupling attention weight between the diagnosis label and the bias feature is extracted in the residual sensitive attention submodule to generate a diagnosis reconstruction vector;

[0057] The correction confidence factor is calculated based on the diagnosis reconstruction vector and the original diagnosis confidence value in the confidence regulation submodule, and the corrected diagnosis label and the corrected diagnosis confidence value are fused and output;

[0058] The corrected diagnosis label and the corrected diagnosis confidence value are written into the diagnosis node attribute field, and the original diagnosis result is kept as a historical reference field;

[0059] Record the trigger round of diagnostic correction operation, the participation mode combination and the correction model version information, form a correction record structure and store it in the diagnosis and treatment data chain.

[0060] Optionally, the step seven is specifically:

[0061] Extract each diagnostic node containing the correction record substructure in the diagnosis and treatment data chain, and arrange it in time sequence to form a diagnostic correction sequence;

[0062] Extract the diagnostic label variation, diagnostic confidence value change and efficacy deviation response information of each round of correction operation in the diagnostic correction sequence, and construct a node state vector;

[0063] Construct each state variation as a node in the evolution graph, and establish a directed edge between adjacent correction nodes, and mark the diagnostic variation amplitude and the confidence correction amplitude as edge weight;

[0064] Record the dynamic confidence score value and the corresponding deviation feature weight distribution in each correction round that triggers the diagnostic correction, as an edge attribute added to the evolution graph structure;

[0065] Record the correction trigger source identification, participation mode combination and correction model version information in each node as node label metadata;

[0066] Organize the diagnostic node state evolution, confidence adjustment trend and efficacy response feedback information in the form of graph structure, and construct an evolution graph;

[0067] Annotate the path in the constructed evolution graph, divide it into stable convergence path, fluctuation path and offset path, and write the path type annotation result into the diagnosis and treatment data chain diagnosis correction graph index domain.

[0068] Optionally, the step eight is specifically:

[0069] Identify the diagnostic node sequence in which diagnostic correction occurs continuously in multiple rounds in the evolution graph, and extract the diagnostic label, diagnostic confidence value and residual information of each diagnostic node to construct a state clustering label set;

[0070] Based on the state clustering label set, perform joint similarity analysis on the diagnostic label, diagnostic confidence value and residual information, and calculate the state variation amplitude of each diagnostic node and the state similarity between adjacent nodes;

[0071] Screen the continuous diagnostic nodes with state variation amplitude less than a preset compression threshold or state similarity greater than a preset similarity threshold as a compression candidate node sequence;

[0072] Select a representative node in the compression candidate node sequence as a key node in the skip diagnosis path, and construct a skip diagnosis edge between the key nodes;

[0073] Embed the skip diagnosis edge into the diagnosis and treatment data chain, and synchronously update the diagnosis order, node reference index and confidence propagation path structure in the diagnosis and treatment data chain;

[0074] Record the path compression operation number, path compression rule identifier and state similarity parameter configuration as the structure attribute field of the diagnosis and treatment data chain for additional storage.

[0075] The beneficial effects of the present application are:

[0076] The present application constructs a multi-source diagnosis and treatment data acquisition, modal feature alignment and multi-task learning model, and solves the problems of data structure heterogeneity, inconsistent sampling frequency and time sequence alignment in the multi-modal sample set. Unified patient identifier association and time consistency mapping are used, combined with modal attention gate and feature fusion mechanism, to realize high-precision alignment and information redundancy suppression of images, tests, physiological signals and electronic medical record texts. In the diagnosis and efficacy prediction link, an improved Transformer diagnosis task sub-module with modal gating mechanism and an efficacy trend prediction sub-module with fusion causal attention mechanism are constructed, the weight distribution and gradient fusion between tasks are optimized by joint loss function, and the synchronous optimal output of diagnosis label and predicted efficacy trend is realized. In the efficacy feedback processing link, a joint review mechanism based on efficacy deviation sequence and diagnosis confidence value is introduced, the diagnosis reliability is dynamically adjusted, and the residual sensitive attention and confidence regulation module is triggered to perform diagnosis correction, effectively improving the matching degree of diagnosis results and real efficacy changes. In the diagnosis and treatment path management link, a multi-dimensional diagnosis evolution graph containing correction records and state changes is constructed, and a skip diagnosis path compression is performed based on state similarity and change amplitude, which significantly reduces redundant nodes and path length. Finally, the AI-driven disease diagnosis and efficacy monitoring analysis system realizes high-precision diagnosis, sensitive efficacy tracking and efficient path optimization under the condition of multi-modal heterogeneous data, and improves the decision accuracy, response speed and explainability in complex clinical environment. BRIEF DESCRIPTION OF DRAWINGS

[0077] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0078] Figure 1 A flowchart of an AI-driven disease diagnosis and efficacy monitoring analysis system is proposed for the present application;

[0079] Figure 2A module diagram of an AI-driven disease diagnosis and efficacy monitoring analysis system is proposed for the present application. DETAILED DESCRIPTION

[0080] The present application will now be further described in greater detail in connection with the accompanying drawings. These drawings are not to scale and are merely schematic representations of the present application, with like structures carrying like reference numerals. They show embodiments which are exemplary of the principles of the application.

[0081] REFERENCE Figure 2 An AI-driven disease diagnosis and efficacy monitoring analysis system comprises:

[0082] A data processing module is configured to collect multi-source diagnosis and treatment data and construct a multi-modal sample set.

[0083] A diagnosis prediction module is configured to generate diagnosis labels, diagnosis confidence values, and predicted efficacy trends through a multi-task learning model.

[0084] An efficacy comparison module is configured to collect actual efficacy data, construct an efficacy observation sequence, and generate an efficacy deviation sequence based on the efficacy observation sequence and the predicted efficacy trend.

[0085] A credibility evaluation module is configured to perform credibility re-evaluation based on the efficacy deviation sequence and the diagnosis confidence values.

[0086] A path evolution module is configured to record continuous rounds of diagnosis correction results, generate a jump diagnosis path, and update a diagnosis and treatment data chain.

