A method for predicting the risk of complications of a chemotherapy infusion port

Abstract

The present application relates to the field of intelligent identification of chemotherapy infusion port complication risk, aiming to solve the problems of multi-source heterogeneous medical data correlation difficulty and risk evolution judgment difficulty. The core scheme is as follows: firstly, the structured fields and nursing texts of electronic medical records are collected and standardized, key risk events are automatically identified and labeled based on evidence-based medicine rules, context units with true description are generated through alignment and fusion of multi-modal time series data, same event representations are normalized, risk anchor point sequence graphs reflecting causal and time sequence relationships are constructed, dominant information is screened by using graph neural network dynamic pruning mechanism, and finally the risk level and key causal link of chemotherapy infusion port complication are output, effectively improving the accuracy and explainability of clinical early warning.

Description

Technical Field

[0001] This invention relates to the field of medical and health information processing and clinical risk prediction technology, and in particular to a method for predicting the risk of complications from chemotherapy infusion ports. Background Technology

[0002] With the widespread use of chemotherapy in cancer treatment, chemotherapy infusion ports are widely used in patients receiving long-term intravenous medication. However, serious complications such as catheter-related infections, thrombosis, and mechanical injury occur frequently. Existing risk prediction methods mainly draw on the time-series modeling paradigms of machine learning and deep learning, utilizing patient electronic medical records and multimodal monitoring data to construct end-to-end risk grading models. Common techniques include recurrent neural networks, long short-term memory networks, temporal convolutional networks, and self-attention mechanisms. Some approaches attempt to introduce multimodal fusion and clinical knowledge graphs to improve representation capabilities.

[0003] However, existing technologies suffer from the following problems: Models rely on continuous acquisition and point-by-point traversal modeling of full-time, high-density data, resulting in high computational and storage resource consumption and making real-time deployment on edge devices or lightweight medical terminals difficult; traditional temporal neural networks have blind spots in noise suppression and key information localization, and lack interpretability and decision traceability; data augmentation and multimodal alignment techniques further increase model complexity; different patients have significantly different treatment plans and disease progression characteristics, making it difficult to unify cross-individual representations and limiting the consistency and generalization of risk prediction. The industry urgently needs a novel modeling technology that can automatically identify key clinical anchors, extract local dynamic features only within a very short window, and encode evolutionary paths through semantic structuring. Summary of the Invention

[0004] This application provides a method for predicting the risk of complications from chemotherapy infusion ports, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0005] This application provides a method for predicting the risk of complications from chemotherapy infusion ports, specifically including:

[0006] S1: Obtain the structured fields of the electronic medical records and the natural language text of the nursing records of patients with chemotherapy infusion ports, and use evidence-based medicine validation rules to perform risk anchor point matching and identification, identify and label risk anchor point events, and generate a set of risk anchor point events.

[0007] S2: Based on the timestamps in the risk anchor event set, automatically associate multi-source local data fragments within a preset time window before and after each risk anchor event to construct an anchor context unit;

[0008] S3: Perform fixed-dimensional encoding operations independently on the anchor context unit, and map the encoding results to a unified semantic space to generate an anchor semantic embedding vector;

[0009] S4: Using the anchor point semantic embedding vector as a node, generate connection edges according to the causal, co-occurrence and time sequence dependencies defined in the clinical knowledge graph, and calculate the edge weights by combining the confidence and time decay factors of historical clinical sample data to construct the anchor point sequence graph;

[0010] S5: Input the anchor point sequence graph into the pre-constructed graph attention network to perform message passing operation, activate the node gating mechanism according to the contribution entropy of each node to the final risk level, shield low-entropy nodes and generate risk evolution feature representation;

[0011] S6: Based on the risk evolution characteristics, perform classification reasoning calculations, output the risk level determination results of chemotherapy infusion port complications, and simultaneously extract the dominant anchor link as an interpretable time logic basis to generate a comprehensive early warning report.

[0012] The method for predicting the risk of complications from chemotherapy infusion ports provided in this application has the following beneficial effects:

[0013] This solution addresses the computational redundancy and high deployment costs of existing chemotherapy port complication prediction models that rely on full-time, high-density data. It proposes a sparse temporal modeling paradigm based on risk anchor events, extracting only pathologically significant contextual units, significantly reducing the dimensionality of input data and the length of model processing. Accurate risk projection can be achieved without real-time streaming monitoring equipment. To address semantic fragmentation and feature drift issues in multi-source heterogeneous data fusion, a clinical knowledge graph-guided anchor sequence graph structure is introduced. This transforms isolated events into a topological network with causal logic, employing a lightweight graph attention mechanism for local message passing and dynamic filtering of key links using node gating, ensuring a linear relationship between inference complexity and the number of effective clinical events. Simultaneously, the solution generates dominant anchor link trajectories while outputting risk levels, providing a key medical event transmission path from initial triggers to final risk states. This endows the prediction results with a highly traceable semantic structure. The entire system requires no data augmentation or large-scale labeled datasets, constructing a low-resource-consumption, high-semantic-fidelity, and closed-loop feedback-supporting intelligent early warning framework based on evidence-based medicine logic. Attached Figure Description

[0014] Figure 1 This is the main flowchart of a method for predicting the risk of complications from chemotherapy infusion ports;

[0015] Figure 2 This is a sub-flowchart of a method for predicting the risk of complications from chemotherapy infusion ports;

[0016] Figure 3 This is another sub-flowchart of a method for predicting the risk of complications from chemotherapy infusion ports. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0019] like Figure 1 As shown, this application provides a method for predicting the risk of complications from chemotherapy infusion ports, specifically including:

[0020] S1: Obtain the structured fields of the electronic medical records and the natural language text of the nursing records of patients with chemotherapy infusion ports, and use evidence-based medicine validation rules to perform risk anchor point matching and identification, identify and label risk anchor point events, and generate a set of risk anchor point events.

[0021] S2: Based on the timestamps in the risk anchor event set, automatically associate multi-source local data fragments within a preset time window before and after each risk anchor event to construct an anchor context unit;

[0022] S3: Perform fixed-dimensional encoding operations independently on the anchor context unit, and map the encoding results to a unified semantic space to generate an anchor semantic embedding vector;

[0023] S4: Using the anchor point semantic embedding vector as a node, generate connection edges according to the causal, co-occurrence and time sequence dependencies defined in the clinical knowledge graph, and calculate the edge weights by combining the confidence and time decay factors of historical clinical sample data to construct the anchor point sequence graph;

[0024] S5: Input the anchor point sequence graph into the pre-constructed graph attention network to perform message passing operation, activate the node gating mechanism according to the contribution entropy of each node to the final risk level, shield low-entropy nodes and generate risk evolution feature representation;

[0025] S6: Based on the risk evolution characteristics, perform classification reasoning calculations, output the risk level determination results of chemotherapy infusion port complications, and simultaneously extract the dominant anchor link as an interpretable time logic basis to generate a comprehensive early warning report.

[0026] Step S1: Obtain the structured fields of the electronic medical record and the natural language text of the nursing record for patients with chemotherapy infusion ports. Then, use evidence-based medicine validation rules to perform risk anchor point matching and identification processing, identify and label risk anchor point events with clear clinical significance, and generate a risk anchor event set. The risk anchor event set includes a unique ID and a type label. Specifically, it includes:

[0027] S1.1: Obtain the structured fields of the electronic medical record and the natural language text of the nursing record of the patient receiving chemotherapy infusion port. Perform field mapping standardization processing on the structured fields of the electronic medical record to generate a structured data stream in a unified format. At the same time, perform medical terminology normalization processing on the natural language text of the nursing record to generate a standardized text corpus, thereby obtaining a multi-source heterogeneous original dataset containing the entire treatment process of the patient.

[0028] The system retrieves structured fields from the electronic medical records of chemotherapy infusion port patients from the hospital information system interface, including patient demographic information, laboratory test values, medical orders, and nursing assessment sheets. At the same time, it captures unstructured nursing records in natural language text to form a raw, multi-source, heterogeneous data input stream.

[0029] Dictionary mapping and unit unification processing are performed on the structured fields of the electronic medical record to convert the heterogeneous coding standards used by different departments into a unified data format that conforms to the HL7 FHIR specification, eliminate field naming ambiguity and dimensional differences, and generate a standardized structured data stream.

