A postoperative infection risk early warning system for craniocerebral trauma patients

Abstract

The application discloses a kind of postoperative infection risk early warning system of craniocerebral trauma patient.The system uses "multimodal perception-privacy computing-interpretable early warning-closed loop intervention" four-layer integrated architecture, including multimodal perception and fusion module, privacy computing hub module, interpretable early warning engine and intelligent closed loop intervention module connected in turn.Communication connection is integrated millimeter wave radar, flexible sensor array and MS@SLET sensor chip in multimodal perception and fusion module, realize multidimensional data high-fidelity acquisition;Privacy computing hub module is based on improved FlexFair federated learning framework, and data privacy protection and group prediction fairness are considered;Interpretable early warning engine realizes accurate early warning and generates structured attribution report by the CNN-Transformer hybrid model of stage perception attention mechanism;Intelligent closed loop intervention module realizes hierarchical early warning, individualized scheme matching and model incremental optimization.The application solves the bottleneck of the prior art, such as insufficient timeliness of early warning, imbalance between data utilization and privacy fairness, uninterpretable decision and early warning intervention fragmentation, shortens the infection index detection time to within 30 seconds, the system response time is less than 200 ms, the infection-related mortality rate is reduced by 65.4%, fully meets the requirements of laws and regulations such as "Personal Information Protection Law", and has very high clinical application value.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical informatics, artificial intelligence and biomedical engineering. Specifically, it is an intelligent early warning system for postoperative infection in patients with craniocerebral trauma. Background Technology

[0002] Traumatic brain injury (TBI) is a common acute and critical illness, and postoperative intracranial infection, surgical site infection, and other related complications are core risk factors that restrict patient clinical prognosis and increase mortality and disability rates. Especially in neurosurgical intensive care settings, postoperative infections are characterized by short incubation periods, rapid disease progression, and narrow intervention windows. Achieving early and accurate warning of postoperative infection risks and timely and standardized intervention is crucial for shortening patient recovery periods, reducing the incidence of adverse outcomes, and improving the success rate of critical care. This has significant practical implications and application value for neurosurgical clinical diagnosis and treatment.

[0003] Currently, clinical prevention and early warning of postoperative infection risk in patients with traumatic brain injury mainly rely on traditional laboratory biochemical tests, subjective judgment based on the clinical experience of medical staff, and basic monitoring using single-parameter physiological monitoring equipment. Existing conventional techniques and similar early warning schemes have many inherent technical limitations in actual clinical application, making it difficult to meet the core requirements of "early, accurate, safe, efficient, and implementable" infection early warning in neurosurgical intensive care settings. Specific technical limitations are as follows:

[0004] Current technology has the following drawbacks:

[0005] 1. Delayed pathogen detection leads to a lack of early warning capabilities.

[0006] Currently, the detection of cerebrospinal fluid infection-related indicators mainly uses traditional methods such as routine microbial culture and smear microscopy. The complete detection cycle takes 24-48 hours, which cannot achieve rapid screening and immediate interpretation of infection-related indicators. At the same time, the existing detection methods are mostly invasive procedures and highly dependent on biochemical markers, making it difficult to achieve continuous dynamic monitoring. This can easily lead to missing the golden window period for early intervention of postoperative infection, resulting in the spread of infection and aggravation of the patient's condition.

[0007] 2. Monitoring dimensions are singular and fragmented, and multimodal information has not been effectively integrated.

[0008] Existing clinical monitoring equipment is mostly a single-function, independent module that can only collect basic vital signs parameters such as heart rate, respiration, and body temperature, or monitor intracranial pressure alone. It cannot simultaneously integrate multi-dimensional core monitoring information such as cerebrospinal fluid pathogenic characteristics, real-time physiological time-series characteristics, and static anatomical characteristics of the brain. The problems of data silos and fragmentation are prominent, and the risk assessment is based on one-sided and single criteria, which directly leads to high false alarm and false negative rates in early warning models, and the accuracy of early warning cannot meet clinical needs.

[0009] 3. Multi-center collaborative modeling is limited, and the contradiction between data privacy and model generalization is prominent.

[0010] Medical data related to traumatic brain injury and infection is highly sensitive medical privacy data. Due to data compliance management and privacy protection requirements, it is difficult for medical centers to achieve secure interoperability and compliant sharing of data, resulting in serious data silos and making it impossible to carry out multi-center collaborative modeling to optimize the model's generalization ability. Traditional centralized modeling models require the aggregation of original patient data from various centers, which poses a great risk of data leakage and compliance issues. Furthermore, existing conventional federated learning frameworks do not take into account the predictive fairness of different patient groups, resulting in significant differences in early warning accuracy between groups and insufficient clinical adaptability of the model.

[0011] 4. The early warning results lack interpretability, making clinical application difficult.

[0012] Existing AI-based infection early warning models are mostly black-box models, capable of outputting only a single risk score. They cannot quantify the contribution weight of each risk factor to the early warning result, nor can they generate reasoning logic and attribution analysis reports that align with clinical diagnosis and treatment thinking. Clinical medical staff find it difficult to intuitively interpret the basis of the early warning and the causes of the risk, and cannot quickly develop personalized intervention plans. The early warning results are disconnected from actual clinical applications, significantly reducing their practicality.

[0013] 5. Lack of closed-loop intervention and iterative optimization mechanisms leads to easy degradation of system performance.

[0014] Existing early warning systems of the same type only have the single function of "data collection-risk warning" and have not built a complete business closed loop of "risk warning-tiered intervention-feedback collection-model optimization". After the warning is triggered, it cannot automatically match and push tiered personalized intervention plans. At the same time, the system lacks a clinical intervention feedback data collection channel and a model incremental optimization mechanism. After long-term operation, the model is prone to performance degradation, catastrophic amnesia and other problems, and cannot adapt to the dynamic changes in clinical scenarios, making it difficult to maintain a high-precision early warning state.

[0015] 6. The lack of an edge-cloud collaborative architecture makes it difficult to balance real-time performance and security.

[0016] Existing early warning systems mostly adopt a centralized cloud data processing mode, with all raw monitoring data being directly uploaded to the cloud for processing. This results in long data transmission links and high interaction latency, which cannot meet the real-time response requirements of neurosurgical intensive care scenarios. Furthermore, they have not carried out dedicated encryption and identity authentication optimizations for medical data transmission scenarios, resulting in insufficient data transmission security and making it difficult to simultaneously meet the real-time requirements of intensive care and the privacy requirements of medical data.

[0017] In response to the aforementioned core deficiencies in existing technologies, there is currently no comprehensive integrated solution in clinical practice or similar technologies, making it impossible to achieve early, accurate, safe, and practical closed-loop early warning of postoperative infections following traumatic brain injury. Summary of the Invention

[0018] The purpose of this invention is to provide a postoperative infection risk warning system for patients with traumatic brain injury.

[0019] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a postoperative infection risk early warning system for patients with traumatic brain injury, comprising:

[0020] The multimodal perception and fusion module communicates with the privacy computing hub module via HTTPS protocol to achieve data interaction. It is used to collect multi-dimensional raw data of patients and complete feature fusion to generate multimodal feature vectors. The edge computing node module is used to perform denoising, normalization and outlier removal preprocessing operations on the data transmitted by the multimodal perception and fusion module to finally generate a standardized multimodal dataset.

[0021] The privacy computing hub module communicates with the interpretable early warning engine module via an encrypted data link. It uses the improved FlexFair federated learning framework to achieve multi-center model collaborative training and generate a global early warning model for postoperative infection in patients with traumatic brain injury. The improved FlexFair federated learning framework optimizes the prediction fairness of different groups by introducing a fairness regularization term into the local training loss function, and at the same time uses differential privacy technology to protect the privacy of the model gradient uploading process.

[0022] The interpretable early warning engine module works in collaboration with the intelligent closed-loop intervention module through a local area network. It is used to calculate the postoperative dynamic infection risk score of patients with traumatic brain injury based on multimodal feature vectors and through a CNN-Transformer hybrid model that integrates stage perception attention mechanism, and generate a structured attribution report corresponding to the infection risk score.

[0023] The intelligent closed-loop intervention module, in conjunction with the interpretable early warning engine module, triggers tiered early warnings based on risk scores and matches recommended intervention plans that include diagnostic and treatment suggestions. It also optimizes the early warning model based on clinical feedback data.

[0024] A method for processing postoperative infection risk warning data in patients with traumatic brain injury includes the following steps:

[0025] S1: Collect multi-dimensional raw data of patients through the multimodal perception and fusion module, and perform feature fusion using an improved multi-head attention mechanism to generate multimodal feature vectors;

[0026] S2: Based on the improved FlexFair federated learning framework, combined with differential privacy technology, multi-center model collaborative training is achieved to generate a global early warning model;

[0027] S3: Input the multimodal feature vector into the global early warning model, calculate the dynamic infection risk score, and generate a structured attribution report;

[0028] S4: Trigger tiered early warnings based on risk scores and output a list of intervention recommendations using a matching algorithm;

[0029] S5: Collect clinical intervention feedback data and incrementally optimize the global early warning model using an improved elastic weight merging algorithm.

