Burn patient energy evaluation system based on AI technology

Abstract

The invention relates to the technical field of burn treatment AI energy evaluation, and discloses a burn patient energy evaluation system based on an AI technology. The system comprises an image analysis unit, a stage marking unit, an energy baseline generation unit, a multi-modal fusion unit and a convolutional neural network unit. An infrared thermal image of a burn area is registered with a standard anatomical partition, a tissue metabolism activeness score of the partition is calculated, and a dynamic energy baseline is generated in combination with a time label of a disease course record. And further fusing the continuously monitored resting metabolism rate and blood biochemical indexes, and generating a partition energy time sequence curve after calibration. And splicing the time sequence curve and the thermal imaging feature map, inputting the spliced time sequence curve and the thermal imaging feature map into a three-dimensional convolutional neural network for spatio-temporal feature extraction, and predicting the energy demand of each anatomical partition in the future. According to the method, regional, dynamic and accurate evaluation and prediction of the energy metabolism state of the burn patient are realized, and a basis is provided for individualized nutrition support.

Description

Technical Field

[0001] This invention relates to the field of AI energy assessment technology for burn treatment, specifically an AI-based energy assessment system for burn patients. Background Technology

[0002] In the clinical treatment of burn patients, precise energy support is crucial to prognosis and rehabilitation. Currently, the widely used clinical method for estimating energy requirements is based on the total burn area and a fixed formula. This method mainly relies on macroscopic calculations based on the proportion of body surface area, which has its limitations. It assumes that the metabolic consumption of all burn areas is homogeneous and static, ignoring the significant differences in tissue metabolic rates between deep and superficial burns, and at different stages of wound healing. It also fails to reflect the spatial heterogeneity of local inflammatory responses and metabolic activities, leading to the risk of overall over- or local under-energy supply programs.

[0003] Existing technologies attempt to improve assessment by introducing infrared thermography to monitor wound temperature or by measuring resting metabolic rate through indirect calorimetry. However, infrared thermography is mostly used for qualitative observation or measurement of overall / average temperature, failing to precisely correlate temperature distribution with human anatomical structures and unable to convert spatial differences in the temperature field into quantitative, regionalized indicators of metabolic status. Indirect calorimetry obtains mixed data of the entire body, losing information on the distribution of energy consumption in different burn areas. Blood biochemical indicators also lack a direct correspondence with local anatomical regions. These monitoring data are isolated from each other, lacking an effective fusion mechanism, making it impossible to construct a spatiotemporal model reflecting the dynamic evolution of metabolic status in different body regions, and hindering the achievement of refined and personalized predictions of future energy needs. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based energy assessment system for burn patients to address the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides an energy assessment system for burn patients based on AI technology, the system comprising: The image analysis unit reads the infrared thermal image of the deep burn area from the burn treatment file of the target patient and identifies the contour boundaries of different temperature ranges in the thermal image; it registers the contour boundaries with standard human anatomical partitions and calculates the tissue metabolic activity score of each partition based on the distribution density of pixels at different temperatures within each partition. The staging unit retrieves the course record text from the burn treatment file, extracts time stage labels after the burn occurrence using natural language processing technology, and associates the time stage labels with each anatomical region. The energy baseline generation unit sets a dynamically changing energy consumption baseline value for each anatomical region based on the tissue metabolic activity score and the associated time stage label. The multimodal fusion unit acquires the resting metabolic rate data and blood biochemical index data continuously monitored in the burn treatment file, calibrates the resting metabolic rate data and blood biochemical index data with the energy consumption baseline value at the corresponding time point, and generates the fused partitioned energy time series curve. The convolutional neural network unit concatenates the temporal curves of the partitioned energy with the corresponding infrared thermal imaging feature maps in the channel dimension, inputs them into a three-dimensional convolutional neural network for joint extraction of spatiotemporal features, and outputs the predicted energy demand value of each anatomical partition in a future specified time period.

[0006] Preferably, the construction of the target patient's burn treatment record includes: The patient identification code is parsed and converted into an index key value that conforms to the medical record storage node query protocol; The index key value is used to concurrently query multiple distributed medical record storage nodes to obtain the patient's basic identity information, records of previous burn debridement surgeries, dressing change records, laboratory test reports, and storage paths of medical image files. Data fragments from different storage nodes are integrated in chronological order to assemble the burn treatment archive with a unified timestamp sequence.

[0007] Preferably, the image analysis unit processes the infrared thermal image of the deep burn area through the following steps: An adaptive threshold segmentation algorithm is applied to separate the infrared thermal image from the background to obtain a pure image of body surface temperature distribution. On the surface temperature distribution image, a region growing algorithm is used to expand from a preset high-temperature seed point until a temperature gradient abrupt change edge is encountered, forming multiple closed contour boundaries. For each of the contour boundaries, label its internal average temperature value, temperature variance, and the pixel area enclosed by the boundary.

[0008] Preferably, the step of registering the contour boundary with standard human anatomical regions and calculating the tissue metabolic activity score of each region based on the distribution density of pixels at different temperatures within each region includes: Load a standard human body surface anatomical mesh model, and register the infrared thermal image onto the human body surface anatomical mesh model through affine transformation; For each predefined anatomical region on the human body surface anatomical mesh model, count all the contour boundaries that fall within its range; The tissue metabolic activity score, which characterizes the metabolic level of a region, is derived based on the average temperature value, temperature variance, and pixel area of ​​the contour boundary falling within the partition.

[0009] Preferably, the method by which the staging unit extracts time stage tags from the medical record text includes: A pre-trained biomedical text entity recognition model was used to scan the medical record text and identify all time-related entities and entities describing the burn status. Based on the identified time entities, an absolute timeline from the date of the burn to the present is constructed, and this timeline is divided into the inflammation phase, the proliferative phase, and the remodeling phase. Through syntactic analysis, the anatomical location corresponding to the entity describing the burn state is associated and bound with the stage on the absolute timeline, thereby assigning one or more time stage labels to each anatomical location.

