Path planning safety quantitative evaluation system based on multi-source medical data

Abstract

The invention relates to the technical field of medical clinical path planning, and discloses a multi-source medical data-based path planning safety quantitative evaluation system, which comprises a multi-source data access module, a data preprocessing module, a feature extraction module, a multi-dimensional quantitative evaluation module, a risk feedback module and a model optimization module, all the modules are in communication connection through a data bus to form a cooperative work framework. According to the path planning safety quantitative evaluation system based on the multi-source medical data, a multi-source data access module comprehensively collects electronic medical record data, image examination data, physiological index data and drug interaction data and then transmits the data to a subsequent module, and key information is integrated to cover multiple evaluation dimensions; the multi-dimensional quantitative evaluation module constructs a multi-source data fusion quantitative model, maps features into risk indexes through clear evaluation dimensions and quantitative rules in combination with weight distribution, replaces experience judgment of medical staff, establishes a unified quantitative standard, and improves the consistency of evaluation results.

Description

Technical Field

[0001] This invention relates to the field of medical clinical pathway planning technology, specifically a pathway planning safety quantification assessment system based on multi-source medical data. Background Technology

[0002] Clinical pathway planning is a core component of the healthcare service system. It refers to a standardized treatment process developed based on the patient's condition, the allocation of medical resources, and treatment guidelines. Its safety is directly related to the patient's treatment outcome, medical risk control, and the efficiency of medical resource utilization, and it is a key support for improving the quality of clinical diagnosis and treatment.

[0003] However, traditional methods for assessing the safety of medical pathway planning have many technical shortcomings and are difficult to meet actual clinical needs:

[0004] 1. Limited data sources lead to incomplete assessment dimensions: Traditional assessments often rely on electronic medical records or single examination data, failing to integrate key information such as drug metabolism data and underlying medical history, which can easily overlook potential risk points;

[0005] 2. Safety assessments are primarily qualitative, with low quantitative accuracy and high subjectivity: they rely on the experience and judgment of medical staff, lack unified quantitative standards, and different medical staff may have different risk assessments of the same case, resulting in poor consistency of assessment results;

[0006] 3. Lack of multi-source data fusion mechanism: The semantic systems of different modal data such as images, physiological indicators, and drug data are very different, and cross-modal information distortion is prone to occur during feature extraction. For example, image features and physiological indicators cannot be effectively correlated, resulting in insufficient risk identification accuracy.

[0007] 4. Lack of real-time feedback mechanism for assessment results: Traditional assessment is a static one-time assessment, which cannot capture the dynamic changes in the patient's condition - such as fluctuations in physiological indicators and adverse drug reactions, resulting in the path planning lagging behind the actual condition and making it difficult to dynamically adapt to the real-time changes in the patient's physical condition.

[0008] Therefore, a quantitative evaluation system for the safety of path planning based on multi-source medical data was proposed. Summary of the Invention

[0009] (a) Technical problems to be solved

[0010] To address the shortcomings of existing technologies, this invention provides a path planning safety quantitative assessment system based on multi-source medical data. It has the advantages of multi-source data fusion, accurate quantitative assessment, and real-time dynamic feedback, and solves the problems of traditional medical path planning safety assessment methods, such as single assessment data source, low quantitative accuracy, difficulty in cross-modal data fusion, and inability to dynamically adapt to changes in the patient's condition.

[0011] (II) Technical Solution

[0012] To achieve the aforementioned goals of multi-source data fusion, accurate quantitative assessment, and real-time dynamic feedback, this invention provides the following technical solution: a path planning safety quantitative assessment system based on multi-source medical data, comprising a multi-source data access module, a data preprocessing module, a feature extraction module, a multi-dimensional quantitative assessment module, a risk feedback module, and a model optimization module. Each module is connected via a data bus to form a collaborative working architecture.

[0013] The multi-source data access module is used to collect electronic medical record data, imaging examination data, physiological indicator data, and drug interaction data;

[0014] The data preprocessing module is used to remove redundant information, unify data format, and correct outliers from the collected multi-source data so that the data meets the requirements of subsequent processing.

[0015] The feature extraction module employs a triple constraint algorithm to extract features such as patient disease severity, underlying disease risk, treatment contraindications, drug interaction correlation, and treatment step suitability from the preprocessed data, ensuring that the feature space overlap rate of different source data is ≥85%.

[0016] The multi-dimensional quantitative assessment module constructs a multi-source data fusion quantitative model, mapping the extracted features into a calculable risk index;

[0017] The risk feedback module generates path adjustment suggestions based on the risk index and provides real-time feedback to the clinical pathway planning terminal.

[0018] The model optimization module is communicatively connected to the multi-dimensional quantitative assessment module and is used to iteratively optimize the quantitative model based on clinical feedback data.

[0019] Preferably, the data acquisition objects of the multi-source data access module include:

[0020] Electronic medical record data includes: basic patient information, past medical history, present medical history, medical records, and medical orders.

[0021] Imaging data: including magnetic resonance imaging data, computed tomography (CT) scan data, ultrasound imaging data, and pathological slide image data;

[0022] Physiological indicators: covering blood pressure, blood sugar, heart rate, blood oxygen saturation, liver and kidney function indicators, and electrolyte concentration data;

[0023] Drug interaction data: covering metabolic pathways, contraindicated combinations, and adverse reaction association data of commonly used clinical drugs.

[0024] Preferably, the execution steps of the data preprocessing module include:

[0025] Step 1: Redundant Information Removal: A duplicate data identification method based on the Jaccard coefficient is used to calculate the ratio of the number of intersection elements to the number of union elements between data entries. When the ratio is greater than 0.8, the data is identified as redundant and deleted.

[0026] Step 2: Data Format Unification: Based on the HL7 FHIR protocol, convert data from different sources into JSON format, with fields including data type identifier, collection timestamp, data value, and data credibility score;

[0027] Step 3, Outlier Correction: Outlier detection based on the 3σ principle is adopted. For data that exceeds the mean ± 3 times the standard deviation, it is corrected by linear interpolation of normal data at adjacent time points to ensure data accuracy ≥ 95%.

