A stricture intestinal obstruction risk prediction method based on multi-modal data fusion
By constructing a multimodal clinical evidence atlas framework and generating a traceable modeling dataset, the problem of difficulty in aligning and fusing multimodal evidence was solved, thereby improving the credibility and acceptability of strangulated intestinal obstruction risk prediction and enhancing the stability and interpretability of the prediction model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN INTEGRATED TRADITIONAL CHINESE & WESTERN MEDICINE HOSPITAL (TIANJIN NANKAI HOSPITAL)
- Filing Date
- 2026-03-05
- Publication Date
- 2026-06-19
AI Technical Summary
In the prediction of strangulated intestinal obstruction risk, existing technologies struggle to align and integrate multimodal evidence, resulting in a lack of traceability and interpretability in the prediction results, thus leading to insufficient acceptability of the results.
By establishing a multimodal clinical evidence atlas skeleton, performing structured and standardized coding of multimodal data, extracting acoustic parameters, generating a clinical structured feature table, constructing co-occurrence interaction features and evidence path features, forming a traceable modeling dataset, constructing a strangulated intestinal obstruction risk prediction model, and outputting a joint evidence summary.
It achieves comparability and availability of multimodal data on the same timeline, reduces interference from time mismatch and semantic drift, improves the credibility and acceptability of risk prediction, and enhances the stability and generalization ability of the model.
Smart Images
Figure CN122245748A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart healthcare technology, and in particular to a method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion. Background Technology
[0002] Emergency assessment of suspected bowel obstruction typically relies on a comprehensive interpretation of medical history, symptoms, signs, laboratory test results, and CT imaging findings. In some cases, bowel sounds are auscultated to reflect the dynamic state. In recent years, the structuring of electronic medical records, the standardization of imaging terminology, and the extraction of digital auscultation and acoustic parameters have gradually matured. Multi-source data fusion and risk stratification methods based on medical informatics have been continuously applied in clinical decision support.
[0003] Existing methods often construct scores or predictions using single-modal or loosely parallel multimodal variables, which are easily affected by data heterogeneity, inconsistencies between unit and time expression, mismatch of consultation time windows, and difficulty in synchronously annotating imaging and bowel sound evidence. The predicted output often lacks a correspondence with examination number, image sequence index, and bowel sound event fragment index, resulting in insufficient traceability and interpretability of the results, making it difficult to support review and quality control, and also limiting the stable implementation of multimodal combined evidence in the early identification of strangulated intestinal obstruction. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion to solve the problems of difficulty in aligning and fusing multimodal evidence for intestinal obstruction and difficulty in retrospectively interpreting prediction results.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0007] This invention provides a method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion, comprising:
[0008] A case index was established for patients suspected of intestinal obstruction, and medical history, symptoms and signs, laboratory test indicators, CT imaging signs and bowel sounds were collected simultaneously to generate multimodal raw data packages and establish a multimodal clinical evidence atlas skeleton.
[0009] The multimodal raw data packets are structured and standardized, and the bowel sound audio is denoised, segmented, and the event is located. Acoustic parameters are extracted, and clinical structured feature tables, imaging feature tables, and bowel sound acoustic feature tables are generated and written into the multimodal clinical evidence map skeleton to form an evidence map node set.
[0010] The three feature tables are aligned in a multimodal manner according to the time window of the visit, and co-occurrence interaction features and evidence path features are constructed in the evidence map node set to generate a fusion feature matrix and pair it with the outcome record corresponding to the case index to form a traceable modeling dataset.
[0011] Logistic regression training was performed on the traceable modeling dataset to construct a risk prediction model for strangulated intestinal obstruction, and a mapping table was established from model parameters to evidence nodes and evidence paths to form a model package.
[0012] The model package is invoked to generate risk probabilities for new patients, and a joint evidence summary is output based on the mapping table.
[0013] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, wherein: the generation of multimodal raw data packets specifically includes:
[0014] Generate a case index for suspected intestinal obstruction patients and bind the patient's identifier and visit timestamp;
[0015] Retrieve electronic medical records and triage records to extract medical history, symptoms and signs, and form a medical history, symptoms and signs record;
[0016] Retrieve test records, extract test indicators, and generate laboratory test records;
[0017] Retrieve CT image records to extract CT image features and form CT image feature records;
[0018] Collect bowel sound audio from standard body surface auscultation sites to form a record of bowel sound audio segments;
[0019] By summarizing medical history, symptoms and signs records, laboratory test records, imaging signs records and audio clip records, and binding them to the case index, a multimodal raw data package is obtained.
[0020] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, the step of establishing a multimodal clinical evidence atlas framework specifically includes:
[0021] Based on the case index, evidence nodes are generated for medical history, symptoms and signs records, laboratory test records, CT imaging signs records and bowel sound audio records, and node type and timestamp attributes are written;
[0022] Establish the relationship between inspection number and timestamp for evidence nodes to form a multimodal clinical evidence atlas skeleton.