[0010] Preferably, the method by which the energy baseline generation unit sets the energy consumption baseline value includes: Establish an energy baseline comparison table that includes different burn stages and different tissue metabolic activity score intervals; For each anatomical region, query its associated time phase label and the calculated tissue metabolic activity score; The initial energy consumption baseline value of the partition under the corresponding conditions is obtained by matching and searching in the energy baseline lookup table. A correction factor that decays over time is introduced to adjust the initial energy consumption baseline value, generating the energy consumption baseline value that dynamically decreases with the recovery process.

[0011] Preferably, the calibration and generation process performed by the multimodal fusion unit includes: Synchronize the timestamps of the resting metabolic rate data and blood biochemical index data with the timestamp of the energy consumption baseline value; For each synchronized time point, the ratio of the resting metabolic rate measurement at that time point to the energy consumption baseline value of the corresponding anatomical region at that time point is calculated and used as the first calibration coefficient; Analyze the concentrations of specific inflammatory markers in the blood biochemistry data at this time point, and obtain a second calibration coefficient based on the concentration values; Multiplying the energy consumption baseline value, the first calibration coefficient, and the second calibration coefficient yields the energy consumption value at that time point after multimodal data calibration. Connecting the calibrated energy consumption values ​​at all time points forms the partitioned energy time series curve.

[0012] Preferably, the feature extraction step performed by the convolutional neural network unit includes: The input to the three-dimensional convolutional neural network is a four-dimensional tensor, whose dimensions represent the number of anatomical partitions, time step, height and width of the infrared thermal imaging feature map, and number of channels, respectively. The number of channels includes the normalized numerical channels of the partition energy time series curve and the original pixel channels of the infrared thermal imaging feature map. The three-dimensional convolutional neural network is composed of multiple three-dimensional convolutional layers, three-dimensional pooling layers, and skip connection structures stacked together, and is used to simultaneously extract the temporal evolution pattern of energy consumption and the spatial distribution pattern of body surface temperature. The final layer of the network flattens the spatiotemporal features, outputting a high-dimensional feature tensor that contains the joint features of all partitions.

[0013] Preferably, the output of the predicted energy demand for each anatomical region within a specified future time period includes: The high-dimensional feature tensor is input into the input layer of the multilayer perceptron; The multilayer perceptron contains multiple fully connected hidden layers, each followed by a nonlinear activation function and a random deactivation layer, for learning a complex nonlinear mapping from joint features to future energy requirements; The number of output layer nodes of the multilayer sensor is consistent with the number of dissected partitions, and each node outputs the predicted energy demand value of the corresponding partition within a specified future time period.

[0014] Preferably, the system further includes a comprehensive reporting unit for performing: The predicted energy requirements of all anatomical regions are summed to obtain the total predicted energy requirements of the patient's whole body. Calculate the estimated basal energy requirements based on the patient's current weight and height; The total predicted energy requirement is added to the estimated basic energy requirement, and the digestibility coefficient is taken into account to obtain the final daily total energy intake value. The daily total energy intake value, the predicted energy demand value for each region, the key parameters on which the calculation is based, and the data source are filled in according to the predefined report template to generate the structured daily total energy intake recommendation report that can be directly read and executed by clinicians.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By identifying the contour boundaries of different temperature ranges in infrared thermographic images of deep burn areas and spatially registering these boundaries with standard human anatomical divisions, tissue metabolic activity scores are calculated within each division based on the distribution density of pixels at different temperatures. This transforms non-uniform temperature field information into anatomically significant, quantitative indicators of local tissue metabolic status. It overcomes the limitations of conventional thermography, which only provides overall or average temperature values, and can quantify the differences in metabolic gradients among surviving tissue, inflammatory zones, and necrotic tissue within deep burn areas. This allows for the identification of scattered high-metabolic regions that conventional body surface area estimation cannot reflect, providing more refined anatomical and physiological evidence for subsequent zoned energy calculations.

[0016] Based on tissue metabolic activity scores and time-stage labels extracted from disease course texts, dynamically changing energy consumption baseline values ​​are set for each anatomical region. Continuously monitored resting metabolic rate data and blood biochemical index data are acquired, and these real-time monitoring data are calibrated against the corresponding time-point region energy baseline values ​​to generate a fused region energy time-series curve. This time-series curve is then concatenated with the corresponding infrared thermal imaging feature map along the channel dimension and input into a 3D convolutional neural network for joint spatiotemporal feature extraction. This achieves deep fusion of multi-source heterogeneous data. Resting metabolic rate and biochemical indicators calibrate the baseline model based on images and text, while the concatenated multimodal data, learned by the 3D convolutional network, can capture the nonlinear evolution of metabolic state over both time and anatomical space. This overcomes the limitations of traditional methods that can only provide whole-body static or overall trend data, enabling the output of differentiated energy demand predictions for each anatomical region within a specified future time period, achieving a leap from overall estimation to local prediction. Attached Figure Description

[0017] Figure 1 This is a timing diagram of the AI-based energy assessment system for burn patients described in this invention. Figure 2 A flowchart for processing infrared thermal images of deep burn areas by an image analysis unit; Figure 3 A flowchart of a method for extracting time stage labels for phased marking units; Figure 4 A time-series graph showing the changes in multimodal calibration coefficients for burn patients; Figure 5 A time-series comparison of the raw and calibrated energy consumption values ​​of burn patients. Detailed Implementation

[0018] The technical solutions of 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.