[0028] Preferably, the triple constraint algorithm of the feature extraction module includes:

[0029] Image spatial transformation steps: A cross-modal data generator is constructed using a generative adversarial network to transform image inspection data from different modalities into image data with unified semantics. The generator contains 6 convolutional layers and 3 deconvolutional layers with a kernel size of 3×3. The activation function is Leaky ReLU. The distribution difference between real data and generated data is measured by the loss function of the generative adversarial network.

[0030] Feature space alignment step: Construct a parameter-shared autoencoder, which contains 4 residual convolutional blocks and a decoder, which contains 4 deconvolutional blocks. The goal is to minimize the L1 norm of the input data and the reconstructed data, so that the feature space overlap rate of different source data is ≥85%.

[0031] Semantic space constraint steps: Introduce a semantic discriminator and use adversarial training to ensure that the Euclidean distance between the semantic feature vectors of different modalities is ≤0.1. The discriminator contains 3 fully connected layers, and the output layer uses the Sigmoid activation function.

[0032] Preferably, the evaluation dimensions and quantitative indicators of the multi-dimensional quantitative evaluation module include:

[0033] Severity of illness dimension: A 5-point scoring system is used, which is calculated based on the frequency of symptom occurrence, the severity of symptoms and the degree of impact on daily life, with a score range of 1-5.

[0034] Underlying disease risk dimension: A risk matrix is ​​constructed based on the type of underlying disease, duration of disease, and control status, with a risk value range of 0-10.

[0035] Treatment contraindications dimension: using binary encoding, the presence of a contraindication is marked as 1, the absence of a contraindication is marked as 0, and the total contraindication value is the sum of all contraindication marks;

[0036] Drug interaction risk dimension: A three-level risk score is constructed based on the degree of conflict in drug metabolism pathways, with 0 points for no conflict, 3 points for mild conflict, and 6 points for severe conflict;

[0037] Treatment step fit dimension: calculated based on the matching degree between treatment steps and patient's physiological state, with a fit range of 0-10 points;

[0038] Each dimension's score is mapped to a risk index through weight allocation, with the total weight being 1, ensuring the consistency of the scoring system.

[0039] Preferably, the method for constructing the multi-source data fusion quantization model includes:

[0040] Step 1: Construct a potential symptom space and use an alternating optimization algorithm to learn the mapping relationship between symptoms, drugs, and treatment steps. The dimension of the potential space is set to 40-60.

[0041] Step 2: Train a multi-label classifier. The classifier uses two fully connected layers with a hidden layer dimension of 256. The activation function is ReLU, and the output layer uses the Softmax activation function. Classification is achieved by calculating the exponential proportion of the input scores of each category.

[0042] Step 3: Calculate the risk index by multiplying the severity score of the illness, the risk value of the underlying disease, the total contraindication value, the drug interaction risk score, and the suitability of the treatment steps by their respective weights and then summing them. The risk index ranges from 0 to 10.

[0043] Preferably, the real-time feedback mechanism of the risk feedback module includes:

[0044] Risk level classification: a risk index of 0-3 (inclusive) is low risk, 3-7 (inclusive) is medium risk, and 7-10 is high risk. The boundary values ​​were determined based on data from 136 clinical patients.

[0045] Risk point identification: By analyzing the contribution of features, the sum of the products of each dimension's score and its corresponding weight is divided by the risk index to obtain the contribution ratio of each dimension to the risk index, thus identifying the core risk points.

[0046] Adjustment suggestion generation: Based on the type of risk point, targeted adjustment suggestions are generated, including drug replacement plans, treatment step sequence adjustment plans, and supplementary examination items plans. The adoption rate of suggestions is ≥85%.

[0047] Preferably, the training parameters of the generative adversarial network are set as follows: batch size = 8, number of iterations = 150,000-250,000 rounds, generator learning rate = 0.0001-0.0003, discriminator learning rate = 0.00001-0.00005, momentum parameter = 0.9, weight decay coefficient = 0.0001, and the training process uses the Adam optimizer to update the parameters according to the correlation between network parameters and learning rate and gradient.

[0048] Preferably, the execution steps of the alternating optimization algorithm include:

[0049] Step 1: Initialize the transformation matrix U, the potential symptom matrix P, and the mapping matrix V. U is initialized as a random matrix, P is obtained by dimensionality reduction using PCA, and V is initialized as an all-one matrix.

[0050] Step 2: Fix P and V, update U using gradient descent, with the step size determined by backtracking linear search;

[0051] Step 3: Fix U and P, and update V using a fast iterative soft thresholding algorithm;

[0052] Step 4: Fix U and V, and update P using gradient descent.

[0053] Step 5: Repeat the iteration until the objective function converges, with the convergence threshold set between 1e-6 and 1e-4.

[0054] Preferably, the optimization mechanism of the model optimization module includes:

[0055] The optimization cycle is 5-10 days, and each optimization uses a mix of new clinical data and historical data at a ratio of 1:3-1:5 for training.

[0056] Historical data uses a sliding window mechanism, retaining data from the past 3-6 months;

[0057] Using the mean square of the predicted risk index and the actual clinical risk index as the loss function, the model's risk point localization accuracy is continuously optimized to ≥90%.

[0058] (III) Beneficial Effects

[0059] Compared with existing technologies, this invention provides a quantitative evaluation system for the safety of path planning based on multi-source medical data, which has the following beneficial effects:

[0060] 1. This path planning safety quantitative assessment system based on multi-source medical data comprehensively collects electronic medical record data, imaging examination data, physiological indicator data and drug interaction data through a multi-source data access module. After being processed by the data preprocessing module, the data is transmitted to subsequent modules, integrating key information to cover multiple assessment dimensions. This solves the problem of incomplete assessment dimensions and omission of potential risk points caused by the single data source in traditional systems.

[0061] 2. This path planning safety quantitative assessment system based on multi-source medical data constructs a multi-source data fusion quantitative model through a multi-dimensional quantitative assessment module. By defining clear assessment dimensions and quantitative rules, and combining weight allocation, it maps features into risk indices, replacing the experience-based judgment of medical staff, establishing unified quantitative standards, improving the consistency of assessment results, and solving the problems of low quantitative accuracy and strong subjectivity.

