Joint assessment method for war injury liquid resuscitation based on multi-modal incomplete data

Abstract

The present application relates to a kind of war injury liquid recovery joint evaluation method based on multi-modal incomplete data, belong to the intersection field of edge computing model and artificial intelligence.Step includes: the construction of multi-modal monitoring data and missing mode;Double-path encoding of numerical path and missing path;State vector construction and capacity-reactivity joint coordinate mapping;Capacity-reactivity decision region division and joint evaluation output;Uncertainty estimation and active check recommendation;Model training and deployment.The advantage is that: not only the accuracy and robustness of capacity and reactivity evaluation are improved at the algorithm level, but also the safety, explainability, scalability and market competitiveness of war injury liquid recovery related products are significantly enhanced at the product level, which directly and substantially improves existing and proposed battlefield treatment projects.

Description

Technical Field

[0001] This invention relates to the intersection of edge computing models and artificial intelligence, particularly to the fields of military intelligent medical equipment, intensive care, battlefield medical treatment and emergency medical treatment, and especially to a joint assessment method for combat injury fluid resuscitation based on multimodal incomplete data. This method is an intelligent method for jointly assessing the patient's volume status and fluid responsiveness under the condition of incomplete multimodal monitoring data. Background Technology

[0002] In actual clinical and battlefield treatment scenarios, the monitoring data available to patients typically include: vital signs: blood pressure (invasive / non-invasive, arterial pressure waveform), heart rate, respiratory rate, blood oxygen saturation, etc.; laboratory indicators: lactate, electrolytes, hemoglobin, blood gas analysis, coagulation function, etc.; fluid intake and output: urine output, fluid resuscitation volume, blood transfusion volume, blood loss, etc.; bedside ultrasound: inferior vena cava (IVC) diameter and its respiratory variability, echocardiographic parameters, number of lung B-lines, etc.; other hemodynamic parameters: cardiac output, stroke volume variability (SVV), pulse pressure variability (PPV), central venous pressure, etc. However, in battlefield or primary care emergency settings, these monitoring data exhibit significant incompleteness and inconsistency: missing modalities, different measurement frequencies, mismatched acquisition times, and unstable signal quality. Existing methods for assessing fluid responsiveness and volume status generally assume relatively complete data, making reliable operation difficult under the aforementioned conditions.

[0003] In existing technologies, the assessment of volumetric state and liquid reactivity mainly falls into the following categories:

[0004] 1. Traditional methods based on single or few indicators: Static indicators such as central venous pressure (CVP) and pulmonary artery wedge pressure are used to assess volume status; dynamic indicators such as pulse pressure variability (PPV), pulse volume variability (SVV), and passive leg raise test (PLR) are used to predict fluid responsiveness. These methods usually rely on single or few hemodynamic signals (arterial pressure waveform, cardiac output, etc.) and are mostly used in ICU environments.

[0005] 2. Solutions based on advanced hemodynamic monitoring equipment: Commercial products such as FloTrac / Vigileo, PiCCO, LiDCO, EV1000, and NICOM estimate cardiac output and blood volume status using technologies such as arterial pressure waveforms, thermodilution, and bioreactance, for fluid management in surgical and critically ill patients. Some devices provide "optimized fluid management guidance" calculated based on waveforms / parameters, but are essentially still based on finite modal analysis.

[0006] 3. Theoretical framework of multimodal hemodynamic monitoring: Some studies have proposed the concept of "multimodal, individualized hemodynamic management", which emphasizes the comprehensive judgment based on multiple information such as blood pressure, cardiac output, oxygen delivery, lactate, and ultrasound. However, these are mostly theoretical and procedural suggestions, lacking a specific algorithmic framework.

[0007] 4. Studies using machine learning to predict fluid responsiveness / volume-related responses: Some studies use machine learning methods such as XGBoost to predict volume responsiveness or urine output response to fluid resuscitation based on ICU electronic medical record data (vital signs, laboratory indicators, etc.). Other studies use machine learning models to predict hemodynamic responses after fluid challenges from arterial pressure waveform data. These studies typically assume relatively complete data, and missing data is usually handled through simple imputation or removal.

[0008] 5. Relevant solutions in published patents: Patent US11445975 proposes using high-resolution electrocardiogram (ECG) signals to construct a prognostic index by extracting features such as the length, slope, and area of ​​P / Q / R / S / T waves for predicting fluid reactivity. This is a single-modal ECG-driven prediction method. Its core is to use ECG morphological changes (including so-called "late potentials") to predict the patient's response to fluid resuscitation, emphasizing high sampling and algorithmic processing of single-modal ECG. It is almost entirely based on a single ECG modality (with optional addition of a small number of physiological variables); it does not treat "which modalities are missing and when" as an independent modeling object, nor does it have a dual-pathway structure; it only outputs a single FR prediction result. Patent US10687781 designs a hemodynamic monitoring device that simultaneously measures respiratory cycle data plus central venous pressure (CVP) or blood flow velocity, and assesses hemodynamic status through correlation analysis. It leans more towards hardware structure plus specific signal combinations (respiratory cycle + CVP / blood flow) for monitoring patient status. Its focus is on device structure and signal acquisition scheme (how to measure certain signals simultaneously). It does not systematically discuss the severe lack of multimodal data, nor does it have an explicit missing mode modeling mechanism. It is more about "monitoring and displaying specific hemodynamic parameters".

[0009] In summary, existing technologies have the following shortcomings in complex battlefield / emergency environments:

[0010] 1. Overly strong assumptions about data completeness: Most methods implicitly assume that key parameters can be stably obtained, which is insufficient for scenarios with incomplete multimodal data (some monitoring is missing or frequently interrupted); missing data is usually only processed by simple imputation (mean imputation, forward imputation), without incorporating the "missing mode itself" into the decision-making logic.

[0011] 2. Volume status and fluid reactivity are usually assessed separately: Many methods only predict whether fluid reactivity (FR) will respond to fluid replenishment, or only focus on fluid overload / volume status. There are few unified quantitative frameworks that output both VS and FR, which is not conducive to truly realizing an individualized strategy that "can replenish and can afford to replenish".

[0012] 3. Lack of uncertainty expression and inspection guidance mechanism: Most existing indicators (such as PPV threshold and ML model output probability) are regarded as "deterministic values" and lack explicit expression of uncertainty caused by modality loss; doctors find it difficult to judge "how reliable the current recommendation is" and "when to do echocardiography / blood gas analysis".

[0013] 4. Insufficient models and system solutions adapted to portable battlefield terminals: Most of the above-mentioned research and equipment are designed for ICU or operating room environments, relying on large equipment, complete monitoring and relatively stable network and power conditions; there are few publicly available algorithms and system architectures for joint evaluation of low-power portable devices and multimodal incomplete data on the front lines of the battlefield. Summary of the Invention

[0014] The purpose of this invention is to provide a joint assessment method for combat injury fluid resuscitation based on multimodal incomplete data, which solves the following core technical problems existing in the prior art: 1. How to achieve synchronous and quantitative assessment of patient volume status and fluid responsiveness when multimodal monitoring data is severely incomplete, irregular and noisy; 2. How to explicitly give uncertainty / confidence level in the assessment results to help doctors identify "high-risk decisions" and guide supplementary examinations when data is insufficient; 3. How to deploy the above method in battlefield portable terminals or low-power monitoring devices to provide a key assessment module for intelligent fluid resuscitation decision support systems.