[0019] Please see Figure 1 This invention provides an AI-based energy assessment system for burn patients. The system includes an image analysis unit, a staging unit, an energy baseline generation unit, a multimodal fusion unit, and a convolutional neural network unit. The image analysis unit reads infrared thermal images of deep burn areas from the burn treatment records of the target patient, identifies the contour boundaries of different temperature ranges in the thermal images, registers these contour boundaries with standard human anatomical divisions, and calculates the tissue metabolic activity score for each division based on the distribution density of pixels at different temperatures within that division. The staging unit retrieves the medical record text from the burn treatment records, extracts time stage labels after the burn occurrence using natural language processing technology, and associates these time stage labels with each anatomical division. The energy baseline generation unit sets a dynamically changing energy consumption baseline value for each anatomical division based on the tissue metabolic activity score and the associated time stage labels. The multimodal fusion unit acquires continuously monitored resting metabolic rate data and blood biochemical index data from the burn treatment records, calibrates the resting metabolic rate data and blood biochemical index data with the energy consumption baseline values ​​at the corresponding time points, and generates a fused time-series energy curve for each division. The convolutional neural network unit concatenates the temporal curves of the energy of each anatomical region with the corresponding infrared thermal imaging feature maps in the channel dimension, inputs them into a three-dimensional convolutional neural network for joint extraction of spatiotemporal features, and outputs the predicted energy demand value of each anatomical region in the future specified time period.

[0020] In one embodiment of the present invention, see [reference] Figure 2When processing infrared thermal images of deep burn areas, the image analysis unit applies an adaptive threshold segmentation algorithm to separate the infrared thermal image from the background, obtaining a pure image of the body surface temperature distribution. On the body surface temperature distribution image, a region growing algorithm is used to expand from a preset high-temperature seed point until a temperature gradient abrupt change edge is encountered, forming multiple closed contour boundaries. The average temperature value, temperature variance, and pixel area enclosed by the boundary are labeled for each contour boundary. The image analysis unit loads a standard human body surface anatomical mesh model to register the contour boundaries with standard human anatomical partitions, and registers the infrared thermal image onto the human body surface anatomical mesh model through affine transformation. For each predefined anatomical partition on the human body surface anatomical mesh model, all contour boundaries falling within its range are counted. Based on the average temperature value, temperature variance, and pixel area of ​​the contour boundaries falling within the partition, a tissue metabolic activity score characterizing the metabolic level of that region is obtained.

[0021] In practical implementation, a specific example can be used for illustration. The infrared thermal image of the deep burn area is a digital image with a resolution of 640x480 pixels, acquired by an infrared thermal imager. Each pixel in the image stores a value representing the body surface temperature. The image analysis unit applies an adaptive threshold segmentation algorithm to separate the infrared thermal image of the deep burn area from the background. The adaptive threshold segmentation algorithm calculates a global threshold based on the temperature value distribution of the image pixels. Pixels with temperature values ​​higher than this threshold are classified as human body surface areas, and pixels with temperature values ​​lower than this threshold are classified as background areas, thereby obtaining a pure body surface temperature distribution image. The pure body surface temperature distribution image is a matrix with the same size as the original image, but the background area pixels are set to zero or specific identifier values. In some embodiments, a region growing algorithm is used to expand from a preset high-temperature seed point on a pure body surface temperature distribution image. The high-temperature seed point is automatically determined by scanning the pure body surface temperature distribution image. All pixels with a temperature value higher than a preset high-temperature threshold are selected as candidates. Then, the geometric center of the regions among these candidate pixels that are interconnected and have an area greater than the area of ​​the minimum seed region is calculated as the final high-temperature seed point. The region growing algorithm iteratively checks the temperature values ​​of its neighboring pixels with each high-temperature seed point as the center. If the difference between the temperature value of the neighboring pixel and the average temperature value of the current region is less than a preset growth tolerance, and the temperature value of the neighboring pixel is higher than a preset low-temperature boundary threshold, then the neighboring pixel is merged into the current region. The region growing algorithm continues until no new pixel meets the merging condition, and finally multiple closed contour boundaries with relatively uniform internal temperature are formed. Optionally, the average temperature value, temperature variance, and pixel area enclosed by the boundary are labeled for each contour boundary. The average temperature value is obtained by summing the temperature values ​​of all valid pixels within the contour boundary and dividing by the total number of pixels. The temperature variance is obtained by calculating the sum of the squares of the differences between the temperature value of each pixel within the contour boundary and the average temperature value, and then dividing by the total number of pixels. The pixel area enclosed by the boundary is obtained directly by counting the total number of pixels contained within the contour boundary.

[0022] The image analysis unit registers the contour boundaries with standard human anatomical divisions by loading a standard human anatomical mesh model. This model is a 3D digital model containing major sections of the torso, limbs, and head. Each anatomical division is defined by a series of polygonal mesh patches in 3D space. The pure body surface temperature distribution image is registered onto the standard human anatomical mesh model through an affine transformation. The parameters of the affine transformation are solved by minimizing the error between the projected coordinates of key points of the human contour in the pure body surface temperature distribution image and the corresponding anatomical landmarks on the standard human anatomical mesh model. For each predefined anatomical division on the standard human anatomical mesh model, all contour boundaries falling within its range are counted. The determination method involves back-projecting the 2D pixel coordinates of each contour boundary onto the 3D space of the standard human anatomical mesh model according to the registered affine transformation, and then calculating which anatomical division's mesh patch defines the space within which these 3D points fall. Based on the average temperature value, temperature variance, and pixel area of ​​the contour boundary falling within the partition, a tissue metabolic activity score characterizing the metabolic level of that region is derived. The calculation of the tissue metabolic activity score integrates the number, temperature level, and area of ​​high-temperature regions within the region, and its calculation relationship is as follows: in: This represents a tissue metabolic activity score calculated for a specific anatomical region. This represents the total number of contour boundaries falling within this anatomical region. Indicates the first The average temperature value of each contour boundary and These are the maximum and minimum temperature values ​​of all body surface pixels in a pure body surface temperature distribution image, used for temperature normalization. Indicates the first Temperature variance of the contour boundary It is a preset upper limit value for variance, used to normalize the variance. and They are respectively assigned to the first The weighting coefficients for the average temperature term and temperature variance term of each profile boundary. Indicates the first The pixel area of ​​the contour boundary. This represents the total projected pixel area of ​​the anatomical region on a standard human body surface anatomical mesh model. It can be understood as a tissue metabolic activity score. A higher value indicates a larger proportion of metabolically active tissue or a higher metabolic intensity within that anatomical region. Optional, a weighting coefficient. and It can be set to a fixed value, for example , This emphasizes the dominant role of average temperature. Optional, upper limit of variance. It can be set according to the statistical distribution of temperature variance in historical infrared thermal imaging data.

