Postoperative risk assessment system for vertebral fracture based on machine learning

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a vertebral fracture postoperative risk assessment system based on machine learning. The system constructs a structured feature set including inspection indexes, vital sign sequences, image measurement parameters, medication records and activity behaviors by collecting multi-source postoperative data of a patient; hierarchical expression of multi-source data is realized by adopting an ordered sparse auto-encoder with an explicit semantic sequence, bidirectional time sequence modeling is performed on a sparse hidden space sequence in combination with BiGRU, and a postoperative overall risk index containing various risk scores is generated. The system displays risk levels and key feature contribution information at a doctor workstation and a nursing terminal through a hospital local area network, and pushes a management prompt under a high-risk condition. Unified modeling and prospective risk assessment of complex and multi-mode postoperative data are achieved, and the method is suitable for vertebral fracture postoperative informatization management.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a machine learning-based postoperative risk assessment system for vertebral fractures. Background Technology

[0002] With the improvement of hospital information infrastructure, HIS, EMR, PACS, LIS, and bedside monitoring systems can continuously generate a large amount of multi-source structured and unstructured data on patients undergoing vertebral fracture surgery, providing an intelligent data foundation for postoperative risk assessment. However, existing intelligent technologies still have significant shortcomings: most intelligent systems currently only predict data from a single modality or fixed time point based on rule engines, specific threshold strategies, or traditional machine learning models, making it difficult to integrate multi-dimensional dynamic features such as the constantly changing vital signs sequence, bone metabolism and inflammatory response indicators, analgesic drug usage patterns, early mobilization ability, and changes in internal fixation imaging throughout the postoperative recovery process. At the same time, although some existing deep learning models can process some time-series information, they generally lack the ability to express the hierarchical structure of clinical data, cross-time window semantics, and multi-scale risk patterns, and cannot distinguish long-term postoperative recovery trends, subacute physiological state changes, and short-term abnormal offset signals in a unified latent space. In addition, the latent variable representations of existing intelligent systems are mostly disordered and uninterpretable, and the models have difficulty maintaining semantic stability under complex nonlinear relationships, which is not conducive to risk source identification and clinical validation. Due to the above limitations, current intelligent technologies are insufficient to support continuous monitoring, early warning, and dynamic evolution modeling of postoperative risks of vertebral fractures. It is necessary to develop a risk assessment technology that can achieve multi-source data fusion, semantic hierarchical coding, nested structured expression, and temporal prediction in order to overcome the bottlenecks of existing intelligent solutions in terms of multimodal fusion capabilities, risk structure modeling capabilities, and interpretability. Summary of the Invention

[0003] This invention proposes a machine learning-based postoperative risk assessment system for vertebral fractures. It aims to address the challenges of highly heterogeneous multi-source data, significant dynamic features, and multi-scale risk evolution in the postoperative recovery process of vertebral fracture patients. The system constructs an intelligent risk prediction framework with semantic hierarchical representation and dynamic temporal modeling capabilities. By building a hierarchical temporal model composed of an ordered sparse autoencoder and BiGRU, the system achieves semantic deconstruction, risk structure representation, and continuous dynamic prediction of postoperative multi-source data. The core of this invention lies in designing a sparse latent space with explicit semantic order. This allows heterogeneous features such as test indicators, vital sign sequences, activity levels, imaging measurements, and medication responses to be progressively encoded in a unified latent variable space according to a progression of "global trend—mesoscale change—short-term shift," overcoming the limitations of traditional deep models such as disordered latent variables, semantic drift, and difficulty in distinguishing multi-scale risk patterns. By introducing BiGRU for bidirectional temporal modeling of the sparse latent space sequence, the system can accurately characterize the nonlinear changes in the postoperative recovery trajectory and identify precursor signals and potential high-risk shifts. Furthermore, this invention maintains the stability of the latent space semantic structure through nested truncation loss and a hierarchical freezing mechanism for latent variables, enabling the prediction results to have good interpretability and traceability, facilitating clinical validation and application. This invention constructs a unified modeling system for complex, multimodal, and strongly temporal data after vertebral fracture surgery, enabling prospective risk identification and precise grading assessment within a hospital information system environment, effectively improving the intelligence and refinement of postoperative management of vertebral fractures.

[0004] This invention provides a machine learning-based postoperative risk assessment system for vertebral fractures, which includes: a postoperative data acquisition module, a preprocessing and feature construction module, a risk assessment module, and a result display and early warning intervention module.

[0005] The postoperative data acquisition module generates raw multi-source postoperative data.

[0006] The preprocessing and feature construction module, deployed on the processor of the risk assessment server, is configured to: perform patient identification desensitization, timestamp alignment, outlier removal, missing value imputation, and numerical normalization on the original multi-source postoperative data; align postoperative data from different sources and at different time granularities to a preset hourly time window to obtain desensitized structured data; and construct a postoperative risk assessment feature set based on the desensitized structured data.

