Chemotherapy adverse reaction prediction and intervention system based on big data

Abstract

The invention relates to the technical field of medical data processing and prediction, and discloses a chemotherapy adverse reaction prediction and intervention system based on big data. The system integrates an unstructured disease course text and structured inspection data of a patient, generates a time-series symptom event, and performs time window alignment and fusion on the time-series symptom event, a medication record and a physical sign monitoring stream to form a multi-dimensional time-series data block. The system is combined with an external medical knowledge base to construct a dynamic association network among symptoms, medicines and physiological indexes, and dynamically calculates the confidence coefficient of an adverse reaction mode according to real-time data. By using a predictive model of the timing attention mechanism, the system can output a continuous curve of patient risk over time. The system automatically matches and generates a personalized intervention instruction sequence containing specific measures and execution time windows according to key time points and modes when the risk curve exceeds a threshold value. According to the invention, dynamic and advanced early warning and accurate intervention of adverse reaction risks of chemotherapy are realized.

Description

Technical Field

[0001] This invention relates to the field of medical data processing and prediction technology, specifically a big data-based system for predicting and intervening in adverse reactions to chemotherapy. Background Technology

[0002] In the current field of medical data analysis, especially in the monitoring of adverse reactions in chemotherapy patients, the primary reliance is on threshold alarms based on structured test reports and physicians' experience-based judgment. Existing solutions typically input these discrete numerical indicators into rule-based engines or traditional machine learning classifiers to output a static risk label. These methods have limitations: they cannot effectively integrate unstructured text containing rich symptom information from patient medical records, and they struggle to handle the complex temporal relationships between continuously monitored vital signs and medication events, resulting in insensitivity to early, subtle signals of adverse reactions and delayed warnings.

[0003] Existing risk inference models are mostly based on fixed medical knowledge or historical statistical data, lacking the ability to dynamically learn and discover correlations based on real-time data streams from individual patients. Their predictions are often isolated conclusions, failing to reveal the specific evolution path and future trends of risk over time. This directly leads to passive and indiscriminate clinical interventions, making it difficult to implement precise and timely preventative measures before risks truly manifest. A technological solution is needed that can deeply integrate multi-source heterogeneous time-series data, dynamically construct individualized risk correlation networks, and predict the continuous evolution trajectory of risks, in order to overcome the limitations of the current passive response model and achieve proactive and precise management of adverse reactions to chemotherapy. Summary of the Invention

[0004] The purpose of this invention is to provide a big data-based system for predicting and intervening in adverse reactions to chemotherapy, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a big data-based system for predicting and intervening in adverse reactions to chemotherapy, the system comprising:

[0006] The data acquisition and preprocessing module is used to acquire initial patient records from the medical data warehouse. The initial patient records contain structured test values ​​and unstructured medical history text. The module performs sentence segmentation, medical terminology extraction and normalization on the unstructured medical history text to generate time series symptom events aligned with the structured test values.

[0007] The multi-dimensional time-series data fusion module is used to align and fuse the time-series symptom events with patient medication records and vital sign monitoring streams within time windows, and to construct a multi-dimensional time-series data block centered on the individual patient.

[0008] The dynamic association network construction module is used to receive entity relationship injection from an external medical knowledge base, establish a dynamic association network between symptoms, drugs, and physiological indicators on the multi-dimensional time-series data block, and calculate the confidence level of potential adverse reaction patterns based on the connection strength of nodes and the co-occurrence frequency of events in the dynamic association network.

[0009] The risk prediction module uses a prediction model based on a time-series attention mechanism to calculate the multi-dimensional time-series data block and outputs the risk evolution curve of adverse reactions.

[0010] The intervention instruction generation module is used to match a predefined intervention protocol library with the time points and patterns when the risk evolution curve exceeds a preset threshold, and generate a personalized intervention instruction sequence that includes the type of intervention measure, execution time window and dose adjustment suggestions.

[0011] Preferably, the data acquisition and preprocessing module performs sentence segmentation, medical terminology extraction, and normalization on the unstructured medical history text to generate time-series symptom events aligned with structured test values, including:

[0012] A sentence segmentation model combining rules and statistics is used to segment long disease texts into independent semantic segments based on punctuation marks, transition words, and medical terminology.

[0013] Within the independent semantic segments, symptom descriptions, body parts, and severity modifiers are identified using a pre-built medical entity dictionary and a conditional random field model.

[0014] The identified symptom descriptions are mapped to a standard medical terminology set, and body parts and severity modifiers are converted into standard codes, which are then combined to form a standardized symptom expression.

[0015] From the independent semantic fragments, relative or absolute time information is extracted using a time expression recognition model, and the standardized symptom expression is anchored to a specific time point.

[0016] The standardized symptom expressions anchored to specific time points are associated and stored with structured test values ​​within the same time window according to patient identifiers and timestamps to form the time-series symptom events.

[0017] Preferably, the dynamic association network construction module receives entity relationship injection from an external medical knowledge base and establishes a dynamic association network between symptoms, drugs, and physiological indicators on the multi-dimensional time-series data blocks, including:

[0018] Periodically download lists of adverse drug reactions, symptom correlations, and normal ranges of physiological parameters from external medical knowledge bases, and convert them into graph-structured data;

[0019] The graph structure data uses drugs, symptoms, and physiological indicators as nodes, and known causal, accompaniment, and contraindication relationships as edges, and assigns initial weights.

[0020] Using specific drugs, symptom events, and physiological indicators appearing in the multi-dimensional time-series data block as activation nodes, a matching query is performed in the graph structure data;

[0021] Based on the actual time sequence and interval of events in the multi-dimensional time-series data block, the weight coefficients of the edges between the activated nodes are dynamically adjusted.

[0022] The subgraph consisting of activated nodes and their relational edges, after adjusting the weight coefficients, is taken as the dynamic association network.

[0023] Preferably, the confidence level of a potential adverse reaction pattern is calculated based on the connection strength of nodes and the co-occurrence frequency of events in the dynamic association network, including:

[0024] Traverse each symptom node in the dynamic association network, find the drug nodes and physiological indicator nodes directly connected to the symptom node, and form an association node group with the symptom node as the core.

[0025] The frequency with which all nodes in the associated node group co-occur in the same chemotherapy cycle for the same patient is statistically analyzed in historical multi-dimensional time-series data blocks.

[0026] By combining the current weight coefficients of the edges connecting each node in the associated node group in the dynamic association network, the initial confidence of the adverse reaction mode constituted by the associated node group is obtained through a weighted calculation model.

[0027] A time decay factor is introduced to assign lower weights to historical co-occurring events that occurred a long time ago, thereby correcting the initial confidence level.

[0028] The corrected confidence level is bound to the associated node group and used as the confidence level for the potential adverse reaction mode.

[0029] Preferably, the risk prediction module uses a prediction model based on a time-series attention mechanism to calculate the multi-dimensional time-series data block and outputs a risk evolution curve for adverse reactions, including:

[0030] The multi-dimensional time-series data block is sliced ​​at fixed time intervals, and each time slice contains feature vector representations of all symptom events, medication records, and test values ​​within the corresponding time period.

[0031] The feature vector sequence arranged in chronological order is input into the encoder, which is composed of a stacked recurrent neural network layer and a temporal self-attention layer.

[0032] The temporal self-attention layer calculates the correlation weight between the feature vector of each time slice and the feature vectors of all time slices in the sequence, thereby generating a context vector containing global temporal dependencies.