[0062] 3. The path planning safety quantitative assessment system based on multi-source medical data adopts a triple constraint algorithm in its feature extraction module. By generating an adversarial network to transform cross-modal data, aligning the feature space with an autoencoder, and constraining semantic vectors with a semantic discriminator, it achieves effective fusion of multi-source heterogeneous data, avoids cross-modal information distortion, improves the accuracy of risk identification, and solves the problem of difficulty in multi-source data fusion.

[0063] 4. This path planning safety quantitative assessment system based on multi-source medical data has a risk feedback module that classifies risk levels based on risk indices, locates core risk points, and generates adjustment suggestions that are fed back to the terminal in real time. The model optimization module iteratively optimizes the quantitative model through clinical data to form a dynamic closed loop, captures changes in the patient's condition, and adapts path planning to the real-time status, thus solving the problem of delayed assessment feedback. Attached Figure Description

[0064] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0065] Figure 2 This is a flowchart of the triple constraint algorithm for the feature extraction module of the present invention;

[0066] Figure 3 This is a schematic diagram of the structure of the multi-dimensional quantitative evaluation model of the present invention;

[0067] Figure 4 This is a schematic diagram of the feedback process of the risk feedback module of the present invention. Detailed Implementation

[0068] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see Figure 1-4 The path planning safety quantitative assessment system based on multi-source medical data includes a multi-source data access module, a data preprocessing module, a feature extraction module, a multi-dimensional quantitative assessment module, a risk feedback module, and a model optimization module. The modules are connected through a data bus to form a collaborative working architecture.

[0070] The multi-source data access module is used to collect electronic medical record data, imaging examination data, physiological indicator data, and drug interaction data;

[0071] The data preprocessing module is used to remove redundant information, unify data format, and correct outliers from the collected multi-source data so that the data meets the requirements of subsequent processing.

[0072] The feature extraction module employs a triple constraint algorithm to extract features such as patient disease severity, underlying disease risk, treatment contraindications, drug interaction correlation, and treatment step suitability from the preprocessed data, ensuring that the feature space overlap rate of different source data is ≥85%.

[0073] The multi-dimensional quantitative assessment module constructs a multi-source data fusion quantitative model, mapping the extracted features into a calculable risk index;

[0074] The risk feedback module generates path adjustment suggestions based on the risk index and provides real-time feedback to the clinical pathway planning terminal.

[0075] The model optimization module is communicatively connected to the multi-dimensional quantitative assessment module and is used to iteratively optimize the quantitative model based on clinical feedback data.

[0076] Example 1: System Overall Architecture Implementation

[0077] This embodiment provides a detailed implementation of the core technical features of the overall system architecture. The modules are connected through a data bus to form a collaborative working architecture.

[0078] First, the hardware was deployed, with each module deployed on an industrial-grade server cluster. The servers were configured with Intel Xeon Gold 6248 processors, 128GB of RAM, a Tesla V100 GPU with 18GB of video memory, and 10TB of storage. The clinical pathway planning terminal used a medical tablet computer, equipped with a Snapdragon 8cx Gen3 processor, 8GB of RAM, and 256GB of storage.

[0079] In terms of software environment setup, CentOS 7.9 was selected as the operating system, PyTorch 1.12 was used as the deep learning framework, MySQL 8.0 was used as the database, and HTTP / 2, FTP and HL7 FHIR were supported as data transmission protocols.

[0080] Each module achieves bidirectional communication through a high-speed data bus. The data bus adopts the PCIe 4.0 interface standard, with a data transmission bandwidth of ≥32GB / s, ensuring that the data transmission delay between modules is ≤50ms.

[0081] When the collaborative workflow is initiated, the multi-source data access module first collects various types of medical data and transmits them to the data preprocessing module in real time. The preprocessed clean data is then transmitted to the feature extraction module for feature extraction. The extracted feature vectors are simultaneously sent to the multi-dimensional quantitative assessment module to calculate the risk index. The risk index and related feature data are transmitted to the risk feedback module to generate adjustment suggestions, and also sent to the model optimization module for model iteration. The model optimization module then back-synchronizes the optimized parameters to the multi-dimensional quantitative assessment module, forming a closed-loop process of data collection, processing, analysis, assessment, feedback, and optimization.

[0082] In this embodiment, standardized hardware deployment and software environment configuration ensure the stable operation of each module. The adoption of a high-speed data bus eliminates data transmission bottlenecks between modules, ensuring timely collaborative responses from each module. The closed-loop architecture of the entire process enables data to flow completely from acquisition to model optimization, avoiding data gaps. This allows for dynamic evaluation of the safety of clinical pathway planning throughout the entire process, solving the problems of isolated links and delayed responses in traditional evaluation systems.

[0083] Example 2:

[0084] This embodiment corresponds to the core technical features of the multi-source data access module, and details the specific implementation method of data acquisition.

[0085] Electronic medical record data acquisition adopts a distributed architecture, synchronizing data in real time through the API interface of the Hospital Information System (HIS). The acquisition frequency is synchronized within 1 second after data generation, and the acquired content includes:

[0086] Basic patient information, including name, gender, age, etc.;

[0087] Past medical history, including the name of the disease, the time of diagnosis and treatment, and the treatment effect;

[0088] The present medical history, including the onset time and symptoms;

[0089] Medical records, including examination items and diagnostic results;

[0090] This includes doctor's orders containing information such as drug name and dosage.

[0091] Data transmission uses SSL / TLS encryption protocol to ensure security.

[0092] Imaging data is acquired from the Medical Imaging Archive and Communication System (PACS), supporting DICOM and JPEG2000 formats. The acquired data includes MRI data, including T1WI and T2WI sequences; CT data, including plain and enhanced scans; ultrasound data, including two-dimensional ultrasound and Doppler ultrasound; and pathological slide images, including HE staining and immunohistochemical staining slide images. The acquisition trigger mechanism is that the data is automatically pushed to the access module after the imaging examination is completed.

[0093] Physiological data are collected through the Laboratory Information System (LIS) and portable monitoring devices. The LIS collects liver and kidney function indicators, electrolyte concentrations, and other data, which are synchronized once per hour. Portable monitoring devices such as blood pressure monitors and blood glucose meters transmit data in real time via Bluetooth 5.0 protocol, with a collection frequency of once per minute. The collected data includes blood pressure, blood glucose, heart rate, blood oxygen saturation, liver and kidney function indicators, and electrolyte concentrations.