[0007] The risk assessment module, stored in memory and executed by the processor, is configured to: load an ordered sparse autoencoder-BiGRU model; process the postoperative risk assessment feature set using the ordered sparse autoencoder-BiGRU model; and output the patient's overall postoperative risk index. The overall postoperative risk index includes scores for internal fixation failure, recurrent vertebral fracture, severe infection, deep vein thrombosis, and readmission. Based on the overall postoperative risk index, patients are classified into low-risk, medium-risk, and high-risk levels. The ordered sparse autoencoder-BiGRU model... The iGRU model is constructed as follows: Based on the BiGRU model, an ordered sparse autoencoder model is introduced into its input. The postoperative risk assessment feature set is hierarchically encoded, semantically decoupled, and sparsified. A semantically fixed and hierarchically defined sparse latent space sequence is generated through explicit sequential control, nested truncation loss optimization, and hierarchical freezing training mechanism of latent variables. This sparse latent space sequence is then used as the input of the BiGRU model to enhance the BiGRU model's ability to model the temporal evolution of postoperative risk and its class separability, thus obtaining the ordered sparse autoencoder-BiGRU model.

[0008] The results display and early warning intervention module and the risk assessment module are connected to the doctor's workstation terminal, nurse's station terminal, and mobile nursing terminal via the hospital's local area network. The configuration is as follows: On the doctor's workstation terminal, the risk scores, risk levels, and key risk factor contribution rankings of individual patients and all postoperative vertebral fracture patients in the ward are displayed in the form of lists, risk curves, and radar charts. When the risk level is high, a postoperative risk warning message is pushed to the nurse's station terminal and mobile nursing terminal, along with standardized management prompts such as nursing workflow reminders, scheduling suggestions for follow-up examination items and monitoring frequencies, and key points for patient status observation.

[0009] Furthermore, a postoperative risk assessment feature set was constructed, including an analgesic intensity feature unit, bone metabolism-inflammation dynamic feature unit, early mobilization ability feature unit, internal fixation stability change feature unit, and hemodynamic vulnerability feature unit.

[0010] The analgesic drug intensity feature unit extracts analgesic drug usage information based on the medical order records and medication execution sheets in the desensitized structured data, calculates the total dose of analgesic drugs in 24 hours, the continuous use time of the analgesia pump, the pain relief efficiency, the difference in heart rate before and after analgesia, the difference in arterial systolic blood pressure before and after analgesia, and the analgesia response stability index, and generates analgesic drug usage intensity feature data.

[0011] The bone metabolism-inflammation dynamic feature unit, based on postoperative test items in desensitized structured data, calculates the changes in serum calcium concentration and serum phosphorus concentration within three time windows, the difference between preoperative and postoperative 25-hydroxyvitamin D concentration, the slope of the trend of β-collagen cross-linked C-terminal peptide, the time point of C-reactive protein peak appearance, and the recovery speed of white blood cell count, generating bone metabolism-inflammation response feature data.

[0012] The early activity ability characteristic unit, based on the activity behavior data recorded in the desensitized structured data nursing record sheets and mobile nursing terminals, calculates the first time out of bed, the number of steps taken at the first time, the distance taken at the first time, the number of times the patient was turned over in 24 hours, the number of times the patient was seated in 24 hours, and the daily activity index, and generates early activity ability characteristic data.

[0013] The internal fixation stability change feature unit calculates screw position offset, screw tilt angle change, vertebral body anterior edge height maintenance ratio, Cobb angle rebound, and bone cement diffusion balance index based on the measurements of immediate postoperative images and 72-hour postoperative images in the desensitized structured data, generating dynamic change feature data of internal fixation stability.

[0014] The hemodynamic vulnerability feature unit calculates heart rate variability RMSSD, blood pressure variability BPV, autonomic nervous function index, blood oxygen recovery rate, and heart rate-blood pressure coupling index based on continuous vital sign monitoring data in desensitized structured data, generating hemodynamic vulnerability feature data.

[0015] Furthermore, the ordered sparse autoencoder-BiGRU model includes a feature processing building unit, a sparse coding generation unit, and a temporal risk assessment unit:

[0016] The feature processing construction unit performs missing value imputation, outlier removal, trend smoothing, unit normalization and multi-source feature alignment on the feature data of the postoperative risk assessment feature set to construct structured postoperative risk assessment data.

[0017] The sparse coding generation unit inputs structured postoperative risk assessment data into an ordered sparse autoencoder model, performs prefix nesting truncation, sparse activation constraints and multi-level nested reconstruction loss optimization, and obtains a sparse latent space sequence.