[0015] The above-mentioned objective of the present invention is achieved through the following technical solution:

[0016] A combined assessment method for combat injury fluid resuscitation based on multimodal incomplete data, implemented in combat medical settings or intensive care environments, is used to jointly quantify the current volume status and fluid responsiveness of combat injury patients under conditions of severe lack of multimodal monitoring, and to provide uncertainty indicators and guidance for further examination. The method includes the following steps:

[0017] Step S1: Construction of multimodal monitoring data and missing patterns;

[0018] Step S2: Dual-path encoding of numerical and missing paths;

[0019] Step S3: State vector construction and capacity-reactivity joint coordinate mapping;

[0020] Step S4: Capacity-Reactivity Decision Area Delineation and Joint Evaluation Output;

[0021] Step S5: Uncertainty estimation and proactive inspection recommendations;

[0022] Step S6: Model training and deployment.

[0023] The construction of multimodal monitoring data and missing patterns in step S1 includes:

[0024] S1.1: Battle damage monitoring data collection:

[0025] Multimodal monitoring data of combat-wounded patients were collected from battlefield monitoring equipment, portable testing equipment and rear testing systems, and aggregated according to modality and time to obtain a raw multimodal data set indexed by timestamp;

[0026] S1.2: Construction of a unified time window:

[0027] A fixed-length time window is constructed around the current evaluation moment, and all modal data are aligned to a unified time axis according to a preset time step. Divide the time window into N consecutive time slices. On a unified timeline, an initial numerical tensor is constructed with modes as columns and time slices as rows. This is used to store the raw measurement values ​​within each time slice;

[0028] S1.3: Modal Availability Mask, Missing Modes, and Quality Score:

[0029] Based on a unified time window and an initial numerical tensor, for each time slice and each mode It determines whether a valid measurement value exists for the mode within the current time slice and evaluates its signal quality; simultaneously, it acquires numerical tensors for characterizing changes in the physiological state of combat-wounded patients within a unified time window. And modal availability mask matrix for characterizing monitoring incompleteness structures. Missing pattern tensor Modal quality score matrix This provides a foundational input for subsequent dual-pathway encoding and joint evaluation of numerical and missing mode pathways.

[0030] The dual-path encoding of the numerical path and the missing path described in step S2 includes:

[0031] S2.1: Numerical Path Coding:

[0032] The numerical tensor obtained in step S1 Modal availability mask matrix Modal quality score matrix Based on this, numerical feature input vectors are constructed using time slices as units;

[0033] For each time slice within the time window The modal measurements corresponding to each time slice are recorded in a preset order. Modal availability and quality rating The vectors are concatenated to form a numerical input vector. ; Time series Input a numerical path encoder to perform time series modeling on the numerical time series and obtain the numerical feature sequence. .in, Indicates the first time window A time slice, Indicates the first The modal measurements corresponding to each time slice Indicates the first Modal availability corresponding to each time slice Indicates the first The quality score corresponding to each time slice. Indicates the first Numerical input vectors corresponding to each time slice Indicates the number of time slices within a time window. Dimensions representing numerical features Represents the set of real numbers. This represents all components in the corresponding vector;

[0034] The numerical path encoder is a temporal neural network with a gating mechanism, and the numerical feature sequence... Only numerical information from valid data is reflected;

[0035] S2.2: Encoding of missing pathways:

[0036] The missing pattern tensor obtained in step S1 Treating it as a set of independent binary time series, construct the missing input vector according to the time slice order. ; Time series Input the missing path encoder to obtain the missing pattern feature sequence. .in, Represents the missing pattern tensor. Indicates the first The missing input vector corresponding to each time slice Represents the missing pattern tensor In the All components corresponding to each time slice This represents the missing pattern feature sequence output by the missing path encoder. Dimensions representing missing pattern features It represents the set of real numbers.

[0037] Missing pathway encoders model patterns of "which modalities are missing and when", and their outputs do not depend on any physiological measurements, and are used to characterize structural information.

[0038] S2.3: Dual-pathway feature fusion:

[0039] For each time slice within the time window The corresponding numerical features Features of missing patterns The input vector is obtained by concatenating the two vectors. ; Time series Input the fusion feature generation network and obtain the fusion feature sequence through feedforward transformation. .in, Indicates the first Numerical feature vectors corresponding to each time slice Indicates the first The missing pattern feature vector corresponding to each time slice Indicates the first The fused input vector corresponding to each time slice Represents the fused feature sequence. Indicates the fusion feature dimension. It represents the set of real numbers.

[0040] Through the above dual-path encoding and fusion, the fused feature sequence is... It also includes information on changes in the physiological status of combat-wounded patients and information on monitoring missing patterns, providing a basic feature representation for subsequent joint assessment of volume and responsiveness.

[0041] Step S3, the construction of the state vector and the capacity-reactivity joint coordinate mapping, includes:

[0042] S3.1: Time aggregation yields the current state vector:

[0043] The fused feature sequence obtained in step S2 As input, the sequence is aggregated over the entire time window using a temporal aggregation network to obtain the state vector at the current evaluation time. The time aggregation network assigns different weights to different time slices, making recent changes and key events more relevant to the state vector. Their contribution was even greater; among them, This represents the state vector at the current evaluation moment. Represents the dimension of the state vector. It represents the set of real numbers.

[0044] S3.2: Capacity-Reactivity Joint Coordinate Mapping:

[0045] In the state vector Based on this, a capacity-reactivity joint mapping function is constructed to map the state vector. Mapping to two-dimensional joint coordinates ,in, The volume coordinate is used to describe the continuous change in the volume of combat-wounded patients from significant hypovolemia to volume overload. As a reactive coordinate system, it is used to describe the strength of the response of combat-wounded patients to fluid resuscitation under standard fluid challenge conditions; a joint coordinate system is established. Let be the capacity-reactivity joint evaluation space of the coordinates. Used to characterize the current volume status of the patient; Used to characterize the current patient's responsiveness to fluid resuscitation.

[0046] The capacity-responsiveness decision region delineation and joint evaluation output described in step S4 includes:

[0047] S4.1: Division of Decision-Making Areas:

[0048] On the capacity-reactivity joint coordinate plane, several decision regions are predefined based on the combat injury recovery strategy, and a corresponding fluid management strategy is assigned to each region; the decision regions include at least:

[0049] First decision region A, capacity coordinates Less than the lower capacity threshold And reactive coordinates Greater than the lower reactivity threshold This corresponds to the "low capacity – high reactivity" state;

[0050] Second decision region B, capacity coordinates Greater than the capacity limit threshold And reactive coordinates Less than the upper limit of reactivity threshold This corresponds to the "high capacity - low reactivity" state;

[0051] The third decision region C is the intermediate region other than region A and region B, corresponding to a state where capacity and responsiveness are not extreme or information is insufficient.

