Sepsis early risk prediction method and system based on multi-modal medical data

Abstract

The application discloses a sepsis early risk prediction method and system based on multi-modal medical data, relates to the technical field of intelligent medical data processing, and comprises the following steps: collecting multi-modal medical data of a patient, including infrared thermal imaging, an electrocardiosignal, serum lactic acid and procalcitonin dynamic data, constructing a standardized multi-modal input matrix after time stamp alignment and noise filtering; based on thermal imaging heterogeneity distribution and QRS amplitude variation coupling analysis, constructing a tissue perfusion-compensation index, and combining a gradient boosting tree to extract space-time characteristics and quantize early nonlinear representation of microcirculation metabolic instability; the application can capture the coupling deviation of the decrease of peripheral microcirculation blood flow distribution uniformity and the increase of sympathetic compensatory ejection fluctuation by fusing the tissue perfusion-compensation index constructed based on the thermal imaging temperature heterogeneity distribution and the QRS wave group amplitude variation, so that early capture and quantitative representation of microcirculation metabolic instability are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data processing technology, and in particular to a method and system for early risk prediction of sepsis based on multimodal medical data. Background Technology

[0002] Sepsis is a host response disorder caused by infection that can rapidly develop into organ dysfunction. In clinical practice, sepsis progresses rapidly, with a very short golden window for treatment. Its early manifestations often lack specificity and are easily confused with other inflammatory states, making differential diagnosis difficult. Currently, the SOFA (Sequential Organ Failure Assessment) score is often used for sepsis risk stratification.

[0003] In emergency triage scenarios, patients in the preseptic stage presenting with non-infectious symptoms (such as confusion) are often delayed in identification due to the lack of typical signs of infection. Furthermore, the delayed warning of conventional vital sign thresholds makes it difficult to detect microcirculatory metabolic abnormalities during the compensatory phase, and it is challenging to determine whether these abnormalities are driven by infection rather than other stressors such as heart failure or trauma. Additionally, single biomarkers (such as serum lactate) are easily affected by liver and kidney function, leading to a lack of dynamic causal discrimination logic in routine multi-indicator cross-comparisons, thus affecting the determination of infection-driven causes. Moreover, the existing SOFA score relies heavily on static laboratory indicators, leading to an underestimation of the risk of impending septic shock, and lacks a dynamic temporal stress-response coupling mechanism to support risk level classification. Therefore, this paper proposes a method and system for early sepsis risk prediction based on multimodal medical data to address the aforementioned problems. Summary of the Invention

[0004] The main technical problem addressed by this invention is to provide a method and system for predicting the early risk of sepsis based on multimodal medical data, in order to solve the problems mentioned in the background.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: Firstly, it provides a method for predicting the early risk of sepsis based on multimodal medical data, comprising the following steps: Step 1: Collect multimodal medical data of patients, including infrared thermal imaging, electrocardiogram signals, serum lactate and procalcitonin dynamic data. After time stamp alignment and noise filtering, construct a standardized multimodal input matrix to provide temporal consistency guarantee for subsequent multi-source feature fusion. Step 2: Using infrared thermal imaging data and electrocardiogram signal data from the standardized multimodal input matrix as input, and based on the coupled analysis of thermal imaging heterogeneity distribution and QRS amplitude variation, a tissue perfusion-compensation index is constructed, and a gradient boosting tree is combined to extract spatiotemporal features, quantifying the early nonlinear characterization of microcirculatory metabolic instability, so as to capture microcirculatory dysfunction signals before the appearance of typical infection signs. Step 3: The tissue perfusion-compensation index, along with relevant data from the standardized multimodal input matrix, is used as input. An isolated forest is introduced to screen out non-infectious stress interference terms, including trauma and fever. After removing confounding biases, the corrected probability of early metabolic instability is output, which effectively reduces false positive warnings caused by non-infectious factors and improves infection specificity. Step 4: The probability of early metabolic instability is embedded as a prior weight into the XGBoost dynamic baseline comparison framework and coupled with a decision tree of nested "yes / no" infectious etiology discrimination rules. Based on the direction of change of the ratio of serum lactate to procalcitonin in two consecutive times, an infection-driven confidence label is generated to provide a dynamic causal discrimination basis for whether metabolic abnormalities are caused by infection. Step 5: Using the probability of early metabolic instability and the confidence label of infection drive as observation nodes, construct a temporal Bayesian network. Dynamically update the posterior distribution of latent variables of vasoactivity-dependent risk through recursive reasoning. Calculate the first-order time difference of the updated posterior distribution of latent variables to obtain the temporal drift rate of the posterior distribution of latent variables. Capture the inflection point of the evolution from compensation to decompensation, initiate early warning before hemodynamic decompensation occurs, and extend the clinical intervention window. Step 6: Based on the posterior distribution of latent variables and their temporal drift rate, analyze the early risk probability of sepsis and map it to four risk levels: low, medium, high, and extremely high. Output dynamic early warning classification results to guide triage intervention, provide quantitative risk stratification basis for emergency triage, and optimize the allocation of treatment resources.