[0094] Drug interaction data is constructed based on a clinical drug database, collecting data on the metabolic pathways of 31 commonly used clinical drugs, including hepatic metabolism and renal excretion; contraindicated combinations, including contraindications for the combined use of pramipexole hydrochloride and alprazolam; and adverse reaction association data, including reactions such as dizziness and nausea caused by drug combination. Specific drugs include pramipexole hydrochloride, carbidopa, vitamin E, donepezil, zopiclone, coenzyme Q10, iperexone hydrochloride, alprazolam, and stinox. The data is updated monthly by crawling authoritative medical drug databases and clinical research literature.

[0095] This embodiment ensures the targeted and efficient collection of data by designing a dedicated acquisition architecture and transmission protocol for different types of data. The acquisition method, which combines real-time synchronization with timed synchronization, not only guarantees the timeliness of dynamically changing data but also avoids invalid and repeated collection of static data. Encrypted transmission ensures data security, and the comprehensive acquisition scope covers common clinical scenarios, enabling the access module to obtain multi-source data required for clinical pathway planning and solving the problem of single data sources in traditional assessment systems.

[0096] Example 3:

[0097] This embodiment focuses on the core technical features of the data preprocessing module, and details the specific steps and operation methods of preprocessing.

[0098] The first step involves removing redundant information using a duplicate data identification method based on the Jaccard coefficient:

[0099] First, feature segmentation is performed on each collected data entry to split the text data in the electronic medical record into keyword groups, extract pixel feature vectors from the image data, retain numerical features from the physiological index data, and extract drug names and related features from the drug interaction data.

[0100] Then calculate any two data entries. The Jaccard coefficient is calculated using the following formula:

[0101]

[0102] in, and These represent two data entries to be compared. This indicates the number of common feature elements between two data entries. This represents the number of features in the union of all features of two data entries. The threshold is set to 0.8. When the calculated Jaccard coefficient is greater than 0.8, it is considered redundant data and is directly deleted from the dataset. The deletion operation is recorded in the log file in real time.

[0103] The second step is to unify the data format, based on the HL7 FHIR protocol, converting heterogeneous data from different sources into the standard JSON format:

[0104] The fields defined in JSON format include:

[0105] data_type is the data type identifier, such as electronic medical records, MRI images, etc.

[0106] collect_time, which is the collection timestamp, in the format YYYY-MM-DD HH:MM:SS;

[0107] data_value, i.e., data value, text data is directly retained as a string, image data storage path and pixel matrix parameters, and numerical data retains the original value;

[0108] Confidence, or data credibility score, ranges from 0 to 1 and is calculated based on the accuracy of data acquisition equipment and data integrity.

[0109] During the conversion process, field mapping and format conversion are performed on the raw data that does not conform to the HL7 FHIR protocol specification. For example, the DICOM format metadata of the image data is converted into the parameters corresponding to the JSON fields.

[0110] The third step involves outlier correction, employing outlier detection based on the 3σ principle:

[0111] First, the mean μ and standard deviation σ of each data sequence are calculated. For time-series data such as physiological indicators, the statistics are calculated based on historical data of the past 24 hours. For static data such as electronic medical records and drug interactions, the statistics are calculated based on similar datasets.

[0112] Then determine if the data value exceeds the limit. If the range is exceeded, it is considered an outlier; correction is made using linear interpolation of normal data from adjacent time points, with the interpolation formula as follows:

[0113]

[0114] in, These are the corrected abnormal data values. This refers to the abnormal data collection time. and These represent the collection times of normal data adjacent to and before the abnormal data. and σ represents the normal data value at the corresponding time point, μ is the data mean, and σ is the data standard deviation.

[0115] In this embodiment, the Jaccard coefficient can accurately measure the degree of repetition of different types of data. Reasonable threshold settings can effectively remove redundant data, reduce data storage pressure and subsequent processing burden. The adoption of the HL7 FHIR protocol realizes the standardization and unification of heterogeneous data. The universality of the JSON format ensures that each subsequent module can directly parse and use it, avoiding data unusability problems caused by format incompatibility. The 3σ principle combined with the outlier correction method of linear interpolation can accurately identify and reasonably correct abnormal data, ensuring the reliability of the preprocessed data and providing a high-quality data foundation for subsequent feature extraction and quantitative evaluation.

[0116] Example 4:

[0117] This embodiment describes the core technical features of the feature extraction module, detailing the algorithm implementation and operational details.

[0118] 1. Network architecture construction

[0119] The generator of the generative adversarial network consists of 6 convolutional layers and 3 deconvolutional layers. The kernel size of each convolutional layer is 3×3, the stride of the convolutional layers is 1, the stride of the deconvolutional layers is 2, and the activation function is Leaky ReLU with a slope of 0.2. The discriminator adopts the PatchGAN architecture, which consists of 5 convolutional layers. The first 4 convolutional layers have 4×4 kernels with a stride of 2, and the last convolutional layer has 1×1 kernels with a stride of 1. The output layer uses the Sigmoid activation function.

[0120] The encoder of the autoencoder contains 4 residual convolutional blocks, each of which consists of 2 convolutional layers, a batch normalization layer, and a ReLU activation function. The decoder contains 4 deconvolutional blocks, each of which contains 1 deconvolutional layer, a batch normalization layer, and a ReLU activation function.

[0121] The semantic discriminator consists of three fully connected layers with hidden layer dimensions of 512 and 256, respectively, and the output layer uses the Sigmoid activation function.

[0122] 2. Training process design

[0123] First, the triple constraint algorithm is initialized and configured. The training parameters of the generative adversarial network are set as follows: batch size = 8, number of iterations = 200,000 rounds, generator learning rate = 0.0002, discriminator learning rate = 0.00003, momentum parameter = 0.9, weight decay coefficient = 0.0001, and the Adam optimizer is used during the training process.

[0124] The first constraint is the image space transformation step:

[0125] The preprocessed MRI, CT and other different modal imaging data are input into the generator. The generator learns the semantic associations of different modal data and transforms them into image data with unified semantics.