[0052] S4.2: Generation of Joint Evaluation Conclusions:

[0053] The joint coordinates of current combat wounded patients Mapped to the above decision region: when joint coordinates If the location falls within region A, the joint assessment conclusion is "insufficient capacity and significantly responsive to fluid resuscitation," indicating that fluid resuscitation is appropriate; when the joint coordinates... If the patient falls into region B, the joint assessment conclusion is "high volume overload and low responsiveness to fluid resuscitation," indicating that continued fluid administration is not advisable, and vasoconstrictor drugs or dehydration strategies should be prioritized; when the joint coordinates... If the condition falls into region C, the joint evaluation conclusion is "intermediate state or insufficient information", requiring further judgment based on the uncertainty parameters.

[0054] The uncertainty estimation and proactive inspection recommendation mentioned in step S5 include:

[0055] S5.1: Calculation of uncertainty parameters, the state vector obtained in step S3 Based on this, an uncertainty estimation network is constructed, using state vectors. Input and output uncertainty parameters During the training phase, the model's predicted results are compared with the actual capacity and reactivity labels to assess the uncertainty parameters. Perform calibration to reduce the uncertainty parameter It can reflect the reliability of the current joint assessment results; among them, This represents the uncertainty parameter corresponding to the current assessment result;

[0056] S5.2: Proactive check recommendation, when joint coordinates Falling into region C or uncertainty parameter Greater than the uncertainty threshold When the assessment results are uncertain or insufficient, the system proactively guides users to supplement key checks to improve the reliability of subsequent assessments. Indicates the uncertainty threshold;

[0057] S5.3: Recommended output of fluid resuscitation strategy:

[0058] When the uncertainty parameter Less than or equal to the uncertainty threshold and joint coordinates When landing in area A or area B, fluid resuscitation recommendations are generated based on the pre-set fluid management strategy template for the corresponding area. The fluid resuscitation recommendations are displayed to medical personnel through a battlefield portable terminal or monitoring system interface to assist in decision-making regarding combat injury fluid management.

[0059] The model training and deployment described in step S6 includes:

[0060] S6.1: Training Dataset and Label Construction. Based on previous combat casualty treatment cases and related simulation experimental data, a training dataset is constructed that includes multimodal monitoring data, actual fluid resuscitation process records, volume status labels, and fluid reactivity labels. When constructing the training dataset, the missing data patterns of the original monitoring data are preserved, and the corresponding numerical tensors are generated according to step S1. Modal availability mask matrix Missing pattern tensor Modal quality score matrix And store it together with the tag;

[0061] S6.2: Joint Loss Function Design. During training, a joint loss function is used to simultaneously optimize the capacity coordinates. Reactivity coordinates and uncertainty parameters ;

[0062] Total loss is ,in , , These are the weighting coefficients; This is a capacity loss term; This is the reactive loss term; Calibrate the loss term for uncertainty;

[0063] S6.3: Training strategies for incomplete multimodal applications:

[0064] In the early stages of training, samples with relatively complete modalities are used first to train the model, allowing it to learn the basic physiological relationship between capacity and responsiveness. In the later stages of training, samples with high missing rates are introduced, and some modalities are randomly masked at the input end according to the actual missing distribution on the battlefield. This allows the model to maintain stable performance under various missing combination conditions, thereby improving its robustness under incomplete multimodal conditions.

[0065] S6.4: Model Deployment and Online Operation

[0066] After training is completed and validation is passed, the parameter-compressed and optimized model is deployed to a battlefield portable terminal or low-power monitoring device. During online operation, the terminal processes the real-time multimodal monitoring data in sequence according to steps S1 to S5, and outputs the capacity-responsiveness joint assessment results, uncertainty assessment results, active inspection suggestions, and fluid resuscitation strategy suggestions.

[0067] This invention employs multimodal monitoring data (vital signs + laboratory data + fluid intake / output + ultrasound + other hemodynamic physiological parameters) and explicitly models the modal missing structure (M / Z / Q), using a holistic framework of "multi-source + high missing data" rather than simply "changing the signal." It introduces dual-pathway encoding of numerical and missing pathways; the missing mode tensor Z is an independent time series used to influence state estimation and uncertainty. The output of this invention is the volume state. +Liquid Reactivity The two-dimensional joint coordinates are used to divide the decision region on this plane, and uncertainty is considered. This invention presents a proactive inspection and recommendation strategy combined with liquid resuscitation for combat injuries, targeting battlefield / extreme environments and lightweight frontline deployment, thus broadening its application scenarios and objectives. The key feature of this invention is that the algorithm framework is not limited to specific hardware, but rather performs incomplete modeling and joint evaluation of multimodal data from various devices. Starting from step S1, modal availability M, missing modes Z, and quality scores Q are explicitly constructed, and the missing structures are modeled through dual-pathway encoding. The output format is output capacity – reactive two-dimensional coordinates + decision region + uncertainty + proactive inspection and recommendation, forming a complete decision support closed loop.

[0068] The beneficial effects of this invention are as follows: The technical solution for joint assessment of combat wound fluid resuscitation capacity and responsiveness based on incomplete multimodal monitoring data proposed in this invention has many practical benefits and enhancements for related products and project platforms, mainly reflected in the following aspects:

[0069] 1. Significantly improves assessment capabilities under incomplete monitoring conditions.

[0070] In battlefield and emergency settings, multimodal monitoring data naturally suffers from severe gaps, discontinuous acquisition, and inconsistent quality. Traditional fluid management products often assume that monitoring parameters are largely complete, and in scenarios with high gap rates, they can only resort to empirical judgment. This invention introduces a modal availability mask matrix M, a missing mode tensor Z, and a modal quality scoring matrix Q, employing a dual-path coding structure of numerical and missing pathways. It explicitly models "measured values" and "missing modes" at the feature level, enabling the product to output stable joint assessment results of capacity status and fluid reactivity even under conditions of significantly incomplete monitoring data. This greatly enhances the product's applicability and robustness in real battlefield and grassroots environments.

[0071] 2. Provides a unified quantitative framework for volumetric states and liquid reactivity.

[0072] Most existing fluid resuscitation products focus on a single indicator, such as reflecting only volume status (e.g., dry / wet, volume overload level) or predicting only fluid responsiveness (whether fluid administration will increase pressure or cardiac output). This invention uses a combined volume-responsiveness coordinate mapping to uniformly map the status of combat-injured patients to a two-dimensional coordinate system. Furthermore, different decision-making regions are pre-defined on this coordinate plane, enabling the product to simultaneously characterize "whether the current volume is in the high or low range" and "whether the patient is sensitive to fluids." This unified coordinate framework drives the generation of fluid resuscitation strategies, rather than relying on multiple disparate indicators for interpretation. This unified quantitative framework facilitates the output of simpler, more intuitive, and comparable assessment results, making it easier for healthcare professionals to understand and track changes in patient status over time.

[0073] 3. Introduce uncertainty and proactive inspection recommendations to enhance decision-making security and interpretability.