[0018] The temporal risk assessment unit inputs the sparse latent space sequence into the BiGRU model and outputs the patient's overall postoperative risk index.

[0019] Furthermore, the sparse coding generation unit includes an explicit sequence control unit, a truncation loss optimization unit, a latent variable freezing control unit, and a latent space generation unit;

[0020] An explicit sequence control unit is introduced into the ordered sparse autoencoder model. The explicit sequence control mechanism includes latent variable sequential encoding and a prefix truncation mechanism. Through latent variable sequential encoding, the structural priority of each latent variable output by the encoder of the ordered sparse autoencoder model is preset from high-order semantics to low-order semantics, restricting the commutativity of different latent variable dimensions, so that each dimension of the latent space maintains a stable and fixed semantic position during training, thereby establishing an explicit semantic order of latent space dimensions. The latent variable dimensions are divided into high-order latent variables and low-order latent variables. Through the prefix truncation mechanism, the latent variable representation of the structured postoperative risk assessment data is truncated stepwise from front to back and from high-order dimensions to low-order dimensions based on the explicit semantic order, so as to constrain the latent variables to be ordered, so that the latent variable sets corresponding to different truncation lengths form a strictly nested hierarchical structure, thereby constructing a semantically fixed and hierarchically progressive nested latent space structure.

[0021] The truncation loss optimization unit constructs a nested truncation loss function based on the nested latent space structure. This function serves as the core optimization objective of the ordered sparse autoencoder model. During training, the ordered sparse autoencoder model is trained by minimizing the nested truncation loss function, thereby achieving consistency reinforcement of the semantic order in the latent space and hierarchical encoding of multi-scale risk structures. The nested truncation loss function is constructed as follows: it is based on the reconstruction loss function used in traditional autoencoders; multi-level truncation reconstruction constraints for prefix truncation structures, truncation consistency constraints to maintain semantic stability at different truncation levels, and a hierarchical sparse regularization mechanism to distinguish the hierarchical responsibilities of latent variables are introduced to structurally optimize the reconstruction loss function. This enables the model to maintain a stable semantic order and hierarchical progression at multiple truncation depths simultaneously, thus constructing the nested truncation loss function.

[0022] The latent variable freezing control unit addresses the uneven training pressure caused by minimizing the nested truncation loss function during the training of an ordered sparse autoencoder model, which leads to frequent reconstruction of low-order latent variables and insufficient gradient updates for high-order latent variables. A stepwise latent variable freezing strategy is introduced to identify latent variable dimensions that have reached semantic convergence by monitoring their gradient trends. This strategy freezes the corresponding weight rows in the encoder and the corresponding dictionary columns in the decoder, ensuring semantic stability during subsequent training phases. Simultaneously, parameter updates are only performed on unfrozen latent variable dimensions to alleviate the uneven gradient distribution caused by prefix truncation.

[0023] After completing the explicit sequential control, nested truncation loss optimization, and hierarchical freezing of latent variables training processes in steps S21 to S23, the latent space generation unit obtains a trained ordered sparse autoencoder model. The trained ordered sparse autoencoder model is then used to encode postoperative risk assessment data for different monitoring time windows, resulting in a semantically fixed and hierarchically defined multi-time-window sparse latent space sequence.

[0024] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0025] Firstly, this invention constructs an ordered sparse autoencoder with explicit semantic order, enabling hierarchical representation of multi-source clinical data in a unified latent space. This addresses the modeling difficulties caused by the complexity of data sources, heterogeneity of time scales, and strong coupling relationships between risk factors after vertebral fracture surgery. The latent space is progressively organized according to a semantic structure of "global recovery trend—mesoscale physiological changes—short-term offset signal," ensuring that key data such as bone metabolism indicators, changes in inflammatory responses, fluctuations in vital signs, imaging measurement parameters, and activity behaviors are stably presented in their corresponding semantic levels. This improves the clarity and interpretability of risk structure representation. This invention thus enhances the ability to deconstruct and understand complex postoperative data patterns, providing technical support for identifying potential sources of risk changes and analyzing postoperative state evolution.

[0026] Secondly, this invention utilizes BiGRU to perform bidirectional temporal modeling of the aforementioned sparse latent space sequences, achieving continuous capture of the postoperative state evolution process and overcoming the shortcomings of traditional models in identifying medium-term change trends and short-term shift patterns. Based on the bidirectional temporal path, this invention can combine historical information with future change trends, comprehensively modeling the global rate of change, local dynamic characteristics, and short-term abnormal fluctuations, thereby significantly improving the ability to identify risk change patterns. Through this technical approach, this invention can identify potential risk shifts in advance, enabling the system to indicate potential risk change trends before obvious abnormalities appear in postoperative data, improving the timeliness and precision of risk assessment.