[0087] REFERENCE Figure 1 In the present embodiment, the modules are implemented through the following steps:

[0088] Step 1: Collect multi-source diagnosis and treatment data of a patient and construct a multi-modal sample set.

[0089] Step 2: Input the multi-modal sample set into a multi-task learning model to generate diagnosis labels, diagnosis confidence values, and predicted efficacy trends. The multi-task learning model comprises a diagnosis task sub-module and an efficacy prediction task sub-module.

[0090] Step 3: Construct a diagnosis and treatment data chain based on the diagnosis labels, diagnosis confidence values, and predicted efficacy trends.

[0091] Step 4: Collect actual efficacy data of the patient within a preset time window to form an efficacy observation sequence. Perform deviation comparison based on the efficacy observation sequence and the predicted efficacy trend to generate an efficacy deviation sequence.

[0092] Step 5: Perform credibility re-evaluation based on the efficacy deviation sequence and the diagnosis confidence values to calculate a dynamic confidence score.

[0093] Step six: if the dynamic confidence score is less than a preset confidence threshold, a diagnostic correction operation is performed;

[0094] Step seven: record the results of multiple rounds of the diagnostic correction to construct an evolution graph;

[0095] Step eight: perform a path compression operation based on the evolution graph to generate a jump diagnostic path and update the diagnosis and treatment data chain.

[0096] The embodiment collects multi-source diagnosis and treatment data of a patient and constructs a multi-modal sample set, so that the system can comprehensively obtain multi-dimensional information sources representing diagnosis state and treatment effect characteristics; a multi-task learning model is used to jointly generate a diagnosis label, a diagnosis confidence value and a predicted treatment trend, so that the diagnosis and treatment cognitive process is cooperatively modeled; a diagnosis and treatment data chain is constructed based on the diagnosis label and the predicted treatment trend, a time evolution relationship and a confidence propagation path are embedded in the chain structure, and the diagnosis and treatment state tracking capability is enhanced; actual treatment effect data is collected and compared with the predicted trend to form a treatment effect deviation sequence, and a feedback channel between diagnosis prediction and actual treatment effect is constructed; a reliability review is performed based on the treatment effect deviation and the original confidence value, and a dynamic confidence score is generated, so that a score adjustment mechanism based on treatment effect feedback is realized at the diagnosis label layer; a diagnostic correction operation is performed according to the dynamic confidence score and a preset confidence threshold, and a dynamic closed-loop diagnosis reliability adaptive mechanism is constructed; the results of multiple rounds of correction are recorded and an evolution graph is generated to support historical structure modeling of the diagnosis path; a path compression operation is performed based on the evolution graph to form a jump diagnostic path and synchronously update the diagnosis and treatment data chain structure, and the decision compactness and structural clarity of the diagnosis and treatment path are further optimized.

[0097] In the embodiment, the step one is specifically:

[0098] Medical image data is collected; the medical image data includes computed tomography images, magnetic resonance imaging images and ultrasonic imaging images; the medical image data is stored in a digital imaging and communication in medicine standard format, an extensible image format or a joint photographic experts group format, and image acquisition device information, scanning sequence parameter information and patient basic information fields are embedded in each image to establish a structured image data storage index;

[0099] Clinical test data is collected; the clinical test data includes blood biochemical index data, immunology test index data and pathogen detection index data; a structured two-dimensional test matrix is constructed according to sampling time, test item type and data source, and the clinical test data is stored in an extensible health interoperability resources format, and the abnormal marker state and reference value interval of each test index are retained;

[0100] Collecting physiological signal data; the physiological signal data includes electrocardiogram signal, respiration signal and blood pressure signal; a sampling frequency of no less than two hundred times per second is used in the collecting process, and a collecting time stamp is recorded synchronously with each signal waveform; the continuous waveform signal is divided into signal segments of a fixed time window length, and the channel source, amplitude range and integrity evaluation index of each signal are recorded;

[0101] Collecting electronic medical record text data; the electronic medical record text data includes medical history description data, chief complaint content data and doctor observation record data; medical terminology extraction, syntax structure analysis and semantic label modeling are performed on the original electronic medical record text data, the text data is converted into a fixed-length text vector representation through a natural language processing method, and each piece of text is labeled with its collection time information and diagnosis and treatment stage information;

[0102] Performing unified patient identification matching processing on the medical image data, clinical test data, physiological signal data and electronic medical record text data; performing asynchronous modality alignment operation based on time stamp information, and using linear interpolation algorithm or sliding window mechanism to perform time consistency mapping on sampling points of different modalities; performing feature extraction and modality conversion operation on various types of original data through a modality feature coding module to obtain image modality feature tensor, numerical modality feature vector, waveform modality feature matrix and text modality embedding vector; uniformly splicing all modality features to construct a data sample set with complete modality label, unified time index and consistent structure;

[0103] The embodiment realizes standardized collection, unified structure modeling and feature layer alignment of heterogeneous diagnosis and treatment data, provides a high-consistency, high-time-resolution and high-semantic-coverage input sample set for multi-task diagnosis models and efficacy monitoring models, and improves the spatio-temporal consistency and modality fusion efficiency of model input on the basis of ensuring the integrity of each modality data.