[0030] Compared with the prior art, the significant advantages of this invention are:

[0031] 1. Breakthrough in Early Warning Timeliness: Addressing the bottleneck of insufficient early warning timeliness in existing technologies, the terahertz metasurface rapid detection unit shortens the time for obtaining infection indicators from "hours" to within 30 seconds, with an end-to-end system response time of <200ms, enabling early warning within the golden window period;

[0032] 2. Data utilization security and fairness: Addressing the bottleneck of imbalance between data utilization and privacy fairness, the FlexFair framework was improved to reduce the difference in prediction error rate among different age groups by 58%. The differential privacy technology passed third-party security audit, balancing data value mining and compliance requirements.

[0033] 3. Enhanced Clinical Trust: Addressing the bottleneck of unexplainable model decisions, structured thought chain reports increase physician adoption rates to over 80%, breaking through the bottleneck of AI "black box" implementation;

[0034] 4. Enhanced efficiency through closed-loop management: Addressing the bottleneck of the disconnect between early warning and intervention, closed-loop management throughout the entire process advances the initiation time of anti-infection treatment by 5.5 hours, reduces the infection-related mortality rate from 28.3% to 9.8%, and ensures long-term performance stability through the incremental update mechanism of the model;

[0035] 5. Technological Integration and Innovation: For the first time, terahertz fast sensing, fair federated learning, and interpretable AI systems are systematically integrated to form a high technological barrier and provide a standardized solution for postoperative infection management in neurosurgery. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0037] Figure 2 This is a schematic diagram illustrating the working principle of the terahertz metasurface biosensor chip of the present invention.

[0038] Figure 3 This is a schematic diagram of the explanatory warning report interface of the present invention.

[0039] Figure 4 This is a schematic diagram of the intelligent closed-loop intervention workflow of the present invention. Detailed Implementation

[0040] This invention addresses various shortcomings of existing technologies by improving core technologies such as multimodal perception fusion, privacy-compliant federated learning, interpretable early warning, edge-cloud collaborative deployment, and closed-loop iterative optimization. It constructs a comprehensive postoperative infection risk early warning system adapted to neurosurgical intensive care scenarios, comprehensively resolving the pain points of existing technologies and filling application gaps in similar technologies. A four-layer integrated architecture of "multimodal perception - privacy computing - interpretable early warning - closed-loop intervention" is proposed, with the core technical solutions as follows:

[0041] 1. Multimodal perception and fusion module (addressing the bottlenecks of insufficient timeliness of early warning and single data collection dimension in existing technologies)

[0042] Integrating three major sensing units and a multimodal feature fusion unit, it achieves high-fidelity acquisition and fusion of multi-dimensional data, solving the problems of narrow data coverage and poor timeliness in traditional single-sensor methods:

[0043] (1) High-density physiological time-series monitoring unit: A 60GHz millimeter-wave radar (non-contact) and a flexible polyimide sensor array (attached) are used to simultaneously collect heart rate, respiratory rate, intracranial pressure (ICP), brain tissue oxygen saturation and surgical incision humidity / pressure signals; continuous wavelet transform is used to extract the abnormal waveform energy ratio R_wave feature in the 0.05-0.15Hz frequency band of the intracranial pressure signal (R_wave feature refers to the ratio of abnormal waveform energy in the 0.05-0.15Hz frequency band of the intracranial pressure signal to the total waveform energy, which is used to characterize the degree of intracranial pressure fluctuation caused by intracranial infection), thereby improving the accuracy of infection-related abnormal identification.

[0044] (2) Terahertz Metasurface Rapid Detection Unit: The core is the MS@SLET sensor chip (MS@SLET sensor chip refers to a surface plasmon-enhanced terahertz (SLET) sensor chip loaded with metal-semiconductor composite nanomaterials (MS)). It adopts a 45° inclined microfluidic channel design and forms a <100μm ultrathin liquid film from a 10μL cerebrospinal fluid sample through capillary compression effect. Based on polarization-induced multimode resonance technology (by adjusting the periodic structure and refractive index of the metasurface, orthogonally polarized terahertz waves are excited to generate resonance peaks of different frequencies, each resonance peak corresponding to a characteristic response of a cerebrospinal fluid infection index, realizing simultaneous detection of multiple indicators), three sets of resonance peaks with a resolution ≥20GHz are generated under orthogonally polarized terahertz wave excitation, which are coupled to white blood cell size, membrane structure, and inclusion characteristics, respectively. Through multi-parameter fitting of resonant frequency shift Δf and amplitude attenuation ΔA, simultaneous quantitative detection of CWBC and RPMN is achieved, with a detection time ≤30 seconds and a detection limit of up to 10 ... .

[0045] (3) Digital twin modeling unit: Based on preoperative CT / MRI images, an individualized cranial anatomical model is constructed using the U-Net three-dimensional reconstruction algorithm. Spatial registration of multi-source sensor data is achieved through affine transformation (x′y′z′=R·xyz+t, where R is the rotation matrix and t is the translation vector) and the Iterative Closest Point (ICP) algorithm, with a registration error ≤0.5mm.

[0046] (4) Feature fusion calculation: An improved multi-head attention mechanism is adopted, which optimizes the dynamic fusion of cross-modal features by introducing a temporal weight factor, thus solving the problem of insufficient sensitivity of traditional multi-head attention to physiological temporal signals. The calculation formula is as follows: , where T is the temporal weight factor matrix and d_k is the key vector dimension; dynamic weighted fusion of physiological temporal features, pathogen detection features, and anatomical features is performed to generate a 256-dimensional unified feature vector.

[0047] 2. Privacy Computing Central Module (Addressing the bottleneck of imbalance between data utilization and privacy fairness in existing technologies)

[0048] A triple protection mechanism of "knowledge graph alignment + improved FlexFair federated learning + differential privacy" is adopted to achieve compliant collaborative training of multi-center data, balancing model performance and group fairness.

[0049] (1) Heterogeneous data alignment: A standardized neurosurgical knowledge graph was constructed, containing mapping relationships of 1200+ medical terms and 500+ examination items. Cosine similarity (Sim(v1,v2)=v1·v2 / (∥v1∥·∥v2∥)) was used to achieve cross-center data semantic alignment. The alignment threshold can be set to 0.85 to solve the problem of heterogeneous multi-center data.

[0050] (2) Fair Federated Training: Lightweight CNN-BiLSTM models are deployed at each center. A fairness regularization term is introduced into the local training loss function. The prediction accuracy and group fairness are balanced by adjusting the hyperparameter λ. After the model is trained, only the encrypted gradients are uploaded. The central server uses a weighted aggregation algorithm. The global model is generated by (where n_i is the number of samples at center i and N_total is the total number of samples); the improved FlexFair federated learning framework further enhances the collaborative training effect of heterogeneous medical data by introducing a cross-center data semantic alignment mechanism.

[0051] (3) Differential privacy protection: An improved Laplace noise injection mechanism is adopted. Where Δf is the gradient sensitivity, A privacy budget (value between 0.5 and 2.0) is allocated; Rényi Differential Privacy (RDP) is used to calculate the cumulative privacy loss to ensure compliance. This system complies with relevant regulations on medical data privacy protection. The medical data collected by this system is stored for a period not exceeding the necessary period for clinical research (such as 3 months after surgery). It adopts a role-based access control (RBAC) mechanism to restrict data access permissions and only authorizes medical staff and researchers to use the data within the authorized scope. This fully complies with the requirements of Article 28 of the Personal Information Protection Law and Article 15 of the Medical Data Security Guidelines.

[0052] 3. Explainable early warning engine (addressing the bottleneck of unexplainable decision-making in existing technology models)

[0053] Based on "stage-aware modeling + quantitative attribution + thought chain reasoning", reliable early warning is achieved, solving the "black box" problem of traditional models and improving clinical trust.

[0054] (1) Multimodal fusion modeling: The CNN-Transformer hybrid architecture is adopted. The CNN branch extracts local features, and the Transformer branch captures temporal dependencies through a stage-aware attention mechanism. The stage-aware attention mechanism dynamically adjusts the attention weights according to the differences in infection risk characteristics in different postoperative stages (hyperacute, acute, and subacute phases), so that the model focuses more on the key risk indicators of each stage.

[0055] (2) Risk score calculation: Output dynamic risk value of 0-100 points, and use Platt calibration algorithm to optimize the score distribution to ensure that the Pearson correlation coefficient between the score and the actual infection probability is ≥0.92.

[0056] (3) Interpretable output: using the improved SHAP algorithm The contribution of features is quantified by S (where S is a subset of features and N is the total feature set); the integrated and fine-tuned medical big language model generates a structured report of "key findings - clinical reasoning - comprehensive conclusions", which conforms to the logic of clinical thinking.