[0027] Thirdly, this invention introduces a nested truncation loss and a hierarchical freezing mechanism for latent variables, ensuring the consistency of the latent space semantic structure during training and inference. This solves the problems of semantic drift of latent variables and instability of risk structure in traditional deep models. While maintaining clear distinctions between global, local, and individual hierarchical features, this mechanism ensures higher consistency and interpretability of the risk index across different patients and monitoring time windows, significantly enhancing the reliability and usability of the system in real-world hospital information system environments. Based on this stable latent space structure and temporal modeling capabilities, this invention can continuously output structured risk levels and key feature contributions, improving the continuity, technical credibility, and information presentation quality of postoperative risk monitoring, thereby providing technical support for hospitals to achieve refined operational management and postoperative risk information management. Attached Figure Description

[0028] Figure 1 A schematic diagram of the overall structure of a machine learning-based postoperative risk assessment system for vertebral fractures provided by the present invention;

[0029] Figure 2 This is a radar chart showing the five risk categories after vertebral fracture surgery as proposed in Example 5. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] Example 1, according to Figure 1 This invention provides a machine learning-based postoperative risk assessment system for vertebral fracture surgery, applied to the postoperative management of patients undergoing vertebral fracture surgery in the orthopedics and spinal surgery departments of comprehensive hospitals. The system is deployed on a risk assessment server in the hospital's data center. The risk assessment server is a rack-mounted industrial server, including a processor, memory, and network interface. It connects to the hospital's local area network and communicates with the Hospital Information System (HIS), Electronic Medical Record System (EMR), Image Archiving and Communication System (PACS), and Laboratory Information System (LIS). The system includes: a postoperative data acquisition module, a preprocessing and feature construction module, a risk assessment module, and a result display and early warning intervention module.

[0032] The postoperative data acquisition module is configured as follows: It establishes data interfaces with the Hospital Information System (HIS), Electronic Medical Record System (EMR), Image Archiving and Communication System (PACS), and Laboratory Information System (LIS) via network interfaces. Using HL7 and DICOM protocols, it automatically acquires basic demographic information, hospitalization information, surgical segment number, internal fixation material type, intraoperative blood loss, surgical duration, preoperative and postoperative imaging parameters, complete blood count, liver and kidney function tests, coagulation function tests, electrolyte levels, and analgesic medication orders from the HIS, EMR, PACS, and LIS. It also acquires continuous heart rate, systolic blood pressure, diastolic blood pressure, oxygen saturation, and body temperature monitoring data through the bedside monitoring equipment connected to the gateway. Pain scores, functional scores, history of osteoporosis, fall history, and activity records are entered through doctor workstation terminals, nurse station terminals, and mobile ward round terminals, forming the original multi-source postoperative data.

[0033] The preprocessing and feature construction module, deployed on the processor of the risk assessment server, is configured to: perform patient identification desensitization, timestamp alignment, outlier removal, missing value imputation, and numerical normalization on the original multi-source postoperative data; align postoperative data from different sources and at different time granularities to a preset hourly time window to obtain desensitized structured data; and construct a postoperative risk assessment feature set based on the desensitized structured data.

[0034] The risk assessment module, stored in memory and executed by the processor, is configured to: load an ordered sparse autoencoder-BiGRU model; process the postoperative risk assessment feature set using the ordered sparse autoencoder-BiGRU model; and output the patient's overall postoperative risk index. The overall postoperative risk index includes scores for internal fixation failure, recurrent vertebral fracture, severe infection, deep vein thrombosis, and readmission. Based on the overall postoperative risk index, patients are classified into low-risk, medium-risk, and high-risk levels. The ordered sparse autoencoder-BiGRU model... The iGRU model is constructed as follows: Based on the BiGRU model, an ordered sparse autoencoder model is introduced into its input. The postoperative risk assessment feature set is hierarchically encoded, semantically decoupled, and sparsified. A semantically fixed and hierarchically defined sparse latent space sequence is generated through explicit sequential control, nested truncation loss optimization, and hierarchical freezing training mechanism of latent variables. This sparse latent space sequence is then used as the input of the BiGRU model to enhance the BiGRU model's ability to model the temporal evolution of postoperative risk and its class separability, thus obtaining the ordered sparse autoencoder-BiGRU model.

[0035] The results display and early warning intervention module and the risk assessment module are connected to the doctor's workstation terminal, nurse's station terminal, and mobile nursing terminal via the hospital's local area network. The configuration is as follows: On the doctor's workstation terminal, the risk scores, risk levels, and key risk factor contribution rankings of individual patients and all postoperative vertebral fracture patients in the ward are displayed in the form of lists, risk curves, and radar charts. When the risk level is high, a postoperative risk warning message is pushed to the nurse's station terminal and mobile nursing terminal, along with standardized management prompts such as nursing workflow reminders, scheduling suggestions for follow-up examination items and monitoring frequencies, and key points for patient status observation.