[0023] In one embodiment of the present invention, see [reference] Figure 3 The staging unit extracts time stage labels from the medical record text using a pre-trained biomedical text entity recognition model. It scans the medical record text to identify all time-related entities and entities describing the burn state. Based on the identified time entities, an absolute timeline from the burn occurrence to the present is constructed and divided into the inflammatory, proliferative, and remodeling phases. Syntactic analysis is used to associate the anatomical locations corresponding to entities describing the burn state with the stages on the absolute timeline, thereby assigning one or more time stage labels to each anatomical location. The energy baseline generation unit establishes an energy baseline reference table containing different burn stages and different tissue metabolic activity score intervals. For each anatomical region, its associated time stage label and calculated tissue metabolic activity score are queried. Matching is performed in the energy baseline reference table to obtain the initial energy consumption baseline value for that region under the corresponding conditions. A time-decreasing correction factor is introduced to adjust the initial energy consumption baseline value, generating an energy consumption baseline value that dynamically decreases with the recovery process.

[0024] In practice, the method for extracting time stage labels from medical record text by the staging unit employs a biomedical text entity recognition model pre-trained on a large medical corpus. Based on the Transformer architecture, the biomedical text entity recognition model can identify time entities and burn status entities in the text. For example, if the medical record text contains the statement "The patient was admitted to the hospital on October 26, 2023, due to flame burns. Currently, redness and swelling are visible on the left upper forearm, and there is a small amount of exudate from the wound," the biomedical text entity recognition model will identify "October 26, 2023" as a time entity and "redness and swelling" and "exudate" as entities describing the burn status. In some embodiments, an absolute timeline from the date of the burn to the present is constructed based on the identified time entities. The absolute timeline takes the date of the burn as the origin and converts the time points of all subsequent recorded events into days relative to the origin. For example, if relative time descriptions such as "3 days after admission" or "2 weeks after injury" are identified, they need to be uniformly converted into absolute timeline representations such as "+3 days" or "+14 days". The absolute timeline is divided into the inflammatory phase, the proliferative phase, and the remodeling phase. The division criteria are based on clinical consensus. For example, 0 to 3 days after injury is divided into the inflammatory phase, 4 to 21 days after injury is divided into the proliferative phase, and 22 days and above after injury is divided into the remodeling phase. Syntactic analysis associates the anatomical location corresponding to the entity describing the burn condition with the stage on the absolute timeline. By parsing the dependency tree of the text, syntactic analysis establishes the modification relationship between "redness and swelling" and "left upper limb forearm", thus associating the anatomical location "left upper limb forearm" with the entity "redness and swelling" describing the burn condition. Then, based on the time point of the text context in which the entity describing the burn condition appears, the stage to which it belongs is determined, thus assigning a time stage label of "inflammatory phase" to the anatomical location "left upper limb forearm".

[0025] The energy baseline generation unit establishes an energy baseline reference table containing different burn stages and different tissue metabolic activity score intervals. This table is a two-dimensional lookup table. One dimension represents the discrete burn stage, including the inflammatory, proliferative, and remodeling stages. The other dimension represents the continuous interval division of the tissue metabolic activity score, such as scores 0-2, 2-4, and 4-6. Each cell stores a preset baseline energy value. In practice, for each anatomical region, the associated time stage label and the calculated tissue metabolic activity score are queried. For example, for the "left upper limb forearm" region, the associated time stage label is "inflammatory stage," and the tissue metabolic activity score obtained from the image analysis unit is 4.5. A matching lookup is performed in the energy baseline reference table to obtain the initial energy consumption baseline value for that region under the corresponding conditions. For example, the baseline energy value stored in the cell corresponding to the intersection of "inflammatory stage" and the score interval "4-6" is 1500 kcal per square meter of body surface area per day. A time-decreasing correction factor is introduced to adjust the initial energy expenditure baseline. This correction factor is a continuous function based on the recovery process, generating a dynamically decreasing energy expenditure baseline. The correction factor can be calculated based on the following relationship: in: Indicates the day after the injury Baseline value of energy consumption that changes dynamically with the weather This represents the initial energy consumption baseline value obtained from the energy baseline lookup table. This represents the preset decay rate constant. This represents the number of days since the burn occurred. It can be understood that, through the above calculations, the energy baseline generation unit generates a time-varying baseline energy requirement curve for each anatomical region. The decay rate constant can also be understood. The value can be adjusted according to the overall severity of the burn; the larger the total burn area, the better. The smaller the value, the longer the high plateau period of energy demand lasts.

[0026] In one embodiment of the present invention, the multimodal fusion unit synchronizes the timestamps of resting metabolic rate data, blood biochemical index data, and energy consumption baseline values ​​during the calibration and generation process. For each synchronized time point, the ratio of the measured resting metabolic rate value at that time point to the energy consumption baseline value of the corresponding anatomical region at that time point is calculated as a first calibration coefficient. The concentrations of specific inflammatory markers in the blood biochemical index data at that time point are analyzed, and a second calibration coefficient is obtained by mapping the concentration values. The energy consumption baseline value, the first calibration coefficient, and the second calibration coefficient are multiplied to obtain the energy consumption value at that time point after multimodal data calibration. Connecting the calibrated energy consumption values ​​of all time points forms the regional energy time-series curve.