[0126] During training, the loss function of the generative adversarial network is used to measure the difference in distribution between the real data and the generated data. The formula for the loss function is:

[0127]

[0128] Where G is the generator and D is the discriminator. For the true data distribution, Let x be a real data sample and z be random noise. The goal is to minimize this loss function so that the generated data distribution is consistent with the real data distribution.

[0129] The second constraint is the feature space alignment step:

[0130] A parameter-shared autoencoder is constructed, taking preprocessed data as input. The encoder extracts features from the data, and the decoder reconstructs the data from the extracted features. The objective is to minimize the L1 norm of the input data and the reconstructed data. The loss function formula is as follows:

[0131]

[0132] Where X represents the input data. To reconstruct the data, the autoencoder parameters are iteratively optimized through backpropagation to maximize the overlap of feature spaces extracted from different source data by the encoder, ensuring a feature space overlap rate of ≥85%.

[0133] The third constraint is the semantic space constraint step:

[0134] A semantic discriminator is introduced, and the feature vectors extracted by the autoencoder are input into the semantic discriminator. Adversarial training is used to ensure that the Euclidean distance between the semantic feature vectors of different modalities is ≤0.1.

[0135] During training, the semantic discriminator learns to distinguish feature vectors from different modalities. The feature extraction network adjusts its parameters through adversarial training, making the semantic discriminator unable to distinguish the modal origin of the feature vectors, thereby achieving alignment of the semantic space.

[0136] 3. Deployment and Operation Process

[0137] After training, the triple constraint algorithm is deployed to the feature extraction module. During runtime, the preprocessed data is first input into the generative adversarial network to complete cross-modal image transformation. The transformed image and the original data are then input into the autoencoder for feature extraction and reconstruction. After the extracted feature vectors are verified for semantic consistency by the semantic discriminator, five target features are output, namely:

[0138] Based on the frequency and severity of symptoms, etc., the severity of the patient's condition is characterized;

[0139] Based on the risk characteristics of underlying diseases, such as the type and course of the disease;

[0140] Based on the presence or absence of contraindications to treatment;

[0141] Drug interaction association characteristics based on drug metabolism pathway conflicts;

[0142] Treatment step fit characteristics based on the matching of treatment steps with physiological state.

[0143] In this embodiment, the cross-modal data generation of generative adversarial networks solves the semantic differences between different modal image data, enabling image data to be transformed into a unified semantic space. The feature space alignment of the autoencoder ensures the fusion of features from different source data, and the semantic space constraint further eliminates the semantic gap between different modal data. The triple constraint algorithm works synergistically to enable the extracted five types of features to comprehensively and accurately reflect the patient's clinical information, solving the problem of cross-modal information distortion in traditional feature extraction.

[0144] Example 5:

[0145] This embodiment details the specific implementation of the evaluation dimension calculation, model construction, and risk index calculation for the core technical features of the multi-dimensional quantitative evaluation module.

[0146] I. Quantitative Calculation of Evaluation Dimensions

[0147] 1. Severity of Illness: A 5-point scoring system was used, based on the frequency of symptom occurrence (weight 0.4), the severity of symptoms (weight 0.3), and the degree of impact on daily life (weight 0.3).

[0148] Symptom frequency is categorized by the number of times it occurs per day: ≤1 time = 1 point, 2-3 times = 2 points, 4-5 times = 3 points, 6-10 times = 4 points, >10 times = 5 points;

[0149] The severity of symptoms is classified according to the extent of impact: minor local impact is scored as 1 point, significant local impact is scored as 2 points, minor systemic impact is scored as 3 points, significant systemic impact is scored as 4 points, and life-threatening impact is scored as 5 points.

[0150] The degree of impact on daily life is divided according to self-care ability: no impact is 1 point, slight impact is 2 points, partial dependence on others is 3 points, most dependence on others is 4 points, and complete dependence on others is 5 points.

[0151] The weighted sum of the three scores is the score for that dimension, with a score range of 1-5.

[0152] 2. Underlying Disease Risk Dimension: A risk matrix is ​​constructed based on the type of underlying disease, duration of illness, and control status; the types of underlying diseases are assigned values ​​according to their severity.

[0153] Mild diseases (such as stage 1 hypertension) are scored 1 point per item, moderate diseases (such as type 2 diabetes) are scored 2 points per item, and severe diseases (such as coronary heart disease) are scored 3 points per item; disease duration > 5 years adds 2 points; control status: uncontrolled adds 3 points, moderately controlled adds 1 point, well controlled adds 0 points; the sum of the scores of each item is the risk value of that dimension, with the risk value ranging from 0 to 10 points.

[0154] 3. Contraindications dimension: Binary encoding is used. A contraindication is marked as 1, and no contraindication is marked as 0. The total contraindication value is the sum of all contraindication marks, and the total contraindication value ranges from 0 to n (where n is the total number of contraindications).

[0155] 4. Drug Interaction Risk Dimension: Based on the degree of conflict in drug metabolism pathways, a 3-level risk score is constructed: 0 points for no metabolic pathway conflict, 3 points for minor conflict (such as decreased metabolic efficiency), and 6 points for severe conflict (such as the production of toxic metabolites).

[0156] 5. Treatment Step Fit Dimension: Based on the matching degree between treatment steps and the patient's physiological state, the score is calculated based on the degree of trauma of the treatment steps, recovery requirements and the patient's physiological indicators (such as heart rate, liver and kidney function). The fit range is 0-10 points, with 10 points for complete fit and 0 points for complete misfit.

[0157] II. Construction and Training of Multi-Source Data Fusion Quantization Model

[0158] The construction of a multi-source data fusion quantization model is performed according to the following steps:

[0159] Step 1: Construct a potential symptom space and use an alternating optimization algorithm to learn the mapping relationship between symptoms and drugs and treatment steps;

[0160] Initialize the transformation matrix U (dimension d×k, where d is the total number of features and k is the latent space dimension, set to 50), the latent symptom matrix P (dimension n×k, where n is the number of patients, obtained through PCA dimensionality reduction), and the mapping matrix V (dimension k×c, where c is the number of assessment dimensions, initialized as an all-1 matrix).