[0036] Example 2, based on Example 1, constructs a postoperative risk assessment feature set including analgesic drug intensity feature unit, bone metabolism-inflammation dynamic feature unit, early mobilization ability feature unit, internal fixation stability change feature unit, and hemodynamic vulnerability feature unit;

[0037] The analgesic drug intensity feature unit extracts analgesic drug usage information based on the medical order records and medication execution sheets in the desensitized structured data, calculates the total dose of analgesic drugs in 24 hours, the continuous use time of the analgesia pump, the pain relief efficiency, the difference in heart rate before and after analgesia, the difference in arterial systolic blood pressure before and after analgesia, and the analgesia response stability index, and generates analgesic drug usage intensity feature data.

[0038] The bone metabolism-inflammation dynamic feature unit, based on postoperative test items in desensitized structured data, calculates the changes in serum calcium concentration and serum phosphorus concentration within three time windows, the difference between preoperative and postoperative 25-hydroxyvitamin D concentration, the slope of the trend of β-collagen cross-linked C-terminal peptide, the time point of C-reactive protein peak appearance, and the recovery speed of white blood cell count, generating bone metabolism-inflammation response feature data.

[0039] The early activity ability characteristic unit, based on the activity behavior data recorded in the desensitized structured data nursing record sheets and mobile nursing terminals, calculates the first time out of bed, the number of steps taken at the first time, the distance taken at the first time, the number of times the patient was turned over in 24 hours, the number of times the patient was seated in 24 hours, and the daily activity index, and generates early activity ability characteristic data.

[0040] The internal fixation stability change feature unit calculates screw position offset, screw tilt angle change, vertebral body anterior edge height maintenance ratio, Cobb angle rebound, and bone cement diffusion balance index based on the measurements of immediate postoperative images and 72-hour postoperative images in the desensitized structured data, generating dynamic change feature data of internal fixation stability.

[0041] The hemodynamic vulnerability feature unit calculates heart rate variability RMSSD, blood pressure variability BPV, autonomic nervous function index, blood oxygen recovery rate, and heart rate-blood pressure coupling index based on continuous vital sign monitoring data in desensitized structured data, generating hemodynamic vulnerability feature data.

[0042] Example 3, based on Example 2, describes an ordered sparse autoencoder-BiGRU model that includes a feature processing construction unit, a sparse coding generation unit, and a temporal risk assessment unit.

[0043] The feature processing construction unit performs missing value imputation, outlier removal, trend smoothing, unit normalization and multi-source feature alignment on the feature data of the postoperative risk assessment feature set to construct structured postoperative risk assessment data.

[0044] The sparse coding generation unit inputs structured postoperative risk assessment data into an ordered sparse autoencoder model, performs prefix nesting truncation, sparse activation constraints and multi-level nested reconstruction loss optimization, and obtains a sparse latent space sequence.

[0045] The temporal risk assessment unit inputs the sparse latent space sequence into the BiGRU model, and uses its forward and backward temporal paths to simultaneously model the dynamic evolution trend of multidimensional states during postoperative recovery, generating a prediction vector containing multiple risk dimensions; and performs normalization processing and inter-task risk correlation correction on the prediction vector, outputting the patient's overall postoperative risk index.

[0046] In conventional technology fields, the BiGRU model includes a feature processing unit, a feature encoding unit, and a temporal modeling unit;

[0047] The feature processing construction unit performs missing value imputation, outlier removal, trend smoothing, unit normalization and multi-source feature alignment on the feature data of the postoperative risk assessment feature set to construct structured postoperative risk assessment data.

[0048] The feature encoding unit takes structured postoperative risk assessment data as input and uses conventional feature encoding methods to vectorize various features to obtain the encoded feature sequence.

[0049] The temporal modeling unit inputs the encoded feature sequence into the BiGRU model and uses its forward and backward temporal paths to simultaneously model the dynamic evolution trend of multidimensional states during postoperative recovery, generating a prediction vector containing multiple risk dimensions. The prediction vector is then normalized and the risk correlation between tasks is corrected to output the patient's overall postoperative risk index.