[0030] For natural language text of nursing records, a pre-trained medical-specific word segmenter is invoked for fine-grained segmentation, identifying and extracting key semantic segments containing clinical symptoms, operational descriptions, and subjective feelings.

[0031] By using a medical ontology database to perform terminology normalization mapping on the extracted semantic fragments, colloquial expressions, abbreviations, and synonyms are converted into standard medical terminology codes, thus constructing a standardized text corpus with clear semantic indexes.

[0032] Establish a timestamp-based multimodal data alignment mechanism, using the patient's admission time as the baseline zero point, to accurately match and associate discrete events in the standardized structured data stream with continuous text records in the standardized text corpus on the timeline.

[0033] Through the above processing method, the original messy data from the previous step is transformed into a multi-source heterogeneous original dataset that is spatiotemporally aligned and semantically standardized, thereby achieving standardization and consistency of the underlying data structure and providing a high-quality data foundation for the accurate identification of subsequent risk anchor events.

[0034] S1.2: Based on the aforementioned multi-source heterogeneous original dataset, a pre-built evidence-based medicine validation rule base is used to perform parallel semantic risk anchor matching and identification processing on the structured data stream and standardized text corpus, extracting candidate event fragments that conform to the characteristics of catheter-related infection, thrombosis, or mechanical injury, thereby generating an initial risk event sequence with potential clinical significance.

[0035] The pre-built evidence-based medicine validation rule base is invoked, and the structured data stream and standardized text corpus are used as parallel input sources to start the multimodal semantic scanning engine.

[0036] A time-series-based sliding window traversal is performed on the structured data stream to extract the dynamic change trajectories of key physiological indicators such as white blood cell count, platelet count, and C-reactive protein.

[0037] Based on the pathological characteristics of catheter-related infections, a threshold judgment logic for indicator mutations is set to identify abnormal data points whose values ​​exceed the normal fluctuation range within a short period of time.

[0038] Perform natural language processing based on clinical entity recognition on a standardized text corpus to locate unstructured text fragments describing redness, swelling, effusion, or increased resistance to flushing at the puncture site.

[0039] Using a predefined semantic mapping table, the extracted text fragments are transformed into standardized clinical event labels, such as "local skin temperature elevation" or "bloody fluid seen upon aspiration".

[0040] Establish a spatiotemporal correlation index between structured outlier data points and unstructured clinical event labels, and merge multi-source evidence within the same time window.

[0041] The clinical risk confidence level of each merged event fragment was calculated using a weighted scoring mechanism, and candidate events with confidence levels higher than the preset baseline were selected.

[0042] The selected candidate events are arranged in chronological order of occurrence to generate an initial sequence of risk events with potential clinical significance.

[0043] S1.3: For each candidate event segment in the initial risk event sequence, a deep feature verification is performed by applying time window association processing combined with physiological indicator threshold judgment logic to screen out effective clinical inflection points that meet specific time constraints and numerical critical conditions, thereby generating a confirmed discrete high-risk anchor event list.

[0044] For candidate event segments in the initial risk event sequence, extract their associated timestamps and corresponding multi-source physiological indicator data to construct a spatiotemporal feature vector to be verified.

[0045] Based on the pathological mechanism types defined in the evidence-based medicine rule base, time window constraint parameters are dynamically configured, specific duration thresholds for the pre-observation period and the post-validation period are set, and the validity boundary of the time dimension is established.

[0046] A sliding window approach is used to perform local scanning of candidate events on the time axis, calculate the volatility and trend slope of vital sign data within the window, and identify local extreme points that meet the characteristics of acute deterioration or abnormal mutation.

[0047] Based on the physiological indicator threshold judgment logic, the values ​​of local extreme points are compared with preset clinical critical value limits, and the indicator deviation score is calculated using the following formula:

[0048]

[0049] Where S is the indicator deviation score. These are the observed extreme values ​​of physiological indicators. This is the normal reference value for this indicator.

[0050] When the deviation score of the indicator exceeds the preset confidence threshold, the candidate event is determined to meet the numerical critical condition and is marked as a potential high-risk clinical inflection point.

[0051] Further verification of the persistence of this inflection point over time is needed to eliminate transient artifacts caused by sensor noise or single measurement errors, ensuring that the event has genuine pathophysiological significance.

[0052] Effective clinical inflection points that meet both time constraints and numerical criticality conditions are extracted from the initial sequence to generate a confirmed list of discretized high-risk anchor events.

[0053] S1.4: Perform a unique identifier assignment operation on each event in the confirmed discretized high-risk anchor event list, and perform classification label injection processing according to the pathological mechanism type of the event, thereby generating a set of risk anchor events to be labeled containing a globally unique ID and standardized type labels.

[0054] Receive a confirmed list of discretized high-risk anchor events, and iterate through each candidate event fragment in the list to initiate the identifier assignment process.

[0055] A composite hash key is constructed based on a global timestamp and a patient's unique ID. The SHA-256 algorithm is used to calculate the digital fingerprint of each event fragment, generating a globally unique 32-bit hexadecimal identifier to ensure the non-confusing nature of events across datasets.

[0056] Analyze the abnormal physiological indicators and nursing operation records in the event segments to extract key semantic feature words, such as "decreased white blood cell count", "high resistance to flushing", or "local redness and swelling".

[0057] The extracted semantic feature words are mapped to a predefined classification system of pathological mechanisms of chemotherapy infusion port complications, which covers three main categories and their subcategories: infection risk, thrombosis risk, and mechanical injury risk.

[0058] Based on the mapping results, inject standardized type label codes, such as "white blood cell count < 1.5 × 10⁻⁶". 9 “ / L” is marked as “INF-01” (pre-infection sign), and “bloody fluid seen on aspiration” is marked as “THR-02” (suspected thrombosis).

[0059] Construct a structured event object containing a globally unique ID, standardized type label, original timestamp, and event description fields to complete the transformation from unstructured text to standardized data entities.

[0060] S1.5: Based on the set of risk anchor events to be labeled, perform the final data integrity verification and format encapsulation process, serialize the event objects with unique IDs and type labels into a standard data structure, thereby generating a set of risk anchor events that can be directly used for subsequent context unit construction.

[0061] Step S2: Based on the timestamps in the risk anchor event set, automatically associate multi-source local data fragments within a preset time window before and after each risk anchor event to construct an anchor context unit. Specifically, this includes:

[0062] S2.1: Obtain the unique ID and corresponding timestamp in the risk anchor event set, and use the dynamic time window configuration strategy to extend a preset duration forward and backward with the timestamp as the center to generate a local perception time interval including the start time and the end time, so as to establish the spatiotemporal boundary conditions for multi-source data extraction.

[0063] The unique ID and corresponding timestamp from the risk anchor event set are read as the baseline trigger point for dynamic window configuration. Based on a pre-set clinical pathological evolution half-life parameter library, the optimal sensing duration coefficient corresponding to the current anchor type is retrieved. For acute mechanical injury anchors, a shorter forward retrospective window is set to capture immediate operational impact; for infection or thrombosis anchors, a longer backward extension window is set to monitor the trend of inflammation spread. Asymmetric time offset processing is used, tracing forward the preset baseline stabilization period duration and extending backward the preset risk manifestation period duration, centered on the anchor timestamp. The start and end times of the local sensing time interval are calculated using the following formula:

[0064]

[0065]

[0066] in, The timestamp of the anchor event. Forward backtracking window duration, For the duration of the backward extension window, This represents the starting moment of the local perception time interval. This is the termination time of the locally perceived time interval. The calculated start and termination execution boundaries are validated to ensure the time interval does not exceed the entire treatment timeline from patient admission to discharge. If the calculated time interval crosses non-monitoring periods or data gaps, a window shrinking mechanism is automatically triggered to extract the largest continuous sub-interval covered by valid data. The validated time interval is encapsulated as a structured object containing start and end markers.

[0067] S2.2: Based on the local perception time interval, concurrent retrieval is performed on the electronic medical record database, infusion pump operation log, infrared thermal imaging equipment and nursing text records, respectively extracting the vital sign snapshot sequence, infusion pump status log fragment, local infrared thermal image ROI (Region of Interest) pixel mean sequence and nurse assessment text keyword vector to generate a multi-source local data fragment set.

[0068] Receive the local sensing time interval generated by S2.1. This interval defines the start and end times centered on the risk anchor event timestamp, serving as the spatiotemporal boundary constraint for multi-source data retrieval.