[0033] The encoder ultimately outputs an enhanced feature representation for each time slice, which integrates current information with contextual information from historical sequences.

[0034] The enhanced feature representation of each time slice is input into the predictor, which is a multilayer perceptron, and outputs the probability of adverse reaction risk at multiple predetermined time points after the time slice.

[0035] By connecting all time slices to predict the risk probability for the same future target time point, the risk evolution curve is formed.

[0036] Preferably, before inputting the time-ordered feature vector sequence into the encoder, the method further includes a feature vector representation generation step:

[0037] For symptom event features, their standard encoding, severity encoding, and duration encoding are concatenated, and then dimensionality is reduced through a fully connected layer to obtain the symptom event feature vector;

[0038] For medication record features, the drug code, dosage, and route of administration code are concatenated and then dimensionality-reduced through another fully connected layer to obtain the medication feature vector;

[0039] For testing numerical features, the index code, the actual value, and the deviation code from the normal range are concatenated and then dimensionality reduced through another fully connected layer to obtain the test feature vector.

[0040] The symptom event feature vector, medication feature vector, and test feature vector within the same time slice are concatenated to form a comprehensive feature vector representation of the time slice.

[0041] Preferably, the intervention instruction generation module matches a predefined intervention protocol library based on the time points and patterns when the risk evolution curve exceeds a preset threshold, including:

[0042] Continuously monitor the risk evolution curve and identify the point in time when the risk probability value first exceeds the dynamic warning threshold, which is recorded as the warning trigger point.

[0043] Extract a segment of the risk evolution curve that traces backward from the warning trigger point within a fixed time window, and the corresponding multi-dimensional time series data block subset within this fixed time window;

[0044] Morphological analysis is performed on the aforementioned risk evolution curve segments to identify the slope, frequency, and pattern of risk escalation, thereby forming risk characteristic labels;

[0045] Using the main active symptoms and drugs corresponding to the risk feature tags and warning trigger points as search keys, a search is performed in the intervention protocol library;

[0046] The intervention protocol library stores the mapping relationship between various risk patterns, symptom-drug combinations and standard intervention protocols, and returns several candidate intervention protocols with the highest matching degree.

[0047] Preferably, the methods for determining and adjusting the dynamic early warning threshold include:

[0048] Obtain historical adverse reaction risk data of the target patient population, and calculate the baseline distribution of the risk probability of the target patient population;

[0049] Based on the individual patient's age, underlying diseases, and previous chemotherapy history, the baseline distribution is shifted to obtain an individualized baseline.

[0050] Calculate the mean and variance of the risk evolution curve within a recent window in real time;

[0051] Combining the individualized baseline, recent mean, and variance, a warning threshold for the current moment is dynamically generated using a threshold calculation function;

[0052] After the system issues an early warning, the parameters of the threshold calculation function are fine-tuned based on whether adverse reactions occur and their severity.

[0053] Preferably, the intervention instruction generation module generates a personalized intervention instruction sequence that includes the intervention type, execution time window, and dosage adjustment recommendations, including:

[0054] Extract the core intervention type and default parameters from the candidate intervention plans with the highest matching degree;

[0055] The warning trigger point, the predicted peak of the risk evolution curve, and the patient's scheduled next medication or examination time are taken as key time nodes.

[0056] Based on the aforementioned key time nodes, a suggested implementation time window is planned for each intervention measure to ensure that there are no conflicts between measures and that the risk period is covered;

[0057] Based on the real-time liver and kidney function test values ​​of the target patients, the drug dosage in the core intervention measures is adjusted and calculated using a pharmacokinetic model to obtain individualized dosage recommendations;

[0058] The intervention type, recommended execution time window, and individualized dosage recommendations are arranged in chronological order and packaged into a personalized intervention instruction sequence that can be parsed by a clinical workstation.

[0059] Preferably, the steps for constructing the prediction model include:

[0060] Historical multi-dimensional time series data blocks are collected as training samples. Each training sample contains time series symptom events, medication records, test numerical feature vectors, and corresponding time labels.

[0061] The feature vectors in each training sample are standardized to eliminate differences in units, and the time labels are converted into relative time series.

[0062] Initialize the model parameters for the temporal attention mechanism, including the hidden layer dimension of the recurrent neural network, the number of attention heads, and the number of nodes in the fully connected layer;

[0063] The standardized feature vector sequence is input into the model for forward propagation, and the hidden state at each time step is calculated.

[0064] Attention weights are calculated at each time step using an attention mechanism, and historical information is fused to generate a context vector.

[0065] The context vector is concatenated with the hidden state at the current time step and input into the fully connected layer to predict the probability of adverse reaction risk.

[0066] The model parameters are adjusted based on the difference between the predicted results and the actual adverse reaction labels using the backpropagation algorithm and gradient descent optimizer.

[0067] Repeat the training process until the model loss function converges, and save the final model parameters for subsequent risk evolution curve calculation.

[0068] Compared with the prior art, the beneficial effects of the present invention are:

[0069] By extracting and fusing time-series symptom events generated from unstructured text with structured data, a multi-dimensional time-series data block centered on the individual patient was constructed. This enables a panoramic and continuous digital description of the patient's condition, providing a complete, unified, and high-quality data foundation for accurate analysis.

[0070] By integrating time-series data blocks, entity relationships from an external medical knowledge base are introduced to construct a dynamic association network between symptoms, medications, and physiological indicators. This network dynamically calculates the confidence level of potential adverse reaction patterns based on the node connection strength and event co-occurrence frequency in real-time data. This mechanism enables the system to automatically discover and quantify subtle association patterns not covered by the static knowledge base or unique to individual patients, realizing the evolution of risk inference from fixed rule matching to dynamic evidence verification.

[0071] The aforementioned data block is processed using a prediction model based on a time-series attention mechanism, directly outputting a continuous risk evolution curve. The attention mechanism enables the model to autonomously focus on the most relevant historical key segments, enhancing the depth of time-series dependency modeling. The risk evolution curve intuitively reveals the dynamic changing trend and potential peak points of risk over future timelines, upgrading the prediction output from static labels to an evolutionary trajectory that includes a time dimension.

[0072] Based on the specific timing and pattern characteristics of exceeding the threshold in the risk evolution curve, a predefined intervention protocol library is automatically matched to generate a personalized intervention instruction sequence containing specific measure types, execution time windows, and dosage adjustment parameters. This enables clinical intervention recommendations to accurately correspond to the specific content and development stage of the risk, achieving a direct transformation from risk warning to an executable intervention plan. Attached Figure Description

[0073] Figure 1 This is a schematic diagram illustrating the working principle of the big data-based chemotherapy adverse reaction prediction and intervention system described in this invention.

[0074] Figure 2 A flowchart for the data acquisition and preprocessing module;

[0075] Figure 3 A flowchart for constructing modules for dynamic relational networks;

[0076] Figure 4 A chart illustrating the individualized adjustment of dynamic early warning thresholds for adverse chemotherapy reactions;

[0077] Figure 5 This is a graph showing the correlation between creatinine clearance rate and the dosage and adjustment factor of chemotherapy intervention drugs. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] Please see Figure 1This invention provides a big data-based system for predicting and intervening in adverse reactions to chemotherapy. The system includes: a data acquisition and preprocessing module that first acquires initial patient records from a medical data warehouse. These records contain structured test values ​​and unstructured medical history text. This module performs sentence segmentation, medical terminology extraction, and normalization on the unstructured medical history text to generate time-series symptom events aligned with the structured test values ​​in time. A multi-dimensional time-series data fusion module receives these time-series symptom events and aligns and fuses them with the patient's medication records and vital sign monitoring streams within time windows, thereby constructing a multi-dimensional time-series data block centered on the individual patient.