[0074] Traditional products often provide "deterministic values ​​or suggestions." When monitoring data is insufficient or the model lacks confidence, they often still output a seemingly clear conclusion, which can easily mislead users and affect the safety of battlefield casualty treatment. This invention outputs uncertainty parameters through an uncertainty estimation network. This information is then used to differentiate between high-confidence and low-confidence assessment scenarios. Combined with the missing pattern tensor Z and the quality score matrix Q, it drives the proactive examination recommendation module, explicitly telling users "which examinations should be prioritized when the current situation is uncertain." In this way, the product no longer just provides the "result," but also "result + how confident you are + what examinations should be done next," significantly improving the safety and interpretability of decision-making and reducing the decision-making pressure on medical staff when information is insufficient.

[0075] 4. Support collaboration between frontline and rear platforms, enhancing system-level value.

[0076] This invention inherently supports a collaborative deployment model of "lightweight frontline model + complete rear-end model," which benefits related products and project platforms in the following ways: Frontline portable terminals can run a lightweight model version, providing rapid, preliminary capacity-responsiveness joint assessment and fluid resuscitation recommendations under resource constraints; the rear-end platform can run the complete large model, performing detailed analysis of the returned data and providing correction suggestions and long-term strategy recommendations; through data feedback and model distribution mechanisms, a closed loop of "frontline data collection - rear-end learning - model update - frontline upgrade" is formed. For products, this means that the overall intelligence level of the battlefield medical system can be improved without significantly increasing frontline hardware costs; it also facilitates continuous optimization of algorithm performance through software upgrades, extending product lifecycles and enhancing product competitiveness.

[0077] 5. Lower the integration threshold and facilitate interfacing with existing monitoring equipment and project platforms.

[0078] The technical solution of this invention fully considers engineering integration requirements during its design: the input side uses structured data and timestamps output from conventional monitoring and testing systems, without requiring replacement of existing monitoring equipment; missing patterns and quality scores are automatically calculated at the software level without intrusion into the hardware; the algorithm module can be encapsulated as an independent service, software library, or embedded model, and can interface with existing battlefield casualty care platforms, fluid resuscitation algorithm platforms, and ICU decision support systems through standard interfaces. This allows product developers and system integrators to incrementally integrate the technical solution of this invention into existing products without large-scale modifications to the existing system architecture, enhancing the overall clinical value and market competitiveness of the solution as a high-value functional module.

[0079] 6. Facilitates the creation of differentiated selling points and subsequent product line expansion.

[0080] This invention differentiates the product in several key ways: a specialized algorithm for incomplete multimodal combat injury data, rather than a simple migration of algorithms from complete ICU data; a volume-reactivity joint coordinate and decision region division, resulting in an intuitive output format that facilitates product interface display and clinical communication; and an uncertainty and proactive review recommendation mechanism that clearly distinguishes the product from traditional solutions in terms of "knowing when it is unreliable." Building on this foundation, companies or project teams can further leverage the technical solutions of this invention to: expand to other types of shock, major trauma, perioperative fluid management, and other clinical scenarios to create a series of products; and add more tasks (such as prognostic assessment and organ function deterioration early warning) to construct a multi-task decision support system based on the same multimodal-missing modeling framework.

[0081] In summary, the technical solution of this invention not only improves the accuracy and robustness of capacity and responsiveness assessment at the algorithm level, but also significantly enhances the safety, interpretability, scalability, and market competitiveness of battlefield wound fluid resuscitation products at the product level, and has a direct and substantial improving effect on existing and planned battlefield medical treatment project platforms. Attached Figure Description

[0082] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate the invention and are used to explain it, but do not constitute an undue limitation of the invention.

[0083] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0084] Figure 2 This is a schematic diagram of the multimodal monitoring data and missing pattern construction process of the present invention;

[0085] Figure 3 This is a schematic diagram of the dual-path coding structure of the numerical path and the missing path of the present invention;

[0086] Figure 4 This is a schematic diagram of the capacity-reactivity joint coordinates and decision region division of the present invention;

[0087] Figure 5 This is a schematic diagram of the uncertainty estimation and proactive inspection recommendation workflow of the present invention;

[0088] Figure 6 This is a schematic diagram of the system architecture for the collaborative deployment of the present invention in front-line portable terminals and back-end platforms (the front-line side emphasizes real-time performance and lightweight inference, while the back-end side emphasizes high-precision evaluation and continuous learning). Detailed Implementation

[0089] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0090] See Figures 1 to 6 As shown, the joint assessment method for combat wound fluid resuscitation based on multimodal incomplete data of the present invention is applied to intelligent decision-making terminals for battlefield wounded fluid resuscitation; the "fluid management module" in ICU / emergency critical care decision support systems; portable multi-parameter monitoring devices + cloud decision-making platforms, etc. The method of the present invention includes the following steps:

[0091] Step S1: Construction of multimodal monitoring data and missing patterns:

[0092] S1.1 Battle damage monitoring data collection:

[0093] Multimodal monitoring data of combat-wounded patients are collected from battlefield monitoring equipment, portable testing equipment, and rear-area testing systems. The multimodal monitoring data includes at least the following:

[0094] Time-series data of vital signs: arterial blood pressure or non-invasive blood pressure, heart rate, respiratory rate, and blood oxygen saturation;

[0095] Laboratory indicators: lactate, electrolytes, hemoglobin, blood gas analysis parameters;

[0096] Intake and output data: urine output, fluid replacement volume, blood transfusion volume, estimated blood loss;

[0097] Bedside ultrasound parameters: inferior vena cava diameter and its respiratory variability, cardiac ultrasound parameters, lung B-line count, etc.

[0098] Other hemodynamic parameters include: pulse volume variation (SVV), pulse pressure variation (PPV), and cardiac output.

[0099] The above-mentioned monitoring data are aggregated according to modality and time to obtain the original multimodal data set indexed by timestamp.

[0100] S1.2, Construction of a unified time window:

[0101] Construct a fixed-length time window around the current evaluation moment, align all modal data to a unified time axis, and divide it into N time slices according to a preset time step.

[0102] A fixed-length time window is constructed around the current evaluation moment. All modal data are aligned to a unified time axis, and the time window is divided into N consecutive time slices {τ1, … , τ} according to a preset time step Δt. N On a unified timeline, construct the initial numerical tensor X with modes as columns and time slices as rows. raw This is used to store the original measurement values ​​within each time slice; for modes that do not have measurement values ​​within a certain time slice, in X... raw The space is marked with a pre-defined null value.

[0103] S1.3 Modal Availability Mask, Missing Modes, and Quality Scoring:

[0104] Based on a unified time window and an initial numerical tensor, for each time slice τ1 and each mode k, it is determined whether the mode has a valid measurement value within the current time slice, and its signal quality is evaluated, specifically including:

[0105] (1) Construct the modal availability mask matrix M. If mode k is in time slice τ i If a valid measurement exists within memory and no severe noise or obvious anomalies are detected, then M(i, k) = 1; if mode k exists in time slice τ i If there is no measurement value or the measurement value is determined to be unreliable, then M(i, k)=0.

[0106] (2) Based on this, construct the missing mode tensor Z, for each time slice τ i For each mode k, define Z(i,k)=1-M(i, k), where Z(i, k)=1 represents the time slice τ of mode k. i When a mode is in a missing state, Z(i, k)=0 indicates that there is valid data for that mode in that time slice; the missing mode tensor Z serves as an independent input describing "which modes are missing and when", and is used for subsequent modeling of missing mode pathways.