[0069] Based on the aforementioned spatiotemporal boundary conditions, a structured query request is initiated to the electronic medical record database to extract all vital sign monitoring records within the interval, including body temperature, heart rate, blood pressure, and blood oxygen saturation values, and to construct a discrete vital sign snapshot sequence.

[0070] Synchronously access the infusion pump operation log storage unit, filter out the device status data matching the time window, parse the flow rate setpoint, blockage alarm code and remaining drug volume field, and generate infusion pump status log fragments reflecting the dynamic process of infusion.

[0071] The infrared thermal imaging device interface is called to obtain the local skin temperature distribution image within the corresponding time period. The Region of Interest (ROI) is located for the infusion port implantation area. The arithmetic mean of the pixel grayscale values ​​within the ROI is calculated to form a sequence of local infrared thermal map ROI pixel mean values.

[0072] Natural language processing retrieval is performed on nursing text records. A pre-trained medical named entity recognition model is used to extract keywords describing local redness, swelling, pain, or effusion. Unstructured text is then transformed into fixed-dimensional nurse assessment text keyword vectors through word embedding mapping technology.

[0073] The extracted vital sign snapshot sequence, infusion pump status log fragment, local infrared thermal image ROI pixel mean sequence, and nurse assessment text keyword vector were logically aggregated and encapsulated into a multi-source local data fragment set containing multimodal information.

[0074] S2.3: Perform time axis alignment processing on the multi-source local data fragment set, use linear interpolation to fill in the missing time points in the vital signs snapshot sequence and the local infrared heat map ROI pixel mean sequence, and perform time-series resampling of the nurse assessment text keyword vector according to the time granularity of the infusion pump status log fragment to generate a time-synchronized multimodal data stream.

[0075] A collection of multi-source local data fragments is obtained, including vital sign snapshot sequences, infusion pump status log fragments, local infrared heat map ROI pixel mean sequences, and nurse assessment text keyword vectors. Timestamp indexes are constructed for the vital sign snapshot sequences and local infrared heat map ROI pixel mean sequences to identify the sampling time distribution of each modality within the local perception time interval. Missing sampling points in the above continuous time-series data are detected to determine the time locations requiring interpolation. Linear interpolation is used to estimate the data values ​​at the missing time points, and the estimated value at the missing point is calculated based on the slope of the values ​​of adjacent valid sampling points. The interpolation result for the missing time t is calculated using the following formula:

[0076]

[0077] Where y(t) is the interpolation result at time t, t1 and t2 are the two most recent valid sampling times before and after the missing time, and y1 and y2 are the observation values ​​corresponding to times t1 and t2, respectively. After filling in all missing points, a continuous and uninterrupted time series sequence of vital signs and heatmap mean values ​​is generated. The time granularity of the infusion pump status log segment is read, and the minimum time interval of the log record is determined as the reference clock frequency. Temporal resampling processing is performed on the keyword vector of the nurse's assessment text, and the non-uniformly distributed text features are mapped to a unified time axis grid according to the reference clock frequency. For multiple text vectors in the same time grid, a weighted average strategy is used to fuse semantic features; for time grids without text records, the nearest neighbor forward filling strategy is used to maintain semantic continuity. The interpolated continuous time series sequence and the resampled discrete text vector sequence are timestamped to ensure that each modality data has a corresponding feature representation at the same time point. The aligned multimodal data are merged and arranged in chronological order to generate a time-synchronized multimodal data stream.

[0078] S2.4: Based on the time-synchronized multimodal data stream, the vital signs snapshot sequence, infusion pump status log fragment, local infrared heat map ROI pixel mean sequence, and nurse assessment text keyword vector are tensor-concatenated and dimension-normalized using the feature fusion operator to eliminate the dimensional differences between different modal data and generate a standardized multimodal feature matrix.

[0079] The system receives a time-synchronized multimodal data stream, including vital sign snapshot sequences, infusion pump status log fragments, local infrared thermogram ROI pixel mean sequences, and nurse assessment text keyword vectors. For continuous numerical data such as blood pressure and heart rate in the vital sign snapshot sequences, Z-Score standardization is used to unify the dimensions, eliminating the order-of-magnitude differences between different physiological indicators and generating a standardized physiological feature sub-matrix. One-hot encoding is performed on the discrete status codes in the infusion pump status log fragments to map non-numerical device operating states into a binary vector space, ensuring their additivity with continuous physiological features in the mathematical space. For the local infrared thermogram ROI pixel mean sequences, Min-Max normalization is applied to map them to the [0,1] interval to eliminate the influence of ambient temperature fluctuations on the absolute value of skin surface temperature and highlight the relative temperature rise trend. For the nurse assessment text keyword vectors, a pre-trained Word2Vec model is used to extract word embedding vectors, which are then aggregated into fixed-dimensional semantic feature vectors through average pooling, realizing the transformation from unstructured text to structured numerical features. A multimodal tensor concatenation operator is constructed to concatenate the standardized physiological feature submatrix, discrete state encoding vector, normalized heatmap feature vector, and text semantic feature vector along the feature dimension, according to a preset feature channel order. The standardized feature values ​​are calculated using the following formula:

[0080]

[0081] Where x is the original feature value, μ is the mean of the feature on the training set, and σ is the standard deviation. norm These are the standardized eigenvalues. The concatenated high-dimensional sparse matrix is ​​normalized using the L2 norm. By calculating the Euclidean norm of the eigenvectors and dividing by this norm, the weight bias caused by the uneven number of feature dimensions is eliminated, generating a standardized multimodal feature matrix with a unified metric.

[0082] S2.5: Perform a structured encapsulation operation on the standardized multimodal feature matrix, bind it with the unique ID and type label of the corresponding risk anchor event, add a data integrity check code, and generate an anchor context unit with a unified data structure.

[0083] like Figure 2As shown, step S3: A fixed-dimensional encoding operation is independently performed on the anchor context unit, and the encoding results are mapped to a unified semantic space to eliminate representational differences of similar risk anchor events among different patients, generating comparable anchor semantic embedding vectors. Specifically, this includes:

[0084] S3.1: Perform timestamp synchronization and dimension normalization on the vital signs snapshot, infusion pump log, local infrared heat map ROI pixel mean, and nurse assessment text keyword vector in the anchor context unit to generate a standardized multimodal feature tensor.

[0085] It receives anchor context units as input data, which include vital sign snapshots, infusion pump logs, mean pixel values ​​of local infrared thermal images (ROIs), and keyword vectors of nurse assessment texts.

[0086] Perform timestamp alignment verification on the heterogeneous multimodal data in the anchor context unit to identify the sampling time deviation of each modal data stream within the local sensing time interval.

[0087] To address the non-uniform sampling points in the vital signs snapshot sequence and the mean pixel sequence of the local infrared thermal image ROI, cubic spline interpolation is used for resampling to map discrete time points to a unified standard time grid, generating a time-synchronized continuous numerical sequence.

[0088] Event-driven timing coding is performed on the infusion pump status log fragments to convert discrete operation commands into fixed-length one-hot encoded vectors, and zero-order hold-fill is performed according to the standard time grid to ensure strict correspondence between control signals and physiological indicators on the time axis.

[0089] Semantic density weighting is applied to the keyword vectors of nurse assessment texts. Based on the weight coefficients of the keywords in a predefined clinical risk dictionary, the unstructured text features are transformed into a dense semantic feature matrix aligned with a temporal grid.

[0090] A multimodal tensor fusion structure is constructed by splicing together the time-synchronized vital sign numerical sequence, infusion pump control vector, infrared thermogram spatial features, and text semantic matrix along the channel dimension to form the original high-dimensional multimodal tensor.

[0091] The Z-Score normalization method is used to perform dimensionality normalization on the data of each channel in the original high-dimensional multimodal tensor. The mean and standard deviation of each feature channel are calculated using the formula: Eliminate the dimensional differences between different modal data, where x is the original feature value, μ is the channel mean, and σ is the channel standard deviation.

[0092] Missing value masking is performed on the normalized multimodal tensor to generate a binary validity mask matrix of the same dimension as the feature tensor, which identifies data gaps caused by equipment failure or record omission.

[0093] The three-dimensional multimodal data is reconstructed into a two-dimensional feature matrix by tensor reshape operation, preserving the topological relationship between time step and feature dimension, and generating a standardized multimodal feature tensor.