[0057] 4. Intelligent closed-loop intervention module (addressing the bottleneck of the existing technology's fragmented early warning and intervention chain)

[0058] This achieves full automation of the early warning-intervention-feedback process, solving the problems of incomplete closed-loop operation and easy forgetting of model updates in traditional solutions.

[0059] (1) Tiered Early Warning: Risk scores are categorized into four levels (blue < 50 points, yellow 50-69 points, orange 70-84 points, red ≥ 85 points). Warnings are simultaneously pushed through multiple channels, including the central monitoring screen, doctors' mobile phones, and nurses' PDAs. A red warning triggers an audible and visual alarm. Specific tiering standards are shown in the table below:

[0060]

[0061] (2) Personalized solution matching: Based on the matching degree algorithm, the optimal intervention suggestion is recommended from the solution library (containing 50+ standardized solution packages). The solution library is synchronized with the latest clinical guidelines in real time (such as the "Expert Consensus on Prevention and Control of Postoperative Infection in Neurosurgery 2024").

[0062] (3) Feedback optimization: The improved elastic weight merging (EWC) algorithm is adopted. By dynamically adjusting the weight decay coefficient, the incremental learning efficiency of clinical feedback data is improved, avoiding the defect of poor adaptability of traditional EWC algorithm to new data. The model is incrementally updated weekly using clinical feedback data (medical order execution, patient outcome, etiological results) to ensure that the accuracy of the model on historical data decreases by ≤3% after the update, thus avoiding catastrophic amnesia.

[0063] Furthermore, the specific implementation of the postoperative infection risk early warning system for patients with traumatic brain injury according to the present invention is as follows:

[0064] The present invention provides a postoperative infection risk warning system for patients with traumatic brain injury, comprising:

[0065] The multimodal perception and fusion module communicates with the privacy computing hub module via HTTPS protocol to achieve data interaction. It is used to collect multi-dimensional raw data of patients and complete feature fusion to generate multimodal feature vectors. The edge computing node module is used to perform denoising, normalization and outlier removal preprocessing operations on the data transmitted by the multimodal perception and fusion module to finally generate a standardized multimodal dataset.

[0066] The privacy-preserving computation central module communicates with the interpretable early warning engine module via an encrypted data link. It employs an improved FlexFair federated learning framework to achieve multi-center model collaborative training, generating a global early warning model for postoperative infection in patients with traumatic brain injury. The improved FlexFair federated learning framework optimizes prediction fairness among different groups by introducing a fairness regularization term into the local training loss function, and simultaneously uses differential privacy technology to protect the privacy of the model gradient upload process. The core improvement of the improved FlexFair federated learning framework is the introduction of a cross-center data semantic alignment mechanism to enhance the collaborative training effect of heterogeneous medical data.

[0067] The interpretable early warning engine module works in collaboration with the intelligent closed-loop intervention module through a local area network. It is used to calculate the postoperative dynamic infection risk score of patients with traumatic brain injury based on multimodal feature vectors and through a CNN-Transformer hybrid model that integrates stage perception attention mechanism, and generate a structured attribution report corresponding to the infection risk score.

[0068] The intelligent closed-loop intervention module, in conjunction with the interpretable early warning engine module, triggers tiered early warnings based on risk scores and matches recommended intervention plans that include diagnostic and treatment suggestions. It also optimizes the early warning model based on clinical feedback data.

[0069] The multimodal perception and fusion module includes:

[0070] Terahertz metasurface biosensor chip for detecting pathogen detection characteristics and cerebrospinal fluid infection marker signals;

[0071] The millimeter-wave radar module is used for non-contact acquisition of physiological temporal characteristics such as heart rate, respiratory rate, and cerebral blood flow velocity, and simultaneously analyzes the physiological parameter changes corresponding to minute displacements of intracranial brain tissue. As a non-contact acquisition component for intracranial physiological parameters, it uses a 60GHz millimeter-wave radar sensor, deployed at a preset position above the patient's bedside to ensure that the radar signal can penetrate the patient's skull vertically. The module integrates a signal amplification unit and a filtering unit for non-contact real-time acquisition of physiological temporal characteristic signals such as heart rate, respiratory rate, intracranial pressure, and cerebral blood flow velocity, and simultaneously analyzes the physiological parameter changes corresponding to minute displacements of intracranial brain tissue. This module works synchronously with the flexible sensor array to achieve dual-dimensional verification of physiological temporal characteristics and avoid the acquisition errors of a single module.

[0072] A flexible sensor array is used for contact-based acquisition of physiological temporal characteristics such as intracranial pressure, brain tissue oxygen saturation, local temperature and humidity of the surgical incision, and biomechanical signals. Abnormal waveforms and numerical features are extracted to characterize physiological fluctuations caused by intracranial infection. As an auxiliary component for contact-based physiological temporal characteristic acquisition, it uses a flexible polyimide substrate and can be closely attached to the area around the patient's head incision and skull defect. The array integrates a temperature and humidity sensor, a flexible intracranial pressure sensor, and a brain tissue oxygen saturation monitoring unit, enabling simultaneous acquisition of physiological temporal characteristic signals such as intracranial pressure, brain tissue oxygen saturation, local temperature and humidity of the surgical incision, and biomechanical signals at multiple points. Abnormal waveforms and numerical features are extracted to characterize physiological fluctuations caused by intracranial infection. This array works in conjunction with a millimeter-wave radar module to achieve complementary acquisition of physiological temporal characteristics and forms a triggering linkage mechanism with a terahertz metasurface biosensor chip.

[0073] The multimodal feature fusion unit employs an improved multi-head attention mechanism to achieve feature fusion of acquired and imported data. This unit receives physiological temporal features acquired by a millimeter-wave radar module and a flexible sensor array, pathogen detection features acquired by a terahertz metasurface biosensor chip, and anatomical static features such as cranial anatomy, defect morphology, and incision location imported from preoperative CT / MRI cranial images. It dynamically weights and fuses these three types of multimodal features, eliminating redundant feature information and generating a 256-dimensional unified feature vector to provide standardized feature input for subsequent model training. The improved multi-head attention mechanism optimizes cross-modal feature dynamic fusion by introducing a temporal weight factor, the calculation formula of which is as follows: Where Q is the query vector, K is the key vector, and V is the value vector. d_k is the standard notation for matrix transpose, T is the temporal weight factor matrix, and d_k is the dimension of the key vector.

[0074] The digital twin modeling unit communicates bidirectionally with the multimodal feature fusion unit, receiving the fused feature vectors to construct a digital twin of the patient after craniocerebral surgery. This allows for real-time mapping of the patient's intracranial physiological state and incision recovery, and simultaneous prediction of the evolution trend of infection risk.

[0075] The multimodal perception and fusion module of this invention adopts a design logic of "separate acquisition and unified fusion of three types of features". Each component has a clear division of labor and clear feature attribution, with no overlap or confusion: the terahertz metasurface biosensor chip specializes in pathogen detection feature acquisition, achieving label-free rapid detection of core indicators of cerebrospinal fluid infection, directly reflecting the pathogen status of infection; the millimeter-wave radar module and flexible sensor array work together to acquire physiological time-series features, covering dynamic physiological parameters such as heart rate, respiration, intracranial pressure, and brain tissue oxygen saturation, reflecting the fluctuation of the patient's postoperative physiological state in real time; after importing preoperative cranial CT / MRI image data, anatomical features such as cranial structure, defect location, and incision morphology are extracted. The three types of features are simultaneously incorporated into the multimodal feature fusion unit, and after dynamic weighted fusion, a unified feature vector is formed, solving the problems of low accuracy and high false alarm rate of traditional single feature acquisition and early warning.

[0076] Among them, anatomical features are static structural features, which only need to be collected and imported once before surgery and do not require real-time monitoring; physiological time sequence features and pathogen detection features are dynamic features, which are collected continuously in real time by the corresponding sensor modules. The three types of features each perform their own functions and together constitute a comprehensive infection risk assessment feature system to ensure the accuracy and reliability of subsequent early warning models.

[0077] The terahertz metasurface biosensor chip is configured to: receive a 10 μL cerebrospinal fluid sample; the chip incorporates a microfluidic structure to compress the cerebrospinal fluid sample into an ultrathin liquid film through capillary compression; the chip surface integrates a polarization-induced multimode metasurface, which generates multiple independent resonance peaks under the excitation of orthogonally polarized terahertz waves; by detecting the frequency shift Δf and amplitude attenuation ΔA of the resonance peaks, a multi-parameter fitting algorithm is used to quickly output pathogen detection characteristic signals corresponding to the white blood cell concentration (CWBC) and multinucleated cell ratio (RPMN), with a detection limit reaching [value missing]. The chip is directly connected to the cerebrospinal fluid drainage tube and adopts a combination of on-demand and trigger-based acquisition modes. It can be triggered by abnormal signals from the flexible sensor array to start detection.