[0023] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, the step of structuring and standardizing the original multimodal data packets specifically involves:
[0024] The medical history, symptoms and signs records are extracted in a structured manner according to predefined fields and the values are uniformly encoded to generate clinical structured entries;
[0025] Standardize the units of measurement and collection time for laboratory test records, and generate standardized test items.
[0026] The CT image feature records are mapped to terms according to the radiology terminology dictionary and the image sequence index is retained to generate structured image feature entries;
[0027] The clinical structured items, standardized test items, and structured imaging features items are summarized to generate structured and standardized coding results.
[0028] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, wherein: the formation of the evidence map node set specifically includes:
[0029] Denoising and segmentation of bowel sound audio and event localization are performed, and the start and end times of events are recorded. Waveform morphology, amplitude changes, frequency characteristics and event intervals are extracted to form an acoustic feature table of bowel sounds.
[0030] The clinical structured entries are mapped to clinical feature fields according to the case index to form a clinical structured feature table, and the structured imaging feature entries are mapped to imaging feature fields according to the case index to form an imaging feature table.
[0031] The clinical structured feature table, imaging feature table, and bowel sound acoustic feature table are written into the multimodal clinical evidence atlas skeleton, and concept nodes are generated for each feature. The source index relationship between concept nodes and evidence nodes is established to obtain the evidence atlas node set.
[0032] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, wherein: the generation of the fusion feature matrix specifically comprises:
[0033] The consultation time window is divided according to the consultation timestamp of the case index, and the time window identifier is written into the three types of feature tables. The three types of feature tables are aligned according to the time window identifier to generate the basic feature vector.
[0034] The co-occurrence relationship between image feature concept nodes and bowel sound event concept nodes is extracted from the evidence map node set according to the same time window relationship, and encoded as co-occurrence interaction features;
[0035] Evidence paths are generated based on the continuous connection relationships between symptom concept nodes, test indicator concept nodes, CT imaging sign concept nodes, and bowel sound event concept nodes, and encoded as evidence path features.
[0036] The basic feature vectors, co-occurrence interaction features, and evidence path features are aggregated to generate a fused feature matrix.
[0037] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, wherein: the formation of a traceable modeling dataset specifically includes:
[0038] Based on the case index, generate outcome records and write the outcome determination timestamp; write the fusion feature matrix into the case index to form feature index records.
[0039] Paired records are generated for outcome records and feature index records and written into the relationship between the event fragment index and source index of the image sequence index, bowel sound audio records, and form a tracing chain;
[0040] The fused feature matrix, paired records, and traceability chains are combined to form a traceable modeling dataset.
[0041] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, the construction of the strangulated intestinal obstruction risk prediction model specifically includes:
[0042] The traceability modeling dataset is divided into training and validation data according to the case index, and paired records and traceability chains are retained;
[0043] Perform logistic regression fitting on the training data to generate a list of model parameters and feature fields;
[0044] Perform model inference on the verification data to generate verification result records and associate them with the traceability chain;
[0045] The model parameters, feature field list, and validation result records are summarized to form a strangulated intestinal obstruction risk prediction model.
[0046] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, wherein: the formation of the model package specifically includes:
[0047] Generate a parameter index for the model parameters based on the feature field list and write it to the parameter index record;
[0048] Based on the parameter index records, retrieve the evidence graph node set to locate the concept nodes and evidence paths, and locate the concept nodes to the evidence nodes through the source index relationship to form a location record;
[0049] The model parameters and location records are summarized to generate a mapping table from model parameters to evidence nodes and evidence paths;
[0050] The model parameters, feature field list, mapping table, and validation result records are encapsulated to form a model package.
[0051] As a preferred embodiment of the strangulated intestinal obstruction risk prediction method based on multimodal data fusion described in this invention, the output of the joint evidence summary specifically includes:
[0052] New multimodal raw data packages are generated for new patients, and the evidence graph node set and fusion feature matrix are updated.
[0053] Read the model parameters and feature field list from the model package, generate inferred feature vectors from the fused feature matrix according to the feature field list, and perform logistic regression inference on the inferred feature vectors to generate risk probabilities;
[0054] The evidence nodes and evidence paths associated with risk probability are located by mapping table, and the image sequence index and bowel sound event fragment index are extracted to output a joint evidence summary.