[0107] (3) Simultaneously, construct the modal quality scoring matrix Q. For the position M(i, k)=1, calculate the modality in time slice τ based on indicators such as sampling frequency, waveform integrity, and number of measurements. i The quality score is Q(i, k), and the value of Q(i, k) ranges from [0, 1]. The larger the value, the higher the signal quality. For the position where M(i, k)=0, Q(i, k) is set to 0.

[0108] (4) After removing the null value markers and combining them with the modal availability mask M, the valid measurements are filled into the numerical tensor X, so that the numerical tensor X is fully aligned with the modal availability mask matrix M, the missing mode tensor Z, and the modal quality score matrix Q in time and modal dimensions.

[0109] Through the above steps, the present invention simultaneously obtains within a unified time window: a numerical tensor X for characterizing changes in the physiological state of combat wound patients, and a modal availability mask matrix M, a missing pattern tensor Z, and a modal quality score matrix Q for characterizing the monitoring of incomplete structures, providing basic input for subsequent dual-pathway encoding and joint evaluation of numerical and missing pattern pathways.

[0110] Step S2: Dual-path encoding of numerical and missing pathways:

[0111] S2.1, Numerical Path Coding:

[0112] Based on the numerical tensor X, modal availability mask matrix M, and modal quality score matrix Q obtained in step S1, a numerical feature input vector is constructed in units of time slices.

[0113] For each time slice τ within the time window i The modal measurement values ​​X(i,·), modal availability M(i,·), and quality score Q(i,·) corresponding to the time slice are concatenated in a preset order to form a numerical input vector x. val (i). The time series {x} val (1), …, x val (N)} Input the numerical path encoder to perform time series modeling on the numerical time series, and obtain the numerical feature sequence H val H val Representing a dimension of N×d val A real matrix, where N represents the number of time slices within the time window, and d val The dimension represents the numerical feature, and the dot represents all components in the corresponding vector.

[0114] The numerical path encoder is a temporal neural network with a gating mechanism. Internally, when calculating the state update at each time step, it masks the contribution of missing modes based on the modal availability M(i,·) of the corresponding time step, thus ensuring that the numerical feature sequence H... val It only reflects numerical information from valid data.

[0115] S2.2 Missing Pathway Encoding:

[0116] The missing pattern tensor obtained in step S1 Treating it as a set of independent binary time series, construct the missing input vector according to the time slice order. Time series Input the missing path encoder to obtain the missing pattern feature sequence. .in, Represents the missing pattern tensor. Indicates the first The missing input vector corresponding to each time slice Represents the missing pattern tensor In the All components corresponding to each time slice This represents the missing pattern feature sequence output by the missing path encoder. Dimensions representing missing pattern features It represents the set of real numbers.

[0117] The missing pathway encoder models the pattern of "which modalities are missing and when". Its output does not depend on any physiological measurement values ​​and is used to characterize structural information that is independent of numerical values, such as monitoring setup, equipment status and acquisition rhythm.

[0118] S2.3, Dual-pathway feature fusion:

[0119] For each time slice within the time window The corresponding numerical features Features of missing patterns The input vector is obtained by concatenating the two vectors. .Will Input the fusion feature generation network and obtain the fusion feature sequence through feedforward transformation. .in, Indicates the first Numerical feature vectors corresponding to each time slice Indicates the first The missing pattern feature vector corresponding to each time slice Indicates the first The fused input vector corresponding to each time slice Represents the fused feature sequence. Indicates the fusion feature dimension. It represents the set of real numbers.

[0120] Through the above dual-path encoding and fusion, the fused feature sequence is... It also includes information on changes in the physiological status of combat-wounded patients and information on monitoring missing patterns, providing a basic feature representation for subsequent joint assessment of volume and responsiveness.

[0121] Step S3: State vector construction and capacity-reactivity joint coordinate mapping:

[0122] S3.1. Time aggregation yields the current state vector:

[0123] The fused feature sequence obtained in step S2 As input, the sequence is aggregated over the entire time window using a temporal aggregation network to obtain the state vector at the current evaluation time. The temporal aggregation network assigns different weights to different time slices, thus influencing the state vector with recent changes and key events. Their contribution was even greater. Among them, This represents the state vector at the current evaluation moment. Represents the dimension of the state vector. It represents the set of real numbers.

[0124] S3.2, Capacity-Reactivity Joint Coordinate Mapping:

[0125] In the state vector Based on this, a capacity-reactivity joint mapping function is constructed to map the state vector. Mapping to two-dimensional continuous coordinates ,in, The volume coordinate is used to describe the continuous change in the volume of combat-wounded patients from significant hypovolemia to volume overload. The metric is a reactivity coordinate used to describe the strength of a combat injury patient's response to fluid resuscitation under standard fluid challenge conditions.

[0126] The mapping process is implemented through a joint coordinate generation network. The joint coordinate generation network uses a state vector... As input, capacity coordinates are output after nonlinear transformation. and reactivity coordinates During training, monotonicity and order constraints related to capacity and reactivity labels are introduced to make the capacity coordinates... With capacity state, reactivity coordinates A stable correspondence is established between the reactivity of the liquid and the reactivity of the liquid.

[0127] Through the above steps, a joint coordinate system was established. The capacity-reactivity joint evaluation space is defined by coordinates.

[0128] Step S4: Capacity-Reactivity Decision Region Delineation and Joint Evaluation Output:

[0129] S4.1 Division of Decision-Making Areas:

[0130] On the capacity-reactivity joint coordinate plane, several decision regions are predefined based on the combat injury recovery strategy, and a corresponding fluid management strategy is assigned to each region. The decision regions include at least:

[0131] First Decision Region A: Capacity Coordinates Less than the lower capacity threshold And reactive coordinates Greater than the lower reactivity threshold This corresponds to the "low capacity – high reactivity" state;

[0132] Second Decision Region B: Capacity Coordinates Greater than the capacity limit threshold And reactive coordinates Less than the upper limit of reactivity threshold This corresponds to the "high capacity - low reactivity" state;

[0133] Third decision region C: The intermediate region other than region A and region B, corresponding to a state where capacity and responsiveness are not extreme or information is insufficient.

[0134] Each threshold , , , The settings are based on clinical experience and statistical results of training data.

[0135] S4.2, Generation of Joint Evaluation Conclusions:

[0136] The joint coordinates of current combat wounded patients Mapped to the decision region above:

[0137] When joint coordinates If the sample falls into region A, the joint assessment conclusion is "insufficient capacity and highly responsive to fluid resuscitation," indicating that fluid resuscitation is appropriate.

[0138] When joint coordinates If the patient falls into zone B, the joint assessment conclusion is "high volume overload and low responsiveness to fluid resuscitation", and it is not advisable to continue fluid resuscitation. The priority should be given to strategies such as vasoconstrictors or dehydration.

[0139] When joint coordinates If the condition falls into region C, the joint evaluation conclusion is "intermediate state or insufficient information", requiring further judgment based on the uncertainty parameters.