[0104] In the embodiment, the multi-task learning model includes a diagnosis task sub-module and an efficacy prediction task sub-module, specifically:

[0105] The multi-modal sample set is sequentially input into a modality embedding layer and a feature compression layer to extract image features, test features, physiological signal features and text semantic features, and perform dimension compression and unified embedding processing on various types of modality features to construct a multi-modal feature set with high representation ability and information condensation ability;

[0106] The multi-modal feature set is input to a fusion feature extraction module, and a modal attention gate mechanism is introduced to perform weight adjustment and redundancy suppression operations on the effectiveness and relevance of each modal feature to form a multi-modal fusion feature representation; the modal attention gate mechanism is jointly modeled based on a gating function and an attention score, and the specific weight distribution is fitted by minimizing an inter-modal redundancy rate function;

[0107] The fusion feature representation is input to a multi-task learning model, which includes a diagnosis task sub-module and a therapeutic effect trend prediction sub-module; wherein the diagnosis task sub-module adopts an improved Transformer with a modal gate mechanism, which is used to output diagnosis labels and diagnosis confidence values; the therapeutic effect trend prediction sub-module adopts a structure that combines a time series convolution network and a causal attention mechanism, which is used to capture long-term dependency relationships and mutation response features, and output a predicted therapeutic effect trend sequence;

[0108] The improved Transformer introduces a modal gate mechanism and embeds a task guide vector based on the original Transformer model, specifically including: using a multi-head self-attention layer to capture local and global dependencies in the fusion feature representation; introducing a modal gating function before each self-attention head to dynamically adjust the activation proportion of the input channels according to the significance of each modal feature; embedding a task guide vector in the position encoding stage to jointly model the context prior information of the target task with the input sequence through embedding; introducing a task-specific channel weight adjustment structure in the feedforward network module; in the multi-layer stacking structure, the feature stability is maintained through residual connection and layer normalization, and finally the feature representation for generating diagnosis labels and confidence is output;

[0109] The modal gating function is fitted by a gating weight generation network containing a multi-layer perceptron structure; the network takes the global statistical vector of each modal feature as input, and generates a normalized modal importance weight vector through linear transformation and nonlinear activation operation;

[0110] The modal gating mechanism is obtained by constructing an independent feature projection network for each input modal feature, performing linear mapping on each modal feature to obtain a corresponding feature representation vector; the feature representation vector is input to a gating calculation module, and based on the trainable parameters in the gating module, a Sigmoid activation operation is performed to generate a gating weight; the gating weight is used to perform weighted adjustment on the feature representation of the corresponding modal to obtain a modal weighted fusion representation; the gating weight is automatically learned through the backpropagation mechanism during training, so that the model can dynamically adjust the contribution of each modal feature to the fusion feature representation according to the quality and expression ability of the modal in different samples;

[0111] The task guide vector acquisition manner is: first, according to the target category of the diagnosis task or the efficacy prediction task, a fixed dimension task label vector is constructed, and the task category is represented by using one-hot encoding or trainable embedding; the task label vector is input into the guide vector generation network, nonlinear mapping is performed through a multilayer perceptron structure to obtain a task guide vector; in the input layer of the Transformer, the task guide vector and the modal fusion feature representation are spliced or cross-projected to form an input sequence matrix; through the task guide vector embedding strategy, the Transformer structure can explicitly obtain the task target information in the feature modeling process;

[0112] Further, the diagnosis label is embedded and coded, spliced with the fusion feature representation to form a condition control feature, and input into the efficacy trend prediction sub-module to perform a controlled prediction task;

[0113] In the joint training process, the diagnosis task sub-module and the efficacy trend prediction sub-module share a unified multi-modal joint feature encoding layer, and generate corresponding task outputs through a task-specific parameter separation structure, avoiding model overfitting;

[0114] The joint loss function of the multi-task learning model includes a cross-entropy loss term of the diagnosis label distribution and a mean square error loss term of the efficacy trend prediction. The joint loss function adjusts the loss sensitivity of each task through confidence weight and trend fluctuation factor, improving the robustness of the model on uncertain samples.

[0115] The joint loss function formula is:

[0116] ;

[0117] Wherein, denotes the joint loss function; denotes the cross-entropy loss of the diagnosis task; denotes the mean square error loss of the predicted efficacy trend; denotes the diagnosis confidence value; denotes the fitting error of the historical efficacy observation trend; the task adjustment weight , are respectively fitted by the following ways:

[0118] ;

[0119] Wherein, denotes the task weight vector; denotes the feature similarity between samples; represents the label confidence distribution; the function f is constructed by sample clustering and distribution learning; the joint loss function constructs a task relevance adjustment factor based on the diagnosis confidence and the efficacy bias trend after each round of training, and performs the following gradient fusion process:

[0120] ;

[0121] wherein, represents the gradient direction; represents the gradient of the loss function of the diagnosis task to the model parameters; represents the gradient of the loss function of the efficacy prediction task to the model parameters; and obtained by a dynamic task weight allocation strategy, and finally fitted on the basis of the task weight vector to obtain the optimal task optimization rhythm.

[0122] Finally, the joint optimization process is performed on the above-mentioned joint loss function to obtain the joint optimal solution path of the diagnosis label, the diagnosis confidence value and the predicted efficacy trend, which is fitted by the joint bias function between the diagnosis label distribution and the efficacy trend fluctuation.

[0123] The embodiment can significantly improve the synergistic efficiency between the diagnosis task and the efficacy prediction task, improve the diagnosis accuracy and the fitting ability of individualized efficacy trend prediction under small sample conditions, and realize high reliability construction and update of the diagnosis and treatment data chain.

[0124] In the embodiment, the step three is specifically:

[0125] The diagnosis label is processed by one-hot encoding to generate a diagnosis label vector, and the diagnosis confidence value is added to the diagnosis label vector as the confidence weight of the diagnosis node; the diagnosis label vector is set as the attribute representation of the initial diagnosis node, as the basic node unit in the diagnosis and treatment graph structure;

[0126] According to the diagnosis label output by the diagnosis task submodule and the efficacy trend sequence generated by the efficacy trend prediction submodule, an efficacy trend node is constructed; the efficacy trend node contains the continuous time interval of the predicted efficacy trend, the trend direction mark and the trend intensity value, and establishes a directional connection relationship with the corresponding diagnosis node, and the numerical value of the predicted time interval and the efficacy change direction is added to the connection attribute;

[0127] The patient identity and the treatment time information are respectively converted into unique identifiers, and a joint encoding operation is performed to generate a diagnosis and treatment primary key index; the diagnosis and treatment primary key index, the diagnosis node and the efficacy trend node are bound through a mapping mechanism to form a unified data structure, ensuring the consistency and traceability of the node index;