[0161] With P and V fixed, update U using gradient descent, with the step size determined by backtracking linear search; with U and P fixed, update V using a fast iterative soft thresholding algorithm; with U and V fixed, update P using gradient descent; repeat the iteration until the objective function converges, with the convergence threshold set to 5e-5.

[0162] Step 2: Train a multi-label classifier. The classifier uses two fully connected layers with a hidden layer dimension of 256. The activation function is ReLU, and the output layer uses the Softmax activation function. The Softmax function formula is:

[0163]

[0164] in, For the first The input score for each class is K, which is the total number of categories. Classification is achieved by calculating the index proportion of the input score for each category. The training dataset uses feature data from 136 clinical patients and their corresponding evaluation dimension scores.

[0165] Step 3: Calculate the risk index by multiplying the scores of each dimension by their corresponding weights and then summing the results. The weights are allocated as follows:

[0166] The risk index is calculated using the following weightings: disease severity (0.3), underlying disease risk (0.25), treatment contraindications (0.2), drug interaction risk (0.15), and treatment procedure suitability (0.1). The total weight is 1, and the risk index ranges from 0 to 10. The formula is as follows:

[0167]

[0168] in, Score the severity of the condition. Based on the risk value of underlying diseases, The total taboo value, For drug interaction risk scoring, For the suitability of treatment steps.

[0169] III. Model Deployment and Operation

[0170] The trained multi-source data fusion and quantification model is deployed to the multi-dimensional quantification assessment module. During runtime, the five types of features output by the feature extraction module are input into the corresponding quantification calculation units of the dimensions to obtain scores for each dimension. The scores are then substituted into the risk index calculation formula to output the final risk index. The entire calculation process takes ≤200ms per patient.

[0171] In this embodiment, the quantitative calculation method for the five assessment dimensions is designed based on clinical diagnosis and treatment guidelines. The weight allocation has been verified by clinical data to ensure the objectivity and rationality of the scores for each dimension. The potential symptom space constructed by the alternating optimization algorithm can effectively establish the association between symptoms and drugs and treatment steps. The training of the multi-label classifier ensures the accurate mapping of features to assessment dimension scores. The weighted calculation method of the risk index integrates risk information from each dimension, transforming abstract safety into calculable specific values, thus solving the problem that traditional assessments are mainly qualitative descriptions with low accuracy.

[0172] Example 6:

[0173] This embodiment details the implementation of the feedback mechanism, focusing on the core technical features of the risk feedback module.

[0174] I. Construction and Training of Risk Level Assessment Mechanism

[0175] The risk level classification boundary values ​​were determined through validation using data from 136 clinical patients. Specifically, a risk index of 0-3 points (inclusive) indicates low risk, 3-7 points (inclusive) indicates medium risk, and 7-10 points indicates high risk.

[0176] Construct a risk level determination function. The input of the function is the risk index output by the multi-dimensional quantitative assessment module, and the output is the corresponding risk level. During the training process, the boundary value threshold is adjusted to make the accuracy of the determination function ≥95%.

[0177] II. Implementation of Risk Point Location Algorithm

[0178] The contribution percentage of each dimension to the risk index is calculated through feature contribution analysis. The calculation formula is as follows:

[0179]

[0180] in, For the first The contribution percentage of each dimension For the first Dimensional scoring The corresponding dimension weights are Risk, which is the risk index.

[0181] The contribution percentages of each dimension are sorted in descending order, and the dimension with the highest contribution percentage is the core risk point. If there are multiple dimensions with a contribution percentage difference of ≤5%, they are all identified as core risk points.

[0182] III. Adjusting the proposed model construction and deployment

[0183] Based on the types of core risk points, an adjustment suggestion generation model is constructed. The model contains four types of suggestion generation units, each corresponding to a different type of core risk point:

[0184] 1. If the core risk point is the risk dimension of drug interaction, generate a drug replacement plan. Based on the principle of complementary drug metabolism pathways, screen alternative drugs from the drug database that have the same therapeutic effect as the original drug and do not conflict with the metabolism pathway. At the same time, indicate the usage, dosage and precautions of the alternative drugs.

[0185] 2. If the core risk point is the treatment step suitability dimension, generate a treatment step sequence adjustment plan, sort the steps based on the patient's physiological tolerance, prioritize the treatment steps with less trauma and faster recovery, and mark the optimal execution interval for each step after adjustment.

[0186] 3. If the core risk point is the severity of the illness or the risk of underlying diseases, generate a supplementary examination plan, add examination items that can further clarify the severity of the illness or the control of underlying diseases, and indicate the timing of the examination and the judgment criteria.

[0187] 4. If the core risk point is the dimension of treatment contraindications, generate contraindication avoidance plan, clearly identify the prohibited treatment methods or drugs, and provide alternative treatment ideas.

[0188] After the adjustment suggestions are generated, they are fed back to the clinical pathway planning terminal in real time via a high-speed data transmission protocol with a feedback delay of ≤50ms. Upon receiving the feedback, the terminal notifies medical staff via pop-up windows and voice prompts, and at the same time stores the suggestions in the patient's medical records.

[0189] In this embodiment, risk level boundary values ​​validated by clinical data ensure the accuracy of risk level classification; quantitative calculation of feature contribution analysis can accurately locate core risk points, avoiding blind risk screening; targeted adjustment suggestions are generated based on clinical diagnosis and treatment logic, ensuring the feasibility and effectiveness of the suggestions; and a real-time feedback mechanism enables rapid response, allowing medical staff to adjust clinical pathways in a timely manner according to risk conditions, solving the problem of delayed feedback of traditional assessment results.

[0190] Example 7:

[0191] This embodiment details the specific implementation of the optimization mechanism, focusing on the core technical features of the model optimization module.

[0192] I. Optimize Data Preparation and Processing

[0193] The optimization cycle is set to 7 days, and the optimization trigger mechanism is: the optimization process is automatically started every 7 days, or the optimization is started in advance when the model prediction deviation rate reported by clinical feedback is ≥5%.