[0055] The beneficial effects of this invention are as follows: By unifying and consistently expressing multimodal data through case indexing, medical history, symptoms, signs, laboratory test indicators, CT imaging features, and bowel sound audio can form a comparable and usable data foundation for the same patient and at the same time of visit, reducing the interference of time mismatch and semantic drift of records from different sources on risk assessment; by establishing source index relationships between features and original evidence through evidence atlas, the risk probability is no longer a black box output, but can be located to the examination number, image sequence index, and bowel sound event fragment index to form a joint evidence summary, which facilitates rapid clinical review, accountability, and quality control, significantly improving the credibility and acceptability of results; by explicitly expressing the co-occurrence relationship of multimodal evidence and the connection relationship of clinical reasoning through co-occurrence interaction features and evidence path features, the model can use cross-modal complementary information to suppress misjudgments caused by single-modal noise and random abnormalities, thereby improving the stability and generalization ability of strangulated intestinal obstruction risk indication. Attached Figure Description
[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0057] Figure 1 This is a flowchart of a method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion.
[0058] Figure 2 This is a flowchart of multimodal data acquisition and preprocessing.
[0059] Figure 3 A flowchart for feature fusion and traceable dataset construction.
[0060] Figure 4 This is a flowchart of the model training and application process. Detailed Implementation
[0061] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0062] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0063] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0064] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion, comprising the following steps:
[0065] S1. Establish a case index for suspected intestinal obstruction patients, and simultaneously collect medical history, symptoms and signs, laboratory test indicators, CT imaging signs and bowel sounds, generate multimodal raw data packages and establish a multimodal clinical evidence atlas skeleton.
[0066] S1.1. Generate a case index for suspected intestinal obstruction patients in the emergency information environment, and write the patient identifier and the consultation timestamp into the case index to obtain the case index and consultation timestamp binding record. The case index and consultation timestamp binding record is used to locate medical history, symptoms and signs records, laboratory test records, CT image signs records and bowel sound audio segment records.
[0067] S1.2. Based on the case index and the timestamp of the visit, retrieve the electronic medical record and triage record, extract the key points of the medical history, the start and end time of symptoms, the chief complaint information such as pain and vomiting, and the physical examination findings, and write them into the medical history, symptoms and signs record according to the fields to obtain the medical history, symptoms and signs record with the case index and the medical history, symptoms and signs record with the case index.
[0068] Based on the case index and the timestamp of the visit, the test records are retrieved, and laboratory test indicators such as white blood cell count, coagulation-related indicators, electrolytes, blood urea nitrogen and D-dimer are extracted and written into the laboratory test records. At the same time, the test number and collection timestamp of the test record are written into the laboratory test records, resulting in laboratory test records with test number and timestamp attributes.
[0069] Based on the case index and the timestamp of the visit, the imaging records are retrieved and the plain or enhanced CT images are reviewed. Ascites, mesenteric and intestinal fluid accumulation, spiral sign, concentric circle sign, small intestinal fecal sign and intestinal wall edema, etc., are extracted and written into the CT imaging sign record. At the same time, the corresponding examination number, examination timestamp and image sequence index of the imaging record are written into the CT imaging sign record, resulting in CT imaging sign records with examination number, timestamp attribute and image sequence index, and CT imaging sign records with examination number, timestamp attribute and image sequence index.
[0070] Based on the case index and the visit timestamp, bowel sounds are collected using an electronic stethoscope or an abdominal acoustic acquisition device at standard body surface auscultation sites. The start and end times of the collection are recorded, and the bowel sound audio is segmented according to the collection time period. The bowel sound audio segments are recorded, and the case index and timestamp attributes are written into the bowel sound audio segment records to obtain bowel sound audio segment records with timestamp attributes.
[0071] S1.3. Summarize the medical history, symptoms and signs records with case index identifiers, laboratory test records with test numbers and timestamp attributes, CT image signs records with test numbers, timestamp attributes and image sequence indexes, and bowel sound audio segment records with timestamp attributes, and write the case index into the summary results to obtain the multimodal raw data package.
[0072] Based on the case index and the multimodal raw data package, evidence nodes are generated for the medical history, symptoms and signs records, laboratory test records, CT image signs records and bowel sound audio segments. The evidence nodes are written with node type, timestamp attribute and case index. The test number association is written for the evidence nodes corresponding to laboratory test records with the same test number and the evidence nodes corresponding to CT image signs records. The timestamp association is written for the evidence nodes with sequential timestamps, thus obtaining the multimodal clinical evidence atlas skeleton.
[0073] S2. The multimodal raw data packets are structured and standardized, and the bowel sound audio is denoised, segmented, and the event is located. Acoustic parameters are extracted, and clinical structured feature tables, imaging feature tables, and bowel sound acoustic feature tables are generated and written into the multimodal clinical evidence map skeleton to form an evidence map node set.
[0074] S2.1. Extract medical history, symptoms and signs records, laboratory test records, CT imaging records and bowel sound audio segments from the multimodal raw data package. Perform structured extraction on the medical history, symptoms and signs records according to predefined fields. Write the key points of the medical history, the start and end time of symptoms, chief complaint information such as pain and vomiting, and physical examination findings into the corresponding fields. Perform unified value encoding and missing value marking on the field values to obtain clinical structured entries.