[0140] The joint assessment conclusions serve as input for the subsequent fluid resuscitation strategy generation and review recommendation modules.

[0141] Step S5: Uncertainty estimation and proactive inspection recommendations:

[0142] S5.1 Calculation of Uncertainty Parameters:

[0143] The state vector obtained in step S3 Based on this, an uncertainty estimation network is constructed, using state vectors. Input and output uncertainty parameters During the training phase, the uncertainty parameters are assessed by comparing the errors between the model's predicted results and the actual capacity and reactivity labels. Perform calibration to reduce the uncertainty parameter It can reflect the reliability of the current joint assessment results.

[0144] According to uncertainty parameters The current assessment is divided into two categories: high confidence and low confidence.

[0145] When the uncertainty parameter Less than or equal to the preset uncertainty threshold The current assessment results are deemed to have a high degree of confidence.

[0146] When the uncertainty parameter Greater than the preset uncertainty threshold The current assessment results are deemed to have insufficient confidence.

[0147] S5.2, Recommended Active Inspection:

[0148] When joint coordinates Falling into region C or uncertainty parameter Greater than the uncertainty threshold At that time, the proactive check recommendation module will be activated.

[0149] Actively check the missing pattern tensor in the root step S1 of the recommendation module Modal quality score matrix This involves identifying the set of monitoring modalities that are missing or of poor quality within the current time window, and constructing a list of candidate inspection items. For each inspection item in the candidate list... A predefined information gain evaluation function is used. Based on historical data and the model's internal representation, the expected change in uncertainty parameters after supplementing the project's measurements is estimated. Expected change The most important examination item is identified as the priority examination item, and medical personnel are prompted to perform this examination first in the battlefield terminal or monitoring system interface.

[0150] Through the above steps, when the evaluation results are uncertain or the information is insufficient, the present invention does not directly give a strong decision on fluid replacement or fluid restriction, but actively guides the supplementation of key examinations to improve the reliability of subsequent evaluations.

[0151] S5.3, Recommended output of fluid resuscitation strategy:

[0152] When the uncertainty parameter Less than or equal to the uncertainty threshold and joint coordinates Upon landing in Zone A or Zone B, a fluid resuscitation recommendation is generated based on the pre-defined fluid management strategy template for the corresponding zone. The fluid resuscitation recommendation should include at least:

[0153] (1) Whether to perform fluid resuscitation or stop fluid administration;

[0154] (2) Recommended range of liquid types;

[0155] (3) Recommended infusion rate and single dose range;

[0156] (4) Recommended next assessment window.

[0157] Fluid resuscitation recommendations should be presented to medical personnel via a battlefield portable terminal or monitoring system interface to assist in decision-making regarding combat wound fluid management.

[0158] Step S6: Model Training and Deployment

[0159] S6.1 Training Dataset and Label Construction:

[0160] Based on previous combat injury treatment cases and related simulation experimental data, a training dataset was constructed that includes multimodal monitoring data, actual fluid resuscitation process records, volume status labels, and fluid responsiveness labels. The volume status labels are determined based on a comprehensive assessment of information such as fluid balance, ultrasound findings, and pulmonary water load, while the fluid responsiveness labels are labeled based on changes in cardiac output or mean arterial pressure before and after a standard fluid challenge.

[0161] When constructing the training dataset, the missing patterns of the original monitoring data are preserved, and the corresponding numerical tensors are generated according to step S1. With modal availability mask matrix Missing pattern tensor Modal quality scoring matrix And it is stored together with the tag.

[0162] S6.2 Joint Loss Function Design:

[0163] During training, a joint loss function is used to simultaneously optimize the capacity coordinates. Reactivity coordinates and uncertainty parameters The joint loss function includes at least:

[0164] (1) Capacity loss item corresponding to the capacity status label Used to constrain capacity coordinates Consistency with the actual capacity status;

[0165] (2) Reactivity loss items corresponding to liquid reactivity labels Used to constrain reactive coordinates Consistency with the reactivity of real liquids;

[0166] (3) Uncertainty calibration loss term corresponding to prediction error Used to constrain uncertainty parameters The relationship between the actual prediction error and the actual prediction error.

[0167] Total loss is ,in , , These are the weighting coefficients. For capacity loss, For reactive loss terms, Calibrate the loss term for uncertainty.

[0168] S6.3 Training strategies for incomplete multimodal applications:

[0169] In the early stages of training, samples with relatively complete modalities are used first to train the model, allowing it to learn the basic physiological relationship between capacity and responsiveness. In the later stages of training, samples with high missing rates are introduced, and some modalities are randomly masked at the input end according to the actual missing distribution on the battlefield. This enables the model to maintain stable performance under various missing combination conditions, thereby improving its robustness under incomplete multimodal conditions.

[0170] S6.4 Model Deployment and Online Operation:

[0171] After training is completed and validation is passed, the parameter-compressed and optimized model is deployed to a battlefield portable terminal or low-power monitoring device. During online operation, the terminal processes the real-time acquired multimodal monitoring data sequentially according to steps S1 to S5, and outputs the capacity-responsiveness joint assessment results, uncertainty assessment results, active inspection suggestions, and fluid resuscitation strategy suggestions.

[0172] Example:

[0173] The joint assessment algorithm for combat wound fluid resuscitation volume and responsiveness based on incomplete multimodal monitoring data proposed in this invention can be deployed as an independent software module or an embedded model in different types of combat wound care and intensive care systems, forming a variety of specific application schemes. Typical application schemes include at least the following three categories.

[0174] I. Portable liquid resuscitation decision terminal applied on the front lines of the battlefield:

[0175] On the front lines, medical personnel typically rely on portable monitoring devices and simple laboratory tests to quickly assess combat-wounded patients. The algorithm model of this invention can be embedded into a portable fluid resuscitation decision terminal at the front lines, and the application process is as follows:

[0176] 1. Data access and preprocessing:

[0177] The portable terminal connects to front-line monitoring equipment, handheld ultrasound equipment, and simple testing tools via wired or wireless means to collect finite modal data such as blood pressure, heart rate, blood oxygen saturation, lactate, urine output, and IVC diameter in real time, and performs unified time window construction and missing pattern coding according to step S1.

[0178] 2. Local joint assessment and strategy generation:

[0179] The terminal incorporates the dual-pathway coding and capacity-responsiveness joint assessment model of this invention, processes real-time data according to steps S2 to S5, and outputs the current capacity-responsiveness joint coordinates of the combat-wounded patient. Uncertainty parameters Recommendations for proactive screening and fluid resuscitation strategies.

[0180] 3. Interface display and interaction:

[0181] The terminal interface displays the current volume status and fluid responsiveness assessment results in a concise manner; suggested fluid resuscitation protocols (whether fluid replacement is necessary, fluid type, rate, and dosage range); and prioritized supplementary examinations when uncertainty is high (e.g., "Prioritize bedside ultrasound assessment"). Healthcare professionals can make final decisions based on the terminal's recommendations and the actual battlefield conditions.

[0182] Through the above approach, the algorithm of this invention enables real-time joint assessment of capacity and responsiveness in frontline environments under conditions of severely incomplete monitoring data, providing intelligent decision support for battlefield liquid resuscitation.