[0078] The privacy computing hub module employs an improved FlexFair federated learning framework with fairness regularization constraints and differential privacy protection. It performs multi-center collaborative training optimization on the standardized multimodal dataset preprocessed by the edge computing node module for the multimodal perception and fusion module. This constructs and generates a global early warning model for postoperative infection in patients with traumatic brain injury, integrating physiological temporal features, pathogen detection features, and anatomical static features. This global early warning model is based on a CNN-Transformer hybrid architecture, embedding a stage-aware attention mechanism. Local training uses a composite loss function that balances prediction accuracy and group fairness. During model training, L2 norm pruning and Laplacian noise injection are applied to the gradient parameters, achieving triple optimization of multi-center data collaborative training, privacy protection, and fair early warning.

[0079] 1. Model Infrastructure. A CNN-Transformer hybrid architecture is adopted, specifically designed to adapt to multimodal heterogeneous features: the CNN branch is used to extract terahertz sensor pathogen detection features and local static features of anatomical static structures, capturing key local features such as cerebrospinal fluid leukocyte concentration, multinucleated cell ratio, and cranial anatomical defects; the Transformer branch embeds a stage-aware attention mechanism, dividing temporal weights according to the postoperative hyperacute, acute, and subacute phases, focusing on capturing the dynamic changes of physiological temporal features such as heart rate, respiratory rate, and intracranial pressure, taking into account both static and dynamic temporal features, which is different from conventional single CNN or Transformer models.

[0080] 2. Improve the core rules of FlexFair federated training. Model training follows the principle of "data not leaving the local machine and multi-center collaborative aggregation." Each collaborating hospital acts as a local client, transmitting only model gradient parameters and not leaking original patient data. The local composite loss function is adopted. Where Lclsk is the cross-entropy classification loss, ensuring the accuracy of the infected / non-infected binary classification prediction; For fairness regularization, patients are divided into groups based on age to reduce the prediction error rate bias between different age groups. λ is set to 0.1~0.5 to balance accuracy and fairness. Differential privacy protection: After local training, L2 norm pruning (pruning threshold 1.0) is performed on the model gradient, and Laplacian noise is injected. The privacy budget ε is controlled between 0.5 and 2.0 to prevent gradient back-inference of the original data, complying with medical data privacy compliance requirements. Global aggregation rule: The cloud-based federated learning server aggregates the anonymized gradient parameters uploaded by each client according to the amount of local data, updates the global model parameters, and converges after 50 iterations or 5 consecutive iterations when the global accuracy is ≥95% and the group error difference is ≤5%, generating the final global early warning model.

[0081] 3. Model Input and Output. The input is a standardized multimodal dataset output by edge computing nodes, covering pathogen detection features from terahertz chip detection, physiological temporal features acquired by millimeter-wave radar and flexible sensors, and anatomical static features extracted from preoperative images; the output is a dynamic infection risk score of 0-100 points, which is matched with four levels of early warning to achieve early quantitative early warning of postoperative infection.

[0082] The privacy computing hub module includes:

[0083] The heterogeneous data alignment submodule utilizes a medical knowledge graph to perform semantic alignment and standardization mapping on non-standardized and heterogeneous data from different medical center systems. It introduces a cross-center data semantic alignment mechanism to construct a standardized neurosurgical knowledge graph containing mapping relationships for over 1200 medical terms and over 500 examination items. Cosine similarity (Sim(v1,v2)=v1·v2 / (∥v1∥·∥v2∥)) is used to achieve cross-center data semantic alignment, with an alignment threshold set to 0.85. Postoperative infection-related data from various medical centers' neurosurgery departments are collected, core fields and medical terms are extracted, and heterogeneous fields from different centers are mapped to a unified concept within the standardized knowledge graph, completing data semantic normalization and ensuring that multi-center data resides in the same semantic space. v1 refers to the medical centers to be aligned. The semantic feature vectors of heterogeneous fields / medical terms in the local neurosurgical knowledge graph refer to the high-dimensional numerical vectors obtained after transformation by word vector models for the field names and medical terms (such as non-standard expressions like "cerebrospinal fluid WBC", "cerebrospinal fluid white blood cell count", and "cerebrospinal fluid white blood cell count") defined by each hospital's local system. These vectors carry the core semantic information of the local terms. v2 refers to the semantic feature vectors of standard terms / unified fields in the neurosurgical standardized knowledge graph. These are the high-dimensional numerical vectors obtained after transformation by the same word vector model for the standard expressions corresponding to unified concepts (such as the unified standard term "cerebrospinal fluid white blood cell concentration (CWBC)") in the pre-constructed standardized medical terminology library. These vectors carry the baseline semantic information of the standard terms. v1·v2 represents the dot product of the two vectors, and ∥v1∥·∥v2∥ represent the L2 norm (modulus) of the two semantic feature vectors, respectively. The cosine similarity between the two sets of vectors is calculated using this formula. The closer the value is to 1, the higher the semantic similarity between the two sets of terms. This is used to determine whether the local heterogeneous terms match the standard terms.

[0084] The federated learning training submodule employs an improved federated learning algorithm that incorporates attention mechanisms and graph convolutional networks. It builds training logic based on the improved FlexFair federated learning framework, adapting to the privacy compliance requirements of multi-center medical data. This allows collaborating medical centers to securely participate in global model collaborative training without leaving their local original data domain or being leaked. During training, an adaptive learning rate optimization strategy is used, with a preset initial learning rate and an iterative decay mechanism. Combined with a local fairness constraint loss function, this achieves coordinated optimization of global model convergence speed, prediction accuracy, and group fairness, ensuring the model's generalization ability and robustness after training on heterogeneous multi-center data.

[0085] The federated learning training submodule, as the core execution unit of the privacy computing hub, implements multi-center collaborative training based on the improved FlexFair federated learning framework. The specific implementation parameters and process are as follows: An improved federated learning algorithm incorporating a temporal attention mechanism and graph convolutional networks is adopted, strictly adhering to the medical data compliance requirements of "data not leaving the local machine and gradients being uploaded with encryption." Each medical center independently conducts model training as a local client. The training optimizer uses an adaptive stochastic gradient descent optimizer with a preset initial learning rate of 0.001. A step-decay strategy is adopted, with the learning rate automatically decreasing by 10% every 10 rounds of global iteration. This adaptive learning rate mechanism avoids the problems of gradient oscillation in the early stage of training and slow convergence in the later stage, accelerating the convergence of the global model. At the same time, combined with the fairness regularization term mentioned above, a composite loss function is constructed to balance the model's prediction accuracy and the prediction fairness of different patient groups. After the global model iterates to the convergence condition, the final global early warning model for postoperative infection in patients with traumatic brain injury is generated.

[0086] The system calls upon the standardized multimodal dataset processed by the heterogeneous data alignment submodule. Each center acts as a client, performing local model training. After training, only the model parameters are uploaded. The cloud server receives the model parameters from each client and aggregates them using a weighted average algorithm to generate global model parameters, which are then broadcast to each client. Each client updates its local model based on the global parameters. This iteration is repeated until the model converges, outputting a global early warning model for postoperative infection in patients with traumatic brain injury. The local training loss function is: L(k) = L_cls(k) + λ·L_fair(k); where L_cls(k) is the cross-entropy loss term. Here, λ is the fairness regularization term, G is the set of patient groups divided by sensitivity attributes, R_g is the prediction error rate of group g, R̄ is the global average error rate, and λ is a hyperparameter that balances accuracy and fairness. The value of λ ranges from 0.1 to 0.5 and is used to adjust the fairness optimization weight.

[0087] The differential privacy protection submodule is used to add random noise conforming to the (ε, δ)-differential privacy definition during the upload of model parameters or gradients to the client side of federated learning to resist privacy inference attacks; an improved Laplacian noise injection mechanism is employed. Where Δf is the gradient sensitivity, A privacy budget (value between 0.5 and 2.0) is allocated; L2 norm clipping is performed on the gradients of each client model after local training to control gradient sensitivity Δf; calibration noise is added to the clipped gradients before uploading to the cloud server; Rényi differential privacy RDP is used to calculate the cumulative privacy loss to ensure that privacy requirements are met. It integrates a privacy budget management unit to track privacy consumption in real time.

[0088] The explainable early warning engine module includes:

[0089] The risk score calculation submodule receives a global early warning model for postoperative infection in traumatic brain injury patients from the privacy computing center module, based on the standardized multimodal dataset output by the edge computing layer. It extracts local features from the multimodal data through a CNN branch, captures the temporal dependencies between different postoperative stages using a stage-aware attention mechanism in the Transformer branch, fuses the local and temporal features, and inputs them into a fully connected layer. The Platt calibration algorithm is used to optimize the score distribution, outputting a dynamic postoperative infection risk score for traumatic brain injury patients ranging from 0 to 100. The stage-aware attention mechanism can dynamically adjust feature weights according to different postoperative stages.