[0094] S3.2: Based on the standardized multimodal feature tensor, the pre-trained cross-modal feature fusion encoder is invoked to perform channel attention weighting operations, mapping discrete vital sign values, continuous pump log sequences, spatial heatmap textures, and unstructured text semantics to a shared hidden layer space to generate a high-dimensional hybrid feature representation vector.

[0095] The system receives a standardized multimodal feature tensor, which includes vital sign values, infusion pump log sequences, infrared thermal textures, and semantic vectors of nursing text. The different modal features of the standardized multimodal feature tensor are mapped to a unified dimensional space, and the global average pooling value for each modal channel is calculated to generate a modality importance descriptor.

[0096] The cross-modal feature fusion encoder includes an attention weight generation network and a feature mapping layer. The attention weight generation network consists of two fully connected layers: the first fully connected layer takes the modality importance descriptor as input, expands the input dimension to a 16-dimensional hidden space, and after activation by the ReLU activation function, outputs a first-layer interaction feature vector to capture the interaction relationship between modalities; the second fully connected layer receives the first-layer interaction feature vector, compresses it back to a 4-dimensional space with the same number of modalities, and after activation by the Sigmoid activation function, generates attention weight coefficients for each modality channel.

[0097] The attention weight coefficients are multiplied element-wise with each modality feature vector in the standardized multimodal feature tensor to generate a recalibrated multimodal feature vector, thus achieving feature recalibration. The multimodal feature vectors are then concatenated to form a fused feature matrix. The feature mapping layer takes the fused feature matrix as input, maps it to a high-dimensional hidden layer space through linear projection, and introduces nonlinear expressive power through the ReLU activation function to generate a high-dimensional hybrid feature representation vector. It should be noted that each modality feature vector in the standardized multimodal feature tensor corresponds to: vital sign numerical encoding vector, infusion pump log sequence encoding vector, infrared thermal texture encoding vector, and nursing text semantic encoding vector, respectively.

[0098] The pre-training process of the cross-modal feature fusion encoder is as follows: A training dataset is constructed based on anchor context units labeled with clear complication outcomes from historical clinical sample data. The optimization objective is to minimize the semantic distance between different anchors for the same patient and maximize the semantic consistency of similar anchors for different patients. A contrastive loss function is used to iteratively optimize the encoder network parameters until the loss function converges. After the encoder pre-training is complete, the clinical event semantic projection matrix is ​​trained by minimizing the cross-entropy loss function, with encoder weights frozen as a condition and chemotherapy infusion port complication type labels as the classification task objective.

[0099] S3.3: For the high-dimensional hybrid feature representation vector, apply the pre-constructed clinical event semantic projection matrix to perform a linear transformation operation, forcibly constrain the feature distribution in the general medical semantic space to the special semantic subspace for chemotherapy infusion port complications, so as to generate a preliminary clinical semantic mapping vector.

[0100] Receive the high-dimensional hybrid feature representation vector generated in the previous step. This high-dimensional hybrid feature representation vector integrates multimodal information such as vital signs, infusion logs, heat map textures and text semantics, and is located in the general medical semantic space.

[0101] The pre-trained clinical event semantic projection matrix is ​​invoked, and matrix multiplication is performed on the high-dimensional hybrid feature representation vector to map the feature distribution in the general medical semantic space to the semantic subspace of chemotherapy infusion port complications, generating a preliminary clinical semantic mapping vector.

[0102]

[0103] Among them, V H W is the high-dimensional mixed feature representation vector of the input. P b represents the weighting coefficients of the semantic projection matrix of clinical events. P For the bias term, Vp init This is the initial clinical semantic mapping vector for the output.

[0104] The row dimension of the clinical event semantic projection matrix is ​​consistent with the dimension of the high-dimensional hybrid feature representation vector, and the column dimension is the dimension of the semantic subspace of chemotherapy infusion port complications. This dimension value is equal to the total number of predefined type labels in the risk anchor event set. The semantic subspace of chemotherapy infusion port complications is a dedicated semantic space that uses the type labels of the risk anchor events as the basis for category discrimination. It is used to constrain general medical semantic features to the exclusive discrimination dimension of chemotherapy infusion port-related complications. The clinical event semantic projection matrix is ​​obtained by optimizing and learning historical clinical sample data during the pre-training stage and fixing the weight values.

[0105] A nonlinear activation function is introduced to rectify the linear transformation results, preserving positive semantic features that are highly correlated with the risk of complications and suppressing the propagation of irrelevant noise signals.

[0106] By using projection transformation, the multimodal features scattered in the general space are forcibly constrained to a low-dimensional manifold that is specialized for catheter infection, thrombosis and mechanical injury.

[0107] By performing a linear transformation operation on the clinical event semantic projection matrix, the high-dimensional mixed feature representation vector from the previous step is transformed into a preliminary clinical semantic mapping vector, achieving accurate mapping from general semantics to a dedicated risk semantic space and eliminating representational bias caused by individual differences across patients.

[0108] S3.4: Calculate the Euclidean distance between the preliminary clinical semantic mapping vector and the center of the predefined standard risk anchor prototype cluster, and optimize the feature distribution through gradient backpropagation to minimize the intra-class variance in order to generate a de-individualized, canonical clinical semantic vector.

[0109] The system receives a preliminary clinical semantic mapping vector transformed by a clinical event semantic projection matrix. This vector carries the initial feature distribution of the dedicated subspace for chemotherapy infusion port complications. A predefined set of standard risk anchor prototype cluster centers is constructed. This set contains baseline semantic vectors for typical complication categories such as infection, thrombosis, and mechanical injury, serving as reference anchors for feature alignment.

[0110] Calculate the Euclidean distance between the current preliminary clinical semantic mapping vector and the centers of each standard risk anchor prototype cluster to quantify the semantic deviation of the current sample from various typical pathological states. The distance between the i-th sample vector and the center of the k-th prototype class is calculated using the following formula:

[0111]

[0112] Where, d i,k v represents the distance between the i-th sample and the k-th class prototype. i For the initial clinical semantic mapping vector, c k It is the center of the prototype cluster of the k-th type of standard risk anchor point.

[0113] A prototype contrastive learning loss function is introduced, aiming to minimize the distance between samples of the same class and their corresponding prototype centers while maximizing the distance between samples of different classes. An optimization objective function is constructed. The semantic encoder network parameters are updated using the gradient backpropagation algorithm to adjust the feature space distribution and compress intra-class variance.

[0114] The distance calculation and parameter update process is iteratively executed until the loss function converges to the preset threshold, so that the same type of risk anchor point exhibits a highly consistent semantic representation among different patients, eliminating feature drift caused by individual physiological differences.

[0115] S3.5: Perform fixed-dimensional truncation and L2 norm normalization on the de-individualized standardized clinical semantic vectors, remove redundant noise components and unify vector magnitude, and finally output anchor semantic embedding vectors with strict comparability.

[0116] like Figure 3 As shown, step S4: Using the anchor point semantic embedding vector as nodes, generate connection edges according to the causal, co-occurrence, and temporal dependencies defined in the clinical knowledge graph, and calculate edge weights by combining historical clinical sample data with statistical confidence and time decay factors to construct an anchor point sequence graph. Specifically, this includes:

[0117] S4.1: Based on the comparable anchor semantic embedding vectors generated in the previous steps, perform graph node initialization processing, map each anchor semantic embedding vector to an independent graph node in the anchor sequence graph, and assign a unique ID and type label consistent with the risk anchor event set to generate an initial graph node set containing node attribute information.

[0118] It receives anchor point semantic embedding vectors with strict comparability from the previous steps and uses them as the basic data input for graph node initialization.

[0119] The anchor semantic embedding vector is subjected to dimension verification and integrity check to ensure that each vector conforms to the preset fixed dimension standard and to remove invalid data fragments caused by encoding anomalies.

[0120] Based on the verified anchor semantic embedding vectors, independent node objects are constructed in the graph computing framework. The numerical sequence of the vectors is mapped to the hidden layer state attributes of the nodes, thus establishing the feature representation basis of the nodes in the graph structure.

[0121] Extract the globally unique ID corresponding to the current embedded vector from the risk anchor event set, assign the ID to the newly created graph node, establish a bidirectional index association between the node and the original clinical event, and ensure the traceability of subsequent inference results.