[0183] II. Application in decision support systems for rear-area treatment centers and intensive care units:

[0184] In rear-area treatment centers and hospital intensive care units, monitoring equipment is more advanced, and multimodal data is richer. The algorithm of this invention can be integrated into decision support systems at rear-area treatment centers or ICUs for continuous optimization of fluid management strategies, specifically including:

[0185] 1. Integration with the monitoring information system:

[0186] The algorithm of this invention is deployed on a back-end server or ICU monitoring information system, and interfaces with bedside monitors, testing systems, PACS / ultrasound workstations to automatically receive various vital signs, laboratory indicators, ultrasound parameters, and intake and output records.

[0187] 2. Continuous capacity-reactive trajectory monitoring:

[0188] The system calls the algorithm of this invention on each war wound patient at fixed time intervals to obtain the capacity-reactivity coordinate trajectory and uncertainty change curve over time; when the assessment results show that the patient has entered a dangerous area such as "high capacity-low reactivity", it automatically issues an early warning to medical staff.

[0189] 3. Liquid Management Strategy Optimization and Retrospective Analysis:

[0190] For procedures that have already been performed involving fluid resuscitation, the system replays the patient's joint coordinates. The changes and corresponding fluid infusion records help medical staff assess the rationality of previous resuscitation strategies, provide experience reference for subsequent similar cases, and provide labeled data for the retraining of the model of this invention.

[0191] Through the above approach, the algorithm of this invention provides continuous volume-responsiveness joint monitoring and strategy optimization support in the rear treatment point and ICU environment, which helps to reduce the risk of fluid overload and inadequate resuscitation.

[0192] III. Cloud-based decision support applied to collaboration between front-line terminals and back-end platforms:

[0193] To balance the low power requirements of front-end terminals with the powerful computing advantages of back-end platforms, the algorithm of this invention can also adopt a collaborative deployment approach of "lightweight front-end model + complete back-end model" to form a cloud-based decision support application solution:

[0194] 1. Lightweight inference for the front lines:

[0195] A lightweight model version based on the core idea of ​​this invention is deployed in a portable terminal at the front line to quickly process the real-time collected finite modal data and output preliminary capacity-responsiveness joint assessment results and fluid resuscitation recommendations to support rapid response at the front line.

[0196] 2. Data feedback and detailed back-end evaluation:

[0197] When communication conditions permit, the frontline terminal transmits raw multimodal monitoring data, lightweight model evaluation results, and information on the treatment of combat-wounded patients back to the rear platform. The rear platform runs the full version of the algorithm of this invention to perform a more refined joint evaluation of the transmitted data, and generates revised fluid management recommendations or long-term treatment plans when necessary, and feeds them back to the frontline or intermediate treatment points through a secure channel.

[0198] 3. Continuous learning and model updates:

[0199] The rear platform centrally stores multimodal data, tags, and intervention results uploaded from various battlefields and treatment points. Based on the algorithm of this invention, the model is retrained and parameters are updated regularly. The updated model can be distributed to the front-line terminals as needed, realizing the continuous optimization and iteration of the battlefield liquid resuscitation decision algorithm.

[0200] Through the aforementioned cloud-based collaborative solution, the algorithm of this invention can not only meet the requirements of the front-line environment for real-time performance and computing resources, but also make full use of the computing power and data resources of the rear platform to improve the accuracy of assessment, providing a unified intelligent decision support capability for large-scale battlefield casualty treatment.

[0201] IV. Integration methods with existing equipment and project platforms:

[0202] The joint assessment method or model for combat injury fluid resuscitation based on multimodal incomplete data proposed in this invention has good integrability in engineering implementation, specifically manifested in:

[0203] 1. Interface level: This invention receives raw measurement values ​​and timestamps from various monitoring devices and information systems through a standardized data interface. It does not force changes to the existing hardware structure, but only adds processing logic for missing modes and multimodal alignment at the software level.

[0204] 2. Deployment level: This invention can be embedded into existing combat casualty care project platforms, fluid resuscitation algorithm platforms or ICU decision support platforms through software upgrades, forming a complementary relationship with existing threshold rules, scoring systems, etc., and providing data-driven capacity-responsiveness joint assessment results without replacing human decision-making.

[0205] 3. Application level: The joint coordinates output by this invention Uncertainty parameters In addition, proactive inspection and recommendations can serve as key input indicators for the upper-level liquid resuscitation strategy recommendation module, injury prognosis assessment module, and resource scheduling module, thereby improving the automation and intelligence level of the entire system under battlefield conditions.

[0206] Through the above-mentioned multiple application schemes, the algorithm or model proposed in this invention can not only run independently in a single device, but also be deeply integrated with existing monitoring equipment, combat casualty treatment platforms and critical care decision systems to form an integrated intelligent assessment and decision support solution for combat casualty fluid resuscitation scenarios.

[0207] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made to the present invention should be included within the scope of protection of the present invention.

Claims

1. A joint assessment method for combat wound fluid resuscitation based on multimodal incomplete data, characterized in that: Includes the following steps: Step S1: Construction of multimodal monitoring data and missing patterns; Step S2: Dual-path encoding of numerical and missing paths; Step S3: State vector construction and capacity-reactivity joint coordinate mapping; Step S4: Capacity-Reactivity Decision Area Delineation and Joint Evaluation Output; Step S5: Uncertainty estimation and proactive inspection recommendations; Step S6: Model training and deployment.

2. The joint assessment method for combat wound fluid resuscitation based on multimodal incomplete data according to claim 1, characterized in that: The construction of multimodal monitoring data and missing patterns in step S1 includes: S1.1: Battle damage monitoring data collection: Multimodal monitoring data of combat-wounded patients were collected from battlefield monitoring equipment, portable testing equipment and rear testing systems, and aggregated according to modality and time to obtain a raw multimodal data set indexed by timestamp; S1.2: Construction of a unified time window: A fixed-length time window is constructed around the current evaluation moment, and all modal data are aligned to a unified time axis according to a preset time step. Divide the time window into N consecutive time slices. On a unified timeline, an initial numerical tensor is constructed with modes as columns and time slices as rows. This is used to store the raw measurement values ​​within each time slice; S1.3: Modal Availability Mask, Missing Modes, and Quality Score: Based on a unified time window and an initial numerical tensor, for each time slice and each mode It determines whether a valid measurement value exists for the mode within the current time slice and evaluates its signal quality; simultaneously, it acquires numerical tensors for characterizing changes in the physiological state of combat-wounded patients within a unified time window. And modal availability mask matrix for characterizing monitoring incompleteness structures. Missing pattern tensor Modal quality score matrix This provides a foundational input for subsequent dual-pathway encoding and joint evaluation of numerical and missing mode pathways.