[0128] Introducing multiple auxiliary attributes between the diagnosis node and the efficacy trend node, the auxiliary attributes including a diagnosis source category number, a model version identifier, and a sample modality combination encoding value; performing additional embedding processing on the auxiliary attribute vector and controlling the attribute as a conditional variable for path generation to supplement the connection edge attribute between nodes;

[0129] Sorting the diagnosis node and the efficacy trend node based on a node generation time and a prediction time interval; performing sequential numbering on all nodes and connection edges using a time window sliding mechanism to generate a time axis-oriented diagnosis and treatment trajectory set; dividing the diagnosis and treatment trajectory set by patient dimension to construct an individual-oriented multi-round diagnosis and treatment trajectory structure;

[0130] Performing uniform format coding on the node attribute vector, edge connection feature, and auxiliary annotation information in each diagnosis and treatment trajectory structure, and constructing a graph structure data storage format to save the diagnosis and treatment trajectory structure in the form of a graph database or a graph structure tensor; ultimately forming a diagnosis and treatment data chain including diagnosis labels, diagnosis confidence values, efficacy trend intervals, auxiliary attributes, and structural connection relationships.

[0131] By introducing diagnosis confidence weights and predicted efficacy trends as graph structure attributes, a trajectory structure expression method oriented to diagnosis and treatment time series logic is established, and the structured utilization efficiency of the diagnosis and treatment data chain in subsequent evaluation, reasoning, and correction processes is improved.

[0132] In the present embodiment, the step four is specifically:

[0133] Collecting multi-dimensional efficacy data covering image indicators, test values, physiological signals, and subjective scores; performing correlation integration according to patient identity and time labels to construct an efficacy observation sequence containing multiple efficacy observation nodes, and uniformly storing them in the diagnosis and treatment data chain;

[0134] Performing a modality standardization process, applying a score table conversion function to image indicators, performing Z-Score standardization on test values, using a band-pass filter to filter high-frequency noise from physiological signals, and establishing a discrete score mapping table for subjective scores; completing scale unification processing between modalities through a uniform value domain mapping function to generate a uniform format efficacy observation vector;

[0135] Based on the efficacy prediction trend node output by the diagnosis task sub-module, extracting the corresponding prediction time interval and efficacy change direction; performing position alignment to one-to-one match the prediction node and the efficacy observation node in the time label; calculating the item-by-item difference between the predicted value and the observed value to construct a residual deviation matrix, wherein the residual value is obtained by fitting a residual function, the residual function takes the observed value and the predicted value as input, and outputs a residual value vector at each time step, and the formula is obtained by least mean square error fitting.

[0136] According to the residual deviation matrix, a time-ordered efficacy deviation sequence is constructed; the residual amplitude is hierarchically clustered and labeled with residual intensity levels, and the deviation trend is marked by a directionality function, which takes the cumulative change trend of the residual vector within a unit time window as input and outputs a set of directionality labels, which are fitted by a residual deviation direction fitting function;

[0137] The residual amplitude, deviation direction, time label, and their matching relationship of each node in the efficacy deviation node set are attribute-labeled, and are synchronously bound with the diagnostic confidence value output by the diagnostic task submodule; a multi-attribute connection relationship between nodes is established in the diagnosis and treatment data chain, and an optional evaluation field is added to mark whether to trigger the diagnostic confidence review operation;

[0138] Perform node topology alignment operation; according to the time topology position of the predicted efficacy trend node in the diagnosis and treatment data chain, complete the index labeling and structure embedding of the efficacy deviation node; ensure that the time sequence relationship between the efficacy deviation node and the diagnosis node, the efficacy trend node is consistent, and the expanded chain diagnosis and treatment structure representation is formed.

[0139] Through residual function fitting, directionality function labeling, and diagnosis and treatment chain structure fusion, multi-dimensional quantitative expression and dynamic correlation modeling effect of the prediction error of efficacy are realized, and the accuracy of diagnosis and evaluation and the precision of efficacy tracking are improved.

[0140] In this embodiment, step five is specifically:

[0141] Extract the residual intensity level and deviation direction information of each efficacy observation node in the efficacy deviation sequence, and construct a deviation feature representation vector using vector splicing; arrange a plurality of deviation feature representation vectors in time sequence to form a deviation feature vector set, and construct a joint review feature representation with the diagnostic confidence value output by the diagnostic task submodule through modal fusion operation; the modal fusion operation includes weight normalization processing of the deviation feature vector, and Laplace smoothing method for numerical stabilization processing of the diagnostic confidence value, and the fusion process is fitted by a modal joint mapping function;

[0142] Input the joint review feature representation into the confidence evaluation module, call the feature deconstruction function to extract the residual amplitude index, directionality distribution weight, and confidence factor value therein, and use them as input variables in the confidence score calculation process; the confidence score calculation function is based on the nonlinear relationship between the deviation amplitude and the confidence factor, and uses a deviation-confidence nonlinear interaction model to construct a score function, which outputs a diagnostic confidence score; the score function is constructed through a bivariate fitting process;

[0143] According to a preset confidence score standard threshold and an adjustment factor for fault-tolerant dynamic adjustment, calibration processing is performed on the above diagnostic confidence score to form a dynamic confidence score; the calibration function adopts a score segmentation mapping function, and residual correction is performed based on a confidence distribution history, and the mapping function is obtained by fitting a confidence error distribution;

[0144] The dynamic confidence score is written into a diagnostic node attribute field in a diagnosis and treatment data chain structure, a confidence mark field of the diagnostic node is updated, score source types, bias feature paths and time index fields are added, and connection attributes between the diagnostic node and an efficacy trend node are synchronously refreshed to mark a confidence update state;

[0145] The dynamic confidence scores of the diagnostic nodes in each round are formed into a time sequence, the sequence is arranged in ascending order of time labels, and a confidence change trajectory is generated; a score slope, a fluctuation amplitude and a trend direction are calculated on the change trajectory; the trend attribute is marked as improvement, decline or fluctuation, and the fluctuation amplitude is marked as stable, slight or severe; and the trend analysis process is modeled by a confidence trajectory change fitting function.