[0080] The dynamic relational network construction module receives entity relationship information from an external medical knowledge base, injects it into the system, and establishes a dynamic relational network between symptoms, drugs, and physiological indicators based on the aforementioned multi-dimensional time-series data blocks. This module calculates the confidence level of potential adverse reaction patterns based on the connection strength of nodes and the co-occurrence frequency of events in the dynamic relational network. The risk prediction module uses a prediction model based on a time-series attention mechanism to calculate the multi-dimensional time-series data blocks, outputting an adverse reaction risk evolution curve characterizing future risk changes. The intervention instruction generation module continuously monitors the risk evolution curve. When the curve value exceeds a preset threshold, the module matches a predefined intervention protocol library based on the time point of exceedance and the risk pattern, generating a personalized intervention instruction sequence that includes the intervention type, execution time window, and dosage adjustment suggestions.

[0081] Example 1: See Figure 2 The data acquisition and preprocessing module generates time-series symptom events through specific steps. A sentence segmentation model combining rules and statistics is used to segment long-course texts into independent semantic segments based on punctuation, transition words, and medical terminology. Within each semantic segment, symptom descriptions, body parts, and severity modifiers are identified using a pre-built medical entity dictionary and a conditional random field model. The identified symptom descriptions are mapped to a standard medical terminology set, and body parts and severity modifiers are converted into standard codes, forming standardized symptom expressions. From the independent semantic segments, a time expression recognition model extracts relative or absolute time information, anchoring the standardized symptom expressions to specific time points. The standardized symptom expressions anchored to specific time points are then associated and stored with structured test values ​​within the same time window, according to patient identifiers and timestamps, forming time-series symptom events.

[0082] In practical implementation, the data acquisition and preprocessing module performs its function. An example of unstructured medical history text from an initial patient record in a medical data warehouse is as follows: "On October 26, 2023, the patient completed the first cycle of chemotherapy. On October 28, the patient complained of nausea and loss of appetite, which was slightly relieved after taking antiemetics. On October 30, a follow-up blood test showed a decrease in white blood cell count. On November 1, the patient developed a fever of 38.5℃, accompanied by a cough." Structured test values ​​include records such as "2023-10-30, white blood cell count, 2.1×10^9 / L." The sentence segmentation model, based on periods, dates, and transition words and medical terminology such as "complaints" and "accompanied," segments the long medical history text into independent semantic segments. For example, it segments "The patient complained of nausea and loss of appetite, which was slightly relieved after taking antiemetics" as an independent semantic segment.

[0083] In some embodiments, within independent semantic segments, symptom descriptions, body parts, and severity modifiers are identified using a pre-built medical entity dictionary and a conditional random field model. For example, in the phrase "the patient complains of nausea and loss of appetite," the symptom descriptions "nausea" and "loss of appetite" are identified; in the phrase "slight relief after taking oral antiemetics," the severity modifier "slight relief" is identified. The pre-built medical entity dictionary contains standard symptom entries such as "nausea," "vomiting," and "fever," and the conditional random field model uses contextual features to perform sequence annotation on non-fixed collocations such as "loss" and "relief."

[0084] In practice, identified symptom descriptions are mapped to a standard medical terminology set. Body parts and severity modifiers are converted into standard codes; for example, "nausea" is mapped to the standard medical term "nausea," "loss of appetite" to "anorexia," and the severity "slight relief" is converted into the code "MILD_RELIEF." These are combined to form standardized symptom expressions, such as "nausea" and "anorexia." The mapping between symptom descriptions and standard terms is achieved by calculating text similarity. The mapping matching score is calculated using the following formula:

[0085]

[0086] in: Represents the matching score. The character set representing candidate standard terms, The set of characters representing the symptom description to be mapped. Represents the size of the intersection. Represents the size of the union. and These are the vector representations of terms and descriptions obtained through word embedding models, respectively. Calculate cosine similarity. and These are character overlap weight and semantic similarity weight, respectively.

[0087] In some embodiments, relative or absolute time information is extracted from independent semantic fragments using a time expression recognition model, anchoring standardized symptom expressions to specific time points. The time expression recognition model, based on rules and neural networks, identifies absolute dates such as "October 28th" and "November 1st," and relative dates such as "3 days after chemotherapy," and converts them into absolute timestamps relative to the start date of chemotherapy. For example, "patient complains of nausea and loss of appetite" is anchored to the time point "2023-10-28."

[0088] Optionally, standardized symptom expressions anchored to specific time points are stored in association with structured test values ​​within the same time window, based on patient identifiers and timestamps. For example, the time point "2023-10-30" is associated with the standardized symptom expression "leukopenia" and the structured test value "white blood cell count: 2.1×10^9 / L". The time point "2023-11-01" is associated with the standardized symptom expressions "fever" and "cough". The resulting time-series symptom event data structure includes patient identifiers, timestamps, standardized symptom terms, severity codes, and is linked to test value records at the same timestamp via database foreign keys.

[0089] It is understandable that the data acquisition and preprocessing module processes unstructured medical history texts starting from the original text "The patient complains of nausea and loss of appetite." After sentence segmentation, independent semantic fragments are obtained. "Nausea" and "loss of appetite" are obtained through entity recognition. Standard terms "nausea" and "anorexia" are obtained through term mapping and normalization. The data is then anchored to "2023-10-28" through time extraction and finally associated with other data records from the same day.

[0090] In practice, the generated time-series symptom events are aligned with the original structured test values ​​in the time dimension, forming the basis of the patient's time-series data. The original medical record text, "A follow-up blood test on October 30th showed a decrease in white blood cell count," is processed to generate an event record with the timestamp "2023-10-30" and the standard symptom "leukopenia." This record is directly correlated with the white blood cell count test result of "2.1 × 10^9 / L" recorded in the laboratory information system at the same timestamp "2023-10-30."

[0091] Example 2: See Figure 3The system periodically downloads lists of adverse drug reactions, symptom correlations, and normal ranges of physiological parameters from an external medical knowledge base, transforming them into graph-structured data. This graph-structured data uses drugs, symptoms, and physiological indicators as nodes, and known causal, comorbid, and contraindication relationships as edges, assigning initial weights. Specific drugs, symptom events, and physiological indicators appearing in multi-dimensional time-series data blocks are used as activation nodes for matching queries within the graph-structured data. The weight coefficients of the edges between activation nodes are dynamically adjusted based on the actual time sequence and intervals of events in the multi-dimensional time-series data blocks. The subgraph formed by the activated nodes and their relational edges after weight adjustment is used as a dynamic association network. Each symptom node in the dynamic association network is traversed to find drug nodes and physiological indicator nodes directly connected to that symptom node, forming a group of associated nodes centered on that symptom node. The frequency of all nodes in the associated node group appearing together in the same chemotherapy cycle for the same patient within historical multi-dimensional time-series data blocks is statistically analyzed. By combining the current weight coefficients of the edges connecting each node within a group of related nodes in a dynamic association network, a weighted calculation model is used to obtain the initial confidence level of the adverse reaction pattern constituted by the group of related nodes. A time decay factor is introduced to assign lower weights to historical co-occurring events that occurred far in the past, thus correcting the initial confidence level. The corrected confidence level is then bound to the group of related nodes as the confidence level of the potential adverse reaction pattern.