[0090] The structured attribution report generation submodule generates a structured attribution report that includes feature quantification contribution and clinical reasoning logic. For the dynamic infection risk score and feature weights output by the risk score calculation submodule, an improved SHAP algorithm is used to quantify the positive / negative contribution of each multimodal feature vector to the risk score, selecting the top N key risk indicators with the highest contribution (N being a positive integer). An integrated medical big data language model based on Llama 2-7B fine-tuning is used, inputting key risk indicators, contribution values, risk scores, and basic patient information to generate a clinical attribution report with a "key findings - clinical reasoning - comprehensive conclusion" structure.

[0091] The structured attribution report generation submodule receives the output results from the risk scoring calculation submodule. Its input indicators are all derived from the 256-dimensional multimodal feature vector output by the multimodal perception and fusion module mentioned above, covering three categories of multidimensional indicators: pathogen detection features acquired by terahertz chips, physiological temporal features acquired by millimeter-wave radar and flexible sensors, and anatomical features extracted from preoperative images. The improved SHAP algorithm is an adaptation improvement for high-dimensional medical features, with the core improvement being the removal of irrelevant interference features and the strengthening of the contribution weight of clinical indicators. The preset parameter N is preferably 5 and can be adaptively adjusted according to clinical needs. The medical big language model used is based on the open-source Llama 2-7B model with fine-tuning. The fine-tuning training dataset consists of clinical cases of postoperative infection in neurosurgery, which is a conventional model fine-tuning method that can be implemented by those skilled in the art without creative effort. The final generated three-segment structured report is fully aligned with clinical diagnosis and treatment thinking, making it easy for medical staff to quickly interpret the early warning basis.

[0092] The specific implementation method of the structured attribution report generation submodule is as follows: For the dynamic infection risk score and the feature weights of each dimension in the multimodal feature vector output by the risk score calculation submodule, an improved SHAP algorithm is used to quantify the positive and negative contributions of each multimodal collection indicator to the risk score, and the top N key risk indicators with preset parameter N are selected; the key risk indicators, corresponding contribution values, dynamic infection risk score, and patient basic information are input into a fine-tuned medical big language model to generate a clinical attribution report with a fixed structure of "key findings - clinical reasoning - comprehensive conclusion," achieving simultaneous optimization of the interpretability and clinical readability of the risk score; the postoperative time is divided into three stages: hyperacute phase (0–24h), acute phase (24–72h), and subacute phase (>72h), and attention weight is calculated using the formula α=softmax(v·tanh(W_h·h). +W_p·p)) dynamically adjusts the feature weights of each stage; where h is the feature representation at time t, p is the stage embedding vector, W_h and W_p are weight matrices, and v is a 32-dimensional learnable parameter vector with a value range of [-0.1, 0.1], which is obtained by model training and optimization after random initialization.

[0093] The intelligent closed-loop intervention module includes:

[0094] (1) Graded early warning push submodule: For the dynamic infection risk score of craniocerebral trauma patients transmitted by the risk score calculation submodule, four levels of early warning are preset: blue <50 points, yellow 50-69 points, orange 70-84 points, and red ≥85 points. After receiving the risk score, the corresponding early warning level is automatically matched, and the early warning information (risk score, early warning level, key risk indicators) is pushed to the doctor's mobile terminal, nurse's workstation, and central monitoring screen through the local area network. The red early warning triggers the sound and light alarm.

[0095] (2) Personalized solution matching submodule: It has a built-in pre-set solution library containing 50+ standardized solution packages, and synchronizes with the "Expert Consensus on Prevention and Control of Postoperative Infection in Neurosurgery 2024" in real time. Based on the patient's dynamic infection risk score, key risk indicators and individual information including age, underlying diseases and surgical type, it constructs a patient feature vector; extracts the feature vector of each standardized solution package in the solution library, and uses the cosine similarity matching algorithm (similarity threshold ≥ 0.85) to select the solution with the highest similarity from the pre-set solution library as the optimal intervention suggestion list and pushes it to the clinical terminal;

[0096] (3) Feedback data collection and incremental optimization submodule: It interfaces with clinical terminals, including electronic medical record systems and nurse workstations, and automatically extracts clinical intervention feedback data (medical order execution details, patient outcomes, etiological results, and intervention effect evaluation) entered by medical staff for intervention plans pushed by the intelligent closed-loop intervention module; after standardizing the collected data (unifying data format, removing invalid data, and supplementing missing fields), it generates a compliant and usable feedback optimization dataset; it transmits the standardized feedback optimization dataset to the privacy computing center core module, and realizes incremental optimization of the global early warning model for postoperative infection of patients with traumatic brain injury based on the improved elastic weight merging EWC algorithm, to prevent catastrophic amnesia of the model and complete the closed-loop iteration of the system.

[0097] The matching algorithm of the intelligent closed-loop intervention module is: MatchScore(S)=α·I(L_curr==L_S)+β·∑w_f·I(f∈S); where α=0.6 and β=0.4 are weight coefficients, L_curr is the current warning level, L_S is the warning level corresponding to scheme S, F is the Top-N risk feature set, w_f is the normalized SHAP weight of feature f, and I is the indicator function (I=1 when the condition in parentheses is true, otherwise I=0).

[0098] An edge-cloud collaborative architecture is adopted: edge nodes are deployed in the neurosurgical intensive care unit, responsible for real-time data acquisition and preprocessing; cloud nodes serve as the hardware operating platform, internally deploying a privacy computing central module and simultaneously running a global early warning model for postoperative infection in traumatic brain injury patients. Data transmission uses HTTPS protocol + national cryptographic SM4 encryption algorithm to achieve end-to-end secure transmission, and identity authentication uses OAuth 2.0 protocol; the data interaction latency between the edge nodes and cloud nodes is ≤50ms. This global model is the core early warning algorithm carrier of the system, receiving standardized multimodal datasets uploaded by edge nodes, performing postoperative infection risk score calculation, multimodal feature attribution analysis, and outputting interpretable early warning results and hierarchical early warning instructions, providing core algorithmic support for subsequent intelligent closed-loop intervention, and does not participate in the front-end raw data acquisition and preprocessing work.

[0099] The system is also equipped with a closed-loop model incremental optimization process. The specific execution steps include: collecting clinical intervention feedback data through the intelligent closed-loop intervention module, performing standardized preprocessing on the feedback data, and transmitting it to the privacy computing center module; the privacy computing center module, based on the improved elastic weight merging algorithm, uses the standardized intervention feedback data to perform incremental iterative optimization of the global early warning model for postoperative infection in patients with traumatic brain injury, preventing catastrophic amnesia during model training and realizing continuous iterative upgrades of the global early warning model.

[0100] The present invention provides a method for processing postoperative infection risk warning data in patients with traumatic brain injury, comprising the following steps:

[0101] S1: Collect multi-dimensional raw data of patients through the multimodal perception and fusion module, and perform feature fusion using an improved multi-head attention mechanism to generate multimodal feature vectors;

[0102] S2: Based on the improved FlexFair federated learning framework, combined with differential privacy technology, multi-center model collaborative training is achieved to generate a global early warning model;

[0103] S3: Input the multimodal feature vector into the global early warning model, calculate the dynamic infection risk score, and generate a structured attribution report;

[0104] S4: Trigger tiered early warnings based on risk scores and output a list of intervention recommendations using a matching algorithm;

[0105] S5: Collect clinical intervention feedback data and incrementally optimize the global early warning model using an improved elastic weight merging algorithm.

[0106] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0107] In the diagram, 1-Edge computing node, 2-Cloud node, 3-Millimeter-wave radar module, 4-Flexible sensor array, 5-Terahertz metasurface biosensor chip, 6-Multimodal feature fusion module (used to receive data from various sensors and complete feature fusion), 7-Privacy computing hub module, 8-Explainable early warning engine, 9-Intelligent closed-loop intervention module, 10-Federated learning server, 11-Hospital client, 12-CNN-Transformer hybrid model, 13-Graded early warning push module, 14-Personalized solution library, 15-Doctor mobile terminal, 16-Nurse workstation, 17-Central monitoring screen, 18-45° inclined plane extrusion microchannel, 19-Polarization-induced multimodal metasurface, 20-Terahertz wave transmitter, 21-Signal receiver, 22-Risk score trend chart, 23-Feature contribution bar chart, 24-Think chain report window.

[0108] Figure 1 The architecture adopts a layered design, clearly presenting the perception layer, intelligence layer, application layer, and closed-loop link from top to bottom. The modules and data flow of each layer are as follows:

[0109] (1) Perception layer: includes millimeter-wave radar module 3, flexible sensor array 4, terahertz metasurface biosensor chip 5, edge computing node 1 and CT / MRI image interface. Physiological time-series signals, cerebrospinal fluid signals and image data collected by each sensor are integrated into multimodal feature fusion module 6 (used to receive data from each sensor and complete feature fusion). Its core function is "multi-dimensional data acquisition and fusion". The edge node deploys a data preprocessing module, which is responsible for the real-time preprocessing of the raw data.