[0146] By constructing a joint review feature representation, confidence score calculation, score calibration and confidence trajectory modeling process, quantitative dynamic expression and structural tracking of the diagnostic confidence level are realized, and the sensitivity and time sequence response capability of the diagnostic confidence evaluation are improved.

[0147] In the embodiment, the diagnostic correction operation is specifically:

[0148] The fusion feature representation, the diagnostic category label and the diagnostic confidence value stored in the diagnostic node attribute field are extracted, a three-dimensional structure diagnostic correction input tensor is constructed, the tensor structure includes a multi-modal feature dimension, a time label dimension and a confidence mapping dimension; the diagnostic correction input tensor and the bias feature vector set corresponding to each efficacy observation node in the efficacy bias sequence are spliced according to the modal dimension to form a diagnostic correction feature set;

[0149] The diagnostic correction feature set is input into a diagnostic correction module, and the diagnostic correction module includes two structured submodules, a residual sensitive attention submodule and a confidence regulation submodule; the attention mechanism calculation is performed in the residual sensitive attention submodule, the coupling response value between the diagnostic category label and the bias feature is extracted, the attention mechanism adopts a label-guided attention mapping function, which is fitted through a label-bias co-occurrence distribution relationship, and a diagnostic reconstruction vector is generated;

[0150] The diagnostic reconstruction vector is input into the confidence regulation submodule, and a correction input pair is constructed by fusing the original diagnostic confidence value; a confidence adjustment function is operated in the confidence regulation submodule to calculate a correction confidence factor, and the confidence adjustment function is fitted by a joint distribution of a diagnostic label and bias information; and a modified diagnostic category label and a modified diagnostic confidence value are output after fusing the correction confidence factor and the original diagnostic category label.

[0151] The modified diagnostic category label and the modified diagnostic confidence value are written into an attribute field of a diagnostic node, and the original diagnostic category label and the original diagnostic confidence value are retained in the attribute field as a historical reference field; the historical reference field includes an initial diagnostic timestamp, an original diagnostic label, an original confidence value, and a correction trigger marker.

[0152] Trigger information of the diagnostic correction operation is recorded, including a trigger round, a bias level range of a trigger condition, a modality combination of an input feature, an executed diagnostic correction module structure configuration, and a model weight version number, a structured correction record field is formed, and the diagnostic node record field in the diagnosis and treatment data chain structure is written; the information flow state in the diagnosis and treatment data chain is updated, and a correction traceability path and a confidence update state field are marked.

[0153] The linkage reconstruction operation of the diagnostic label and the confidence is completed by introducing the residual sensitive attention mechanism and the confidence regulation structure, the pertinence and convergence of the diagnostic correction are enhanced, the error adaptive correction and dynamic confidence enhancement of the diagnostic result are realized, and the diagnosis stability and the efficacy alignment ability of the system in a complex case scenario are effectively improved.

[0154] In the embodiment, the step seven is specifically:

[0155] All diagnostic nodes containing the completed diagnostic correction record field in the diagnosis and treatment data chain structure are extracted; the diagnostic nodes are arranged in ascending order according to the timestamp field of the diagnostic nodes to generate a diagnostic correction sequence with a time sequence attribute;

[0156] The diagnostic category label, the diagnostic confidence value, and the efficacy bias feature information recorded in the diagnostic node corresponding to each round of diagnostic correction operation are extracted one by one in the diagnostic correction sequence; the node state variation value is calculated according to the label variation, the confidence variation, and the efficacy bias response amplitude, and the node state vector is fitted by a state variation function;

[0157] Each corrected diagnostic node is taken as a node element in a graph structure, a directed edge is established between adjacent two corrected nodes, and the difference between the node state vectors is taken as an edge weight value; the edge weight includes two components of a diagnostic label variation amplitude and a diagnostic confidence value correction amplitude, wherein the diagnostic label variation amplitude is fitted by a label difference measurement function, and the diagnostic confidence value correction amplitude is fitted by a confidence variation function.

[0158] The dynamic confidence score corresponding to each round of correction operation, the weight distribution information of each feature dimension in the bias feature vector are recorded, which are embedded in the graph structure as additional attributes of the edge, a multi-attribute edge weight expression is constructed to represent the decision basis and bias source of the correction trigger;

[0159] In each diagnosis node, the correction trigger source identification is further written, including trigger source type mark, participation mode combination encoding and model version information of the diagnosis correction module; the above information is embedded in the evolution graph node attribute field as the metadata label of the node;

[0160] The diagnosis node state evolution process is expressed by a graph data structure, the diagnosis category change, the diagnosis confidence adjustment trend and the response feedback information related to the efficacy bias are organized, and an evolution graph is constructed;

[0161] The path type recognition operation is performed on the constructed evolution graph, and all paths in the graph are divided based on the label change trend and the confidence fluctuation range; the path type is divided into three categories: stable convergence path, fluctuation path and offset path; the stable convergence path corresponds to the node sequence with continuous confidence improvement and clear label convergence trend; the fluctuation path corresponds to the path with up and down fluctuation of diagnosis confidence exceeding the set threshold; the offset path corresponds to the path with continuous label inversion or confidence decline; the path type is obtained by joint fitting of the confidence change rate function and the label consistency index function;

[0162] The path type recognition result is written into the evolution graph structure as a label field of the graph structure, and is simultaneously written into the diagnosis correction graph index domain of the diagnosis and treatment data chain to support subsequent path compression operation and rapid identification of abnormal diagnosis chain.

[0163] By introducing the multi-dimensional diagnosis state variation graphing process, the continuity, bias and confidence evolution trend of the diagnosis decision correction process are systematically described, the structure expression ability of the diagnosis and treatment data chain and the visualization depth of the correction path are improved, and the graph computing basis is provided for diagnosis path compression and unstable node identification.