[0027] In practice, the multimodal fusion unit synchronizes the timestamps of resting metabolic rate data, blood biochemical index data, and energy consumption baseline values ​​during the calibration and generation process. For example, the resting metabolic rate data comes from an indirect calorimetry device, recorded at 09:00 on the 5th day post-injury, with a value of 1800 kcal per day. The blood biochemical index data comes from the laboratory information system, sampled at 08:30 on the 5th day post-injury, including a C-reactive protein concentration of 12 mg / L. The energy consumption baseline value is output by the energy baseline generation unit, with its corresponding time point being the 5th day post-injury. The multimodal fusion unit aligns all data times to the common time point of "the 5th day post-injury," allowing data within a preset time window to be considered synchronous. The resting metabolic rate data, blood biochemical index data, and energy consumption baseline value are then used for subsequent calculations after the timestamps are aligned. For each synchronized time point, the ratio of the measured resting metabolic rate at that time point to the baseline energy consumption value of the corresponding anatomical region at that time point is used as the first calibration coefficient. For example, for the "anterior trunk" anatomical region, the baseline energy consumption value on the 5th day after injury is 1400 kcal per square meter of body surface area per day. The patient's body surface area is 1.8 square meters, so the baseline total energy requirement for this region is 2520 kcal / day. The measured whole-body resting metabolic rate is 1800 kcal / day. Therefore, the first calibration coefficient for the "anterior trunk" anatomical region is 1800 divided by 2520, and the calculation result is approximately 0.714. The concentrations of specific inflammatory markers, such as C-reactive protein (CRP), in the blood biochemical data at this time point are analyzed. A second calibration coefficient is obtained by mapping the concentration values. The mapping relationship is defined by a preset piecewise linear function. For example, when the CRP concentration is below 5 mg / L, the second calibration coefficient is 1.0. When the CRP concentration is between 5 and 20 mg / L, the second calibration coefficient increases linearly from 1.0 to 1.3. When the CRP concentration is above 20 mg / L, the second calibration coefficient remains constant at 1.3. According to this rule, the second calibration coefficient corresponding to a CRP concentration of 12 mg / L is 1.15.

[0028] Multiplying the baseline energy consumption value, the first calibration coefficient, and the second calibration coefficient yields the energy consumption value at that time point after multimodal data calibration. Connecting the calibrated energy consumption values ​​at all time points forms the partitioned energy time-series curve. The calibrated energy consumption value can be calculated based on the following relationship: in: Indicates at a point in time For the Energy consumption values ​​of each anatomical region after multimodal data calibration. Indicates at a point in time No. Baseline values ​​of dynamic changes in energy consumption for each anatomical region. Indicates at a point in time The first calibration coefficient is derived from the resting metabolic rate measurement. Indicates at a point in time The second calibration coefficient is derived from the C-reactive protein concentration. It can be understood that, through the above calculations, the original baseline energy expenditure, extrapolated based on anatomical divisions and burn staging theory, is calibrated with actual systemic metabolic measurements and biochemical data reflecting systemic inflammatory status.

[0029] In some embodiments, for the "anterior trunk" anatomical region on day 5 post-injury, the energy consumption baseline value of 2520 kcal / day, a first calibration coefficient of 0.714, and a second calibration coefficient of 1.15 are multiplied to obtain a calibrated energy consumption value of approximately 2067 kcal / day. This calculation process is repeated for the same anatomical region at different time points, and the results are connected in chronological order to form a region energy time-series curve reflecting the dynamic changes in energy demand for that region. It can be understood that the region energy time-series curve integrates theoretical models, measured metabolic rates, and biochemical inflammation information, providing input data closer to reality for subsequent predictions. In specific implementations, the theoretical model is represented by the energy baseline generation unit setting a dynamically changing energy consumption baseline value for each anatomical region. This baseline value is generated based on tissue metabolic activity scores and disease course stage labels, reflecting the theoretical expectation of wound metabolism. The measured metabolic rate data is used to correct the theoretical baseline by incorporating the actual measured values ​​through a first calibration coefficient. This coefficient is the ratio of the resting metabolic rate measurement to the energy consumption baseline value of the corresponding anatomical region, proportionally distributing the measured energy consumption to each region to ensure the data accurately reflects the patient's current overall metabolic level. Optionally, the inflammatory markers used in the blood biochemistry data to map the second calibration coefficient are not limited to C-reactive protein, but may also include interleukin-6 and procalcitonin.

[0030] See Figure 4This is a time-series graph showing the changes in multimodal calibration coefficients for burn patients, primarily illustrating the dynamic trends of the first and second calibration coefficients over time after injury. The second calibration coefficient is in the range of 1.1-1.3 in the early post-injury period (1-10 days), gradually decreasing to 1.0 over time, reflecting the pattern of high CRP concentration during the inflammatory phase and subsequent inflammation resolution. The first calibration coefficient gradually increases from 0.6 to 0.9, reflecting the trend of resting metabolic rate gradually recovering with the recovery process. This graph is used in the multimodal fusion phase, visually presenting the dynamic changes in calibration coefficients and helping clinicians understand the multidimensional calibration process of "theory + measurement + inflammation" in energy requirement calculation. It is a core parameter visualization tool for energy assessment of burn patients.