[0006] Preferably, in step 1, the process of constructing the standardized multimodal input matrix is ​​as follows: Simultaneously acquire the patient's surface infrared thermal imaging temperature distribution matrix and continuous ECG lead QRS complex amplitude time-series signal, and obtain dynamic data of serum lactate concentration and procalcitonin obtained from bedside rapid detection to ensure a unified starting point for multi-source data on the time axis and avoid information misalignment; Using the peak detection time of the ECG R wave as the global synchronization benchmark, timestamp hard alignment is performed on each data stream, and median filtering is applied to the surface infrared thermal imaging temperature distribution matrix to suppress spatial salt-and-pepper noise. Adaptive wavelet threshold denoising is performed on the continuous ECG lead QRS complex amplitude time sequence signal to eliminate motion artifacts and electromagnetic interference during the acquisition process and restore the true physiological waveform. Missing values ​​in serum lactate concentration and procalcitonin dynamic data were imputed using cubic spline interpolation. The preprocessed medical data of each modality were arranged according to a uniform time grid and combined to construct a standardized multimodal input matrix, bridging the time granularity difference between biochemical detection and continuous physiological signals, and establishing a basis for synchronous analysis.

[0007] Preferably, in step 2, the process of constructing the tissue perfusion-compensation index and extracting spatiotemporal features using a combined gradient boosting tree is as follows: The ratio of the spatial standard deviation to the mean of the surface infrared thermal imaging temperature distribution matrix is ​​calculated to obtain the thermal imaging heterogeneity coefficient. At the same time, the coefficient of variation of the QRS amplitude of the continuous electrocardiogram QRS complex amplitude time sequence signal within the sliding time window is calculated to quantify the amplitude of stroke volume compensation driven by myocardial electrical activity and to quantify the synergistic relationship between peripheral perfusion heterogeneity and cardiac compensation fluctuation. By multiplying the thermal imaging heterogeneity coefficient with the QRS amplitude variation coefficient and performing normalization mapping, a tissue perfusion-compensation index is constructed to characterize the dynamic shift in the coupling relationship between peripheral microcirculation blood flow distribution and sympathetic tension, reflecting the imbalance trend between microcirculation and systemic compensation. Using the tissue perfusion-compensation index time series combined with the principal components of thermal imaging and the amplitude change rate as the input feature set, the cross-modal spatiotemporal interaction features are extracted through the gradient boosting tree model, and the early nonlinear quantitative characterization vector of microcirculation metabolic instability is output, thus extracting complex spatiotemporal coupling features that cannot be reflected by simple linear indicators.

[0008] Preferably, in step 3, the process of outputting the corrected probability of early metabolic instability is as follows: Using tissue perfusion-compensation index, time-domain index of heart rate variability, absolute value of body surface temperature, and patient trauma and surgical history identifiers as input dimensions, an isolated forest model was trained on a historical sample set to construct an isolation discriminant hyperplane for non-infectious stress events and to construct a feature isolation boundary that distinguishes between infectious and non-infectious stress. The current patient's input dimension features are input into the trained isolated forest model to obtain abnormal scores and determine whether they belong to the category of non-infectious stress interference items such as trauma, acute heart failure or heatstroke, and identify non-infectious high stress states that are easily confused with sepsis. Based on the preset abnormal scoring threshold, the early nonlinear quantitative characterization vector is weighted and mapped and inversely probabilistically corrected. After removing the confounding bias, it is transformed into an early metabolic instability probability value after interference isolation correction, thereby reducing the interference of non-infectious factors on the assessment of metabolic instability.

[0009] Preferably, when the abnormal score of the isolated forest model is higher than a preset threshold, it is determined that the current patient has a significant non-infectious stress interference item, and the confidence weight of the probability of early metabolic instability is reduced; when the abnormal score is lower than the preset threshold, the original probability output is retained to achieve probabilistic separation correction between infectious metabolic abnormalities and non-infectious stress states.

[0010] Preferably, in step 4, the process of generating infection-driven credibility tags is as follows: The corrected probability of early metabolic instability is used as a prior weight and embedded into the XGBoost dynamic baseline comparison framework. This framework maintains a steady-state baseline feature model for each patient based on their own historical data, highlighting metabolic abnormalities that deviate from their own homeostasis with reference to the patient's own baseline. In the XGBoost decision path, a hard rule decision tree containing "yes / no" infectious etiology is coupled and nested. Based on the change direction and consistency of the ratio of serum lactate to procalcitonin obtained from two consecutive tests, it is determined whether the current metabolic abnormality tends to be infection-driven, and the biochemical dynamic directionality is incorporated into the infectious etiology discrimination logic. Based on the results of the ratio change direction determination and the XGBoost deviation analysis, dynamically updated infection drive confidence labels are generated to distinguish between three types of states: high confidence infection drive, low confidence infection drive, and unobserved state. This provides a confidence stratification of the possibility of infection drive and reduces misdiagnosis of etiology.