[0164] In the embodiment, the step eight is specifically:

[0165] All diagnosis node subgraphs containing continuous multi-round diagnosis correction records are identified in the evolution graph structure; all diagnosis nodes in the subgraph are extracted, the diagnosis label, the diagnosis confidence value and the efficacy bias residual information recorded in the node are read, a three-dimensional state vector group is formed; a state clustering label set is constructed according to the vector dimension, and a state label field is written in the diagnosis node;

[0166] Based on the state clustering label set, a joint similarity analysis operation is performed to calculate the label distance between diagnostic labels, the feature difference between the change gradient and the residual information of the diagnostic confidence value, respectively; the similarity measurement results in three dimensions are normalized to construct the state similarity value between the state change amplitude value and the adjacent nodes; the above two indicators are fitted by the state clustering similarity function and the confidence residual function;

[0167] The path compression condition parameters are set, wherein the continuous diagnostic node sequence with a state change amplitude less than a path compression threshold or a state similarity greater than a similarity determination threshold is filtered as a compression candidate node sequence; the node sequence meeting the condition is subgraph divided and marked as a compression candidate sub-path;

[0168] In each group of compression candidate sub-paths, a representative node is extracted based on the state vector clustering center, and the representative node is the node with the minimum distance to the clustering center in the state space; the representative node is taken as a key node in the jump diagnostic path, and a jump diagnostic edge is established between adjacent key nodes; the edge weight of the jump diagnostic edge is fitted by the state vector change and the confidence improvement value between the front and rear key nodes;

[0169] The jump diagnostic edge is written into the diagnosis and treatment data chain structure, and the timestamp field, node reference index structure and confidence propagation path field of the diagnostic node are updated in the diagnosis and treatment data chain; a nested linked list structure containing a jump edge pointer field and an original path backtracking field is constructed to support jump path tracking and original path rollback reconstruction;

[0170] After completing the compression path writing, the number of this round of path compression operation, the path compression rule identifier and the state similarity parameter configuration adopted are recorded; these meta information are attached to the metadata area of the diagnosis and treatment data chain structure as a structure attribute field, and a path compression version number is generated to support multi-version diagnostic path management.

[0171] Through the jump path construction and path compression parameter synchronous maintenance operation, the diagnosis and treatment data chain structure is greatly reduced in path length and information redundancy on the basis of retaining key evolution information, effectively enhancing the visual expressiveness of the diagnostic correction sequence and the execution efficiency of the subsequent diagnostic optimization algorithm.

[0172] Embodiment 1:

[0173] To verify the feasibility of the application in implementation, the application is applied to the outpatient and follow-up scene of the rheumatology and immunology department of a certain third-grade class-A hospital for three months. The department has long been faced with problems such as difficulty in diagnosing autoimmune diseases, lag in response evaluation of curative effect, and difficulty in quantifying misdiagnosis correction. The patient population is diverse and the disease is complex. Traditional diagnosis mainly relies on subjective judgment of doctors and periodic review, which is particularly unfavorable for chronic diseases. There are problems such as unstable diagnostic confidence, inaccurate curative effect prediction, and lack of data support for diagnosis correction.

[0174] The hospital selects 468 systemic lupus erythematosus patients as a pilot population, collects the patient's historical outpatient data, image data, genetic test reports, laboratory examination records and drug reaction information to construct a multi-modal diagnosis and treatment data sample set, and an average of about 53 records are formed for each patient. In the data preprocessing stage, modal alignment, format unification and feature normalization operations are performed, and a window period of 14 days per round is set to form a continuous 6-round curative effect observation sequence.

[0175] Based on the method proposed in the application, an improved multi-task learning model is used to output initial diagnosis labels, diagnosis confidence values and future three-round curative effect trend prediction results, and the diagnosis output results are constructed into a diagnosis and treatment data chain to support real-time evolution record of the state in follow-up. In the patient follow-up window, the actual curative effect observation data are collected and compared with the model prediction results to obtain a curative effect deviation sequence and calculate a dynamic confidence score accordingly. If the score is lower than the system set confidence threshold 0.75, the system built-in diagnosis correction module is triggered to correct the current label.

[0176] Supporting multi-modal correction strategy, the most suitable correction model version is selected in combination with the modal response of the deviation source. The diagnosis and treatment trajectory of the patient who has correction behavior for more than three rounds is automatically recorded in the evolution graph structure, and the key path nodes are extracted for diagnosis path compression. In the path compression process, the system automatically identifies nodes with high state similarity or low change amplitude as jump connection candidates, and constructs a jump-type diagnosis path to reduce redundant records and improve system response speed.

[0177] After three months of pilot application, the system triggered diagnosis correction 1824 times in total, of which 513 times were high deviation correction with curative effect deviation absolute value greater than 0.4, and the confidence correction amplitude was significantly more than 125% of the original score. At the same time, the system cumulatively constructed 312 evolution graphs, and the diagnosis path compression reached 785 times, greatly improving the compactness and interpretability of the diagnosis chain. The following is a part of the diagnosis correction and path compression statistics table generated by the system during the pilot period.

[0178] Table 1 Patient diagnosis correction statistical data table

[0179] Field P045 P102 P287 P331 Initial diagnosis label SLE_Moderate activity SLE_Active SLE_Stable SLE_Mild remission Corrected label SLE_Mild remission SLE_Stable SLE_Moderate activity SLE_Stable Correction round 3 2 4 3 Dynamic confidence score 0.683 0.702 0.641 0.676 Bias source modality Lab + text Image + text Lab Multi-modal joint

[0180] From Table 1, it can be seen that the system can accurately identify the deviation source mode while the dynamic score is decreasing, and correct according to the label change trend. Among them, the P287 patient correction round reaches 4 rounds, and the system dynamic score is 0.641, indicating that there is obvious uncertainty in the diagnosis, and the system effectively realizes the label regression to the active period state through multiple rounds of correction, which is consistent with the subsequent manual evaluation results, reflecting the high sensitivity and accuracy of the system.