[0122] Read the pre-defined standard type labels from the risk anchor event set, such as infection warning, thrombosis risk, or mechanical failure, encode the labels as discrete classification attributes, and inject them into the node metadata to enrich the semantic information dimension of the node.

[0123] The node objects containing hidden layer state attributes, globally unique IDs, and classification attributes are encapsulated into standard graph data structure units to form an initial graph node set, providing structured input for subsequent edge generation based on clinical knowledge graphs.

[0124] S4.2: Use a pre-built clinical knowledge graph to perform semantic association analysis on the initial graph node set. Based on the causal relationships, co-occurrence relationships and time sequence dependency rules defined in the clinical knowledge graph, scan and identify the potential connection paths between any two graph nodes in the initial graph node set, and generate a candidate connection edge set.

[0125] The clinical knowledge graph refers to a pre-constructed structured knowledge base with clinical events related to chemotherapy port-in-the-loop complications as nodes and causal relationships, co-occurrence relationships, and time-series dependencies between events as edges. It is obtained by: extracting complication event types such as infection warning, thrombosis risk, and mechanical failure, along with their timestamps, based on complete electronic medical record cohort data of patients who underwent chemotherapy port-in-the-loop implantation within the past five years; calculating the co-occurrence frequency and conditional probability among these complication events through statistical analysis; and summarizing causal rules, co-occurrence rules, and time-series dependency rules between events by combining them with pathological causal paths defined in medical guidelines; and encoding and storing these rules in the form of a directed graph to form the clinical knowledge graph.

[0126] Obtain the initial graph node set, extract the standardized type label and unique ID of each node, and construct the semantic query index to be associated.

[0127] The system invokes a pre-built clinical knowledge graph, retrieves ontology concepts and their attribute constraints that match the current node's type label, and locates relevant entity subgraphs within the graph.

[0128] Based on the causal rules defined in the clinical knowledge graph, the system traverses any two node type combinations in the initial graph node set to identify whether there is a pathological logical path where a preceding event leads to a subsequent event.

[0129] Based on the co-occurrence relationship rules, the frequency patterns of different anchor events occurring simultaneously in the same treatment cycle in the historical cohort are statistically analyzed, and node pairs with significant statistical correlation are marked.

[0130] Apply time-dependent rules to calculate whether the timestamp difference between nodes conforms to a reasonable window period of clinical physiological evolution, and eliminate invalid connections with inverted times or excessively long intervals.

[0131] Map node pairs that satisfy any of the causal, co-occurrence, or temporal rules to directed edges, and record the source node ID, target node ID, and relationship type of the edges to form the original candidate connection edge set.

[0132] S4.3: Calculate the statistical confidence of the candidate connection edge set based on historical clinical sample data, extract the co-occurrence frequency and conditional probability of each candidate connection edge in historical cases, calculate the time decay factor in combination with the timestamp of the current risk anchor event, and perform a weighted fusion operation on the statistical confidence and the time decay factor to generate the dynamic edge weight value of each candidate connection edge.

[0133] Complete electronic medical record cohort data of chemotherapy port implantation patients from the past five years were retrieved from the clinical data center to establish a historical knowledge base including patient ID, anchor event type, occurrence timestamp, and complication outcome label.

[0134] For each directed edge in the candidate connection edge set, define the event type of the starting node as A and the event type of the ending node as B, and count the frequency N(A,B) of event B occurring within a specified time window after event A occurs in the historical queue.

[0135] Calculate the total frequency N(A) of event A in the historical queue, and use the maximum likelihood estimation method to calculate the conditional probability P(B|A) from event A to event B, which serves as the basic statistical confidence index reflecting the strength of pathological causality.

[0136] The basic statistical confidence score is calculated using the following formula. stat :

[0137]

[0138] Where N(A,B) is the co-occurrence frequency of events A and B, and N(A) is the total frequency of event A.

[0139] Obtain the timestamp of the current patient's risk anchor event and retrieve the time interval Δt between this event and the last similar or related preceding event.

[0140] An exponential decay function is introduced to simulate the natural decline of clinical risk over time, and a time decay factor is calculated to reflect the high contribution of recent events to current risk.

[0141] The time decay factor is calculated using the following formula. time :

[0142]

[0143] Among them, Factor time λ is the time decay factor, λ is the decay coefficient, which is set to 0.1 / hour based on the empirical value of the complication half-life, and Δt is the time interval.

[0144] By nonlinearly weighting and fusing the basic statistical confidence level with the time decay factor, dynamic edge weights are constructed to balance general pathological patterns and individual temporal characteristics.

[0145] The dynamic edge weights are calculated using the following formula:

[0146]

[0147] Here, α and β are adjustment hyperparameters, set to 0.6 and 0.4 respectively, to emphasize historical statistical patterns while retaining timeliness sensitivity. edge For dynamic edge weights, Score stat Based on statistical confidence level, Factor time This is the time decay factor.

[0148] S4.4: Perform threshold filtering on the candidate connection edge set based on the dynamic edge weight value, remove redundant candidate connection edges whose dynamic edge weight value is lower than the preset confidence threshold, retain the valid connection edges with high confidence, and generate a sparse connection edge set optimized by pruning.

[0149] Obtain the candidate connection edge set containing dynamic edge weight values ​​generated in the previous steps. Each edge in the set is associated with a weighted fusion result of statistical confidence and time decay factor.

[0150] A preset confidence threshold parameter is set, which is determined based on the minimum effective support of the complication occurrence pathway in historical clinical sample data, and is used to define the significance boundary of the risk transmission relationship.

[0151] Iterate through each edge in the candidate connection edge set, extract its corresponding dynamic edge weight value, and compare it with a preset threshold value.

[0152] For candidate connection edges whose dynamic edge weight values ​​are less than the preset confidence threshold, they are determined to be low-confidence redundant connections. A logical elimination operation is performed to remove them from the current processing set in order to block the risk transmission path of weak correlation or noise interference.

[0153] For candidate connection edges whose dynamic edge weight values ​​are greater than or equal to a preset confidence threshold, they are determined to be high-confidence valid connections, their node index relationships and weight attributes are retained, and they are included in the subset of valid connection edges.

[0154] Perform topology integrity checks on the retained valid connection subset to ensure that all nodes connected by the retained edges exist in the initial graph node set, preventing graph structure breaks caused by isolated nodes or dangling edges.

[0155] The validated subset of valid connection edges is grouped and sorted according to the source node ID to generate a sparse adjacency list, forming a sparse connection edge set optimized by pruning.

[0156] S4.5: Perform topological assembly of the initial graph node set and sparse connection edge set, establish directed connection relationships between nodes and bind corresponding dynamic edge weight values, and construct an anchor sequence graph for encoding risk transmission logic.

[0157] Obtain the pruned and optimized set of sparse edges and the initial graph node set as the topological foundation for constructing the anchor sequence graph. Traverse each valid edge in the sparse edge set, parsing its source node ID and target node ID, and locate the corresponding source and target node objects in the initial graph node set. Establish directed pointer links from the source node to the target node, establishing the temporal order and causal logic direction of risk transmission, forming a preliminary graph topology skeleton. Read the dynamic edge weight value corresponding to each valid edge, bind this numerical attribute to the newly generated directed edge, and assign a quantitative indicator representing the intensity of risk transmission to this edge. Encapsulate the attributes of each node in the initial graph node set, integrating the node's unique ID, type label, and anchor semantic embedding vector generated in the previous steps into a node feature dictionary, ensuring the integrity and structure of node information. Perform graph connectivity verification, detecting and removing isolated nodes caused by threshold filtering, ensuring that all nodes in the anchor sequence graph have at least one incoming or outgoing edge, maintaining the continuity of the risk evolution path. The set of nodes containing a dictionary of node features is serialized and assembled with a set of directed edges bound with dynamic weights to generate a standard graph data structure object, which fully encodes the risk transmission logic of chemotherapy infusion port complications.

[0158] Step S5: Input the anchor sequence graph into the pre-constructed graph attention network to perform message passing operation. At the same time, during the inference phase, dynamically activate the node gating mechanism based on the contribution entropy of each node to the final risk level, shield low-entropy nodes, and generate risk evolution feature representation.