[0011] Preferably, if the ratio of lactate to procalcitonin shows an increasing trend in the same direction for two consecutive times and the absolute concentrations of both increase synchronously, then the decision tree output is a highly reliable infection-driven result, strengthening the dynamic evidence chain of infectious etiology. If the ratio changes in the opposite direction or procalcitonin remains below the preset threshold, the decision tree output is low-confidence infection-driven, incorporating biochemical dynamics into the infection etiology discrimination logic, excluding metabolic abnormalities caused by simple stress, and reducing the risk of misjudgment.

[0012] Preferably, in step 5, the process of capturing the inflection point of the evolution from the compensatory period to the decompensatory period is as follows: Using the probability of early metabolic instability and the confidence label of infection drive as two observation nodes, and defining vasoactivity-dependent risk as a latent variable, a discrete-time temporal Bayesian network structure containing a state transition probability matrix is ​​constructed to establish a dynamic assessment framework for vascular tension reserve that can evolve over time. The sequential importance resampling particle filter algorithm is adopted. At each time interval, the latent variable posterior distribution is recursively updated based on the newly acquired observation node data at a uniform time grid point to estimate the probability density function of the current vascular activity dependence risk. Under uncertainty conditions, the vascular dependence risk estimate is continuously calibrated. The first-order time difference of the updated latent variable posterior distribution is calculated to obtain the temporal drift rate of the latent variable state. When the temporal drift rate exceeds the drift threshold for two consecutive time phases, it is marked as the critical inflection point of the evolution from compensation to decompensation, and the critical transition moment of hemodynamics from compensation to decompensation is identified. The drift threshold is calculated in real time by multiplying the standard deviation of the posterior distribution of the latent variable by a preset adaptive coefficient, so that the inflection point detection threshold is dynamically adjusted with the estimated confidence level. When the time series drift rate first exceeds the drift threshold, an inflection point warning flag is activated. If the time series drift rate in the next time phase continues to be higher than the drift threshold, the inflection point is confirmed. This is to avoid misjudgment of the inflection point caused by instantaneous disturbances, prevent instantaneous disturbances from triggering false inflection points, and improve clinical reliability.

[0013] Preferably, in step 6, the process of outputting dynamic early warning and grading results to guide triage and intervention is as follows: By integrating three types of information—the mean of the posterior distribution of the current latent variables, the time-series drift rate, and the infection-driven confidence label—a continuous value estimate of the early risk probability of sepsis is obtained through multivariate logistic regression mapping, thus condensing multidimensional pathophysiological signals into an intuitive continuous risk quantification index. Based on the pre-defined risk probability interval thresholds, the early risk probability of sepsis is mapped to four risk levels: low, medium, high, and very high. Each level corresponds to a different frequency of emergency monitoring and triggering conditions for intervention preparation, providing a differentiated basis for treatment priority in clinical decision-making. The system generates dynamic early warning classification results that include risk level, dominant risk driving factors, and inflection point indicators. These results are then output to the emergency triage decision support system in the form of structured data messages for the formulation of triage intervention strategies. This provides the triage system with directly accessible structured decision support information.

[0014] Secondly, a sepsis early risk prediction system based on multimodal medical data is also provided to implement the aforementioned sepsis early risk prediction method based on multimodal medical data. The system includes a visualization platform, which is communicatively connected to the following modules: The data acquisition and alignment module is used to simultaneously acquire multimodal medical data, including surface infrared thermal imaging, electrocardiogram signals, and bedside biochemical test data. It performs hard timestamp alignment with the peak time of the electrocardiogram R wave as the global benchmark. After median filtering, wavelet denoising, and spline interpolation, a standardized multimodal input matrix is ​​constructed according to a 5-minute uniform grid, providing a time-consistent multi-source input basis for downstream analysis modules. The metabolic instability quantification module is used to calculate the thermal imaging spatial heterogeneity coefficient and QRS amplitude variation coefficient and fuse them into a tissue perfusion-compensation index. The spatiotemporal interaction features are extracted through gradient boosting tree and the nonlinear characterization vector of microcirculatory metabolic instability is output. At the same time, the isolated forest is used to screen out non-infectious stress interference and correct the output early metabolic instability probability, so as to achieve the initial separation of infectious and non-infectious factors. The microcirculatory dysfunction signal can be captured in the early stage of infection, while reducing the interference of non-infectious factors. The infection-driven discrimination module is used to embed the corrected early metabolic instability probability as a prior weight into the XGBoost dynamic baseline comparison framework, calculate the feature deviation, and simultaneously nest a hard rule decision tree based on the change direction of the lactate to procalcitonin ratio. It jointly infers and generates three types of infection-driven labels and their confidence levels: high confidence, low confidence, and unobserved. This provides a dynamic discrimination basis for etiology attribution and a dynamic and quantifiable discrimination basis for sepsis etiology attribution. The risk evolution early warning module is used to construct a temporal Bayesian network with instability probability and infection label as observation nodes, and recursively update vasoactivity-dependent risk latent variables through particle filtering. It captures the decompensation evolution inflection point by combining the latent variable drift rate and a two-stage inflection point confirmation mechanism. It integrates three types of information and maps them to the early risk probability of sepsis through logistic regression. It outputs structured early warning messages to the triage system according to four risk levels, captures the evolution inflection point before hemodynamic decompensation occurs, and extends the effective intervention window.