[0110] (2) Intelligent layer: Cloud node 2 includes privacy computing hub module 7 (built-in federated learning server 10 and multiple hospital clients 11, labeled "improved FlexFair federated learning + differential privacy") and interpretable warning engine 8 (built-in CNN-Transformer hybrid model 12, labeled "stage perception attention + interpretable thinking chain"), to realize model training and risk warning calculation;

[0111] (3) Application layer: includes intelligent closed-loop intervention module 9, connected to graded early warning push module 13, graded early warning push module 13 outputs to doctor mobile terminal 15, nurse workstation 16, central monitoring screen 17, and is connected to personalized solution library 14. The core function is "four-level early warning + personalized solution matching".

[0112] (4) Closed-loop link: The arrow marked "feedback data" is drawn from "clinical intervention execution" and points to privacy computing central module 7 and interpretable early warning engine 8, marked "incremental learning optimization", forming a complete closed loop of "monitoring-early warning-intervention-feedback-optimization".

[0113] Figure 2 From left to right, the overall chip structure, a magnified view of the core structure, and the signal detection chain are shown below:

[0114] (1) Overall structure of the chip on the left: The syringe squeezes the microchannel to inject cerebrospinal fluid sample (labeled "10μL cerebrospinal fluid") into the 45° inclined plane, and the end of the microchannel is connected to the polarization-induced multimode metasurface (labeled "polarization-induced multimode metasurface").

[0115] (2) Enlarged view of the intermediate core structure: clearly shows the process of cerebrospinal fluid forming an ultrathin liquid film of <100μm covering the metastructure surface under the pressure of the inclined plane, labeled "capillary pressure effect";

[0116] (3) Right signal detection link: The terahertz wave transmitter emits orthogonally polarized terahertz waves, which are received by the signal receiver after passing through the ultrathin liquid film sample, and output two waveforms: Figure A is the resonance frequency shift curve (labeled "corresponding to white blood cell concentration"), and Figure B is the amplitude attenuation curve (labeled "corresponding to multinucleated cell ratio").

[0117] (4) The core technical parameters are marked at the bottom: "Detection time ≤ 30 seconds" and "markerless rapid detection", which clearly shows the performance advantages of the chip.

[0118] Figure 3 The system software interface is displayed by function, as follows:

[0119] (1) Top: Patient basic information column, which includes core information such as name, gender, age, and postoperative time, to facilitate doctors to quickly locate patients;

[0120] (2) Left side: Risk score trend chart, with the horizontal axis representing postoperative time and the vertical axis representing risk score. The current time point is clearly marked "92 points (red warning)", which intuitively shows the trend of risk change;

[0121] (3) Middle section: Feature contribution bar chart, horizontally displaying the top 5 risk features and their corresponding contribution values, namely "cerebrospinal fluid white blood cell concentration (+52)", "intracranial pressure fluctuation (+28)", "body temperature (+8)", "respiratory rate variability (+4)" and "incision moisture (+2)", with "SHAP value quantification contribution" to clarify the degree of influence of each feature on the risk score;

[0122] (4) Right side: Mind Chain Report window, displaying a structured natural language report, the content of which is "High fever (38.5℃) occurred 42 hours after surgery, and the cerebrospinal fluid white blood cell concentration was significantly increased." The presence of predominantly multinucleated cells (88%), combined with the abnormal intracranial pressure fluctuation pattern, strongly suggests an acute bacterial central nervous system infection, most likely caused by retrograde infection from the surgical incision. Immediate emergency intervention is recommended. [This is marked "Clinical Thinking Chain Reasoning," and the report content aligns with clinical diagnostic and treatment logic.]

[0123] (5) Bottom: Function button area, with three core function buttons: "View 3D heat map", "Export consultation report" and "Confirm warning", to meet the actual clinical application needs.

[0124] Figure 4 The entire closed-loop logic is illustrated in flowchart form. The specific process and nodes are explained below:

[0125] (1) Process start: "Receive red alert and attribution report", clarify the process triggering conditions;

[0126] (2) First judgment node: "Doctor confirms infection risk?" ① Yes: Enter "Initiate personalized intervention package", which includes three parallel sub-steps: "Emergency cerebrospinal fluid re-examination + drug sensitivity test", "Empirical anti-infection treatment", and "Intensive monitoring and nursing"; ② No: Enter "Mark false positive / other diagnoses" to ensure that the process covers different clinical judgment scenarios;

[0127] (3) Feedback data collection: After the two branches merge, enter "Record clinical treatment process and patient outcome", marked "Feedback data collection", and clarify the core content of data collection;

[0128] (4) Second judgment node: "Has the model update cycle (weekly) been reached?" ① Yes: Enter "Improve EWC algorithm incremental learning" (marked "Prevent catastrophic forgetting"), which will then lead to "Update global federated model"; ② No: Return to "Continuous monitoring" to ensure the regularity and continuity of model updates;

[0129] (5) Closed loop formation: The dashed arrow from "update global federated model" points to the beginning of the process and marks "improve future early warning accuracy", forming a complete closed loop of "early warning-intervention-feedback-optimization-re-early warning" to achieve continuous improvement of system performance.

[0130] (a) System Deployment Environment

[0131] Hardware configuration: The edge computing box uses an Intel Core i7-12700H processor, integrates 16GB DDR5 memory and 1TB SSD storage, and connects to a multimodal perception and fusion module; the cloud uses two NVIDIA A100 GPU servers (80GB video memory), configured with encryption acceleration cards, and builds a distributed storage cluster.

[0132] Software environment: The edge device runs Ubuntu 22.04 LTS and deploys a lightweight Python 3.9 data acquisition program; the cloud device is based on CentOS 8, using the federated learning framework FATE v2.0 with integrated and improved FlexFair modules, an interpretable engine based on a finely tuned Llama 2-7B model, and a Milvus 2.4 vector database; data transmission uses HTTPS protocol + SM4 encryption, and authentication uses OAuth 2.0 protocol to ensure data transmission and storage security; the data storage period for medical data collected by this system is set to 3 months post-surgery, and RBAC mechanism is used to restrict data access permissions, complying with medical data compliance requirements.

[0133] (II) Example of a complete workflow

[0134] Taking a 45-year-old male patient who underwent decompressive craniectomy after severe traumatic brain injury as an example, the detailed procedure is as follows (this study protocol has been approved by the Ethics Committee of Nanjing Drum Tower Hospital, approval number: 20250117):

[0135] 1. System initialization (0h postoperative): Deploy the multimodal sensing and fusion module in the neurosurgical intensive care unit. The flexible sensor array 4 is attached around the head incision, and the terahertz detection port 5 is connected to the cerebrospinal fluid drainage tube. The system loads the digital twin model reconstructed from the patient's preoperative CT images and connects to the hospital's HIS / LIS system to obtain basic medical record data (the data has been desensitized and meets privacy protection requirements).

[0136] 2. Early monitoring (18 hours post-surgery): The system monitors abnormal intracranial pressure fluctuations in real time through millimeter-wave radar module 3 and flexible sensor array 4, with R_wave=0.35 (threshold 0.3), for 12 minutes; feature fusion is completed through the multimodal feature fusion module using an improved multi-head attention mechanism (corresponding to the feature fusion function of the "multimodal perception and fusion module"). After fusion, the data is input into the early warning model, with a risk score of 58, triggering a yellow warning. Nurse workstation 16 prompts "Closely monitor intracranial pressure changes and record body temperature every 30 minutes".

[0137] 3. Key indicator detection (42 hours post-surgery): The patient's body temperature rose to 38.5℃. Following instructions, the nurse collected 10μL of cerebrospinal fluid through a drainage tube and injected it into the terahertz metasurface biosensor chip 5; the system output the results 30 seconds later. RPMN=88% (corresponding to the detection function of the terahertz chip).

[0138] 4. Explainable Early Warning Generation (42h+1min post-surgery): The Explainable Early Warning Engine 8 calls the CNN-Transformer hybrid model 12, updating the risk score to 92 points (red warning) through a stage-based perceptual attention mechanism; the improved SHAP algorithm is used to quantify feature contribution (corresponding to the "feature quantification function of the Explainable Early Warning Engine"), and the results show: elevated cerebrospinal fluid white blood cell count contributes +52 points, and intracranial pressure fluctuation contributes +28 points; the thought chain report is generated: "High fever (38.5℃) occurred 42 hours post-surgery, and the concentration of white blood cells in the cerebrospinal fluid was significantly elevated." The presence of multinucleated cells (88%), combined with abnormal intracranial pressure fluctuations, strongly suggests an acute bacterial central nervous system infection, most likely caused by retrograde infection from the surgical incision. Immediate emergency intervention is recommended. The warning information is simultaneously pushed to the attending physician's mobile phone (15) and the central monitoring screen (17) via the tiered warning push module (9).