[0075] To further explain, the predefined fields are compiled from the fixed information items that can be repeated in the electronic medical record and triage record and correspond to the medical history, symptoms and signs records. The predefined fields include the medical history key points field, the symptom start and end time field, the pain field, the vomiting field, and the physical examination findings field. The field names and value types of the predefined fields are consistent with the standard input items in the electronic medical record and triage record. The structured extraction parses the corresponding input items or text fragments in the electronic medical record and triage record item by item according to the predefined fields, and writes the parsed values into the corresponding fields of the medical history, symptoms and signs records. The symptom start and end time field uses the consultation timestamp to unify the expression to the same time. Fields that cannot be parsed are marked with missing values, thereby obtaining the clinical structured entries.
[0076] Based on the laboratory test records, the units of measurement and the expression of collection time for the test indicators were standardized. Laboratory test indicators such as white blood cell count, coagulation-related indicators, electrolytes, blood urea nitrogen and D-dimer were written into standardized fields, and standardized value encoding was performed on numerical test indicators to obtain standardized test items.
[0077] To further explain, the standardized value encoding expression for numerical inspection indicators is as follows:
[0078] ;
[0079] in, This represents the raw values of the test indicators in the laboratory test records. This represents the mean of the set of values corresponding to the same inspection indicator in the structured and standardized coding results. This represents the standard deviation of the set of values corresponding to the same inspection indicator in the structured and standardized coding results. This indicates the standardized values in the standardized test items, used to unify the expression of values for test indicators with different dimensions.
[0080] S2.2. The CT image feature records are mapped to terms such as ascites, mesenteric and intestinal fluid accumulation, spiral sign, concentric circle sign, small intestinal fecal sign and intestinal wall edema according to the imaging terminology dictionary, and the image sequence index, examination number and timestamp attributes are retained to obtain structured image feature entries.
[0081] To further explain, the medical imaging terminology dictionary is compiled from the names of CT imaging features such as ascites, mesenteric and intestinal fluid accumulation, spiral sign, concentric sign, small intestinal fecal sign and intestinal wall edema that appear in CT imaging feature records. Standardized terminology entries and synonym entries are set for each CT imaging feature. The terminology entries in the medical imaging terminology dictionary are consistent with the terminology used in CT imaging feature records, and are used to merge different expressions into the same standardized terminology.
[0082] The terminology mapping matches the sign names in the CT image sign record item by item according to the imaging terminology dictionary. The matching rules adopt the priority matching of standardized term entries and the auxiliary matching of synonym expression entries. When the matching is successful, the matched standardized term is written into the structured image sign entry and the original sign name is retained. When the matching fails, the original sign name is directly written into the spare field of the structured image sign entry and the image sequence index, examination number and timestamp attributes are retained, thus obtaining the structured image sign entry.
[0083] The clinical structured entries, standardized test entries, and structured imaging signs entries are compiled and written into the case index to form structured and standardized coding results.
[0084] Based on the recorded bowel sound audio segments, the bowel sound audio is denoised and the denoised bowel sound audio segments are retained. The denoised bowel sound audio segments are segmented and the events are located. The start and end times of the events are recorded and bowel sound event segments are recorded. The waveform morphology, amplitude changes, frequency characteristics and event intervals of the bowel sound event segments are extracted segment by segment to form a bowel sound acoustic feature table.
[0085] To further explain, the root mean square expression for the amplitude change is:
[0086] ;
[0087] in, This indicates the first segment of the bowel sound event record. The amplitude of each sampling point This indicates the number of sampling points recorded in the bowel sound event segment. The root mean square amplitude in the bowel sound acoustic characteristic table is used to characterize the amplitude intensity of a recorded bowel sound event segment.
[0088] Segmented processing divides the bowel sound audio segments into continuous time periods based on the timestamp attribute. Within each time period, a short-time energy sequence is calculated for the bowel sound audio amplitude sequence. Continuous intervals where the short-time energy sequence exceeds the short-time energy level of the background segment are identified as bowel sound event segment records. Event localization uses the start time of the continuous interval as the event start time and the end time as the event end time, and assigns an event sequence number to each continuous interval, thus obtaining bowel sound event segment records with event start and end times.
[0089] S2.3. Based on the structured and standardized coding results, the clinical structured entries and standardized test entries are mapped to clinical feature fields according to the case index to form a clinical structured feature table. The structured imaging sign entries are mapped to imaging feature fields according to the case index to form an imaging feature table.
[0090] Based on the multimodal clinical evidence map skeleton, the clinical structured feature table, imaging feature table, and bowel sound acoustic feature table are written into the multimodal clinical evidence map skeleton one by one. Concept nodes are generated for each feature and case index and timestamp attributes are written into them. Based on the evidence nodes, the source index relationship between concept nodes and evidence nodes is established in the multimodal clinical evidence map skeleton. Based on the timestamp attributes, the same visit time window relationship between concept nodes is established in the multimodal clinical evidence map skeleton, resulting in the evidence map node set.