[0015] The beneficial effects of this invention are: 1. This invention provides a method and system for early risk prediction of sepsis based on multimodal medical data. By fusing the heterogeneous distribution of body surface infrared thermal imaging temperature with the amplitude variation of electrocardiogram QRS complex, a tissue perfusion-compensation index is constructed. This index can capture the coupling shift between the decrease in the uniformity of peripheral microcirculation blood flow distribution and the enhancement of sympathetic compensatory stroke fluctuations, thereby achieving early capture and quantitative characterization of microcirculation metabolic instability.

[0016] 2. This invention provides a method and system for early risk prediction of sepsis based on multimodal medical data. It constructs an isolation discriminant hyperplane for non-infectious stress events using an isolated forest model. Based on multidimensional features such as tissue perfusion-compensation index, heart rate variability, body surface temperature, and history of trauma and surgery, it automatically identifies non-infectious interference items such as trauma, acute heart failure, and heatstroke, and performs attenuation correction on the probability of metabolic instability, thereby reducing false positive warnings caused by stress interference and realizing dynamic separation and probabilistic correction between infectious causes and non-infectious stress states.

[0017] 3. This invention provides a method and system for early risk prediction of sepsis based on multimodal medical data. It embeds the corrected probability of metabolic instability as a priori weight into a dynamic baseline comparison framework, and nests a hard rule decision tree based on the direction of change in the ratio of serum lactate to procalcitonin. It judges the infection driving tendency by the temporal directionality of continuous detection rather than a single static value, which overcomes the limitation of single biomarkers being interfered with by liver and kidney function. It incorporates the dynamic evolution law of biochemical indicators into the causal reasoning chain and strengthens the credibility of etiological attribution. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the workflow of the sepsis early risk prediction method based on multimodal medical data of the present invention; Figure 2 This is a schematic diagram of the method flow for the early risk prediction method of sepsis based on multimodal medical data according to the present invention. Detailed Implementation