[0092] In its implementation, the dynamic association network construction module receives entity relationship injections from an external medical knowledge base. This external knowledge base provides a list of adverse drug reactions, symptom associations, and normal ranges for physiological parameters. An example of knowledge entries periodically downloaded from the external medical knowledge base is as follows: the drug "docetaxel" is known to cause "neutropenia" and "peripheral neuropathy"; the symptoms "vomiting" and "nausea" are associated; and the lower limit of the normal value for the physiological indicator "absolute neutrophil count" is "1.5 × 10^9 / L". The dynamic association network construction module transforms this knowledge into graph-structured data, using specific drug names, standard symptom terms, and standard physiological indicator names as nodes. Edges are established between nodes based on the knowledge entries, indicating "causing," "associating with," or "below range" relationships. Each edge is assigned an initial weight representing the strength of the relationship; for example, the initial weight of the edge "docetaxel -> causes -> neutropenia" is set to 0.8.

[0093] In some embodiments, the multidimensional time-series data block constructed by the multidimensional time-series data fusion module contains specific events extracted from patient records. Taking a patient receiving chemotherapy with "docetaxel" combined with "carboplatin" as an example, the specific drug nodes appearing in the multidimensional time-series data block include "docetaxel" and "carboplatin," the specific symptom event nodes include "vomiting" and "neutropenia," and the specific physiological indicator nodes include the measurement value of "absolute neutrophil count" as "0.8 × 10^9 / L." The dynamic association network construction module uses "docetaxel," "carboplatin," "vomiting," "neutropenia," and "absolute neutrophil count" as activation nodes to perform matching queries in the graph structure data downloaded and transformed from an external medical knowledge base. The query activates the corresponding nodes in the graph structure data and the edges connecting these nodes, such as the "docetaxel-neutropenia" edge, the "carboplatin-vomiting" edge, and the "neutropenia-neutropenia-absolute neutrophil count below normal range" edge.

[0094] In practice, the weight coefficients of the edges between activated nodes are dynamically adjusted based on the actual time sequence and intervals of events in the multi-dimensional time-series data block. For example, in the multi-dimensional time-series data block, records show that a patient developed neutropenia on day 7 after using docetaxel, while vomiting occurred on day 1 after carboplatin administration. The dynamic association network construction module, based on the fact that the docetaxel administration time and the neutropenia event occurred close together, might strengthen the weight coefficient of the "docetaxel-neutropenia" edge, for example, increasing it from the initial 0.8 to 0.9. Simultaneously, because the vomiting and neutropenia events occur at relatively different times and have no direct pharmacological correlation, the system will not automatically create or strengthen the edge between vomiting and neutropenia.

[0095] Optionally, when calculating the confidence level of potential adverse reaction patterns, the dynamic association network construction module traverses each symptom node in the dynamic association network. For example, when traversing to the symptom node "neutropenia," it searches for drug nodes "docetaxel" and "carboplatin," and physiological indicator nodes "absolute neutrophil count," which are directly connected to the "neutropenia" node in the dynamic association network, forming an association node group {docetaxel, carboplatin, absolute neutrophil count} with "neutropenia" as the core. The frequency of all nodes in the association node group {docetaxel, carboplatin, absolute neutrophil count} co-occurring in the same chemotherapy cycle for the same patient is then statistically analyzed in historical multi-dimensional time-series data blocks. For example, assuming that among 100 patients who used the "docetaxel" and "carboplatin" regimens, 85 patients had records of "docetaxel," "carboplatin," and "absolute neutrophil count" being lower than normal in their time-series data co-occurring in the same chemotherapy cycle, then the co-occurrence frequency of the association node group is 0.85.

[0096] It can be understood that, by combining the current weight coefficients of the edges connecting each node within a group of related nodes in a dynamic network, a weighted calculation model can be used to obtain the initial confidence level of the adverse reaction pattern constituted by the group of related nodes. The weighted calculation model can be expressed as:

[0097]

[0098] in: Represents the initial confidence level of the associated node group i. This represents the set of edges that exist between all nodes in the associated node group i. This represents the current dynamic weight coefficient of edge j. The initial weight coefficients of edge k are represented. This represents the co-occurrence frequency of the associated node group i in historical data. For the associated node group {docetaxel, carboplatin, absolute neutrophil count}, assuming its edge set includes "docetaxel - neutropenia" with a current weight of 0.9 and an initial weight of 0.8, "carboplatin - neutropenia" with a current weight of 0.7 and an initial weight of 0.7, and "neutropenia - absolute neutrophil count below range" with a current weight of 0.95 and an initial weight of 0.9, and a historical co-occurrence frequency F of 0.85, then the initial confidence level is... Calculated as .

[0099] In practice, a time decay factor is introduced to assign lower weights to historical co-occurring events that occurred a long time ago, thus correcting the initial confidence level. The time decay factor ensures that historical co-occurring events occurring a year ago contribute less to the confidence calculation than events occurring a week ago. The corrected confidence level is then linked to associated node groups, serving as the confidence level for the potential adverse reaction pattern of "docetaxel and carboplatin combined chemotherapy causing neutropenia accompanied by a decrease in absolute neutrophil count." The dynamic association network construction module performs this process on all identified symptom nodes, calculating a confidence value for each symptom-centric associated node group, quantifying its likelihood as a potential adverse reaction pattern.

[0100] Example 3: Multi-dimensional time-series data blocks are sliced ​​at fixed time intervals. Each time slice contains feature vector representations of all symptom events, medication records, and test values ​​within the corresponding time period. For symptom event features, their standard codes, severity codes, and duration codes are concatenated and dimensionality reduced using a fully connected layer to obtain the symptom event feature vector. For medication record features, drug codes, dosage codes, and administration route codes are concatenated and dimensionality reduced using another fully connected layer to obtain the medication feature vector. For test value features, indicator codes, actual values, and deviations from the normal range codes are concatenated and dimensionality reduced using yet another fully connected layer to obtain the test feature vector. The symptom event feature vector, medication feature vector, and test feature vector within the same time slice are concatenated to form a comprehensive feature vector representation for that time slice. The chronologically ordered sequence of feature vectors is input into an encoder, which consists of a stacked recurrent neural network layer and a temporal self-attention layer. The temporal self-attention layer calculates the correlation weights between the feature vector of each time slice and the feature vectors of all time slices in the sequence, thereby generating a context vector containing global temporal dependencies. The encoder ultimately outputs an enhanced feature representation for each time slice, which integrates current information with contextual information from historical sequences. This enhanced feature representation is then input into a predictor, a multilayer perceptron, which outputs the probability of adverse reactions at multiple predetermined future time points after that time slice. All time slices are concatenated to predict the risk probability at the same target future time point, forming a risk evolution curve. The construction steps of the prediction model include: collecting historical multi-dimensional time-series data blocks as training samples; each training sample containing time-series symptom events, medication records, and test numerical feature vectors, along with corresponding time labels; standardizing the feature vectors in each training sample to eliminate dimensional differences and converting the time labels into relative time series; initializing the model parameters of the temporal attention mechanism, including the hidden layer dimension of the recurrent neural network, the number of attention heads, and the number of nodes in the fully connected layer; inputting the standardized feature vector sequence into the model for forward propagation to calculate the hidden state at each time step; calculating the attention weights at each time step using the attention mechanism and fusing historical information to generate a context vector; concatenating the context vector with the hidden state of the current time step and inputting it into the fully connected layer for adverse reaction risk probability prediction. The model parameters are adjusted based on the difference between the predicted results and the actual adverse reaction labels using the backpropagation algorithm and gradient descent optimizer. The training process is repeated until the model loss function converges, and the final model parameters are saved for subsequent risk evolution curve calculations.