[0091] S3. Perform multimodal alignment of the three feature tables according to the consultation time window, and construct co-occurrence interaction features and evidence path features in the evidence map node set to generate a fusion feature matrix and pair it with the outcome record corresponding to the case index to form a traceable modeling dataset.
[0092] S3.1. Based on the case index and the record binding time stamp, determine the time stamp of the visit. Write the time window identifier for each record in the clinical structured feature table, the imaging feature table, and the bowel sound acoustic feature table according to the time difference between the record time stamp and the visit time stamp. In the evidence map node set, retain the same visit time window relationship for concept nodes with the same time window identifier. This will result in the clinical structured feature table, the imaging feature table, the bowel sound acoustic feature table, and the evidence map node set with the time window identifier.
[0093] The expression for determining the appointment time window identifier is:
[0094] ;
[0095] in, Indicates the time window identifier. The timestamps represent the records in the clinical structured features table, imaging features table, and bowel sound acoustic features table. This indicates that the case index is bound to the visit timestamp in the record. This indicates the duration of the consultation window. This indicates a floor operation, used to assign the recorded timestamp to the corresponding medical visit time window.
[0096] S3.2. Align the clinical structured feature table, the imaging feature table, and the bowel sound acoustic feature table with the time window identifier according to the time window identifier. Merge the clinical feature field, imaging feature field, and acoustic parameter field under the same time window identifier according to the field name to generate a time window-level feature vector. Concatenate the time window-level feature vectors corresponding to multiple time window identifiers under the case index in chronological order to generate a basic feature vector.
[0097] In the evidence map node set with time window identifiers, the co-occurrence combinations of image sign concept nodes and bowel sound event concept nodes are located according to the same time window relationship. The co-occurrence combinations are encoded according to the image sign concept node type and the bowel sound event concept node type to generate co-occurrence interaction features. The co-occurrence combinations are associated with the image sequence index, the event segment index of the bowel sound audio record, and the source index relationship to obtain the co-occurrence interaction features corresponding to the case index.
[0098] The existence expression for co-occurrence interaction features is:
[0099] ;
[0100] in, This indicates the existence value of the co-occurrence interaction feature, which can be either 0 or 1. This indicates the existence value of image feature concept nodes within the same time window. This indicates the existence value of the bowel sound event concept node within the same time window, used to characterize the co-occurrence relationship between the imaging feature concept node and the bowel sound event concept node within the same medical visit time window.
[0101] It should be noted that when The value is 1 when at least one image feature concept node exists within the same time window, and 0 when no image feature concept node exists within the same time window. The value of c is 1 when at least one bowel sound event concept node exists within the same time window, and 0 when no bowel sound event concept node exists within the same time window. and When both values are 1, the value is 1. and When at least one of the values is 0, the value is 0.
[0102] S3.3. Locate the symptom concept node, test indicator concept node, CT image sign concept node, and bowel sound event concept node in the evidence map node set with time window identifiers. Generate a continuous connection relationship from the symptom concept node to the test indicator concept node to the CT image sign concept node to the bowel sound event concept node according to the same consultation time window relationship and timestamp association relationship. Encode the continuous connection relationship according to the node type sequence and time window identifier to generate evidence path features. Associate the evidence path features with the image sequence index, the event segment index of the bowel sound audio record, and the source index relationship to obtain the evidence path features corresponding to the case index.
[0103] The basic feature vector, co-occurrence interaction feature, and evidence path feature corresponding to the case index are merged according to the case index to form a fusion feature matrix, and the fusion feature matrix is written into the case index to form a feature index record.
[0104] Based on the case index, the discharge diagnosis record is retrieved, and the judgment result of strangulated intestinal obstruction in the discharge diagnosis conclusion is extracted and written into the outcome record. Based on the case index, the intraoperative observation record is retrieved, and the judgment result of intestinal ischemia-necrosis, intestinal volvulus, and strangulation related manifestations of incarcerated hernia in the intraoperative observation conclusion is extracted and written into the outcome record. Based on the case index, the imaging follow-up record is retrieved, and the judgment result of intestinal wall ischemia or strangulated intestinal obstruction in the imaging follow-up conclusion is extracted and written into the outcome record. The outcome record is written with the outcome judgment timestamp and bound to the case index to obtain the outcome record corresponding to the case index. Paired records are generated for the outcome record corresponding to the case index and the feature index record and written into the relationship between the image sequence index, the event segment index of the bowel sound audio record and the source index to form a traceability chain. The fused feature matrix, paired records and traceability chain are summarized to form a traceable modeling dataset.