3. The joint assessment method for combat wound fluid resuscitation based on multimodal incomplete data according to claim 1, characterized in that: The dual-path encoding of the numerical path and the missing path described in step S2 includes: S2.1: Numerical Path Coding: The numerical tensor obtained in step S1 Modal availability mask matrix Modal quality score matrix Based on this, numerical feature input vectors are constructed using time slices as units; For each time slice within the time window The modal measurements corresponding to each time slice are recorded in a preset order. Modal availability and quality rating The vectors are concatenated to form a numerical input vector. ; Time series Input a numerical path encoder to perform time series modeling on the numerical time series and obtain the numerical feature sequence. ;in, Indicates the number of time slices within a time window. Dimensions representing numerical features Represents the set of real numbers. This represents all components in the corresponding vector; S2.2: Encoding of missing pathways: The missing pattern tensor obtained in step S1 Construct missing input vectors in time slice order ; Time series Input the missing path encoder to obtain the missing pattern feature sequence. ; in, Represents the missing pattern tensor. Indicates the first The missing input vector corresponding to each time slice Represents the missing pattern tensor In the All components corresponding to each time slice This represents the missing pattern feature sequence output by the missing path encoder. Dimensions representing missing pattern features Represents the set of real numbers; S2.3: Dual-pathway feature fusion: For each time slice within the time window The corresponding numerical features Features of missing patterns The input vector is obtained by concatenating the two vectors. ; Time series Input the fusion feature generation network and obtain the fusion feature sequence through feedforward transformation. ; in, Indicates the first Numerical feature vectors corresponding to each time slice Indicates the first The missing pattern feature vector corresponding to each time slice Indicates the first The fused input vector corresponding to each time slice Represents the fused feature sequence. Indicates the fusion feature dimension. It represents the set of real numbers.

4. The joint assessment method for combat wound fluid resuscitation based on multimodal incomplete data according to claim 1, characterized in that: Step S3, the construction of the state vector and the capacity-reactivity joint coordinate mapping, includes: S3.1: Time aggregation yields the current state vector: The fused feature sequence obtained in step S2 As input, the fused feature sequence is aggregated through a temporal aggregation network over the entire time window to obtain the state vector at the current evaluation time. The time aggregation network assigns different weights to different time slices to ensure that recent changes and key events influence the state vector. Their contribution was even greater; among them, This represents the state vector at the current evaluation moment. Represents the dimension of the state vector. Represents the set of real numbers; S3.2: Capacity-Reactivity Joint Coordinate Mapping: In the state vector Based on this, a capacity-reactivity joint mapping function is constructed to map the state vector. Mapping to two-dimensional joint coordinates ,in, Capacity coordinates; For reactive coordinates; establish a two-dimensional joint coordinate system. The capacity-reactivity joint evaluation space is defined by coordinates.

5. The joint assessment method for combat wound fluid resuscitation based on multimodal incomplete data according to claim 1, characterized in that: The capacity-responsiveness decision region delineation and joint evaluation output described in step S4 includes: S4.1: Division of Decision-Making Areas: On the capacity-reactivity joint coordinate plane, several decision regions are predefined based on the combat injury recovery strategy, and a corresponding fluid management strategy is assigned to each region; the decision regions include at least: First decision region A, capacity coordinates Less than the lower capacity threshold And reactive coordinates Greater than the lower reactivity threshold This corresponds to the "low capacity – high reactivity" state; Second decision region B, capacity coordinates Greater than the capacity limit threshold And reactive coordinates Less than the upper limit of reactivity threshold This corresponds to the "high capacity - low reactivity" state; The third decision region C is the intermediate region other than region A and region B, corresponding to a state where capacity and responsiveness are not extreme or information is insufficient; in, This indicates the lower limit threshold for capacity. Indicates the upper limit threshold of capacity. Indicates the lower limit threshold of reactivity. Indicates the upper limit threshold of reactivity; S4.2: Generation of Joint Evaluation Conclusions: The two-dimensional joint coordinates of the current combat wounded patients Mapped to the above decision region: when the two-dimensional joint coordinates When falling into region A, the joint assessment conclusion is "insufficient capacity and significantly responsive to fluid resuscitation," indicating that fluid resuscitation is appropriate; when the two-dimensional joint coordinates When falling into region B, the joint assessment conclusion is "high volume overload and low responsiveness to fluid resuscitation," indicating that continued fluid administration is not advisable, and vasoconstrictor drugs or dehydration strategies should be prioritized; when the two-dimensional joint coordinates When the condition falls into region C, the joint evaluation conclusion is "intermediate state or insufficient information", and further judgment is required by combining uncertainty parameters.

6. The joint assessment method for combat wound fluid resuscitation based on multimodal incomplete data according to claim 1, characterized in that: The uncertainty estimation and proactive inspection recommendation mentioned in step S5 include: S5.1: Calculation of uncertainty parameters, the state vector obtained in step S3 Based on this, an uncertainty estimation network is constructed, using state vectors. As input, output uncertainty parameters During the training phase, the uncertainty parameters are assessed by comparing the errors between the model's predicted results and the actual capacity and reactivity labels. Perform calibration to adjust the uncertainty parameter. It can reflect the reliability of the current joint assessment results; among them, This represents the uncertainty parameter corresponding to the current assessment result; S5.2: Actively check recommendations when two-dimensional joint coordinates Falling into region C, or uncertainty parameter Greater than the uncertainty threshold When the assessment results are uncertain or insufficient, the system proactively guides users to supplement key checks to improve the reliability of subsequent assessments. Indicates the uncertainty threshold; S5.3: Recommended output of fluid resuscitation strategy: When the uncertainty parameter Less than or equal to the uncertainty threshold and two-dimensional joint coordinates When landing in area A or area B, a fluid resuscitation suggestion is generated based on the fluid management strategy template pre-set for the corresponding area. The fluid resuscitation suggestion is displayed to medical personnel through a battlefield portable terminal or monitoring system interface to assist in decision-making regarding combat injury fluid management.

7. The joint assessment method for combat wound fluid resuscitation based on multimodal incomplete data according to claim 1, characterized in that: The model training and deployment described in step S6 includes: S6.1: Training dataset and label construction: Based on previous combat injury treatment cases and related simulation experimental data, a training dataset containing multimodal monitoring data, actual fluid resuscitation process records, volume status labels and fluid reactivity labels is constructed. S6.2: Joint Loss Function Design. During training, a joint loss function is used to simultaneously optimize the capacity coordinates. Reactivity coordinates and uncertainty parameters ; Total loss is ,in, Indicates the total loss. , , These are the weighting coefficients. For capacity loss, For reactive loss terms, Calibrate the loss term for uncertainty; S6.3: Training strategies for incomplete multimodal applications: In the early stages of training, samples with relatively complete modalities are used first to train the model, allowing it to learn the basic physiological relationship between capacity and responsiveness. In the later stages of training, samples with high missing rates are introduced, and some modalities are randomly masked at the input end according to the actual missing distribution on the battlefield. This allows the model to maintain stable performance under various missing combination conditions, thereby improving its robustness under incomplete multimodal conditions. S6.4: Model Deployment and Online Operation After training is completed and validation is passed, the parameter-compressed and optimized model is deployed to a battlefield portable terminal or low-power monitoring device. During online operation, the terminal processes the real-time multimodal monitoring data in sequence according to steps S1 to S5, and outputs the capacity-responsiveness joint assessment results, uncertainty assessment results, active inspection suggestions, and fluid resuscitation strategy suggestions.

Images