[0139] 5. Intelligent Intervention Matching (42h + 3min postoperatively): The intelligent closed-loop intervention module, based on the combination of "red alert + sharp rise in cerebrospinal fluid indicators", matches the "extremely high risk - bacterial intracranial infection" protocol package from the personalized protocol library 14 through a matching degree algorithm. Recommendations: ① Immediately retest cerebrospinal fluid and perform bacterial culture + drug sensitivity test; ② Empirically administer vancomycin (15mg / kg loading dose) + meropenem (2g q8h); ③ Increase the frequency of intracranial pressure monitoring to once every 15 minutes and strictly maintain aseptic care for drainage tubes.

[0140] 6. Clinical Implementation and Feedback (42h-96h post-surgery): The physician adopts and implements the treatment plan. At 72h post-surgery, the cerebrospinal fluid culture result is Staphylococcus aureus (sensitive to vancomycin), and the treatment is adjusted to targeted anti-infection therapy. At 96h post-surgery, the patient's body temperature returns to normal, intracranial pressure stabilizes, and the infection is under control. The system automatically records all process data (early warning time, plan adoption status, treatment effect, etiological results) as feedback data for model optimization.

[0141] 7. Model Incremental Optimization (7 days post-surgery): The system triggers weekly incremental updates, using an improved elastic weight merging algorithm (corresponding to the "model optimization function of the intelligent closed-loop intervention module") to integrate the data from this case, optimizing the model's recognition weight for the "high fever + surge in cerebrospinal fluid white blood cell count" pattern. After the update, the model's recognition sensitivity for this type of pattern is improved by 4%; during the update process, it is ensured that the accuracy of historical data decreases by ≤3%.

[0142] (III) Data for verifying the effect

[0143] This system underwent multi-center retrospective validation (data from 3200 post-traumatic brain injury patients in 5 tertiary hospitals, all of which were anonymized and approved by the ethics committees of each center), and the results are as follows:

[0144] 1. Efficiency indicators: Terahertz detection time is stable at 22-30 seconds, and the system's average early warning response time is 186ms, meeting the real-time requirements at the bedside;

[0145] 2. Accuracy metrics: AUC=0.983 on the independent test set, early warning sensitivity of 96.1% and specificity of 90.5% in the early (within 72h) period, which are significantly better than the traditional logistic regression model (AUC=0.868) and the existing early warning system (AUC=0.897).

[0146] 3. Fairness indicators: The difference in prediction error rate between patients aged <50 and ≥50 years decreased from 18.2% to 7.6%, a reduction of 58% (corresponding to the optimization effect of the fairness regularization term);

[0147] 4. Clinical value: Simulated intervention showed that the average time to initiate anti-infective treatment was shortened by 5.5 hours, and the infection-related mortality rate decreased from 28.3% to 9.8%, a reduction of 65.4%.

[0148] 5. Compliance: Passed the Level 3 certification of the National Information Security Protection System, and the privacy computing module passed the third-party differential privacy audit (when ϵ=1.0, the privacy loss meets the security requirements).

[0149] In summary, the solution of the present invention solves the bottleneck of the prior art:

[0150] 1. Breakthrough in bedside rapid detection technology of infection indicators, achieving second-level, high-sensitivity, label-free detection of cerebrospinal fluid leukocyte concentration and polymorphonuclear cell ratio;

[0151] 2. Solve the triangular problem of "privacy protection - model performance - group fairness" in multi-center data utilization, and realize collaborative value mining of multi-center data under the premise of compliance;

[0152] 3. Construct an interpretable early warning mechanism that aligns with clinical thinking habits, generate transparent and verifiable decision-making evidence, and enhance clinical trust.

[0153] 4. Construct a fully intelligent closed loop to achieve precise matching between early warning and clinical intervention, as well as continuous and stable optimization of the model.

Claims

1. A postoperative infection risk early warning system for a craniocerebral trauma patient, characterized in that include: The multimodal perception and fusion module communicates with the privacy computing hub module via the HTTPS protocol to achieve data interaction. It is used to collect multi-dimensional raw data of patients and complete feature fusion to generate multimodal feature vectors. in The edge computing node module is used to perform preprocessing operations such as denoising, normalization, and outlier removal on the data transmitted by the multimodal perception and fusion module, and finally generate a standardized multimodal dataset. The privacy computing hub module communicates with the interpretable early warning engine module via an encrypted data link. It uses the improved FlexFair federated learning framework to achieve multi-center model collaborative training and generate a global early warning model for postoperative infection in patients with traumatic brain injury. The improved FlexFair federated learning framework optimizes the prediction fairness of different groups by introducing a fairness regularization term into the local training loss function, and at the same time uses differential privacy technology to protect the privacy of the model gradient uploading process. The interpretable early warning engine module works in collaboration with the intelligent closed-loop intervention module through a local area network. It is used to calculate the postoperative dynamic infection risk score of patients with traumatic brain injury based on multimodal feature vectors and through a CNN-Transformer hybrid model that integrates stage perception attention mechanism, and generate a structured attribution report corresponding to the infection risk score. The intelligent closed-loop intervention module, in conjunction with the interpretable early warning engine module, triggers tiered early warnings based on risk scores and matches recommended intervention plans that include diagnostic and treatment suggestions. It also optimizes the early warning model based on clinical feedback data.

2. The system of claim 1, wherein, The multimodal sensing and fusion module includes: Terahertz metasurface biosensor chip for detecting pathogen detection characteristics and cerebrospinal fluid infection marker signals; The millimeter-wave radar module is used for non-contact acquisition of physiological time-series characteristics such as heart rate, respiratory rate, and cerebral blood flow velocity, and synchronous analysis of physiological parameter changes corresponding to minute displacements of intracranial brain tissue. A flexible sensor array is used to collect physiological time-series characteristics such as intracranial pressure, brain tissue oxygen saturation, local temperature and humidity of surgical incision, and biomechanical signals through contact, and to extract abnormal waveforms and numerical features to characterize the fluctuations in physiological indicators caused by intracranial infection. The multimodal feature fusion unit employs an improved multi-head attention mechanism to achieve feature fusion of acquired and imported data; The digital twin modeling unit communicates bidirectionally with the multimodal feature fusion unit, receiving the fused feature vectors to construct a digital twin of the patient after craniocerebral surgery. This allows for real-time mapping of the patient's intracranial physiological state and incision recovery, and simultaneous prediction of the evolution trend of infection risk.

3. The system according to claim 2, characterized in that, The terahertz metasurface biosensor chip is configured to: receive a 10 μL cerebrospinal fluid sample; the chip incorporates a microfluidic structure to compress the cerebrospinal fluid sample into an ultrathin liquid film through capillary compression; the chip surface integrates a polarization-induced multimode metasurface, which generates multiple independent resonance peaks under the excitation of orthogonally polarized terahertz waves; by detecting the frequency shift Δf and amplitude attenuation ΔA of the resonance peaks, a multi-parameter fitting algorithm is used to quickly output pathogen detection characteristic signals corresponding to the white blood cell concentration (CWBC) and multinucleated cell ratio (RPMN), with a detection limit reaching [value missing]. The chip is directly connected to the cerebrospinal fluid drainage tube and adopts a combination of on-demand and trigger-based acquisition modes. It can be triggered to start detection by abnormal signals from the flexible sensor array.

4. The system according to claim 2, characterized in that: The multimodal feature fusion unit receives pathogen detection features, physiological temporal features, and anatomical static features (cranial anatomy, defect morphology, and incision location) acquired by the acquisition component, as well as imported anatomical static features from preoperative CT / MRI images. It dynamically weights and fuses these three types of multimodal features to generate a 256-dimensional unified feature vector, providing standardized feature input for subsequent model training. The improved multi-head attention mechanism optimizes cross-modal feature dynamic fusion by introducing a temporal weight factor, the calculation formula of which is as follows: , where T is the time-series weight factor matrix, d_k is the key vector dimension, Q is the query vector, K is the key vector, and V is the value vector.

5. The system according to claim 1, characterized in that: The privacy computing hub module adopts the improved FlexFair federated learning framework with fairness regularization constraints and differential privacy protection. It conducts multi-center collaborative training optimization on the standardized multimodal dataset preprocessed by the edge computing node module of the multimodal perception and fusion module, and constructs and generates a global early warning model for postoperative infection of craniocerebral trauma patients that integrates physiological temporal features, pathogen detection features and anatomical static features. This global early warning model is based on a CNN-Transformer hybrid architecture and incorporates a stage-aware attention mechanism. Local training employs a composite loss function that balances prediction accuracy and group fairness. During model training, gradient parameters are pruned using the L2 norm and Laplacian noise is injected to achieve triple optimization of multi-center data collaborative training, privacy protection, and fair early warning.