[0105] S4. Perform logistic regression training on the traceable modeling dataset to construct a strangulated intestinal obstruction risk prediction model, and establish a mapping table from model parameters to evidence nodes and evidence paths to form a model package.
[0106] S4.1. Extract the fusion feature matrix, paired records, and traceability chains from the traceable modeling dataset, and divide the traceable modeling dataset according to the case index to obtain training data and validation data. The training data and validation data retain the corresponding paired records and corresponding traceability chains, respectively, to obtain training data, validation data, training paired records, validation paired records, training traceability chains, and validation traceability chains. The training data is used for logistic regression fitting to generate model parameters and a list of feature fields, and the validation data is used for logistic regression inference to generate validation result records.
[0107] The fusion feature matrix is extracted from the training data as feature matrix samples, and the corresponding outcome records of the case index are extracted from the training paired records as label samples. The feature matrix samples and label samples are aligned according to the case index to form a training sample set. Each sample in the training sample set consists of a row of feature values corresponding to the case index in the fusion feature matrix and an outcome record corresponding to the case index in the training paired records. The outcome record is written with a binary label value, and the strangulated intestinal obstruction outcome judgment result is used as the source of the binary label value. The binary label value is 1 when the outcome record contains a strangulated intestinal obstruction outcome judgment result of affirmative, and 0 when the outcome record contains a strangulated intestinal obstruction outcome judgment result of negative. Logistic regression is performed on the training sample set to obtain model parameters, and a list of feature fields is extracted from the field name set of the fusion feature matrix and written into the corresponding field names of the model parameters to obtain the model parameters and the list of feature fields.
[0108] To further explain, logistic regression fitting involves arranging the training sample set into a feature matrix according to the list of feature fields, arranging the binary label values into a label vector, and performing iterative updates on the model parameters under the likelihood function of logistic regression until the change in likelihood values between two adjacent iterations satisfies the convergence condition, thereby obtaining a set of model parameter values that correspond one-to-one with the list of feature fields.
[0109] The iterative update employs either the first-order gradient method or the second-order Newton method based on the likelihood function, and outputs a set of model parameter values and intercept parameter values that correspond one-to-one with the field names in the feature field list.
[0110] Furthermore, the convergence condition is jointly determined by the training objective of the training sample set and the iterative update process. Before the logistic regression fitting begins, the maximum number of iterations and the convergence threshold are set. At the end of each iteration, the likelihood value of the current iteration is calculated and the difference is taken with the likelihood value of the previous iteration. When the absolute value of the likelihood difference is lower than the convergence threshold for several consecutive iterations, convergence is determined and the iteration stops. When the maximum number of iterations is reached and the condition is not met consecutively, convergence is determined and the iteration stops, and the set of model parameter values corresponding to the last iteration is output. The convergence threshold and the maximum number of iterations are determined by the size of the training data and the number of feature fields. The larger the size of the training data and the number of feature fields, the larger the maximum number of iterations and the smaller the convergence threshold, thus ensuring that the change of the model parameter value set tends to be stable in repeated iterations.
[0111] S4.2. Extract the fusion feature matrix from the validation data and generate validation feature vectors according to the feature field list. Perform logistic regression inference on the validation feature vectors to obtain the risk probability. Write the risk probability and case index into the validation result record. Associate the validation result record with the validation traceability chain according to the case index to obtain the validation result record associated with the traceability chain.
[0112] To further explain, the expression for the risk probability obtained by performing logistic regression inference is as follows:
[0113] ;
[0114] in, Indicates the probability of strangulated intestinal obstruction. Represents an exponential function. This represents the linear prediction value of logistic regression. The linear prediction value is calculated by the model parameters and the inferred feature vector according to the correspondence of the feature field list. It is used to convert the linear prediction value into a risk probability.
[0115] To further explain, the calculation of the linear prediction value involves aligning the field names of the validation field value sequence with the field names of the model parameters according to the feature field list, multiplying the validation field value corresponding to each field name with the model parameter value corresponding to the field name, summing the product results corresponding to all field names, and adding the intercept parameter value to the summation result to obtain the linear prediction value.
[0116] The model parameters, feature field list, and verification result records associated with the traceability chain are summarized to form a strangulated intestinal obstruction risk prediction model. The strangulated intestinal obstruction risk prediction model is used to generate parameter index records and to establish a mapping table from model parameters to evidence nodes and evidence paths.
[0117] S4.3. Generate parameter indexes for each model parameter based on the feature field list, and bind each parameter index to the corresponding field name in the feature field list to obtain parameter index records.
[0118] Based on the parameter index records, the evidence graph node set is retrieved one by one, and the concept nodes that match the field names in the feature field list are located. At the same time, the node type sequence and time window identifier encoding results of the evidence path features corresponding to the field names are located in the evidence graph node set, so as to obtain the concept node location results and the evidence path location results.