[0031] In one embodiment of the present invention, the input of the three-dimensional convolutional neural network (3D Convolutional Neural Network) in the feature extraction step is a four-dimensional tensor, whose dimensions represent the number of anatomical partitions, the time step, the height and width of the infrared thermal imaging feature map, and the number of channels, respectively. The number of channels includes the normalized numerical channels of the partition energy time-series curves and the original pixel channels of the infrared thermal imaging feature map. The 3D Convolutional Neural Network is composed of multiple stacked 3D convolutional layers, 3D pooling layers, and skip connection structures, used to simultaneously extract the temporal evolution pattern of energy consumption and the spatial distribution pattern of body surface temperature. The last layer of the network flattens the spatiotemporal features, outputting a high-dimensional feature tensor containing the joint features of all partitions. When outputting the predicted energy demand value for each anatomical partition within a specified future time period, the high-dimensional feature tensor is input to the input layer of a multilayer perceptron. The multilayer perceptron contains multiple fully connected hidden layers, each followed by a nonlinear activation function and a random deactivation layer, used to learn the complex nonlinear mapping from joint features to future energy demand. The number of nodes in the output layer of the multilayer perceptron is consistent with the number of anatomical partitions, and each node outputs the predicted energy demand value for the corresponding partition within a specified future time period.

[0032] In practical implementation, the input of the 3D convolutional neural network (CNN) to the feature extraction step is a four-dimensional tensor. The dimensions of this four-dimensional tensor represent the number of anatomical partitions, the time step, the height and width of the infrared thermal imaging feature map, and the number of channels. For example, when processing data from a patient with 18 predefined anatomical partitions, and selecting data from the five most recent time points, the infrared thermal imaging feature map corresponding to each anatomical partition at each time point is preprocessed and uniformly scaled to 128x128 pixels. Therefore, the specific shape of the input four-dimensional tensor is [18, 5, 128, 128, 2], where the number 2 channels includes the normalized number of partition energy time-series curves. The value channel and the original pixel channel of the infrared thermal imaging feature map, and the normalized value channel of the partition energy time series curve are scalar values ​​formed by scaling the energy consumption value sequence of each anatomical partition at the most recent 5 time points by dividing by a preset normalization base. This scalar value is copied and filled into a 128x128 matrix to form the data of the first channel. The original pixel channel of the infrared thermal imaging feature map is a 128x128 pixel grayscale temperature image registered and scaled according to the corresponding time point and corresponding anatomical partition. Its pixel value represents the normalized temperature information and constitutes the second channel. Refer to Table 1. The data structure of the input four-dimensional tensor can be understood through Table 1.

[0033] Table 1: Input Four-Dimensional Tensor Structure Table The 3D convolutional neural network consists of multiple stacked 3D convolutional layers, 3D pooling layers, and skip connection structures. The 3D convolutional layers use 3x3x3 3D ​​convolutional kernels to perform sliding computations in the temporal, spatial height, and spatial width dimensions to extract local spatiotemporal features. Each 3D convolutional layer is followed by a 3D batch normalization layer and a ReLU activation function layer. The 3D pooling layer uses 2x2x2 3D max pooling to reduce the temporal and spatial resolution of the feature map while preserving significant features. The skip connection structure adds the output feature map of the shallow 3D convolutional layer to the output feature map of the deep 3D convolutional layer in the channel dimension through an identity mapping. This is used to simultaneously extract the temporal evolution pattern of energy consumption and the spatial distribution pattern of body surface temperature and alleviate the gradient vanishing problem. The last layer of the network flattens the spatiotemporal features, outputting a high-dimensional feature tensor containing the joint features of all partitions. The flattening operation is to transform the multi-dimensional feature matrix output by the last three-dimensional convolutional layer into a one-dimensional vector by rearranging it. The high-dimensional feature tensor is this one-dimensional vector with a length of 18x2x16x16x64=589824.

[0034] A high-dimensional feature tensor is input into the input layer of a multilayer perceptron (MLP). The MLP contains multiple fully connected hidden layers. Each fully connected hidden layer performs a linear transformation on the input feature vector and adds a bias, and introduces non-linear mapping capability through a non-linear activation function. Each hidden layer is followed by a non-linear activation function and a random deactivation layer. The non-linear activation function used is the ReLU function. The number of nodes in the output layer of the MLP is the same as the number of dissected partitions. The output layer uses a linear activation function, and each node outputs the predicted energy demand value for the corresponding partition in a specified future time period. The calculation process of the output value can be described according to the following relationship: in: The output of the multilayer perceptron represents the first... The projected energy demand for each anatomical region over a specified future period is typically expressed in kilocalories per day. This indicates that the output layer corresponds to the first... The weight row vector of each node, This represents a feature column vector composed of high-dimensional feature tensors. This represents the dot product operation of vectors. This indicates that the output layer corresponds to the first... The bias scalar of each node. It can be understood that through the cascaded processing of a 3D convolutional neural network and a multilayer perceptron, the system completes end-to-end calculations from multi-source temporal spatial data to the quantitative energy requirements of each anatomical region in the future. In some embodiments, if the specified future period is from the 10th to the 17th day post-injury, the 18 output nodes of the multilayer perceptron will respectively provide the predicted average daily energy requirements of the 18 anatomical regions over the next 8 days. Optionally, the specific number of 3D convolutional layers in the 3D convolutional neural network can be set to 4 layers, with the number of convolutional kernels being 32, 64, 128, and 256 respectively.

[0035] In one embodiment of the present invention, the construction of the target patient's burn treatment file parses the patient's identity code and converts it into an index key value conforming to the medical record storage node query protocol. Multiple distributed medical record storage nodes are queried concurrently using the index key value to obtain the patient's basic identity information, records of previous burn debridement surgeries, dressing change records, laboratory test reports, and the storage paths of medical image files. Data fragments from different storage nodes are integrated in chronological order and assembled into a burn treatment file with a unified timestamp sequence. The comprehensive reporting unit sums the predicted energy requirements of all anatomical regions to obtain the patient's total predicted energy requirements. The baseline energy requirement estimate is calculated based on the patient's current weight and height. The total predicted energy requirement is added to the baseline energy requirement estimate, and the digestibility coefficient is considered to obtain the final daily total energy intake value. The daily total energy intake value, the predicted energy requirements for each region, the key parameters used in the calculation, and the data sources are filled according to a predefined report template to generate a structured daily total energy intake recommendation report that clinicians can directly read and execute.