[0181] Table 2 Comparison table of compression effect of jump diagnosis path

[0182] Field R001 R032 R087 R213 Original node number 7 5 9 6 Key node number after jump 3 2 4 3 State similarity mean 0.916 0.894 0.936 0.882 Path compression rate 57.1% 60.0% 55.6% 50.0% Average confidence growth rate +13.7% +11.2% +16.4% +12.9%

[0183] As can be seen from Table 2, the average number of nodes of the compressed path is reduced by more than 50%, and the average state similarity is generally greater than 0.88, indicating that the compression does not affect the expression of the core information of the path. The average confidence growth rate is significantly improved, and the confidence of some paths increases by more than 15% after compression, which reflects the ability of the system to enhance the reliability of the judgment in the path optimization.

[0184] This embodiment fully verifies the practical value of the present application in the real diagnosis and treatment scene. The system has superior performance in improving the reliability of diagnosis, compressing the diagnosis path and enhancing the accuracy of therapeutic response, and significantly optimizes the decision efficiency and intelligent level of the original diagnosis and treatment process.

[0185] The above describes only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An AI-driven disease diagnosis and efficacy monitoring and analysis system, characterized in that, include: The data processing module is used to collect multi-source diagnostic and treatment data and construct a multimodal sample set; The diagnostic prediction module is used to generate diagnostic labels, diagnostic confidence values, and predictive efficacy trends through a multi-task learning model. The efficacy comparison module is used to collect actual efficacy data, construct efficacy observation sequences, and generate efficacy deviation sequences based on the efficacy observation sequences and predicted efficacy trends. The credibility assessment module is used to perform credibility feedback based on the efficacy deviation sequence and diagnostic confidence value; The path evolution module is used to record the results of multiple rounds of diagnostic corrections, generate leapfrog diagnostic paths, and update the diagnostic and treatment data chain.

2. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 1, characterized in that, The following steps are used to achieve inter-module communication: Step 1: Collect multi-source diagnosis and treatment data from patients to construct a multimodal sample set; Step 2: Input the multimodal sample set into the multi-task learning model to generate diagnostic labels, diagnostic confidence values, and predicted efficacy trends; the multi-task learning model includes a diagnostic task sub-module and an efficacy prediction task sub-module; Step 3: Construct a diagnosis and treatment data chain based on the diagnostic labels, diagnostic confidence values, and predicted efficacy trends; Step 4: Collect actual efficacy data of patients within a preset time window to form an efficacy observation sequence; and perform a deviation comparison between the efficacy observation sequence and the predicted efficacy trend to generate an efficacy deviation sequence. Step 5: Perform a confidence back-evaluation based on the efficacy deviation sequence and diagnostic confidence value, and calculate the dynamic confidence score; Step 6: If the dynamic confidence score is less than the preset confidence threshold, then perform a diagnostic correction operation; Step 7: Record the results of the diagnostic corrections over multiple consecutive rounds and construct an evolutionary graph; Step 8: Perform path compression operation based on the evolution graph to generate a skip-type diagnostic path and update the diagnostic data chain.

3. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 2, characterized in that, Step one specifically involves: Acquire medical imaging data; the medical imaging data includes computed tomography images, magnetic resonance imaging images and ultrasound imaging images, and is stored in a standard image format; Collect clinical laboratory data; the clinical laboratory data includes blood biochemical indicators, immunological detection indicators and pathogen detection indicators, and is stored in a structured laboratory report data format; Collect physiological signal data; The physiological signal data includes electrocardiogram signals, respiratory signals, and blood pressure signals. The sampling frequency is associated with time stamps and a continuous waveform sequence is recorded. Collect electronic medical record text data; the electronic medical record text data includes medical history description, chief complaint content and doctor's observation record, which are preprocessed by natural language processing methods and transformed into structured text embedding; The medical imaging data, clinical laboratory data, physiological signal data, and electronic medical record text data are subjected to unified patient identification association, time alignment processing, and modality mapping processing to construct a multimodal sample set with a unified format.

4. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 3, characterized in that, The multi-task learning model includes a diagnostic task sub-module and an efficacy prediction task sub-module, specifically: The diagnostic task submodule is based on an improved Transformer architecture. It utilizes fused feature representations to perform feature attention extraction and temporal correlation modeling, and outputs disease diagnostic labels and diagnostic confidence values. The diagnostic labels represent structured disease classification results, and the diagnostic confidence values ​​represent the reliability of the model's classification judgment. The improved Transformer structure includes: a modal gating mechanism before each self-attention head, the modal gating mechanism being composed of a gating function generated based on multimodal feature statistical indicators; the gating function being obtained by fitting modal distribution features and saliency weight parameters in the training samples; embedding a task guidance vector in the position encoding stage, the task guidance vector being generated by the structural features of the diagnostic task label space, and jointly constructing a position encoding representation with the input sequence through embedding; setting a task-specific channel weight adjustment structure in the feedforward neural network module; the Transformer structure is constructed in a multi-layer stacked form, with residual connection structures and layer normalization operations introduced between layers; The efficacy trend prediction submodule adopts a structure that integrates time-series convolutional networks and causal attention mechanisms to model historical treatment response trajectories and extract long-term dependencies and stage-specific mutation signals; under the conditional control of the diagnostic task output results, the submodule generates a predicted efficacy trend sequence. The parameters are updated using a joint loss function of a multi-task learning model. The joint loss function consists of a cross-entropy loss term for the diagnostic task and a mean squared error loss term for the efficacy trend prediction task. Furthermore, the diagnostic confidence value and the sample efficacy trend fluctuation index are introduced as weighting factors to perform differentiated adjustment of the loss weights for each task.

5. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 4, characterized in that, Step three specifically involves: The diagnostic label is used as the initial diagnostic node for encoding, and the diagnostic confidence value is added as the weight attribute of the initial diagnostic node. Based on the predicted efficacy trend corresponding to the diagnostic label, construct efficacy trend nodes, establish directional connection relationships, and mark the predicted time interval and the direction of efficacy change. Patient identification and consultation time information are mapped to a primary key index for diagnosis and treatment, and a unified data structure is formed with diagnosis nodes and efficacy trend nodes. Auxiliary attributes are introduced between the diagnostic node and the efficacy trend node, including the source of diagnosis, model version number and sample modality combination identifier, and are additionally coded as conditions for generating the diagnosis and treatment path; The diagnostic nodes, efficacy trend nodes, and connection relationships are sorted by time to form a multi-round diagnosis and treatment trajectory structure. The node information, connection attributes, and auxiliary annotations in each trajectory structure are stored in a graph structure to form a diagnostic data chain.

6. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 5, characterized in that, Step four specifically involves: Multidimensional efficacy data covering imaging indicators, test values, physiological signals and subjective scores are collected, and integrated according to patient identification and time tags to construct an efficacy observation sequence containing multiple efficacy observation nodes. The efficacy observation sequence is processed to achieve a uniform scale, and the position is aligned according to the time interval and direction of change marked in the predicted efficacy trend. Calculate the difference between each efficacy prediction node and the corresponding efficacy observation node, establish a residual deviation matrix, and extract the residual magnitude and direction markers; A therapeutic effect deviation sequence is constructed based on the residual deviation matrix, and the residual intensity level and deviation direction of each therapeutic effect observation node are marked. The residual information of each efficacy observation node in the efficacy deviation sequence is stored synchronously with the diagnostic confidence value, and an optional evaluation label field is set. The efficacy deviation sequence is embedded into the diagnosis and treatment data chain structure, consistent with the index and time topology order of the predicted efficacy trend nodes.

7. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 6, characterized in that, Step five specifically involves: Extract the residual intensity level and deviation direction corresponding to each efficacy observation node in the efficacy deviation sequence, and construct a set of deviation feature vectors; The set of deviation feature vectors is modally fused with the diagnostic confidence values ​​to form a joint feedback feature representation. The joint feedback feature representation is input into the confidence assessment module, which extracts the deviation magnitude, deviation direction weight, and diagnostic confidence factor from the representation, and performs the confidence score calculation process. By combining the preset confidence scoring criteria and dynamic adjustment factors, the confidence score calculation results are calibrated, and a dynamic confidence score is output. The dynamic confidence score is written into the diagnostic node attribute field of the diagnosis and treatment data chain structure, and the evaluation tag field is updated synchronously. The time series of multiple rounds of dynamic confidence scores are sorted to construct a confidence change trajectory and label trend attributes and fluctuation amplitude.

8. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 7, characterized in that, The diagnostic correction operation specifically includes: Extract the fusion feature representation, diagnostic label, and diagnostic confidence value associated with the diagnostic node to construct the diagnostic correction input tensor; The diagnostic correction input tensor is concatenated with the set of deviation feature vectors in the therapeutic effect deviation sequence to form a diagnostic correction feature set; Input the diagnostic correction feature set into the diagnostic correction module, which includes a residual-sensitive attention submodule and a confidence modulation submodule. Extract the coupled attention weights between diagnostic labels and bias features in the residual-sensitive attention submodule to generate a diagnostic reconstruction vector; In the confidence adjustment submodule, the corrected confidence factor is calculated based on the diagnostic reconstruction vector and the original diagnostic confidence value, and the corrected diagnostic label and the corrected diagnostic confidence value are merged and output. Write the corrected diagnostic label and corrected diagnostic confidence value into the attribute field of the diagnostic node, and retain the original diagnostic result as a historical reference field; Record the trigger round, participating modal combination, and correction model version information corresponding to the diagnostic correction operation, form a correction record structure, and store it in the diagnosis and treatment data chain.

9. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 8, characterized in that, Step seven specifically involves: Extract each diagnostic node containing the correction record substructure from the diagnosis and treatment data chain, and arrange them in chronological order to form a diagnostic correction sequence; Extract diagnostic label changes, diagnostic confidence value changes, and efficacy deviation response information from each round of correction operations in the diagnostic correction sequence to construct node state vectors; Each state change is constructed as a node in the evolution graph, and directed edges are established between adjacent correction nodes. The edge weights are labeled to diagnose the change magnitude and confidence correction magnitude. Record the dynamic confidence score values ​​and corresponding bias feature weight distributions that trigger diagnostic corrections in each correction round, and attach them as edge attributes to the evolution graph structure; Record the correction trigger source identifier, participating modal combination, and correction model version information in each node as node tag metadata; An evolutionary graph is constructed by organizing the state evolution, confidence adjustment trend, and efficacy response feedback information of each diagnostic node in a graph structure. The paths in the completed evolution graph are labeled with their types, divided into stable convergent paths, fluctuating paths, and offset paths, and the path type labeling results are written into the diagnostic correction graph index field in the diagnostic data chain.

10. The AI-driven disease diagnosis and efficacy monitoring and analysis system according to claim 9, characterized in that, Step eight specifically involves: In the evolution graph, a sequence of diagnostic nodes that undergo diagnostic correction in multiple consecutive rounds is identified, and the diagnostic labels, diagnostic confidence values ​​and residual information of each diagnostic node are extracted to construct a state clustering label set. Based on the state clustering label set, joint similarity analysis is performed on the diagnostic labels, diagnostic confidence values ​​and residual information to calculate the state change amplitude of each diagnostic node and the state similarity between adjacent nodes. Continuous diagnostic nodes whose state change amplitude is less than a preset compression threshold, or whose state similarity is greater than a preset similarity threshold, are selected as compression candidate node sequences. Representative nodes are selected from the compressed candidate node sequence as key nodes in the skip diagnostic path, and skip diagnostic edges are constructed between the key nodes. The skip diagnostic edge is embedded into the diagnostic data chain, and the diagnostic order, node reference index and confidence propagation path structure in the diagnostic data chain are updated synchronously. Record the path compression operation number, path compression rule identifier, and status similarity parameter configuration, and store them as additional structural attribute fields of the diagnosis and treatment data chain.

Images