[0119] The concept node location result is used to locate the evidence node through the source index relationship. The evidence path location result is written into the image sequence index and the event fragment index of the bowel sound audio record corresponding to the evidence path to obtain the location record.
[0120] The model parameters and location records are summarized according to the parameter index correspondence to generate a mapping table from model parameters to evidence nodes and evidence paths. The model parameters, feature field list, model parameter to evidence node and evidence path mapping table and verification result records associated with the traceability chain are encapsulated to form a model package.
[0121] S5. Call the model package to generate risk probabilities for new patients and output a joint evidence summary based on the mapping table.
[0122] S5.1. In the emergency information environment, perform the case index establishment and multimodal acquisition process for new patients to obtain the multimodal raw data package and the multimodal clinical evidence map skeleton corresponding to the new patients, and retain the examination number association relationship and timestamp association relationship in the multimodal clinical evidence map skeleton.
[0123] The process involves structured and standardized encoding, bowel sound audio denoising and segmentation, event localization, acoustic parameter extraction, and feature table generation for the multimodal raw data packets corresponding to the new patients. This yields a clinical structured feature table, an imaging feature table, and a bowel sound acoustic feature table for the new patients. These feature tables are then written into the multimodal clinical evidence map skeleton for the new patients to form the evidence map node set for the new patients.
[0124] S5.2. Perform the procedures from S3.1 to S3.3 on the clinical structured feature table, the imaging feature table, and the bowel sound acoustic feature table corresponding to the new patient, including writing the consultation time window identifier, multimodal alignment, co-occurrence interaction feature construction, and evidence path feature construction, to obtain the fusion feature matrix corresponding to the new patient. Then, write the fusion feature matrix into the case index to form the feature index record corresponding to the new patient.
[0125] The model parameters and feature field list are read from the model package. The field value sequence is extracted from the fusion feature matrix corresponding to the new patient according to the field name order of the feature field list to form the inferred feature value sequence. Logistic regression is performed on the inferred feature value sequence to generate the risk probability corresponding to the new patient. The risk probability and case index are written into the risk probability record.
[0126] The model parameter entries corresponding to the risk probability records are located by mapping the feature field list. The evidence node location results and evidence path location results are located by the model parameter entries. The image sequence index and the event fragment index of the bowel sound audio record are extracted based on the evidence node location results. The node type sequence, time window identifier and the relationship of the same visit time window are extracted based on the evidence path location results. The image sequence index, the event fragment index of the bowel sound audio record, the node type sequence and the time window identifier are combined to generate a joint evidence summary.
[0127] In summary, this invention achieves a unified and consistent representation of multimodal data through a case index, enabling medical history, symptoms, signs, laboratory test indicators, CT imaging findings, and bowel sound audio to form a comparable and usable data foundation for the same patient and at the same time of visit. This reduces the interference of temporal mismatch and semantic drift from different sources on risk assessment. Furthermore, by establishing source index relationships between features and original evidence through an evidence graph, risk probability is no longer a black-box output but can be located at the examination number, image sequence index, and bowel sound event fragment index, forming a joint evidence summary. This facilitates rapid clinical review, accountability, and quality control, significantly improving the credibility and acceptability of results. Finally, by explicitly expressing the co-occurrence relationships of multimodal evidence and the sequential relationships of clinical reasoning through co-occurrence interaction features and evidence path features, the model can utilize complementary cross-modal information to suppress misjudgments caused by single-modal noise and random anomalies, thereby improving the stability and generalization ability of strangulated intestinal obstruction risk indications.
[0128] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion, characterized in that: include, A case index was established for patients suspected of intestinal obstruction, and medical history, symptoms and signs, laboratory test indicators, CT imaging signs and bowel sounds were collected simultaneously to generate multimodal raw data packages and establish a multimodal clinical evidence atlas skeleton. The multimodal raw data packets are structured and standardized, and the bowel sound audio is denoised, segmented, and the event is located. Acoustic parameters are extracted, and clinical structured feature tables, imaging feature tables, and bowel sound acoustic feature tables are generated and written into the multimodal clinical evidence map skeleton to form an evidence map node set. The three feature tables are aligned in a multimodal manner according to the time window of the visit, and co-occurrence interaction features and evidence path features are constructed in the evidence map node set to generate a fusion feature matrix and pair it with the outcome record corresponding to the case index to form a traceable modeling dataset. Logistic regression training was performed on the traceable modeling dataset to construct a risk prediction model for strangulated intestinal obstruction, and a mapping table was established from model parameters to evidence nodes and evidence paths to form a model package. The model package is invoked to generate risk probabilities for new patients, and a joint evidence summary is output based on the mapping table.
2. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 1, characterized in that: The generation of the multimodal raw data packet specifically involves: Generate a case index for suspected intestinal obstruction patients and bind the patient's identifier and visit timestamp; Retrieve electronic medical records and triage records to extract medical history, symptoms and signs, and form a medical history, symptoms and signs record; Retrieve test records, extract test indicators, and generate laboratory test records; Retrieve CT image records to extract CT image features and form CT image feature records; Collect bowel sound audio from standard body surface auscultation sites to form a record of bowel sound audio segments; By summarizing medical history, symptoms and signs records, laboratory test records, imaging signs records and audio clip records, and binding them to the case index, a multimodal raw data package is obtained.
3. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 1, characterized in that: The establishment of the multimodal clinical evidence atlas framework specifically involves: Based on the case index, evidence nodes are generated for medical history, symptoms and signs records, laboratory test records, CT imaging signs records and bowel sound audio records, and node type and timestamp attributes are written; Establish the relationship between inspection number and timestamp for evidence nodes to form a multimodal clinical evidence atlas skeleton.
4. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 1, characterized in that: The structured and standardized encoding of the multimodal raw data packets specifically involves: The medical history, symptoms and signs records are extracted in a structured manner according to predefined fields and the values are uniformly encoded to generate clinical structured entries; Standardize the units of measurement and collection time for laboratory test records, and generate standardized test items. The CT image feature records are mapped to terms according to the radiology terminology dictionary and the image sequence index is retained to generate structured image feature entries; The clinical structured items, standardized test items, and structured imaging features items are summarized to generate structured and standardized coding results.
5. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 4, characterized in that: The set of nodes forming the evidence graph is specifically as follows: Denoising and segmentation of bowel sound audio and event localization are performed, and the start and end times of events are recorded. Waveform morphology, amplitude changes, frequency characteristics and event intervals are extracted to form an acoustic feature table of bowel sounds. The clinical structured entries are mapped to clinical feature fields according to the case index to form a clinical structured feature table, and the structured imaging feature entries are mapped to imaging feature fields according to the case index to form an imaging feature table. The clinical structured feature table, imaging feature table, and bowel sound acoustic feature table are written into the multimodal clinical evidence atlas skeleton, and concept nodes are generated for each feature. The source index relationship between concept nodes and evidence nodes is established to obtain the evidence atlas node set.
6. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 1, characterized in that: The generation of the fusion feature matrix is specifically as follows: The consultation time window is divided according to the consultation timestamp of the case index, and the time window identifier is written into the three types of feature tables. The three types of feature tables are aligned according to the time window identifier to generate the basic feature vector. The co-occurrence relationship between image feature concept nodes and bowel sound event concept nodes is extracted from the evidence map node set according to the same time window relationship, and encoded as co-occurrence interaction features; Evidence paths are generated based on the continuous connection relationships between symptom concept nodes, test indicator concept nodes, CT imaging sign concept nodes, and bowel sound event concept nodes, and encoded as evidence path features. The basic feature vectors, co-occurrence interaction features, and evidence path features are aggregated to generate a fused feature matrix.
7. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 1, characterized in that: The formation of the traceable modeling dataset specifically involves: Based on the case index, generate outcome records and write the outcome determination timestamp; write the fusion feature matrix into the case index to form feature index records. Paired records are generated for outcome records and feature index records and written into the relationship between the event fragment index and source index of the image sequence index, bowel sound audio records, and form a tracing chain; The fused feature matrix, paired records, and traceability chains are combined to form a traceable modeling dataset.
8. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 1, characterized in that: The construction of the strangulated intestinal obstruction risk prediction model specifically involves: The traceability modeling dataset is divided into training and validation data according to the case index, and paired records and traceability chains are retained; Perform logistic regression fitting on the training data to generate a list of model parameters and feature fields; Perform model inference on the verification data to generate verification result records and associate them with the traceability chain; The model parameters, feature field list, and validation result records are summarized to form a strangulated intestinal obstruction risk prediction model.
9. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 1, characterized in that: The formation of the model package specifically includes: Generate a parameter index for the model parameters based on the feature field list and write it to the parameter index record; Based on the parameter index records, retrieve the evidence graph node set to locate the concept nodes and evidence paths, and locate the concept nodes to the evidence nodes through the source index relationship to form a location record; The model parameters and location records are summarized to generate a mapping table from model parameters to evidence nodes and evidence paths; The model parameters, feature field list, mapping table, and validation result records are encapsulated to form a model package.
10. The method for predicting the risk of strangulated intestinal obstruction based on multimodal data fusion as described in claim 1, characterized in that: The output of the joint evidence digest is specifically as follows: New multimodal raw data packages are generated for new patients, and the evidence graph node set and fusion feature matrix are updated. Read the model parameters and feature field list from the model package, generate inferred feature vectors from the fused feature matrix according to the feature field list, and perform logistic regression inference on the inferred feature vectors to generate risk probabilities; The evidence nodes and evidence paths associated with risk probability are located by mapping table, and the image sequence index and bowel sound event fragment index are extracted to output a joint evidence summary.