[0036] In practice, the construction of the target patient's burn treatment file involves parsing the patient's identification code, which is a unique identifier assigned by the hospital information system, such as "PAT-2023-001". This code is then converted into an index key value that conforms to the medical record storage node query protocol. The conversion process is based on the encoding rules agreed upon by the distributed medical record storage nodes, adding a fixed prefix and check code before the patient's identification code to generate an index key value in the format "QUERY_KEY_PAT-2023-001_VERIFY". In some embodiments, multiple distributed medical record storage nodes are queried concurrently using index key values. These distributed medical record storage nodes include surgical record storage nodes, laboratory report storage nodes, and image archive storage nodes. The query operation simultaneously sends requests carrying index key values ​​to these nodes to obtain the storage paths of the patient's basic identity information, records of previous burn debridement surgeries, dressing change records, laboratory test reports, and medical image files. For example, the storage path " / surgery / 20231026_debridement.pdf" is returned from the surgical record storage node, and the storage path " / lab / 20231027_blood_test.xml" is returned from the laboratory report storage node. Data fragments from different storage nodes are integrated in chronological order. The chronological order is based on the timestamp information contained in each record or file itself. Data from the surgical record storage node, laboratory report storage node, and image archive storage node are sorted according to their timestamps and assembled into a burn treatment archive with a unified timestamp sequence. Logically, the burn treatment archive is represented as an ordered list with a one-to-one correspondence between timestamps and data entries.

[0037] The integrated reporting unit sums the predicted energy requirements of all anatomical regions. These predicted energy requirements, output by the convolutional neural network unit, represent the average daily energy requirement for each anatomical region over a specified future period. For example, the predicted values ​​for 15 anatomical regions are 210, 185, ..., 95 kcal / day, respectively. Summing these values ​​yields the patient's total predicted energy requirement, which is 2150 kcal / day. In practice, the basal energy requirement estimate is calculated based on the patient's current weight and height using the Harris-Benedict formula. This formula requires the patient's current weight and height data; for example, a patient weighing 70 kg and 175 cm tall is substituted into the Harris-Benedict formula. The total predicted energy requirement is then added to the basal energy requirement estimate, taking into account the digestibility factor (DEF), a preset value less than 1 based on the patient's gastrointestinal function. This results in the final daily total energy intake, calculated based on the following relationship: in: This represents the final total daily energy intake value. This indicates the patient's total predicted energy requirements. This represents the estimated basal energy requirement calculated based on the patient's current weight and height. This represents the digestion and absorption efficiency coefficient. It can be understood as the total predicted energy demand. This reflects the additional energy demand resulting from the metabolism of burn wounds, while the estimated basal energy demand... This reflects the energy required to maintain basic life functions. Optional, the digestibility coefficient. The value typically ranges from 0.85 to 0.95, with the specific value set by the clinician based on the patient's intestinal tolerance.

[0038] In some embodiments, if the calculated total predicted energy requirement is 2150 kcal / day, the estimated basal energy requirement is 1650 kcal / day, and the preset digestibility coefficient is 0.9, then the daily total energy intake is (2150 + 1650) / 0.9, which is approximately 4222 kcal / day. It can be understood that this calculation process combines the predicted additional requirement with the basal requirement and considers efficiency losses during nutrient delivery. The final daily total energy intake is the target amount that needs to be achieved through oral or enteral nutritional support. The daily total energy intake, predicted energy requirements for each region, key parameters used in the calculation, and data sources are filled into a predefined report template. The predefined report template is a structured electronic document framework containing fields such as patient information, total energy recommendations, regional energy analysis, and detailed calculation parameters, generating an institutionalized daily total energy intake recommendation report that clinicians can directly read and implement.

[0039] See Figure 5 This is a time-series comparison of the raw and calibrated energy expenditure values ​​of burn patients, primarily demonstrating the effect of multimodal calibration on energy assessment results. The raw energy values ​​fluctuated significantly (e.g., close to 1800 kcal / day at day 15), while the calibrated values ​​more closely reflect actual clinical recovery patterns. In the early post-injury period (days 1-5), the calibrated values ​​were lower than the raw values, gradually converging later, reflecting a reasonable adjustment to energy requirements after inflammation subsides. This graph is used in the multimodal fusion phase, visually presenting the difference between the raw energy estimate and the calibrated results, helping clinicians understand the optimizing role of multimodal data in energy assessment, and serving as a core reference tool for accurate assessment of energy needs in burn patients.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based energy assessment system for burn patients, characterized in that, The system includes: The image analysis unit reads the infrared thermal image of the deep burn area from the burn treatment file of the target patient and identifies the contour boundaries of different temperature ranges in the thermal image; it registers the contour boundaries with standard human anatomical partitions and calculates the tissue metabolic activity score of each partition based on the distribution density of pixels at different temperatures within each partition. The staging unit retrieves the course record text from the burn treatment file, extracts time stage labels after the burn occurrence using natural language processing technology, and associates the time stage labels with each anatomical region. The energy baseline generation unit sets a dynamically changing energy consumption baseline value for each anatomical region based on the tissue metabolic activity score and the associated time stage label. The multimodal fusion unit acquires the resting metabolic rate data and blood biochemical index data continuously monitored in the burn treatment file, calibrates the resting metabolic rate data and blood biochemical index data with the energy consumption baseline value at the corresponding time point, and generates the fused partitioned energy time series curve. The convolutional neural network unit concatenates the temporal curves of the partitioned energy with the corresponding infrared thermal imaging feature maps in the channel dimension, inputs them into a three-dimensional convolutional neural network for joint extraction of spatiotemporal features, and outputs the predicted energy demand value of each anatomical partition in a future specified time period.