[0019] The preferred embodiments of the present invention will now be described in detail so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0020] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a method for predicting the early risk of sepsis based on multimodal medical data, comprising the following steps: Multimodal medical data from patients was collected, including infrared thermography, electrocardiogram (ECG) signals, and dynamic data on serum lactate and procalcitonin. After timestamp alignment and noise filtering, a standardized multimodal input matrix was constructed to ensure temporal consistency for subsequent multi-source feature fusion. Simultaneously, the patient's surface infrared thermography temperature distribution matrix and continuous ECG QRS complex amplitude temporal signals were acquired, along with dynamic data on serum lactate concentration and procalcitonin obtained from bedside rapid testing. This ensured a unified starting point for the multi-source data on the timeline, avoiding information misalignment. The ECG R-wave peak detection time was used as the global synchronization benchmark to verify the data streams. Hard alignment of timestamps was performed, and median filtering was applied to the surface infrared thermal imaging temperature distribution matrix to suppress spatial salt-and-pepper noise. Adaptive wavelet threshold denoising was performed on the continuous ECG lead QRS complex amplitude time series signal to eliminate motion artifacts and electromagnetic interference during the acquisition process and restore the true physiological waveform. Cubic spline interpolation was used to impute missing values ​​in the dynamic data of serum lactate concentration and procalcitonin. The preprocessed medical data of each modality were arranged according to a unified time grid and combined to construct a standardized multimodal input matrix to bridge the time granularity difference between biochemical detection and continuous physiological signals and to build a basis for synchronous analysis. In some embodiments, the infrared thermal imaging temperature distribution matrix of the patient's body surface and the amplitude timing signal of the QRS complex in continuous ECG leads are acquired simultaneously, and dynamic data of serum lactate concentration and procalcitonin obtained from bedside rapid detection are acquired; the time stamp hard alignment operation of each data stream is performed with the ECG R wave peak detection time as the global synchronization reference; specifically, the infrared thermal imaging device uses a 320×240 pixel detection array, the frame rate is set to 1 frame / second, and the temperature resolution is better than 0.05℃; the ECG signal sampling rate is 500Hz, and the QRS complex in lead II is selected as the main analysis channel; during the alignment process, each frame of thermal imaging is bound to the nearest neighbor ECG R wave time; a 3×3 window median filter is applied to the body surface infrared thermal imaging temperature distribution matrix, iterated twice, to suppress salt-and-pepper noise caused by slight patient movement or poor sensor contact, while retaining the temperature heterogeneity caused by peripheral microcirculation blood flow redistribution. The system incorporates edge information; adaptive wavelet thresholding is applied to the QRS complex amplitude time-series signal of continuous ECG leads, using the db4 wavelet basis function for 4-level multi-resolution decomposition. Thresholds for each level are automatically set based on unbiased risk estimation criteria. After soft thresholding, the signal is reconstructed to eliminate electromyographic interference and baseline drift, preserving the true physiological fluctuations of the QRS amplitude. Missing values ​​in serum lactate concentration and procalcitonin dynamic data are imputed using cubic spline interpolation. The interpolation nodes are set to two adjacent detection times, and natural boundary constraints are used to ensure the continuity of the first derivative of the interpolated curve at the detection points. Preprocessed medical data for each modality are arranged in a 5-minute uniform time grid. For thermal imaging and ECG data, the mode or median of all valid sampled values ​​within the grid is taken as the representative value for that grid point. For biochemical data, the interpolated time-series curve is sampled at that grid point, and then combined to construct a standardized multimodal input matrix. In some embodiments, the row index of the standardized multimodal input matrix corresponds to a time grid sequence, and the column index is divided into three main blocks: thermal spatial feature column, electrocardiogram amplitude feature column, and biochemical time series feature column. The thermal spatial feature column stores the vector of the temperature matrix of each grid point after median filtering, totaling 10240 temperature values. The electrocardiogram amplitude feature column stores the mean, standard deviation, and maximum and minimum amplitude of the QRS complex amplitude within the grid point, totaling four statistics. The biochemical time series feature column stores the interpolated values ​​of lactate concentration and procalcitonin concentration at the grid point, totaling two values. For thermal imaging data, if multiple effective thermal images exist within each time grid point, the median temperature between frames is taken as the representative of the spatial temperature distribution of that grid point to resist instantaneous motion artifacts. For electrocardiogram data, if the number of QRS complexes detected within a 5-minute grid point is less than 30, then the grid point is... Grid points are marked as low signal-to-noise ratio grid points, and subsequent feature extraction uses a weighted moving average of adjacent grid points for filling. After hard alignment of timestamps, all modal data use the ECG R-wave time as a unified reference clock, and the data in each column are fully aligned on the time axis. Missing time points are not filled forward or backward. Only when all three modal data in a certain grid point are valid are the data in that row marked as valid samples and included in subsequent index calculations. During preprocessing, the window size of the thermal imaging median filter is fixed at 3×3 and is not adjusted according to the patient's body size. The number of decomposition layers for ECG wavelet denoising is fixed at 4. The node spacing of the cubic spline interpolation of biochemical data is automatically determined based on the actual blood collection timestamp. The interpolated curve does not exceed ±15% of the detection value range to maintain physiological rationality. After the standardized multimodal input matrix is ​​constructed, it is stored as a floating-point two-dimensional array.