[0194] In the data preparation phase, new clinical data (data from ≥100 patients, including complete multi-source data, feature data, risk index predictions, and actual clinical risk assessment results) were collected. Historical data were collected using a sliding window mechanism, retaining clinical data from the past 6 months. New clinical data and historical data were mixed in a 1:4 ratio as the optimized training dataset.

[0195] Data preprocessing was performed according to the standard procedure in Example 3 to ensure that the data quality met the requirements for optimized training.

[0196] 2. Model optimization training process

[0197] The model optimization process employs gradient descent, using the mean of the squares of the predicted risk index and the actual clinical risk index as the loss function. The loss function formula is as follows:

[0198]

[0199] Where N is the sample size. To predict the risk index for the model, This is the actual clinical risk index;

[0200] During optimization training, the mixed dataset is input into the multi-source data fusion quantization model. The gradient of the loss function with respect to the model parameters is calculated through backpropagation. The weight parameters of the model and the latent space mapping relationship are adjusted. The training is iterated until the loss function converges (the convergence threshold is set to 1e-6).

[0201] The Adam optimizer was used during training, with a learning rate of 0.0001 and a momentum parameter of 0.9.

[0202] III. Optimize Model Validation and Deployment

[0203] The optimized model was validated using newly collected clinical data from 20 cases that had not participated in the optimization training. The validation metrics included the accuracy of risk index prediction and the accuracy of risk point localization. If the prediction accuracy was ≥92% and the risk point localization accuracy was ≥90%, the optimized model parameters were synchronized to the multi-dimensional quantitative evaluation module to replace the original parameters. If the validation metrics were not met, the data mixing ratio (within the range of 1:3-1:5) and the learning rate (within the range of 0.0001-0.001) were readjusted, and optimization training was performed again until the validation metrics were met.

[0204] The parameter synchronization process uses an incremental update method to ensure that the multi-dimensional quantitative evaluation module operates normally during optimization.

[0205] In this embodiment, a fixed optimization cycle combined with triggered optimization ensures that the model can adapt to changes in clinical data in a timely manner. The sliding window mechanism retains recent effective data and avoids interference from outdated historical data on model optimization. A reasonable data mixing ratio ensures the diversity and representativeness of training data. Using the sum of squares and mean as the loss function can accurately measure the model's prediction bias. Iterative optimization using gradient descent continuously adjusts the model parameters to the optimal state. A rigorous validation process ensures the performance of the optimized model, maintains the model's high prediction accuracy, and solves the problem of traditional model performance decaying over time.

[0206] In summary, this path planning safety quantitative assessment system based on multi-source medical data comprehensively collects electronic medical record data, imaging examination data, physiological indicator data, and drug interaction data through a multi-source data access module. After being processed by the data preprocessing module, the data is transmitted to subsequent modules, integrating key information to cover multiple assessment dimensions. This solves the problem of incomplete assessment dimensions and omission of potential risk points caused by the single data source in traditional systems.

[0207] Furthermore, this path planning safety quantitative assessment system based on multi-source medical data constructs a multi-source data fusion quantitative model through a multi-dimensional quantitative assessment module. By defining clear assessment dimensions and quantitative rules, and combining weight allocation, it maps features into risk indices, replacing the experience-based judgment of medical staff, establishing unified quantitative standards, improving the consistency of assessment results, and solving the problems of low quantitative accuracy and strong subjectivity.

[0208] Furthermore, this path planning safety quantification assessment system based on multi-source medical data employs a triple constraint algorithm in its feature extraction module. By generating adversarial networks to transform cross-modal data, aligning the feature space with an autoencoder, and constraining semantic vectors with a semantic discriminator, it achieves effective fusion of multi-source heterogeneous data, avoids cross-modal information distortion, improves risk identification accuracy, and solves the problem of difficulty in multi-source data fusion.

[0209] Furthermore, this path planning safety quantitative assessment system based on multi-source medical data has a risk feedback module that classifies risk levels based on risk indices, locates core risk points, and generates adjustment suggestions that are fed back to the terminal in real time. The model optimization module iteratively optimizes the quantitative model through clinical data to form a dynamic closed loop, captures changes in the patient's condition, and adapts path planning to the real-time status, thus solving the problem of delayed assessment feedback.

[0210] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the prior art. The computer software programs or protocols involved in these functional modules are technologies known to those skilled in the art and are not improvements to this system. The improvement of this system lies in the interaction or connection between the modules, that is, in improving the overall structure of the system to solve the corresponding technical problems that this system aims to address.

[0211] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A path planning safety quantitative assessment system based on multi-source medical data, characterized in that, It includes a multi-source data access module, a data preprocessing module, a feature extraction module, a multi-dimensional quantitative assessment module, a risk feedback module, and a model optimization module. These modules are connected through a data bus to form a collaborative working architecture. The multi-source data access module is used to collect electronic medical record data, imaging examination data, physiological indicator data, and drug interaction data; The data preprocessing module is used to remove redundant information, unify data format, and correct outliers from the collected multi-source data so that the data meets the requirements of subsequent processing. The feature extraction module employs a triple constraint algorithm to extract features such as patient disease severity, underlying disease risk, treatment contraindications, drug interaction correlation, and treatment step suitability from the preprocessed data, ensuring that the feature space overlap rate of different source data is ≥85%. The multi-dimensional quantitative assessment module constructs a multi-source data fusion quantitative model, mapping the extracted features into a calculable risk index; The risk feedback module generates path adjustment suggestions based on the risk index and provides real-time feedback to the clinical pathway planning terminal. The model optimization module is communicatively connected to the multi-dimensional quantitative assessment module and is used to iteratively optimize the quantitative model based on clinical feedback data.

2. The path planning safety quantitative assessment system based on multi-source medical data according to claim 1, characterized in that, The data acquisition objects of the multi-source data access module include: Electronic medical record data includes: basic patient information, past medical history, present medical history, medical records, and medical orders. Imaging data: including magnetic resonance imaging data, computed tomography (CT) scan data, ultrasound imaging data, and pathological slide image data; Physiological indicators: covering blood pressure, blood sugar, heart rate, blood oxygen saturation, liver and kidney function indicators, and electrolyte concentration data; Drug interaction data: covering metabolic pathways, contraindicated combinations, and adverse reaction association data of commonly used clinical drugs.