[0159] The graph attention network refers to a graph neural network model based on a multi-head self-attention mechanism, used for message passing and node feature updating of the anchor sequence graph. The network consists of multiple parallel attention heads. Each attention head maps node features to query vectors, key vectors, and value vectors through a learnable linear transformation matrix, calculates the attention coefficients between the target node and its neighboring nodes, and uses edge weights as bias terms to adjust the attention score. Finally, it generates a node update feature vector by weighted summation and aggregation of neighboring node information. Its training method is as follows: based on anchor sequence graphs with known complication outcomes from historical clinical sample data, with the goal of minimizing risk prediction loss, the linear transformation matrix, edge weight bias parameters, and Softmax normalization parameters in the attention mechanism are trained end-to-end using a backpropagation algorithm, enabling the network to learn the relative importance of different anchor nodes in the risk evolution path.

[0160] Specifically, it includes: S5.1: Based on the anchor semantic embedding vector and edge weight matrix in the anchor sequence graph, construct and process the topological data of the generated graph structure based on the sparse adjacency list to eliminate redundant zero-value connections and establish causal transmission paths between nodes, and output a sparse graph topological representation containing effective connection relationships.

[0161] It receives the set of anchor semantic embedding vectors and the corresponding dynamic edge weight matrix from the anchor sequence graph as the initial input object for constructing graph structure topology data.

[0162] The dynamic edge weight matrix is ​​subjected to sparsity threshold determination. Connections with weight values ​​lower than a preset noise threshold are marked as invalid zero-value connections, and a binary adjacency mask matrix is ​​generated to identify valid causal transmission paths.

[0163] Based on the binarized adjacency mask matrix, the connection relationships between nodes are indexed and encoded using a compressed sparse row format, storing only the source node ID, target node ID, and corresponding weight values ​​of non-zero weight edges, thus eliminating redundant zero-value storage space.

[0164] The anchor semantic embedding vectors are rearranged according to their unique IDs and mapped to the compressed sparse adjacency index to ensure that the graph attention network can directly access the features of neighboring nodes through the index during message passing.

[0165] The completed lightweight graph structure is subjected to connectivity verification, and isolated nodes or broken subgraphs caused by pruning are removed to ensure the continuity and integrity of the risk evolution logic in the graph topology.

[0166] By constructing a sparse adjacency list, the dense graph data generated in the previous step is transformed into a sparse graph topological representation containing effective connections. This achieves a linear reduction in computational complexity and establishes causal transmission paths between nodes, providing a structured foundation for efficient reasoning in subsequent graph attention networks.

[0167] For example, in a chemotherapy patient case with 50 risk anchor nodes, a preset noise threshold of 0.05 was used. It was determined that 420 out of 480 potential edges in the original fully connected graph had weights below 0.05 and were marked as invalid. The remaining 60 high-confidence edges were stored using a compressed sparse row format, reducing the storage space from 2500 floating-point numbers in the original dense matrix to 180 integer indices and 60 floating-point weights. Verification revealed no isolated nodes; all anchor nodes were located within the main connected components. This sparse graph topology representation reduced the computational cost of subsequent message passing by 84%, significantly improving the inference speed of the model on mobile devices, while preserving the critical causal path of increased ductal resistance due to platelet decrease.

[0168] S5.2: For each target node and its set of neighboring nodes in the sparse graph topological representation, perform feature aggregation operation under the multi-head self-attention mechanism, generate normalized attention coefficients by calculating the dot product similarity between the query vector and the key vector, and output the node update feature vector that integrates local context information.

[0169] The target node refers to the anchor node currently performing message passing and feature aggregation operations in the sparse graph topology representation. In step S5.2, the system traverses each node in the sparse graph topology representation, takes that node as the current processing object, extracts its first-order neighbor node set, calculates the attention coefficients between the node and each neighbor node through a multi-head self-attention mechanism, and then aggregates the neighbor information to generate the updated feature vector of the node. The target node is the central node in the message passing process of the graph attention network, and its identity dynamically switches during the traversal process. Each node, when acting as a target node, receives semantic information from its neighbor nodes, thereby realizing the layer-by-layer propagation of global topology information.

[0170] It receives the anchor semantic embedding vector and adjacency relationship from the sparse graph topological representation. For the target node and its first-order neighbor set currently being traversed, it initializes the linear projection parameters of the multi-head attention mechanism. For the original feature vectors of the target node and each neighbor node, it applies independent linear transformation matrices to generate query vector, key vector, and value vector, thereby achieving decoupled mapping of the feature space.

[0171] The algorithm calculates the dot product similarity between the query vector of the target node and the key vectors of all neighboring nodes, quantifies the semantic association strength between nodes, generates an unnormalized attention score matrix, and captures key dependencies in the local topology.

[0172] Edge weights are introduced as bias terms into the attention score calculation, and the attention coefficient is adjusted using the following formula:

[0173]

[0174] Among them, e ij Let a be the normalized attention coefficient of node i to its neighbor j. ij The sum of the original dot product score and the edge weights is given by LeakyReLU, which is a non-linear activation function. i Let i be the set of neighbors of node i.

[0175] The attention score is normalized using the Softmax function to ensure that the sum of the attention coefficients of all neighboring nodes is 1, forming a probability distribution of attention weights that highlight the contribution of highly relevant neighboring nodes.

[0176] Based on the normalized attention weights, a weighted summation operation is performed on the value vectors of neighboring nodes to aggregate local context information and generate an intermediate representation vector containing the semantic features of neighboring nodes, thereby realizing information fusion in the message passing process.

[0177] The intermediate representation vectors of each head in the multi-head attention mechanism are concatenated or averaged, and then restored to the original feature dimension through a linear projection layer. The resulting node update feature vector, which incorporates local context information, is output, thereby enhancing the robustness and richness of feature representation.

[0178] S5.3: Based on the mapping relationship between the node update feature vector and the global risk classification hyperplane, the feature distribution dispersion of each node is quantified using the Shannon entropy calculation formula to generate a node contribution entropy value that characterizes the degree of influence of a single node on the final judgment result.

[0179] The global risk classification hyperplane refers to a classification projection module consisting of a fully connected mapping layer and a Softmax normalization function, used to project the node update feature vector from the high-dimensional feature space to the probability space of complication risk categories. It is obtained as follows: during the model training phase, based on known complication outcome labels from historical clinical sample data, a linear transformation matrix and bias vector are learned through backpropagation. This transformation matrix defines the mapping direction from the feature space to the category score space, thus constituting the global risk classification hyperplane.

[0180] The node update feature vector is obtained after aggregation by the multi-head self-attention mechanism. This vector integrates the semantic information of local neighbor nodes and the influence of edge weights, and serves as the basic data input for evaluating the importance of nodes.

[0181] A global risk classification hyperplane mapping layer is constructed to project the updated feature vectors of high-dimensional nodes onto the probability space of complication risk categories, and the initial classification score of each node belonging to each type of complication risk is calculated.

[0182] The initial classification scores are normalized using Softmax to generate a distribution vector representing the probability of each node belonging to each risk category, ensuring that the sum of the probabilities is 1, thus forming a discrete risk probability distribution.

[0183] Based on the principles of information theory, the Shannon entropy formula is used to quantify the uncertainty of the risk probability distribution, thereby measuring the contribution of node characteristics to the final risk judgment. The higher the uncertainty, the more critical the risk evolution information contained in the node.

[0184] The specific calculation formula is as follows:

[0185]

[0186] in, This represents the contribution entropy value of the i-th anchor node, where C is the total number of complication risk categories. Let be the probability value of the i-th node belonging to the c-th risk class.

[0187] Traverse all nodes in the anchor point sequence graph, calculate the corresponding node contribution entropy value for each node, and generate an entropy value sequence that maps one-to-one with the node ID for subsequent gating and filtering.

[0188] S5.4: Based on the comparison result between the node contribution entropy value and the preset dynamic threshold, perform a binarization mask generation operation to construct an anchor point importance gating signal. Use the anchor point importance gating signal to perform zero-masking processing on the node update feature vector corresponding to the low entropy value node, and output the gating node feature set after preliminary screening.

[0189] Obtain the set of contribution entropy values ​​of each node output from step S5.3 and the preset dynamic threshold parameters as input data for the gating mechanism.

[0190] An element-wise traversal operation is performed on the set of node contribution entropy values, and the contribution entropy value of each node is compared with the dynamic threshold.

[0191] A binary mask vector is constructed based on the comparison results. When the node contribution entropy value is greater than or equal to the dynamic threshold, the mask bit corresponding to the node is set to 1, indicating that the node's features are preserved.