2. The AI-based energy assessment system for burn patients according to claim 1, characterized in that, The construction of the target patient's burn treatment record includes: The patient identification code is parsed and converted into an index key value that conforms to the medical record storage node query protocol; The index key value is used to concurrently query multiple distributed medical record storage nodes to obtain the patient's basic identity information, records of previous burn debridement surgeries, dressing change records, laboratory test reports, and storage paths of medical image files. Data fragments from different storage nodes are integrated in chronological order to assemble the burn treatment archive with a unified timestamp sequence.

3. The AI-based energy assessment system for burn patients according to claim 1, characterized in that, The image analysis unit processes the infrared thermal image of the deep burn area through the following steps: An adaptive threshold segmentation algorithm is applied to separate the infrared thermal image from the background to obtain a pure image of body surface temperature distribution. On the surface temperature distribution image, a region growing algorithm is used to expand from a preset high-temperature seed point until a temperature gradient abrupt change edge is encountered, forming multiple closed contour boundaries. For each of the contour boundaries, label its internal average temperature value, temperature variance, and the pixel area enclosed by the boundary.

4. The AI-based energy assessment system for burn patients according to claim 3, characterized in that, The process involves registering the contour boundary with standard human anatomical regions, and calculating the tissue metabolic activity score of each region based on the distribution density of pixels at different temperatures within that region. This includes: Load a standard human body surface anatomical mesh model, and register the infrared thermal image onto the human body surface anatomical mesh model through affine transformation; For each predefined anatomical region on the human body surface anatomical mesh model, count all the contour boundaries that fall within its range; The tissue metabolic activity score, which characterizes the metabolic level of a region, is derived based on the average temperature value, temperature variance, and pixel area of ​​the contour boundary falling within the partition.

5. The AI-based energy assessment system for burn patients according to claim 1, characterized in that, The method by which the staging unit extracts time stage tags from the medical record text includes: A pre-trained biomedical text entity recognition model was used to scan the medical record text and identify all time-related entities and entities describing the burn status. Based on the identified time entities, an absolute timeline from the date of the burn to the present is constructed, and this timeline is divided into the inflammation phase, the proliferative phase, and the remodeling phase. Through syntactic analysis, the anatomical location corresponding to the entity describing the burn state is associated and bound with the stage on the absolute timeline, thereby assigning one or more time stage labels to each anatomical location.

6. The AI-based energy assessment system for burn patients according to claim 1, characterized in that, The method for setting the energy consumption baseline value by the energy baseline generation unit includes: Establish an energy baseline comparison table that includes different burn stages and different tissue metabolic activity score intervals; For each anatomical region, query its associated time phase label and the calculated tissue metabolic activity score; The initial energy consumption baseline value of the partition under the corresponding conditions is obtained by matching and searching in the energy baseline lookup table. A correction factor that decays over time is introduced to adjust the initial energy consumption baseline value, generating the energy consumption baseline value that dynamically decreases with the recovery process.

7. The AI-based energy assessment system for burn patients according to claim 1, characterized in that, The multimodal fusion unit performs the calibration and generation process, which includes: Synchronize the timestamps of the resting metabolic rate data and blood biochemical index data with the timestamp of the energy consumption baseline value; For each synchronized time point, the ratio of the resting metabolic rate measurement at that time point to the energy consumption baseline value of the corresponding anatomical region at that time point is calculated and used as the first calibration coefficient; Analyze the concentrations of specific inflammatory markers in the blood biochemistry data at this time point, and obtain a second calibration coefficient based on the concentration values; Multiplying the energy consumption baseline value, the first calibration coefficient, and the second calibration coefficient yields the energy consumption value at that time point after multimodal data calibration. Connecting the calibrated energy consumption values ​​at all time points forms the partitioned energy time series curve.

8. The AI-based energy assessment system for burn patients according to claim 1, characterized in that, The feature extraction steps performed by the convolutional neural network unit include: The input to the three-dimensional convolutional neural network is a four-dimensional tensor, whose dimensions represent the number of anatomical partitions, time step, height and width of the infrared thermal imaging feature map, and number of channels, respectively. The number of channels includes the normalized numerical channels of the partition energy time series curve and the original pixel channels of the infrared thermal imaging feature map. The three-dimensional convolutional neural network is composed of multiple three-dimensional convolutional layers, three-dimensional pooling layers, and skip connection structures stacked together, and is used to simultaneously extract the temporal evolution pattern of energy consumption and the spatial distribution pattern of body surface temperature. The final layer of the network flattens the spatiotemporal features, outputting a high-dimensional feature tensor that contains the joint features of all partitions.

9. The AI-based energy assessment system for burn patients according to claim 1, characterized in that, The output includes the predicted energy demand for each anatomical region over a specified future time period, including: The high-dimensional feature tensor is input into the input layer of the multilayer perceptron; The multilayer perceptron contains multiple fully connected hidden layers, each followed by a nonlinear activation function and a random deactivation layer, for learning a complex nonlinear mapping from joint features to future energy requirements; The number of output layer nodes of the multilayer sensor is consistent with the number of dissected partitions, and each node outputs the predicted energy demand value of the corresponding partition within a specified future time period.

10. The AI-based energy assessment system for burn patients according to claim 1, characterized in that, The system also includes a comprehensive reporting unit for performing: The predicted energy requirements of all anatomical regions are summed to obtain the total predicted energy requirements of the patient's whole body. Calculate the estimated basal energy requirements based on the patient's current weight and height; The total predicted energy requirement is added to the estimated basic energy requirement, and the digestibility coefficient is taken into account to obtain the final daily total energy intake value. The daily total energy intake value, the predicted energy demand value for each region, the key parameters on which the calculation is based, and the data source are filled in according to the predefined report template to generate the structured daily total energy intake recommendation report that can be directly read and executed by clinicians.

Images