3. The path planning safety quantitative assessment system based on multi-source medical data according to claim 1, characterized in that, The execution steps of the data preprocessing module include: Step 1: Redundant Information Removal: A duplicate data identification method based on the Jaccard coefficient is used to calculate the ratio of the number of intersection elements to the number of union elements between data entries. When the ratio is greater than 0.8, the data is identified as redundant and deleted. Step 2: Data Format Unification: Based on the HL7 FHIR protocol, convert data from different sources into JSON format, with fields including data type identifier, collection timestamp, data value, and data credibility score; Step 3, Outlier Correction: Outlier detection based on the 3σ principle is adopted. For data that exceeds the mean ± 3 times the standard deviation, it is corrected by linear interpolation of normal data at adjacent time points to ensure data accuracy ≥ 95%.

4. The path planning safety quantitative evaluation system based on multi-source medical data according to claim 1, characterized in that, The triple constraint algorithm of the feature extraction module includes: Image spatial transformation steps: A cross-modal data generator is constructed using a generative adversarial network to transform image inspection data from different modalities into image data with unified semantics. The generator contains 6 convolutional layers and 3 deconvolutional layers with a kernel size of 3×3. The activation function is Leaky ReLU. The distribution difference between real data and generated data is measured by the loss function of the generative adversarial network. Feature space alignment step: Construct a parameter-shared autoencoder, which contains 4 residual convolutional blocks and a decoder, which contains 4 deconvolutional blocks. The goal is to minimize the L1 norm of the input data and the reconstructed data, so that the feature space overlap rate of different source data is ≥85%. Semantic space constraint steps: Introduce a semantic discriminator and use adversarial training to ensure that the Euclidean distance between the semantic feature vectors of different modalities is ≤0.

1. The discriminator contains 3 fully connected layers, and the output layer uses the Sigmoid activation function.

5. The path planning safety quantitative assessment system based on multi-source medical data according to claim 1, characterized in that, The evaluation dimensions and quantitative indicators of the multi-dimensional quantitative evaluation module include: Severity of illness dimension: A 5-point scoring system is used, which is calculated based on the frequency of symptom occurrence, the severity of symptoms and the degree of impact on daily life, with a score range of 1-5. Underlying disease risk dimension: A risk matrix is ​​constructed based on the type of underlying disease, duration of disease, and control status, with a risk value range of 0-10. Treatment contraindications dimension: using binary encoding, the presence of a contraindication is marked as 1, the absence of a contraindication is marked as 0, and the total contraindication value is the sum of all contraindication marks; Drug interaction risk dimension: A three-level risk score is constructed based on the degree of conflict in drug metabolism pathways, with 0 points for no conflict, 3 points for mild conflict, and 6 points for severe conflict; Treatment step fit dimension: calculated based on the matching degree between treatment steps and patient's physiological state, with a fit range of 0-10 points; Each dimension's score is mapped to a risk index through weight allocation, with the total weight being 1, ensuring the consistency of the scoring system.

6. The path planning safety quantitative assessment system based on multi-source medical data according to claim 1, characterized in that, The method for constructing the multi-source data fusion quantization model includes: Step 1: Construct a potential symptom space and use an alternating optimization algorithm to learn the mapping relationship between symptoms, drugs, and treatment steps. The dimension of the potential space is set to 40-60. Step 2: Train a multi-label classifier. The classifier uses two fully connected layers with a hidden layer dimension of 256. The activation function is ReLU, and the output layer uses the Softmax activation function. Classification is achieved by calculating the exponential proportion of the input scores of each category. Step 3: Calculate the risk index by multiplying the severity score of the illness, the risk value of the underlying disease, the total contraindication value, the drug interaction risk score, and the suitability of the treatment steps by their respective weights and then summing them. The risk index ranges from 0 to 10.

7. The path planning safety quantitative assessment system based on multi-source medical data according to claim 1, characterized in that, The real-time feedback mechanism of the risk feedback module includes: Risk level classification: a risk index of 0-3 (inclusive) is low risk, 3-7 (inclusive) is medium risk, and 7-10 is high risk. The boundary values ​​were determined based on data from 136 clinical patients. Risk point identification: By analyzing the contribution of features, the sum of the products of each dimension's score and its corresponding weight is divided by the risk index to obtain the contribution ratio of each dimension to the risk index, thus identifying the core risk points. Adjustment suggestion generation: Based on the type of risk point, targeted adjustment suggestions are generated, including drug replacement plans, treatment step sequence adjustment plans, and supplementary examination items plans. The adoption rate of suggestions is ≥85%.

8. The path planning safety quantitative evaluation system based on multi-source medical data according to claim 4, characterized in that, The training parameters of the generative adversarial network are set as follows: batch size = 8, number of iterations = 150,000-250,000 rounds, generator learning rate = 0.0001-0.0003, discriminator learning rate = 0.00001-0.00005, momentum parameter = 0.9, weight decay coefficient = 0.0001. The training process uses the Adam optimizer to update the parameters according to the correlation between network parameters and learning rate and gradient.

9. The path planning safety quantitative assessment system based on multi-source medical data according to claim 6, characterized in that, The execution steps of the alternating optimization algorithm include: Step 1: Initialize the transformation matrix U, the potential symptom matrix P, and the mapping matrix V. U is initialized as a random matrix, P is obtained by dimensionality reduction using PCA, and V is initialized as an all-one matrix. Step 2: Fix P and V, update U using gradient descent, with the step size determined by backtracking linear search; Step 3: Fix U and P, and update V using a fast iterative soft thresholding algorithm; Step 4: Fix U and V, and update P using gradient descent. Step 5: Repeat the iteration until the objective function converges, with the convergence threshold set between 1e-6 and 1e-4.

10. The path planning safety quantitative assessment system based on multi-source medical data according to claim 1, characterized in that, The optimization mechanism of the model optimization module includes: The optimization cycle is 5-10 days, and each optimization uses a mix of new clinical data and historical data at a ratio of 1:3-1:5 for training. Historical data uses a sliding window mechanism, retaining data from the past 3-6 months; Using the mean square of the predicted risk index and the actual clinical risk index as the loss function, the model's risk point localization accuracy is continuously optimized to ≥90%.

Images