[0101] In practical implementation, the risk prediction module processes the multidimensional time-series data blocks output by the multidimensional time-series data fusion module. These blocks contain symptom events, medication records, and test values ​​recorded by the patient in chronological order. The multidimensional time-series data blocks are sliced ​​at fixed time intervals, such as one day. Each time slice contains feature vector representations of all symptom events, medication records, and test values ​​occurring within that day. For a specific time slice, the generation of its feature vector representation follows these steps: For the symptom event "nausea" recorded within that time slice, its standard code "NAUSEA", severity code "MODERATE", and duration code "PERSISTENT" are concatenated, and dimensionality reduction is achieved through a fully connected layer to obtain the symptom event feature vector. For the medication record "docetaxel 100mg for injection" within the same time slice, the drug code "DOCETAXEL", dosage "100", and route of administration code "INTRAVENOUS" are concatenated, and dimensionality reduction is achieved through another fully connected layer to obtain the medication feature vector. For the test value "absolute neutrophil count 1.2 × 10^9 / L" within the same time slice, the indicator code "NEUTROPHIL_ABS", the actual value "1.2", and the deviation from the normal range code "BELOW_NORMAL" are concatenated. This concatenation is then performed through another fully connected layer to reduce the dimensionality, resulting in the test feature vector. The symptom event feature vector, medication feature vector, and test feature vector within the same time slice are then concatenated to form a comprehensive feature vector representation for that time slice.

[0102] In some embodiments, a sequence of composite feature vectors arranged chronologically is input to an encoder, which consists of a recurrent neural network layer and a temporal self-attention layer stacked together. The recurrent neural network layer processes the composite feature vector for each time slice sequentially and outputs the hidden state for each time slice. The temporal self-attention layer receives the sequence of hidden states for all time slices and calculates the correlation weight between the hidden state of each time slice and the hidden states of all time slices in the sequence. The correlation weight indicates the importance of historical and future time point information to the current time point. The temporal self-attention layer performs a weighted summation of all hidden states based on the calculated correlation weights, thereby generating a context vector containing global temporal dependencies. The encoder finally outputs an enhanced feature representation for each time slice, which is formed by fusing the original hidden state of the current time slice with the corresponding context vector. The fusion operation can be such as concatenation or addition. The enhanced feature representation integrates current information and context information from the historical sequence.

[0103] In practice, the enhanced feature representation of each time slice is input into the predictor, which is a multilayer perceptron. The predictor outputs the probability of adverse reactions at multiple predetermined time points after this time slice. The predetermined time points can be set as the 1st, 3rd, and 7th day in the future. The output layer of the predictor uses a sigmoid activation function to map the network output to a risk probability value between 0 and 1. Connecting the risk probability predictions of all time slices for the same target time point in the future forms a risk evolution curve. For example, connecting the risk probability predicted by the 1st time slice for the 3rd day in the future, the 2nd time slice for the 3rd day in the future, and the 3rd time slice for the 3rd day in the future forms a risk evolution curve describing the change of the risk probability of the 3rd day in the future over time.

[0104] Optionally, the construction of the prediction model requires collecting historical multi-dimensional time-series data blocks as training samples. Each training sample contains time-series symptom events, medication records, and test numerical feature vectors within a complete chemotherapy cycle, as well as the true label of whether an adverse reaction occurred at each time point. The feature vectors in each training sample are standardized to eliminate dimensional differences, for example, by normalizing all numerical features to the [0,1] interval and converting the time labels into relative time series with respect to the start date of chemotherapy. The model parameters of the time-series attention mechanism are initialized, including setting the hidden layer dimension of the recurrent neural network to 128, the number of attention heads to 4, and the number of nodes in the fully connected layer to [64,32].

[0105] Understandably, during training, the standardized feature vector sequence is input into the model for forward propagation, calculating the hidden state at each time step. Attention weights are calculated at each time step using an attention mechanism, and historical information is fused to generate a context vector. This context vector is then concatenated with the hidden state at the current time step and input into a fully connected layer for adverse reaction risk probability prediction. Using backpropagation and a gradient descent optimizer, the model parameters are adjusted based on the difference between the predicted results and the actual adverse reaction labels; this difference is calculated using a binary cross-entropy loss function. The formula for updating the model parameters is:

[0106]

[0107] in: Represents the updated model parameters. This represents the model parameters before the update. Represents the learning rate. Represents the loss function Regarding parameters The gradient is calculated. The training process is repeated until the model loss function converges, and the final model parameters are saved for subsequent risk evolution curve calculation.

[0108] In practice, for a given input multidimensional time-series data block, the model ultimately outputs multiple risk evolution curves, each corresponding to a future prediction time point. The risk evolution curve depicts the dynamic assessment of the probability of adverse reactions occurring at a future time point from the current moment to that point, with the probability value continuously updated as new data is input and time progresses.

[0109] Example 4: Continuously monitor the risk evolution curve and identify the point in time when the risk probability value first exceeds the dynamic warning threshold, marking it as the warning trigger point. Extract a segment of the risk evolution curve within a fixed time window tracing back from the warning trigger point, along with the corresponding multi-dimensional time-series data subset within this fixed time window. Perform morphological analysis on the risk evolution curve segment to identify the slope, fluctuation frequency, and pattern of risk increase, forming risk feature labels. Use the risk feature labels and the main active symptoms and medications corresponding to the warning trigger point as query keys to search the intervention protocol library. The intervention protocol library stores the mapping relationship between various risk patterns, symptom-drug combinations, and standard intervention protocols, returning several candidate intervention protocols with the highest matching degree. The method for determining and adjusting the dynamic warning threshold includes: obtaining historical adverse reaction risk data of the target patient population and calculating the baseline distribution of the risk probability of the target patient population. Based on the individual target patient's age, underlying diseases, and previous chemotherapy history, shift the baseline distribution to obtain an individualized baseline. Calculate the mean and variance of the risk evolution curve within a recent window in real time. By combining an individualized baseline, recent mean, and variance, a threshold calculation function is used to dynamically generate the warning threshold for the current moment. When the system issues a warning, the parameters of the threshold calculation function are fine-tuned based on whether adverse reactions occur subsequently and their severity.

[0110] In practice, the intervention instruction generation module continuously monitors the adverse reaction risk evolution curve output by the risk prediction module. The risk evolution curve is plotted with time on the horizontal axis and risk probability value on the vertical axis. The intervention instruction generation module identifies the point in time when the risk probability value first exceeds the dynamic warning threshold and records this point as the warning trigger point. The warning trigger point is identified by comparing the instantaneous value of the risk evolution curve with the value of the dynamic warning threshold in real time. When the instantaneous value crosses the dynamic warning threshold from below, the current time point is recorded as the warning trigger point. A segment of the risk evolution curve within a fixed time window (e.g., 72 hours) is extracted from the warning trigger point. Simultaneously, a subset of multi-dimensional time-series data blocks corresponding to this 72-hour time window is extracted. This subset of multi-dimensional time-series data blocks includes all symptom events, medication records, and test values ​​within this period.