[0192] When the node contribution entropy value is less than the dynamic threshold, the mask bit corresponding to the node is set to 0, indicating that the node's features are masked, thereby generating an anchor point importance gating signal vector.

[0193] The Hadamard Product operation is performed using the generated anchor importance gating signal vector and the set of node update feature vectors output from step S5.2.

[0194] For nodes with a mask bit of 0, the values ​​of all dimensions of the corresponding node's updated feature vector are forcibly set to zero, thus achieving feature masking of low-entropy nodes.

[0195] For nodes with a mask bit of 1, the corresponding node's updated feature vector retains its original value, thus achieving feature preservation of high information gain anchor points.

[0196] S5.5: Perform graph structure reconnectivity verification and feature dimension compression on the feature set of the gated nodes, remove isolated edges generated by the shielded nodes and reorganize the feature sequences of the remaining high-gain nodes to generate risk evolution feature representations for subsequent classification reasoning.

[0197] The node feature set after being shielded by the anchor importance gating signal and the topological connection information of the original anchor sequence graph are obtained as input data for graph structure reconnectivity verification.

[0198] Traverse all directed edges in the original graph and check the state identifiers of the source and target nodes connected to each edge in the current gating node feature set.

[0199] If at least one of the source or target nodes of an edge is marked as having low information gain and is subject to zero-masking, then the edge is determined to be an isolated edge or an invalid propagation path.

[0200] Remove all connections that are determined to be isolated edges from the adjacency list, sever the false semantic associations between low-contribution nodes and high-gain nodes, and prevent noise from propagating through the remaining edges.

[0201] The remaining unmasked high-gain nodes and their retained valid connection edges are reindexed and mapped to construct a compact subgraph topology, eliminating index holes caused by node removal.

[0202] Extract the updated feature vectors of all active nodes in the subgraph, arrange them in the order of the re-indexed nodes, and form a continuous feature matrix.

[0203] Global average pooling is used to aggregate the feature vectors of all active nodes in the subgraph, compressing the variable-length node sequence into a fixed-dimensional graph-level representation vector.

[0204] The aggregated graph-level representation vector is nonlinearly transformed and its dimensions compressed by a fully connected projection layer, and then mapped to a latent space dimension that matches the input layer of the classifier.

[0205] By verifying the reconnectivity of the graph structure and compressing the feature dimensions, the feature set of the gated nodes in the previous step is transformed into a risk evolution feature representation after dynamic pruning with fixed dimensions, thereby achieving efficient utilization of computing resources and lossless preservation of key risk information.

[0206] Step S6: Based on the risk evolution feature representation, perform classification reasoning calculations, output the risk level determination result of chemotherapy infusion port complications, and simultaneously extract the dominant anchor link as an interpretable temporal logic basis to generate a comprehensive early warning report. The comprehensive early warning report includes the risk level and the dominant anchor link. Specifically, it includes:

[0207] S6.1: Perform fully connected layer mapping on the risk evolution feature representation to convert the high-dimensional sparse node aggregation vector output by the graph attention network into a fixed-dimensional complication risk probability distribution vector, thereby obtaining the initial risk confidence set.

[0208] The risk evolution feature representation output from step S5 is received. This representation consists of the high-gain node feature sequence and its corresponding sparse topology that are retained after being screened by the gating mechanism.

[0209] A global pooling operation is performed on the high-gain node feature sequence, and attention-weighted summation is used to aggregate the variable-length node sequence into a fixed-dimensional graph-level representation vector, thus eliminating the problem of inconsistent node numbers caused by differences in patient treatment cycles.

[0210] A multi-layer fully connected mapping network is constructed, and the graph-level representation vector is used as input layer data. Through alternating linear transformation and nonlinear activation function processing, high-order abstract features are gradually extracted.

[0211] The first fully connected layer maps the input vector to the hidden layer space and uses the ReLU activation function to introduce nonlinear factors, thereby enhancing the model's ability to fit complex complication patterns.

[0212] The second fully connected layer further compresses the feature dimension and outputs an initial classification score equal to the number of complication categories, forming an initial risk score vector.

[0213] Dropout regularization is used to randomly drop some neuron connections between fully connected layers to prevent overfitting and improve generalization performance.

[0214] By performing linear projection processing before Softmax, the initial classification scores are transformed into an unnormalized set of initial risk confidence scores, reflecting the relative risk propensity of each complication category.

[0215] For example, the graph-level representation vector dimension is set to 128, the output dimension of the first fully connected layer is 64, and the output dimension of the second layer is 5 (corresponding to five types of complications: infection, thrombosis, occlusion, displacement, and others). The input vector is transformed by the weight matrix W1 biased by b1, then activated by ReLU, and further transformed by the weight matrix W2 biased by b2 to obtain a 5-dimensional initial risk confidence set [2.1, -0.5, 1.3, -1.2, 0.8]. This process transforms the sparse graph structure into a compact risk distribution index, realizing the mapping from topological features to probability space, and providing a quantitative basis for subsequent risk grading.

[0216] S6.2: Perform soft maximum normalization operation based on the initial risk confidence set to eliminate the influence of different dimensions on probability assessment and determine the relative likelihood value of each complication category, thereby generating a standardized complication risk probability spectrum.

[0217] Receive the initial risk confidence set output from step S6.1, which contains the original logistic regression scores or unnormalized probability values ​​corresponding to complication categories such as infection, thrombosis, and mechanical injury.

[0218] For each element in the initial risk confidence set, perform an exponential operation to map the real numbers in the interval from negative infinity to positive infinity to the positive real number space, ensuring that the likelihood values ​​of all categories are non-negative and eliminating the interference of negative values ​​on the probability distribution calculation.

[0219] Calculate the sum of all category indexed values ​​to construct a normalized denominator, which represents the total energy level of all possible complications in the current patient state, and is used for subsequent scaling.

[0220] The indexed values ​​for each category are normalized using the soft maximum function formula, and the standardized complication risk probability spectrum is calculated using the following formula:

[0221]

[0222] Among them, P i Let z be the standardized risk probability of the i-th type of complication. i Let be the initial risk confidence level for the i-th type of complication, C be the total number of complication categories, and e be the natural constant.

[0223] This formula uses the monotonically increasing property of the exponential function to amplify the difference between high-scoring and low-scoring categories, and forces the sum of the probability values ​​of all categories to equal 1 by dividing by the total, thus forming a strict probability distribution.

[0224] A standardized complication risk probability spectrum containing relative likelihood values ​​for each complication category is generated. This probability spectrum eliminates the influence of different feature dimensions on the assessment results, making various risks directly comparable.

[0225] S6.3: The standardized complication risk probability spectrum is processed by interval matching using the preset clinical risk threshold grading rules to discretize the continuous probability values ​​into risk level labels with clear clinical guidance significance, thereby outputting the risk level judgment result of chemotherapy infusion port complications.

[0226] Obtain a standardized probability spectrum of complication risks, which includes continuous probability values ​​for complication categories such as infection, thrombosis, and mechanical injury.

[0227] The system reads a pre-defined clinical risk threshold classification rule base, which defines the probability interval boundaries for low risk, medium risk, high risk, and critical risk corresponding to different complication types.

[0228] Perform an independent interval matching operation on the probability value of each type of complication in the risk probability spectrum to map continuous values ​​to a discrete risk level label space.

[0229] For a single complication category, if its probability value falls into multiple overlapping risk intervals, the risk level with the highest confidence level is selected as the determination result for that category based on the principle of maximum membership.

[0230] By considering the risk levels of all complication categories, a risk transmission priority strategy is adopted to determine the final overall risk level, with critical risk having the highest priority, followed by high risk, medium risk, and low risk.

[0231] Generate structured assessment results that include specific complication types and their corresponding risk levels, clearly identifying the dominant risk sources.

[0232] S6.4: Based on the risk level determination result, reverse trace the high-weight message passing path in the graph attention network to identify the node sequence with the highest entropy contribution to the final risk and the causal edges connected to it, thereby extracting the dominant anchor link as an interpretable temporal logic basis.

[0233] The risk level determination result output by S6.3 and the high-gain node feature set retained in S5 are received as the initial input data for reverse tracking.

[0234] Based on the normalized attention coefficient matrix cached during the S5.2 computation of the graph attention network, the weight values ​​of the directed edges connecting high-gain nodes and the final classification output layer are extracted to construct a benchmark for measuring inverse influence.