6. The system according to claim 5, characterized in that, The privacy computing hub module includes: The heterogeneous data alignment submodule utilizes a medical knowledge graph to perform semantic alignment and standardization mapping on non-standardized and heterogeneous data from different medical center systems. It introduces a cross-center data semantic alignment mechanism to construct a standardized neurosurgical knowledge graph containing mapping relationships for over 1200 medical terms and over 500 examination items. Cosine similarity (Sim(v1,v2)=v1·v2 / (∥v1∥·∥v2∥)) is used to achieve cross-center data semantic alignment. Postoperative infection-related data from neurosurgery departments of various medical centers are collected, core fields and medical terms are extracted, and heterogeneous fields from different centers are mapped to a unified concept within the standardized knowledge graph, thus achieving data semantic normalization. This ensures that multi-center data resides in the same semantic space; v1 refers to the semantic feature vectors of heterogeneous fields / medical terms in each medical center to be aligned, i.e., the high-dimensional numerical vectors obtained after transformation of the custom field names and medical terms in each hospital's local system by the word vector model, carrying the core semantic information of the local terms; v2 refers to the semantic feature vectors of standard terms / unified fields in the neurosurgical standardized knowledge graph, i.e., the high-dimensional numerical vectors obtained after transformation of the standard expressions corresponding to unified concepts in the pre-built standardized medical terminology library by the same word vector model, carrying the baseline semantic information of the standard terms; v1·v2 represents the dot product of the two vectors, and ∥v1∥·∥v2∥ represent the L2 norms of the two semantic feature vectors respectively; The federated learning training submodule adopts an improved federated learning algorithm that incorporates attention mechanisms and graph convolutional networks. It constructs training logic based on the improved FlexFair federated learning framework, adapts to the privacy compliance requirements of multi-center medical data, and enables each cooperating medical center to securely participate in the global model collaborative training without leaving the domain or leaking its local original data. During training, an adaptive learning rate optimization strategy is adopted, with a preset initial learning rate and an iterative decay mechanism configured. Combined with a local fairness constraint loss function, this achieves synergistic optimization of global model convergence speed, prediction accuracy, and group fairness, ensuring the model's generalization ability and robustness after training on multi-center heterogeneous data. The differential privacy protection submodule is used to add privacy protection parameters during the client-side model parameter or gradient upload process in federated learning. Differential privacy defines random noise to defend against privacy inference attacks; An improved Laplace noise injection mechanism is adopted. Where Δf is the gradient sensitivity, To ensure privacy, the gradients of each client-side model after local training are pruned using the L2 norm to control gradient sensitivity Δf. Calibration noise is added to the pruned gradients before uploading to the cloud server. Rényi Differential Privacy RDP is used to calculate the cumulative privacy loss, ensuring that privacy requirements are met. It integrates a privacy budget management unit to track privacy consumption in real time.

7. The system according to claim 1, characterized in that, The explainable early warning engine module includes: The risk score calculation submodule receives a global early warning model for postoperative infection in traumatic brain injury patients from the privacy computing center module, based on the standardized multimodal dataset output by the edge computing layer. It extracts local features from the multimodal data through a CNN branch, captures the temporal dependencies between different postoperative stages using a stage-aware attention mechanism in the Transformer branch, fuses the local and temporal features, and inputs them into a fully connected layer. The Platt calibration algorithm is used to optimize the score distribution, outputting a dynamic postoperative infection risk score for traumatic brain injury patients ranging from 0 to 100. The stage-aware attention mechanism can dynamically adjust feature weights according to different postoperative stages. The structured attribution report generation submodule generates a structured attribution report that includes feature quantification contribution and clinical reasoning logic. For the dynamic infection risk score and feature weights output by the risk score calculation submodule, an improved SHAP algorithm is used to quantify the positive / negative contribution of each multimodal feature vector to the risk score, selecting the top N key risk indicators with the highest contribution (N being a positive integer). An integrated medical big data language model based on Llama 2-7B fine-tuning is used, inputting key risk indicators, contribution values, risk scores, and basic patient information to generate a clinical attribution report with a "key findings - clinical reasoning - comprehensive conclusion" structure.

8. The system according to claim 7, characterized in that, The specific implementation method of the structured attribution report generation submodule is as follows: For the dynamic infection risk score and the feature weights of each dimension in the multimodal feature vector output by the risk score calculation submodule, the improved SHAP algorithm is used to quantify the positive and negative contributions of each multimodal collection indicator to the risk score, and the top N key risk indicators with preset parameter N are selected; the key risk indicators, corresponding contribution values, dynamic infection risk score and patient basic information are input into the fine-tuned medical big language model to generate a clinical attribution report with a fixed structure of "key findings-clinical reasoning-comprehensive conclusion", so as to simultaneously optimize the interpretability and clinical readability of the risk score; the postoperative time is divided into three stages: hyperacute phase (0-24h), acute phase (24-72h), and subacute phase (>72h), and the feature weights of each stage are dynamically adjusted by the attention weight calculation formula α=softmax(v·tanh(W_h·h + W_p·p)); Where h is the feature representation at time t, p is the stage embedding vector, W_h and W_p are weight matrices, and v is a 32-dimensional learnable parameter vector with a value range of [-0.1, 0.1], which is obtained by model training and optimization after random initialization.

9. The system according to claim 1, characterized in that, The intelligent closed-loop intervention module includes: (1) Graded early warning push submodule: For the dynamic infection risk score of craniocerebral trauma patients transmitted by the risk score calculation submodule, four levels of early warning are preset: blue <50 points, yellow 50-69 points, orange 70-84 points, and red ≥85 points. After receiving the risk score, the corresponding early warning level is automatically matched, and the early warning information (risk score, early warning level, key risk indicators) is pushed to the doctor's mobile terminal, nurse's workstation, and central monitoring screen through the local area network. The red early warning triggers the sound and light alarm. (2) Personalized solution matching submodule: It has a built-in pre-set solution library containing 50+ standardized solution packages, and synchronizes with the "Expert Consensus on Prevention and Control of Postoperative Infection in Neurosurgery 2024" in real time. Based on the patient's dynamic infection risk score, key risk indicators and individual information including age, underlying diseases and surgical type, it constructs a patient feature vector; extracts the feature vector of each standardized solution package in the solution library, and uses the cosine similarity matching algorithm (similarity threshold ≥ 0.85) to select the solution with the highest similarity from the pre-set solution library as the optimal intervention suggestion list and pushes it to the clinical terminal; (3) Feedback data collection and incremental optimization submodule: It interfaces with clinical terminals, including electronic medical record system and nurse workstation, and automatically extracts clinical intervention feedback data entered by medical staff for intervention plans pushed by intelligent closed-loop intervention module; after standardizing the collected data, it generates a compliant and usable feedback optimization dataset; and transmits the standardized feedback optimization dataset to the privacy computing center core module to achieve incremental optimization of the global early warning model of postoperative infection in patients with traumatic brain injury based on the improved elastic weight merging EWC algorithm.

10. The system according to claim 1, characterized in that: An edge-cloud collaborative architecture is adopted: edge nodes are deployed in the neurosurgical intensive care unit and are responsible for real-time data acquisition and preprocessing; the cloud nodes deploy a privacy computing central module and run a global early warning model for postoperative infection in patients with traumatic brain injury. Data transmission adopts HTTPS protocol + national cryptographic SM4 encryption algorithm to achieve end-to-end secure transmission, and identity authentication adopts OAuth 2.0 protocol; the data interaction latency between the edge nodes and the cloud nodes is ≤50ms.

11. The system according to claim 1, characterized in that, The system is also equipped with a closed-loop model incremental optimization process. The specific execution steps include: collecting clinical intervention feedback data through the intelligent closed-loop intervention module, performing standardized preprocessing on the feedback data, and transmitting it to the privacy computing center module; the privacy computing center module, based on the improved elastic weight merging algorithm, uses the standardized intervention feedback data to perform incremental iterative optimization of the global early warning model for postoperative infection in patients with traumatic brain injury, preventing catastrophic amnesia during model training and realizing continuous iterative upgrades of the global early warning model.

12. A method for processing postoperative infection risk warning data in patients with traumatic brain injury, applied to the system described in any one of claims 1-13, characterized in that, Includes the following steps: S1: Collect multi-dimensional raw data of patients through the multimodal perception and fusion module, and perform feature fusion using an improved multi-head attention mechanism to generate multimodal feature vectors; S2: Based on the improved FlexFair federated learning framework, combined with differential privacy technology, multi-center model collaborative training is achieved to generate a global early warning model; S3: Input the multimodal feature vector into the global early warning model, calculate the dynamic infection risk score, and generate a structured attribution report; S4: Trigger tiered early warnings based on risk scores and output a list of intervention recommendations using a matching algorithm; S5: Collect clinical intervention feedback data and incrementally optimize the global early warning model using an improved elastic weight merging algorithm.

Images