[0111] In some embodiments, morphological analysis is performed on the extracted risk evolution curve segments. This analysis includes calculating the average slope of the curve segments to identify the rate of risk increase and calculating the frequency of curve fluctuations to identify instability. In a specific implementation, the frequency of curve fluctuations is calculated by analyzing risk evolution curve segments extracted within a fixed time window tracing back from the warning trigger point. These curve segments consist of continuous risk probability values ​​output by the risk prediction module. The system identifies continuous change patterns in the risk probability values ​​within the curve segments, detecting turning points where the value direction changes, such as the instantaneous position from rising to falling or from falling to rising. The number of occurrences of these turning points within the fixed time window is counted, and the number is divided by the time window length to obtain the fluctuation frequency value per unit time. The curve pattern is identified as belonging to the categories of "continuous rise," "rapid surge and fall," or "stepwise rise." The results of the morphological analysis are quantified into one or more risk characteristic labels, such as "slope greater than 0.1 / day," "high fluctuation frequency," and "continuous rise pattern." A search is performed in a predefined intervention protocol library using the risk characteristic labels and the main active symptoms and main medications corresponding to the warning trigger point as query keys. The main active symptoms are determined by analyzing the number and severity of symptom events within a time window near the warning trigger point, and the main medications used are determined by analyzing medication records within the same time window.

[0112] In practical implementation, referring to Table 1, a predefined intervention protocol library stores the mapping relationships between various risk patterns, symptom-drug combinations, and standard intervention protocols. An example of the intervention protocol library's structure is shown in the table below. The intervention protocol library returns several candidate intervention protocols with the highest match degree to the query key. The match degree calculation comprehensively compares the similarity of risk feature tags, the overlap of symptoms, and the consistency of medications.

[0113] Table 1: List of Entries in the Intervention Protocol Library

[0114] Risk pattern characteristics Main symptoms Involving pharmaceuticals Standard intervention protocol The slope is >0.1 / day and continues to rise. Neutropenia, fever Dorcetase, Carbopol Option A: White blood cell count boosting therapy, prophylactic anti-infection treatment. Rapid rise followed by fall, high volatility Nausea and vomiting Cisplatin Option B: Strengthen antiemetics and provide fluid support. Slope > 0.05 / day, step-like increase Fatigue, abnormal liver function Methotrexate Option C: Liver protection treatment, dosage adjustment

[0115] Optionally, the method for determining and adjusting the dynamic early warning threshold includes the following steps: Obtain historical adverse reaction risk data for the target patient population. The target patient population can be defined based on cancer type, chemotherapy regimen, age group, etc., to calculate the baseline distribution of risk probability for the target patient population. The baseline distribution can be the mean and standard deviation. Based on the individual target patient's age, underlying diseases, and previous chemotherapy history, shift the baseline distribution. This shift is achieved through a personalized adjustment factor, resulting in an individualized baseline. For example, an elderly patient with a history of severe liver disease may have an individualized baseline that is higher than the baseline for their population.

[0116] This can be understood as calculating the mean and variance of the risk evolution curve within a recent window, such as the last 24 hours. Combining an individualized baseline, the recent mean and variance, a threshold calculation function dynamically generates the warning threshold for the current moment. An example of the threshold calculation function is as follows:

[0117]

[0118] in: Represents the dynamic early warning threshold. Represents an individualized baseline. This represents the mean of the risk evolution curve within the recent window. The standard deviation represents the recent window of the risk evolution curve. and This is the adjustment coefficient. When the system issues an early warning, the adjustment coefficient in the threshold calculation function is adjusted based on whether adverse reactions actually occur and their severity. and Feedback-based fine-tuning is implemented: if a false alarm occurs, the coefficient is adjusted upwards appropriately; if a false alarm occurs, the coefficient is adjusted downwards appropriately.

[0119] In practice, after the matching process is completed, the intervention instruction generation module obtains one or more candidate intervention plans. The candidate intervention plan with the highest matching degree will be used in subsequent intervention instruction generation steps. The risk evolution curve segment, the extracted risk feature labels, the matched intervention plan, and the dynamic early warning threshold value at the time of triggering are all recorded in the system log for auditing and analysis.

[0120] See Figure 4In the personalized adjustment analysis of dynamic warning thresholds, the differences in the distribution of dynamic warning thresholds (risk probabilities) corresponding to individualized baselines and multiple adjustment coefficients (α, β) under different patient population categories are presented intuitively. Specifically, patient populations are divided into gradients of "young people with no underlying diseases", "young and middle-aged people with underlying diseases", "elderly people with moderate underlying diseases", and "elderly people with severe underlying diseases". The horizontal axis represents the population category, and the vertical axis represents the risk probability corresponding to the dynamic warning threshold. The black curve in the figure represents the individualized baseline. As patients age and the severity of underlying diseases increases, the baseline risk probability rises linearly (from 0.10 for young people without underlying diseases to 0.26 for elderly people with severe underlying diseases). The figure also shows the threshold curves corresponding to three sets of adjustment coefficients (α=0.6, β=0.2; α=0.8, β=0.3; α=1.0, β=0.4): as α and β increase, the warning threshold for the same population increases significantly, and all threshold curves follow the trend of the individualized baseline. The warning threshold for the elderly with severe underlying diseases (e.g., 0.39 when α=1.0, β=0.4) is much higher than that for the young people without underlying diseases (e.g., only 0.17 when α=0.6, β=0.2). The visualization results demonstrate the design logic of "individualized baseline + adjustment coefficient" in the threshold calculation function: by adjusting the α and β parameters and combining them with differences in the baseline risk of the population, personalized adaptation of the warning threshold is achieved, providing differentiated judgment criteria for risk warnings for patients with different characteristics.

[0121] Example 5: The process of generating a personalized intervention instruction sequence using the intervention instruction generation module includes the following specific steps. From the candidate intervention plans with the highest matching degree, the core intervention measure type and default parameters are extracted. The warning trigger point, the predicted peak time of the risk evolution curve, and the patient's scheduled next medication or examination time are designated as key time nodes. Based on these key time nodes, a suggested execution time window for each intervention measure is planned to ensure no conflicts between measures and coverage of the risk period. Based on the target patient's real-time liver and kidney function test values, the drug dosage in the core intervention measures is adjusted and calculated using a pharmacokinetic model to obtain individualized dosage recommendations. The intervention measure type, suggested execution time window, and individualized dosage recommendations are arranged in chronological order and encapsulated into a personalized intervention instruction sequence that can be parsed by the clinical workstation.

[0122] In practical implementation, the intervention instruction generation module generates a personalized intervention instruction sequence based on the matched candidate intervention protocols. From the candidate intervention protocols with the highest matching degree, the core intervention type and default parameters are extracted. For example, from "Program A: White blood cell boosting therapy, prophylactic anti-infection," the core interventions "subcutaneous injection of recombinant human granulocyte colony-stimulating factor" and "oral levofloxacin" are extracted, along with the default parameters "Recombinant human granulocyte colony-stimulating factor dosage: 5 μg / kg body weight" and "Levofloxacin dosage: 500 mg once daily." The warning trigger point, the predicted peak time of the risk evolution curve, and the patient's scheduled next medication or examination time are used as key time nodes. For example, the warning trigger point is identified as "2023-11-05 14:30," the risk evolution curve predicts that the risk of neutropenia will peak on "2023-11-07," and the patient's scheduled next chemotherapy treatment time is "2023-11-10."