[0021] Infrared thermography data and electrocardiogram (ECG) signal data from a standardized multimodal input matrix are used as inputs. Based on the coupled analysis of thermography heterogeneity distribution and QRS amplitude variation, a tissue perfusion-compensation index is constructed, and gradient boosting trees are combined to extract spatiotemporal features, quantifying the early nonlinear characterization of microcirculatory metabolic instability. This allows for the capture of microcirculatory dysfunction signals before the appearance of typical infection signs. The ratio of the spatial standard deviation to the mean of the surface infrared thermography temperature distribution matrix is ​​calculated to obtain the thermography heterogeneity coefficient. Simultaneously, the QRS amplitude variation coefficient of the continuous ECG lead QRS complex amplitude time-series signal within the sliding time window is calculated to quantify the stroke volume compensation wave driven by myocardial electrical activity. The dynamic amplitude is used to quantify the synergistic relationship between peripheral perfusion heterogeneity and cardiac compensatory fluctuations. The thermographic heterogeneity coefficient is multiplied with the QRS amplitude variation coefficient and normalized to construct a tissue perfusion-compensation index. This index is used to characterize the dynamic shift in the coupling relationship between peripheral microcirculation blood flow distribution and sympathetic nerve tension, reflecting the imbalance trend between microcirculation and systemic compensation. The tissue perfusion-compensation index time series is combined with the principal components of thermographic imaging and amplitude variation rate as input feature set. Cross-modal spatiotemporal interaction features are extracted through a gradient boosting tree model. The output is an early nonlinear quantitative characterization vector of microcirculatory metabolic instability, extracting complex spatiotemporal coupling features that cannot be reflected by simple linear indicators. In some embodiments, the ratio of the spatial standard deviation to the mean of the surface infrared thermal imaging temperature distribution matrix is ​​calculated to obtain the thermal imaging heterogeneity coefficient. Simultaneously, the coefficient of variation of QRS amplitude in continuous ECG leads within a 5-minute sliding time window is calculated to quantify the amplitude fluctuations in stroke volume compensation driven by myocardial electrical activity. Specifically, the thermal imaging heterogeneity coefficient is calculated based on all pixel values ​​of the effective temperature matrix within each time grid point. If the number of effective pixels in a grid point is less than 80% of the total number of pixels, the grid point is marked as invalid. The coefficient of variation of QRS amplitude is calculated using the ratio of the standard deviation to the mean of the amplitude values ​​of all detected QRS groups within the current 5-minute window. The window sliding step size is consistent with the unified time grid point, which is 5 minutes, and each window must contain at least 30 effective QRS groups; otherwise, the coefficient of variation of that window will vary. The coefficients of variation for each window were obtained by linear interpolation. The thermal imaging heterogeneity coefficient was multiplied by the QRS amplitude coefficient of variation and then mapped to the 0-1 interval using a quantile normalization method based on the historical distribution of normal populations. This constructed a tissue perfusion-compensation index. An increase in the tissue perfusion-compensation index indicates a decrease in the uniformity of peripheral microcirculation blood flow distribution and an increase in sympathetic compensatory pulse fluctuations, reflecting a shift in the perfusion-compensation coupling relationship toward instability. The first five principal components of the tissue perfusion-compensation index time series combined with the dimensionality-reduced thermal imaging temperature matrix and the QRS amplitude change rate were used as the input feature set. Cross-modal spatiotemporal interaction features were extracted using a gradient boosting tree model. The gradient boosting tree model was set with 100 trees, a maximum depth of 5 layers, a learning rate of 0.1, and a feature subsampling ratio of 0.8. The output was an early nonlinear quantitative characterization vector of microcirculation metabolic instability. In some embodiments, the normalization mapping of the tissue perfusion-compensation index employs a quantile mapping method. Historical normal population data are derived from age-matched healthy subjects without infection or cardiovascular disease recorded at previous health check-up centers, totaling 2000 samples. The product of the thermal imaging heterogeneity coefficient and the QRS amplitude variation coefficient for each sample is calculated, and the 95th quantile is used as the upper limit of normalization, and the 5th quantile as the lower limit. When the current patient's product value is below the lower limit, it is mapped to 0; when it is above the upper limit, it is mapped to 1; and when it is between the two, it is linearly mapped to the 0-1 interval. The gradient boosting tree model uses five-fold cross-validation during the training phase, and the training sample set contains at least 500... Data from 0 clinically diagnosed emergency patients were used. Positive samples were patients ultimately diagnosed with sepsis, and negative samples were patients with confirmed non-infectious systemic inflammatory response syndrome. The ratio of the two types of samples was controlled at 1:1.2. During model training, the top few features with a cumulative contribution of more than 90% were selected as retained features based on feature importance, and the remaining features were set to zero to reduce computational redundancy. The output frequency of the early nonlinear quantization representation vector of microcirculatory metabolic instability was consistent with the unified time grid, that is, a set of 128-dimensional vectors was generated every 5 minutes. If the current grid point was marked as a low signal-to-noise ratio grid point or an invalid grid point, no representation vector was output for that grid point. Output continued after the next valid grid point was restored.