[0235] For each retained risk anchor node, calculate its feature contribution to the root node through all outgoing edges, and aggregate the local attention weights on the path using a chain rule to quantify the marginal effect of a single node on the global risk assessment:

[0236]

[0237] Wherein, C(v) i Let α be the contribution entropy value of node i. ij Let be the attention coefficient from node i to neighbor node j, and N(i) be the set of neighbors of node i.

[0238] Based on the calculated node contribution entropy values, a descending sorting operation is performed to select the Top-K key nodes whose contribution ranks in the top K percentiles, forming a candidate set of core risk sources.

[0239] By traversing the connection relationships between nodes in the core risk source candidate set, retrieving the causal dependency edges defined in the clinical knowledge graph, and connecting nodes with high contribution entropy and strong causal connections, the temporal logic chain of risk evolution is reconstructed.

[0240] The reconstructed logical chain is tested for connectivity, isolated nodes and broken paths are eliminated, and the continuity of the dominant anchor link in terms of time sequence and the rationality of the pathological mechanism are ensured.

[0241] S6.5: The risk level assessment results of the chemotherapy infusion port complications are structurally encapsulated with the dominant anchor link to generate a comprehensive early warning report containing quantitative risk rating and qualitative evolution path, thereby completing the final decision support information output for clinicians.

[0242] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0243] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0244] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the risk of complications from chemotherapy infusion ports, characterized in that, include: S1: Obtain the structured fields of the electronic medical records and the natural language text of the nursing records of patients with chemotherapy infusion ports, and use evidence-based medicine validation rules to perform risk anchor point matching and identification, identify and label risk anchor point events, and generate a set of risk anchor point events. S2: Based on the timestamps in the risk anchor event set, automatically associate multi-source local data fragments within a preset time window before and after each risk anchor event to construct an anchor context unit; S3: Perform fixed-dimensional encoding operations independently on the anchor context unit, and map the encoding results to a unified semantic space to generate an anchor semantic embedding vector; S4: Using the anchor point semantic embedding vector as a node, generate connection edges according to the causal, co-occurrence and time sequence dependencies defined in the clinical knowledge graph, and calculate the edge weights by combining the confidence and time decay factors of historical clinical sample data to construct the anchor point sequence graph; S5: Input the anchor point sequence graph into the pre-constructed graph attention network to perform message passing operation, activate the node gating mechanism according to the contribution entropy of each node to the final risk level, shield low-entropy nodes and generate risk evolution feature representation; S6: Based on the risk evolution characteristics, perform classification reasoning calculations, output the risk level determination results of chemotherapy infusion port complications, and simultaneously extract the dominant anchor link as an interpretable time logic basis to generate a comprehensive early warning report.

2. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 1, characterized in that, The multi-source local data fragments include: vital sign snapshots, infusion pump logs, mean pixel values ​​of local infrared thermal images (ROI), and keyword vectors of nurse assessment texts.

3. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 2, characterized in that, Step S3 specifically includes: The vital signs snapshot, infusion pump log, local infrared heat map ROI pixel mean, and nurse assessment text keyword vector in the anchor context unit are time-stamped and dimension-normalized to generate a standardized multimodal feature tensor. Based on the standardized multimodal feature tensor, a pre-trained cross-modal feature fusion encoder is invoked to perform channel attention weighting operations to generate a high-dimensional hybrid feature representation vector; For the high-dimensional hybrid feature representation vector, a linear transformation operation is performed using a pre-constructed clinical event semantic projection matrix to generate a preliminary clinical semantic mapping vector; Calculate the Euclidean distance between the preliminary clinical semantic mapping vector and the center of the predefined standard risk anchor prototype cluster, and optimize the feature distribution through gradient backpropagation to minimize the intra-class variance, thereby generating a standardized clinical semantic vector. The standardized clinical semantic vector is subjected to fixed-dimensional truncation and L2 norm normalization to output the anchor semantic embedding vector.

4. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 3, characterized in that, The cross-modal feature fusion encoder includes an attention weight generation network and a feature mapping layer.

5. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 4, characterized in that, The attention weight generation network consists of two fully connected layers. Based on the standardized multimodal feature tensor, a pre-trained cross-modal feature fusion encoder is invoked to perform channel attention weighting operations, generating a high-dimensional hybrid feature representation vector, including: The different modal features of the standardized multimodal feature tensor are mapped to a unified dimensional space and the global average pooling value of each modal channel is calculated to generate a modal importance descriptor. The first fully connected network takes the modality importance descriptor as input, expands the input dimension to a 16-dimensional hidden layer space, and after activation by the ReLU activation function, outputs the first layer interaction feature vector. The second fully connected network receives the interaction feature vector from the first layer, compresses the first layer interaction feature vector back into a 4-dimensional space with the same number of modalities, and generates the attention weight coefficients for each modal channel after activation by the Sigmoid activation function. The attention weight coefficients are multiplied element-wise with each modality feature vector in the standardized multimodal feature tensor to generate a recalibrated multimodal feature vector. Perform a concatenation operation on the recalibrated multimodal feature vectors to form a fused feature matrix; The feature mapping layer takes the fused feature matrix as input, maps the fused feature matrix to a high-dimensional hidden layer space through linear projection, and activates it through the ReLU activation function to generate the high-dimensional hybrid feature representation vector.

6. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 4, characterized in that, The row dimension of the clinical event semantic projection matrix is ​​consistent with the dimension of the high-dimensional hybrid feature representation vector, and the column dimension is the dimension of the semantic subspace of chemotherapy infusion port complications; the clinical event semantic projection matrix is ​​obtained by optimizing and learning the historical clinical sample data during the pre-training stage and fixing the weight values.

7. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 1, characterized in that, The risk anchor event set contains a unique ID and a type label.

8. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 7, characterized in that, Step S4 specifically includes: Based on the anchor semantic embedding vector, a graph node initialization process is performed, mapping each anchor semantic embedding vector to an independent graph node in the anchor sequence graph, and assigning it a unique ID and type label to generate an initial graph node set. Using a pre-constructed clinical knowledge graph, semantic association analysis is performed on the initial graph node set. Based on the causal relationships, co-occurrence relationships, and time sequence dependency rules defined in the clinical knowledge graph, potential connection paths between any two graph nodes in the initial graph node set are scanned and identified, generating a candidate connection edge set. The statistical confidence level of the candidate connection edge set is calculated based on historical clinical sample data. The co-occurrence frequency and conditional probability of each candidate connection edge in historical cases are extracted. The time decay factor is calculated by combining the timestamp of the current risk anchor event. The statistical confidence level and the time decay factor are weighted and fused to generate the dynamic edge weight value of each candidate connection edge. Based on the dynamic edge weight value, a threshold filtering process is performed on the candidate connection edge set to remove redundant candidate connection edges whose dynamic edge weight value is lower than a preset threshold, thereby generating a sparse connection edge set. The initial graph node set and the sparse connection edge set are topologically assembled to establish directed connection relationships between nodes and bind corresponding dynamic edge weight values ​​to construct the anchor point sequence graph.

9. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 1, characterized in that, Step S5 specifically includes: Based on the anchor semantic embedding vector and edge weight matrix in the anchor sequence graph, the graph structure topology data is generated by constructing and processing based on the sparse adjacency list, and the sparse graph topology representation is output. For each target node and its set of neighboring nodes in the sparse graph topology representation, feature aggregation operation under multi-head self-attention mechanism is performed. Normalized attention coefficients are generated by calculating the dot product similarity between the query vector and the key vector, and the node update feature vector is output. Based on the mapping relationship between the node update feature vector and the global risk classification hyperplane, the feature distribution dispersion of each node is quantified using the Shannon entropy calculation formula to generate the node contribution entropy value. Based on the comparison result between the node contribution entropy value and the preset dynamic threshold, a binarization mask generation operation is performed to construct an anchor point importance gating signal. The anchor point importance gating signal is used to perform zero-masking processing on the node update feature vector corresponding to the low entropy value node, and the gating node feature set after preliminary screening is output. The graph structure reconnectivity verification and feature dimension compression processing are performed on the feature set of the gated nodes. Isolated edges generated by the shielded nodes are removed and the feature sequences of the remaining high-gain nodes are recombined to generate the risk evolution feature representation.

10. The method for predicting the risk of complications from chemotherapy infusion ports according to claim 1, characterized in that, The comprehensive early warning report includes risk level and dominant anchor link.

Images