[0050] Example 4, based on Example 3, in which the sparse coding generation unit includes an explicit sequence control unit, a truncation loss optimization unit, a latent variable freezing control unit, and a latent space generation unit;

[0051] An explicit sequence control unit is introduced into the ordered sparse autoencoder model, which includes latent variable sequential encoding and a prefix truncation mechanism. Through latent variable sequential encoding, the structural priority of each latent variable in the encoder output of the ordered sparse autoencoder model is preset from high-order semantics to low-order semantics, restricting the commutativity of different latent variable dimensions. This ensures that each dimension of the latent space maintains a stable and fixed semantic position during training, thereby establishing an explicit semantic order of the latent space dimensions. The latent variable dimensions are divided into high-order latent variables (corresponding to global and stable semantics) and low-order latent variables. The quantities (corresponding to locality and fine-grained semantics) are as follows: high-order latent variables are used to characterize the global, long-term, and physiological stability trends during the postoperative recovery process; low-order latent variables are used to characterize individual patient differences, local short-term changes, and abnormal risk shift patterns. Through a prefix truncation mechanism, the latent variable representation of structured postoperative risk assessment data is truncated from front to back and from high-order to low-order dimensions based on the explicit semantic order, so as to impose ordered constraints on the latent variables, so that the latent variable sets corresponding to different truncation lengths form a strictly nested hierarchical structure, thereby constructing a semantically fixed and hierarchically progressive nested latent space structure.

[0052] The truncation loss optimization unit, based on the nested latent space structure, constructs a nested truncation loss function as the core optimization objective of the ordered sparse autoencoder model. During the training process of the ordered sparse autoencoder model, the nested truncation loss function is minimized to train the model, thereby achieving consistency enhancement of the semantic order of the latent space and hierarchical encoding of multi-scale risk structures. Specifically, this includes: 1. Constructing reconstruction constraints for different truncation lengths: Constructing reconstruction constraints for different truncation lengths through the nested truncation loss function, so that high-order latent variables converge first and maintain stable semantics under the condition of high-frequency participation in reconstruction, while low-order latent variables undertake the fine-grained representation task of individual differences and short-term disturbances in the subsequent truncation stage, thus ensuring that the hierarchical structure of the latent space remains fixed and reliable; 2. Performing stepwise sparse optimization to form multi-scale expressions: By performing stepwise sparse optimization on the set of latent variables with different truncation lengths, the global postoperative recovery trend, subacute physiological change pattern, and patient-specific risk offset information are formed in the latent space in a multi-level expression from coarse to fine and from stable to specific.

[0053] The nested truncation loss function is constructed as follows: Based on the reconstruction loss function used in traditional autoencoders, multi-level truncation reconstruction constraints for prefix truncation structures, truncation consistency constraints to maintain semantic stability at different truncation levels, and hierarchical sparse regularization mechanisms to distinguish the hierarchical responsibilities of latent variables are introduced to structurally optimize the reconstruction loss function, enabling the model to maintain a stable semantic order and hierarchical progression relationship at multiple truncation depths at the same time, thus constructing the nested truncation loss function.

[0054] Nested truncation loss function formula:

[0055] ;

[0056] in, Represents the decoder (dictionary matrix). Indicates encoder, This indicates a nested truncation loss function; Indicates the cut-off length From 1 to Add them one by one; Indicates the cut-off length is Sampling probability weights at time; This represents the input data matrix, which in this invention is post-operative risk assessment data. This indicates that the encoder processes the input data. The output of this is the latent space representation; Indicates the encoding result Perform Top-m sparse selection operation. Indicates the prefix truncation operator; Denotes the square of the Frobenius norm;

[0057] The loss function is used to ensure that the model can simultaneously capture the following in postoperative recovery: global trends (such as bone metabolism recovery and the downward trend of chronic inflammation), mesoscale physiological changes (such as improved mobility and pain fluctuations), and individualized short-term risk signals (such as early signs of infection and sudden increase in the risk of thrombosis).

[0058] The latent variable freezing control unit addresses the uneven training pressure distribution caused by minimizing the nested truncation loss function during the training of an ordered sparse autoencoder model. This leads to frequent reconstruction of low-order latent variables and insufficient gradient updates for high-order latent variables. A progressive latent variable freezing strategy is introduced to identify latent variable dimensions that have reached semantic convergence by monitoring their gradient trends. The corresponding weight rows in the encoder and the corresponding dictionary columns in the decoder are then frozen to maintain semantic stability during subsequent training. Simultaneously, parameter updates are performed only on unfrozen latent variable dimensions to alleviate the uneven gradient distribution caused by prefix truncation. This ensures that all latent variable dimensions receive sufficient and balanced training opportunities, thereby preserving the semantic order of latent variables, solidifying the hierarchical structure, and enhancing the overall interpretability of the model. This strategy effectively avoids the uneven training pressure distribution caused by progressive truncation, making the functions of different latent variables clearer and more interpretable.

[0059] After completing the explicit sequential control, nested truncation loss optimization, and hierarchical freezing of latent variables training processes in steps S21 to S23, the latent space generation unit obtains a trained ordered sparse autoencoder model. The trained ordered sparse autoencoder model is then used to encode postoperative risk assessment data for different monitoring time windows, resulting in a semantically fixed and hierarchically defined multi-time-window sparse latent space sequence.