[0022] Using the tissue perfusion-compensation index, along with relevant data from the standardized multimodal input matrix, as input, an isolated forest model is introduced to filter out non-infectious stress interference terms, including trauma and fever. After removing confounding biases, the corrected probability of early metabolic instability is output, effectively reducing false positive warnings caused by non-infectious factors and improving infection specificity. Using the tissue perfusion-compensation index, time-domain indicators of heart rate variability, absolute body surface temperature, and patient trauma and surgical history identifiers as input dimensions, an isolated forest model is trained on a historical sample set to construct an isolation hyperplane for non-infectious stress events and to build a feature isolation boundary distinguishing between infectious and non-infectious stress. The current patient's input dimensions are then input into the trained isolated forest model to obtain anomaly scores and determine whether they belong to trauma, acute heart failure, or fever. Within the category of non-infectious stress interference items such as sepsis, this study identifies non-infectious high-stress states that are easily confused with sepsis. Based on a preset abnormal score threshold, the early nonlinear quantitative representation vector is weighted and inversely probabilistically corrected. After removing confounding biases, it is transformed into an early metabolic instability probability value corrected for interference isolation, thus reducing the interference of non-infectious factors on the assessment of metabolic instability. When the abnormal score of the isolated forest model is higher than the preset threshold, it is determined that the current patient has significant non-infectious stress interference items, and the confidence weight of the early metabolic instability probability is reduced. When the abnormal score is lower than the preset threshold, the original probability output is retained to achieve probabilistic separation and correction between infectious metabolic abnormalities and non-infectious stress states. In some embodiments, the tissue perfusion-compensation index, heart rate variability time-domain index, absolute body surface temperature, and patient trauma and surgical history identifiers are used as input dimensions to train an isolated forest model on a historical sample set to construct an isolation discriminant hyperplane for non-infectious stress events. Specifically, the tissue perfusion-compensation index is set to 0 to 1, the heart rate variability time-domain index uses the standard deviation of RR intervals within a 5-minute window, the absolute body surface temperature is the global spatial mean of the thermal imaging matrix, and the trauma and surgical history identifiers are binary variables (set to 1 if there is a surgical or trauma record on the same day, otherwise set to 0). The four-dimensional features are displayed at each effective time grid point (5 minutes). The data is extracted from the sample and input into a pre-trained isolated forest model. This isolated forest model contains 100 isolated trees, each with a subsample size of 256 samples. The contamination rate parameter is automatically set based on the actual proportion of non-infectious stress events in the training set. The training set consists of historical data from at least 5,000 clinically diagnosed emergency patients. When the model runs, for the current patient's four-dimensional feature vector, the path length from the root node to the leaf node of each isolated tree is calculated independently. The path lengths of all trees are averaged and normalized to obtain an anomaly score (the higher the value, the more the feature pattern deviates from the normal population distribution). In some embodiments, the early nonlinear quantitative representation vector is weighted and inversely probabilistically corrected based on a preset abnormal score threshold. After removing confounding biases, it is transformed into an early metabolic instability probability value corrected by interference isolation. The preset abnormal score threshold is the 95th percentile of the abnormal scores in the training set. This percentile is determined and fixed during the model training phase through five-fold cross-validation. When the abnormal score output by the isolated forest model is higher than the preset threshold, it is determined that the current patient has significant non-infectious stress interference, including but not limited to trauma, acute heart failure, or heatstroke. At this time, the initial instability probability obtained by logistic regression mapping of the early nonlinear representation vector of microcirculatory metabolic instability output by the gradient boosting tree model is attenuated and corrected. The correction magnitude is linearly positively correlated with the magnitude of the abnormal score exceeding the threshold. The maximum attenuation coefficient is set to 0.6, that is, in extreme abnormal cases, 40% of the original probability is retained as the corrected probability. When the abnormal score is lower than or equal to the preset threshold, it is determined that there is no significant non-infectious stress interference, and the original probability output is retained. The corrected probability value is the early metabolic instability probability corrected by interference isolation, which is output to the subsequent process at a frequency of one grid point every 5 minutes.

[0023] The probability of early metabolic instability is embedded as a prior weight into the XGBoost dynamic baseline comparison framework, coupled with a decision tree that nests "yes / no" infectious etiology discrimination rules. Based on the direction of change of the ratio of serum lactate to procalcitonin in two consecutive times, an infection-driven confidence label is generated, providing a dynamic causal discrimination basis for whether metabolic abnormalities originate from infection.

[0024] Using the probability of early metabolic instability and the confidence label of infection drive as observation nodes, a temporal Bayesian network is constructed. The latent variable posterior distribution of vasoactivity-dependent risk is dynamically updated through recursive reasoning. The first-order time difference of the updated latent variable posterior distribution is calculated to obtain the temporal drift rate of the latent variable posterior distribution. The inflection point of the evolution from compensation to decompensation is captured, and an early warning is initiated before hemodynamic decompensation occurs, thus extending the clinical intervention window.

[0025] Based on the posterior distribution of latent variables and their temporal drift rate, the early risk probability of sepsis is analyzed and mapped to four risk levels: low, medium, high, and extremely high. Dynamic early warning and classification results are output to guide triage intervention, providing a quantitative risk stratification basis for emergency triage and optimizing the allocation of treatment resources.