[0123] In some embodiments, a recommended implementation time window for each intervention is planned based on key time nodes. For "subcutaneous injection of recombinant human granulocyte colony-stimulating factor," the recommended implementation time window may be planned to begin immediately after the warning trigger point, i.e., "2023-11-05 15:00," and continue until after the risk prediction peak, such as "2023-11-08." For "oral levofloxacin," the recommended implementation time window may be planned to begin on the evening of "2023-11-05" and continue until "2023-11-09."

[0124] In practice, based on the target patient's real-time liver and kidney function test values, a pharmacokinetic model is used to adjust the drug dosage in the core intervention, resulting in individualized dosage recommendations. Assuming the target patient's latest serum creatinine clearance is 45 mL / min, indicating moderate renal insufficiency, the pharmacokinetic model adjusts the default dosage of the intervention drug "levofloxacin" based on its clearance rate and renal function. The pharmacokinetic model uses the following formula to calculate the individualized dosage adjustment coefficient:

[0125]

[0126] in: Representative dose adjustment factor, Drug-specific parameters representing the dependence of drug clearance on renal function. This represents the patient's current measured creatinine clearance rate. This represents a normal creatinine clearance reference value. According to... The calculated value is adjusted to the default dose to obtain an individualized dose recommendation.

[0127] Optionally, the intervention type, recommended execution time window, and individualized dosage recommendation can be arranged chronologically and packaged into a personalized intervention instruction sequence that can be parsed by the clinical workstation. The packaging format can be a structured data object or a standard clinical message format. For example, the generated personalized intervention instruction sequence contains two instructions ordered by time: the first instruction is {Intervention type: subcutaneous injection of recombinant human granulocyte colony-stimulating factor, execution time window: 15:00 on 2023-11-05 to 08:00 on 2023-11-08, recommended dosage: 5 μg / kg body weight, once daily}; the second instruction is {Intervention type: oral levofloxacin, execution time window: 20:00 on 2023-11-05 to 20:00 on 2023-11-09, recommended dosage: 250 mg, once daily}.

[0128] Understandably, the final output of a personalized intervention instruction sequence is an executable plan containing specific operations, clear timing, and precise dosage. This personalized intervention instruction sequence is transmitted to the clinician's workstation or the nurse's mobile terminal for review and execution by medical personnel. The intervention instruction generation module runs continuously throughout the chemotherapy cycle. When a new warning trigger point appears on the risk evolution curve, it will trigger the matching and generation process again, outputting an updated personalized intervention instruction sequence to achieve dynamic intervention management of the patient's adverse reaction risk.

[0129] See Figure 5During the intervention instruction generation phase, the correlation between creatinine clearance rate ranges and levofloxacin dose, colony-stimulating factor adjustment coefficient, and fluid replacement volume adjustment coefficient is quantified using this graph. Specifically, the horizontal axis is divided based on creatinine clearance rate ranges (mL / min), covering five gradients: ≥80, 60-79, 40-59, 20-39, and <20. The left vertical axis corresponds to the levofloxacin dose (mg), and the right vertical axis corresponds to the adjustment coefficient. When creatinine clearance is ≥80%, levofloxacin is maintained at an initial dose of 500 mg, with both the colony-stimulating factor (CSF) adjustment coefficient and the fluid resuscitation adjustment coefficient remaining at a baseline level of 1.0. As creatinine clearance decreases to the 60-79 range, the levofloxacin dose remains at 500 mg, but the fluid resuscitation adjustment coefficient begins to increase. When creatinine clearance enters the 40-59 range (moderate renal insufficiency), the levofloxacin dose is reduced to 250 mg, the CSF adjustment coefficient drops below 1.0, and the fluid resuscitation adjustment coefficient further increases. In the 20-39 range, the levofloxacin dose is reduced to approximately 125 mg, the CSF adjustment coefficient is approximately 0.7, and the fluid resuscitation adjustment coefficient approaches 1.6. When creatinine clearance is <20%, the levofloxacin dose is reduced to approximately 75 mg, the CSF adjustment coefficient drops to 0.6, and the fluid resuscitation adjustment coefficient reaches a peak of 1.8. By observing the interconnected changes of multiple indicators, the influence of renal function level on the dosage of adjuvant chemotherapy drugs and the adjustment coefficient of intervention measures is presented intuitively, providing a visualized parameter correlation basis for individualized dosage calculation in the intervention instruction generation module.

[0130] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0131] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A big data based chemotherapy adverse reaction prediction and intervention system, characterized in that, The method comprises the following modules: a data acquisition and preprocessing module, configured to acquire an initial treatment patient record from a medical data warehouse, the initial treatment patient record comprising structured test values and unstructured medical history text, and to perform sentence segmentation, medical term extraction and normalization processing on the unstructured medical history text to generate time series symptom events aligned with the structured test values; a multi-dimensional time series data fusion module, configured to perform time window alignment and fusion of the time series symptom events with patient medication records and vital sign monitoring data to construct a multi-dimensional time series data block centered on an individual patient; a dynamic association network construction module, configured to receive entity relationship injection from an external medical knowledge base, to construct a dynamic association network among symptoms, drugs and physiological indicators on the multi-dimensional time series data block, and to calculate confidence of a potential adverse reaction pattern according to connection strength of nodes in the dynamic association network and event co-occurrence frequency; a risk prediction module, configured to calculate the multi-dimensional time series data block by using a prediction model based on a time series attention mechanism, and to output a risk evolution curve of an adverse reaction; an intervention instruction generation module, configured to generate a personalized intervention instruction sequence comprising intervention measure type, execution time window and dosage adjustment suggestion according to a time point and a pattern at which the risk evolution curve exceeds a preset threshold, and to match a predefined intervention protocol library.

2. The big data based chemotherapy adverse reaction prediction and intervention system according to claim 1, characterized in that, The data acquisition and preprocessing module performs sentence segmentation, medical term extraction and normalization processing on the unstructured medical history text to generate time series symptom events aligned with the structured test values, comprising: a sentence segmentation model combining rules and statistics is used to divide long medical history text into independent semantic segments according to punctuation marks, transition words and medical term features; in the independent semantic segments, a pre-constructed medical entity dictionary and a conditional random field model are used to identify symptom descriptions, body parts and severity modifiers; the identified symptom descriptions are mapped to a standard medical term set, and the body parts and severity modifiers are converted into standard codes to form standardized symptom expressions; from the independent semantic segments, relative time or absolute time information is extracted by using a time expression recognition model, and the standardized symptom expressions are anchored to specific time points; the standardized symptom expressions anchored to specific time points are stored in association with the structured test values in the same time window according to patient identification and time stamp to form the time series symptom events.