[0060] Example 5, according to Figure 2 This embodiment is based on Embodiment 4. In this embodiment,

[0061] The risk assessment module, stored in memory and executed by the processor, is configured to: load the ordered sparse autoencoder-BiGRU model, process the postoperative risk assessment feature set through the ordered sparse autoencoder-BiGRU model, and output the patient's overall postoperative risk index; the patient's overall postoperative risk index includes the risk score of internal fixation failure, the risk score of recurrent vertebral fracture, the risk score of severe infection, the risk score of deep vein thrombosis, and the risk score of readmission.

[0062] The overall postoperative risk index of patient A is shown in Table 1:

[0063] Table 1

[0064] ;

[0065] The system further performs unified normalization and correlation correction on the above five-dimensional risk vectors, resulting in a postoperative overall risk index R_total=0.67;

[0066] Patient A's R_total=0.67, and the system automatically categorized him as a high-risk patient.

[0067] The results display and early warning intervention module and the risk assessment module are connected to the doctor's workstation terminal, nurse's station terminal, and mobile nursing terminal via the hospital's local area network. The configuration is as follows: On the doctor's workstation terminal, the risk scores, risk levels, and key risk factor contribution rankings of individual patients and all postoperative vertebral fracture patients in the ward are displayed in the form of lists, risk curves, and radar charts. When the risk level is high, a postoperative risk warning message is pushed to the nurse's station terminal and mobile nursing terminal, along with standardized management prompts such as nursing workflow reminders, scheduling suggestions for follow-up examination items and monitoring frequencies, and key points for patient status observation.

[0068] The doctor's workstation interface automatically displays the risk analysis results for patient A:

[0069] Figure 2 The radar chart for five risk categories after vertebral fracture surgery: This radar chart contains five risk dimensions, which correspond to different postoperative risk directions calculated by the system: risk of internal fixation failure, risk of recurrent vertebral fracture, risk of serious infection, risk of deep vein thrombosis, and risk of readmission.

[0070] Key feature contribution ranking (TOP5):

[0071] The magnitude of change in serum calcium (contribution: 0.21).

[0072] CRP peak time delayed (contribution: 0.18).

[0073] Screw tilt angle change (contribution: 0.17);

[0074] Heart rate variability was low (contribution: 0.16);

[0075] The first time the patient got out of bed was relatively late (contribution: 0.12);

[0076] The system also displays the overall risk distribution of the patient's ward:

[0077] High risk: 6 people;

[0078] Medium risk: 12 people;

[0079] Low risk: 24 people.

[0080] The system automatically pushes the following content to the nurse station terminal:

[0081] [Risk Warning: Patient A (Hospital Number: 0183956)]

[0082] Current risk level: High risk (0.67);

[0083] Key risk dimensions: risk of refracture, risk of readmission;

[0084] The risk factors contributing the most to the risk were: serum calcium fluctuations, delayed CRP peak, and changes in screw tilt angle.

[0085] System management prompt:

[0086] It is recommended to increase the frequency of recording vital signs and activity data (once every 4 hours);

[0087] Post-operative follow-up appointments: Follow-up imaging on the 3rd day post-surgery;

[0088] Pay attention to the quality of recording information on the patient's standing and walking stability;

[0089] Provide nursing process prompts: assist in recording the number of times the patient turns over, the walking distance, and key points of postural stability.

[0090] The mobile message received by nursing staff is as follows:

[0091] [Postoperative Risk Warning] Patient A (Bed No. 23)

[0092] Risk level: High risk;

[0093] Recent risk trend: trending upward;

[0094] Key areas of focus: quality of activity and behavior records, and monitoring of vital sign fluctuations;

[0095] Key indicators to watch as suggested by the system:

[0096] Changes in the daily activity index;

[0097] Temporary deviations in blood pressure and heart rate;

[0098] Records of rolling over and getting out of bed.

[0099] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A machine learning-based postoperative risk assessment system for vertebral fractures, characterized in that, The system includes: a postoperative data acquisition module, a preprocessing and feature construction module, a risk assessment module, and a risk assessment server; the risk assessment server includes a processor and memory. The postoperative data acquisition module generates raw multi-source postoperative data. The preprocessing and feature construction module, deployed on the processor of the risk assessment server, is configured to: perform preprocessing on the original multi-source postoperative data to obtain desensitized structured data; and construct a postoperative risk assessment feature set based on the desensitized structured data. The risk assessment module, stored in memory and executed by the processor, is configured to: load the ordered sparse autoencoder-BiGRU model, process the postoperative risk assessment feature set through the ordered sparse autoencoder-BiGRU model, and output the patient's overall postoperative risk index.

2. The machine learning-based postoperative risk assessment system for vertebral fractures according to claim 1, characterized in that: The postoperative risk assessment feature set was constructed, including the analgesic intensity feature unit, the bone metabolism-inflammation dynamic feature unit, the early mobilization ability feature unit, the internal fixation stability change feature unit, and the hemodynamic vulnerability feature unit.

3. The machine learning-based postoperative risk assessment system for vertebral fractures according to claim 1, characterized in that: The ordered sparse autoencoder-BiGRU model is constructed as follows: based on the BiGRU model, an ordered sparse autoencoder model is introduced into its input end to enhance the temporal modeling ability and class separability of the BiGRU model for the evolution of postoperative risk, thus constructing the ordered sparse autoencoder-BiGRU model.

4. The machine learning-based postoperative risk assessment system for vertebral fractures according to claim 2, characterized in that: The analgesic drug intensity feature unit, based on the medical order records and medication execution sheets in the desensitized structured data, calculates the total dose of analgesic drugs, the continuous use time of the analgesia pump, the pain relief efficiency, the difference in heart rate before and after analgesia, the difference in arterial systolic blood pressure before and after analgesia, and the analgesia response stability index, and generates analgesic drug use intensity feature data. The bone metabolism-inflammation dynamic feature unit, based on postoperative test items in desensitized structured data, calculates the change range of serum calcium concentration, the change range of serum phosphorus concentration, the difference between preoperative and postoperative 25-hydroxyvitamin D concentration, the slope of the trend of β-collagen cross-linked C-terminal peptide, the time point of C-reactive protein peak appearance, and the recovery speed of white blood cell count, generating bone metabolism-inflammation response feature data. Early activity ability characteristic unit, based on nursing record sheets and activity behavior data recorded by mobile nursing terminals in desensitized structured data, generates early activity ability characteristic data; The internal fixation stability change feature unit calculates screw position offset, screw tilt angle change, vertebral body anterior edge height maintenance ratio, Cobb angle rebound, and bone cement diffusion balance index based on the measured values ​​of immediate postoperative images and postoperative images in the desensitized structured data, generating dynamic change feature data of internal fixation stability. The hemodynamic vulnerability feature unit calculates heart rate variability RMSSD, blood pressure variability BPV, autonomic nervous function index, blood oxygen recovery rate, and heart rate-blood pressure coupling index based on continuous vital sign monitoring data in desensitized structured data, generating hemodynamic vulnerability feature data.

5. The machine learning-based postoperative risk assessment system for vertebral fractures according to claim 3, characterized in that: The ordered sparse autoencoder-BiGRU model includes a feature processing building unit, a sparse coding generation unit, and a temporal risk assessment unit. The feature processing construction unit processes the feature data of the postoperative risk assessment feature set to construct structured postoperative risk assessment data. The sparse coding generation unit inputs structured postoperative risk assessment data into an ordered sparse autoencoder model, performs prefix nesting truncation, sparse activation constraints and multi-level nested reconstruction loss optimization, and obtains a sparse latent space sequence. The temporal risk assessment unit inputs the sparse latent space sequence into the BiGRU model and outputs the patient's overall postoperative risk index.

6. The machine learning-based postoperative risk assessment system for vertebral fractures according to claim 5, characterized in that: The sparse coding generation unit includes an explicit sequence control unit, a truncation loss optimization unit, a latent variable freezing control unit, and a latent space generation unit; An explicit sequence control unit is introduced into the ordered sparse autoencoder model. This explicit sequence control mechanism includes latent variable sequential encoding and a prefix truncation mechanism. Through latent variable sequential encoding, the dimensions of each latent variable output by the encoder of the ordered sparse autoencoder model are pre-prioritized, restricting the commutativity of different latent variable dimensions and establishing an explicit semantic order. The prefix truncation mechanism, based on the explicit semantic order, performs step-by-step truncation of the latent variable representation of the structured postoperative risk assessment data from front to back and from higher to lower order dimensions, constructing a nested latent space structure. The latent variable dimensions are divided into high-order latent variables and low-order latent variables. The truncation loss optimization unit constructs a nested truncation loss function based on the nested latent space structure, which serves as the core optimization objective of the ordered sparse autoencoder model. The ordered sparse autoencoder model is trained by minimizing the nested truncation loss function. The latent variable freezing control unit addresses the issue that minimizing the nested truncation loss function during the training of the ordered sparse autoencoder model leads to frequent reconstruction of low-order latent variables and insufficient gradient updates for high-order latent variables. A step-by-step latent variable freezing strategy is introduced to identify latent variable dimensions that have reached semantic convergence by monitoring the gradient change trends of each latent variable dimension and then performing freezing control. Parameter updates are only performed on unfrozen latent variable dimensions to alleviate the uneven gradient distribution problem caused by prefix truncation. The latent space generation unit obtains a trained sparse autoencoder model after completing the training process from step E22 to step E23; and outputs a sparse latent space sequence using the trained sparse autoencoder model.

Images