3. The big data based chemotherapy adverse reaction prediction and intervention system according to claim 2, characterized in that, The dynamic association network construction module receives entity relationship injection from an external medical knowledge base to construct a dynamic association network among symptoms, drugs and physiological indicators on the multi-dimensional time series data block, comprising: drug adverse reaction lists, symptom association relationships and physiological parameter normal range knowledge are periodically downloaded from the external medical knowledge base and converted into graph structure data; the graph structure data takes drugs, symptoms and physiological indicators as nodes, and known causal, concomitant and contraindicated relationships as edges, and is assigned with initial weights; specific drugs, symptom events and physiological indicators appearing in the multi-dimensional time series data block are taken as active nodes to perform matching queries in the graph structure data; According to the real-time sequence and interval of the event occurrence in the multi-dimensional time-series data block, the weight coefficient of the edge between the activated nodes is dynamically adjusted; The subgraph composed of the activated nodes and the relationship edges after adjusting the weight coefficient is taken as the dynamic correlation network.

4. The big data based chemotherapy adverse reaction prediction and intervention system according to claim 3, characterized in that, According to the connection strength and event co-occurrence frequency of the nodes in the dynamic correlation network, the confidence of the potential adverse reaction mode is calculated, including: Traverse each symptom node in the dynamic correlation network, find the drug nodes and physiological indicator nodes directly connected to the symptom node, and form a group of correlation nodes with the symptom node as the core; Statistically, in the historical multi-dimensional time-series data block, the frequency of the co-occurrence of all nodes in the group of correlation nodes in the same patient in the same chemotherapy cycle is calculated; Combined with the current weight coefficient of the edge connecting each node in the group of correlation nodes in the dynamic correlation network, the initial confidence of the adverse reaction mode composed of the group of correlation nodes is obtained through a weighted calculation model; Introducing a time decay factor, giving a lower weight to the historical co-occurrence events that occur far away from the current time, and correcting the initial confidence; The corrected confidence is bound to the group of correlation nodes as the confidence of the potential adverse reaction mode.

5. The big data based chemotherapy adverse reaction prediction and intervention system according to claim 1, characterized in that, The risk prediction module adopts a prediction model based on a time-series attention mechanism to calculate the multi-dimensional time-series data block, and outputs the risk evolution curve of adverse reactions, including: The multi-dimensional time-series data block is sliced according to fixed time intervals, and each time slice contains the feature vector representation of all symptom events, medication records, and test values in the corresponding period; The sequence of feature vectors arranged in time sequence is input into an encoder, and the encoder is composed of a recurrent neural network layer and a time self-attention layer. The time self-attention layer calculates the correlation weight of the feature vector of each time slice with all time slice feature vectors in the sequence, and generates a context vector containing global time-series dependencies accordingly; The encoder finally outputs the enhanced feature representation of each time slice, which integrates the current information and the context information of the historical sequence; The enhanced feature representation of each time slice is input into a predictor, which is a multi-layer perceptron, to output the adverse reaction risk probability at multiple predetermined future time points after the time slice; Connecting the risk probability predictions of all time slices for the same target time point forms the risk evolution curve.

6. The big data based chemotherapy adverse reaction prediction and intervention system according to claim 5, characterized in that, Before inputting the sequence of feature vectors arranged in time sequence into the encoder, it also includes the generation step of feature vector representation: For symptom event features, concatenate their standard encoding, severity encoding, and duration encoding, and reduce the dimension through a fully connected layer to obtain a symptom event feature vector; For medication record features, concatenate drug encoding, dose, and administration route encoding, and reduce the dimension through another fully connected layer to obtain a medication feature vector; For test value features, concatenate index encoding, actual value, and deviation degree encoding relative to the normal value range, and reduce the dimension through another fully connected layer to obtain a test feature vector; The symptom event feature vector, the medication feature vector, and the test feature vector in the same time slice are spliced to form a comprehensive feature vector representation of the time slice.

7. The big data based chemotherapy adverse reaction prediction and intervention system according to claim 1, characterized in that, The intervention instruction generation module matches a pre-defined intervention protocol library according to the time point and mode at which the risk evolution curve exceeds the preset threshold, including: The risk evolution curve is continuously monitored to identify a time point at which a risk probability value first exceeds a dynamic early warning threshold, which is recorded as an early warning trigger point; A risk evolution curve segment within a fixed time window backtracking from the early warning trigger point is extracted, as well as a corresponding multi-dimensional time series data block subset within the fixed time window; The risk evolution curve segment is subjected to morphological analysis to identify the slope, fluctuation frequency and mode of risk increase, forming a risk feature label; The risk feature label and the main active symptoms and drugs corresponding to the early warning trigger point are used as query keys to search the intervention protocol library; The intervention protocol library stores the mapping relationship between various risk modes, symptom-drug combinations and standard intervention schemes, and returns a number of candidate intervention schemes with the highest matching degree.

8. The big data based chemotherapy adverse reaction prediction and intervention system according to claim 7, characterized in that, The determination and adjustment method of the dynamic early warning threshold includes: Historical adverse reaction risk data of the target patient's population is obtained, and the baseline distribution of the risk probability of the target patient's population is calculated; The baseline distribution is shifted according to the age, underlying disease, and past history of chemotherapy of the target patient to obtain an individualized baseline; The mean and variance of the risk evolution curve within a recent window are calculated in real time; The early warning threshold at the current time is dynamically generated by a threshold calculation function in combination with the individualized baseline, recent mean and variance; When the system issues an early warning, the parameters of the threshold calculation function are adjusted in a feedback manner according to whether an adverse reaction actually occurs and its severity.

9. The big data based chemotherapy adverse reaction prediction and intervention system according to claim 7, characterized in that, The intervention instruction generation module generates a personalized intervention instruction sequence including intervention measure types, execution time windows and dose adjustment suggestions, including: The core intervention measure type and default parameters are extracted from the candidate intervention scheme with the highest matching degree; The early warning trigger point, the time point of the predicted peak of the risk evolution curve, and the patient's next scheduled medication or examination time point are used as key time nodes; The suggested execution time window of each intervention measure is planned according to the key time nodes to ensure that the measures do not conflict and cover the risk period; The individualized dose suggestion is obtained by adjusting and calculating the drug dose in the core intervention measure using a pharmacokinetic model according to the target patient's real-time liver and kidney function test values; The intervention measure types, suggested execution time windows, and individualized dose suggestions are arranged in chronological order and encapsulated as the personalized intervention instruction sequence that can be parsed by a clinical workstation. 10.The construction method of a big data-based chemotherapy adverse reaction prediction and intervention system according to claim 1, characterized in that, The construction steps of the prediction model include: Historical multi-dimensional time series data blocks are collected as training samples, each training sample including time series symptom event, medication record and test value feature vectors, and corresponding time labels; The feature vectors in each training sample are standardized to eliminate dimensional differences, and the time labels are converted into relative time series; Initialize the model parameters of the temporal attention mechanism, including the hidden layer dimensions of the recurrent neural network, the number of attention heads, and the node count of the fully connected layer. Input the normalized feature vector sequence into the model for forward propagation, and calculate the hidden state at each time step. Calculate the attention weight at each time step through the attention mechanism, and fuse the historical information to generate a context vector. Concatenate the context vector with the hidden state at the current time step, and input it into the fully connected layer for adverse reaction risk probability prediction. Use the backpropagation algorithm and gradient descent optimizer to adjust the model parameters based on the difference between the predicted results and the true adverse reaction labels. Repeat the training process until the model loss function converges, and save the final model parameters for subsequent risk evolution curve calculation.

Images