[0026] Example 2, as Figure 1 , Figure 2 As shown, based on Example 1, this invention provides a technical solution: the process of generating infection-driven confidence labels is as follows: the corrected probability of early metabolic instability is used as a priori weight and embedded into the XGBoost dynamic baseline comparison framework. This framework maintains a steady-state baseline feature model for each patient based on their own historical data. Using the patient's own baseline as a reference, metabolic abnormalities deviating from their own homeostasis are highlighted. In the XGBoost decision path, a hard rule decision tree containing "yes / no" infectious etiology discrimination is coupled and nested. Based on the direction and consistency of the change in the ratio of serum lactate to procalcitonin obtained from two consecutive tests, it is determined whether the current metabolic abnormality tends to be infection-driven, and the biochemical dynamic directionality is incorporated into the infectious etiology discrimination logic. Based on the results of the ratio change direction judgment and XGBoost deviation analysis, dynamically updated infection drive confidence labels are generated to distinguish between three states: high confidence infection drive, low confidence infection drive, and unobserved state. This provides a confidence stratification of the possibility of infection drive, reducing misjudgment of etiology. If the ratio of lactate to procalcitonin shows an increasing trend in the same direction for two consecutive times and the absolute concentrations of both increase synchronously, the decision tree output is high confidence infection drive, strengthening the dynamic evidence chain of infectious etiology. If the ratio change direction is contrary or procalcitonin is consistently below the preset threshold, the decision tree output is low confidence infection drive. This incorporates the dynamic directionality of biochemistry into the logic of infectious etiology discrimination, excludes metabolic abnormalities caused by simple stress, and reduces the risk of misjudgment. In some embodiments, the corrected probability of early metabolic instability is used as a priori weight and embedded in the XGBoost dynamic baseline contrast framework. This framework maintains a steady-state baseline feature model for each patient based on their own historical data. Specifically, it constructs a feature mean vector and covariance matrix using the 24-hour historical data prior to the first effective time grid point (5 minutes) after the patient's admission to the emergency department (or, if no historical data is available, the median value of the first 6 effective grid points after admission is used as the initial baseline value) and continuously updates the model using a sliding window (with a window width of 288 grid points, corresponding to 24 hours) as new effective grid points are input. The baseline is updated to adaptively reflect the patient's own baseline state drift. The XGBoost model is set with 50 trees, a maximum depth of 4 layers, a learning rate of 0.05, and a binary log loss objective function. The input features are a 128-dimensional microcirculatory metabolic instability representation vector and the corrected instability probability. During model operation, the Mahalanobis distance between the current feature vector and the steady-state baseline is calculated for each effective grid point as a quantitative indicator of deviation. At the same time, the corrected metabolic instability probability is used as the sample weight, so that the phases with large deviations from the baseline and high instability probabilities have higher influence weight in the model decision. In some embodiments, a hard-rule decision tree containing "yes / no" infectious etiology discrimination is coupled and nested in the XGBoost decision path. Based on the direction and consistency of the change in the ratio of serum lactate to procalcitonin obtained from two consecutive tests, it is determined whether the current metabolic abnormality is infection-driven. Specifically, the upper limit of normal reference for lactate is set at 2.0 mmol / L, and the upper limit of normal reference for procalcitonin is set at 0.5 ng / mL. The time interval between two consecutive tests does not exceed 2 hours. If it exceeds 2 hours, only the comparison result of a single test value with its respective threshold is used as the judgment basis, without directional discrimination. The discrimination logic of the hard-rule decision tree is as follows: if the ratio of lactate to procalcitonin (lactate concentration divided by procalcitonin concentration) shows a unidirectional increasing trend in two consecutive tests (i.e., the second ratio is greater than the first ratio), and the absolute value of lactate in both tests is greater than 2.0 mmol / L and the absolute value of procalcitonin is greater than 0.5 ng / mL. If the ratio changes in two consecutive times are opposite (the second ratio is less than or equal to the first ratio), or if procalcitonin remains below 0.5 ng / mL and lactate increases but its absolute value does not exceed 2.0 mmol / L, the decision tree output is a high-confidence infection-driven state. If none of the above conditions are met, for example, if only a single test result is available or either lactate or procalcitonin is missing in two tests, the decision tree output is an unobservable state. This hard-rule decision tree does not participate in the XGBoost gradient optimization process. Instead, after XGBoost outputs the infection propensity probability, the rule output is used as a correction factor to recalibrate the probability. In some embodiments, a dynamically updated infection-driven confidence label is generated by jointly inferring the result of the ratio change direction determination and XGBoost deviation analysis to distinguish between three states: high-confidence infection-driven, low-confidence infection-driven, and unobserved. Specifically, the XGBoost model outputs the infection tendency probability value at each effective grid point, while the hard rule decision tree outputs three discrete states. During joint inference, if the hard rule output is high-confidence infection-driven and the infection tendency probability output by XGBoost is greater than 0.55, the final infection-driven confidence label is determined to be high-confidence (coded as 1), and the confidence level is taken as the rule confidence level (set to 0.85) and the XGBoost confidence level. The weighted average of the output probabilities is calculated with weights of 0.6 and 0.4. If the hard rule output is a low-confidence infection-driven label or the XGBoost output probability is less than 0.30, the final label is determined to be low-confidence (encoded as 0). All other cases are determined to be unobservable (encoded as -1). After the label is generated, it is cached together with the confidence score. If there is no new biochemical detection data input within two consecutive valid grid points (10 minutes), the most recent valid label is used, but the confidence score is reduced by 5% for each grid point until new biochemical detection data arrives and is recalculated. The final output infection-driven confidence score label is transmitted to the temporal Bayesian network at a frequency of one grid point every 5 minutes.

[0027] The process of capturing the inflection point from compensation to decompensation is as follows: Using the probability of early metabolic instability and the confidence label of infection-driven events as two observation nodes, defining vasoactivity-dependent risk as a latent variable, a discrete-time temporal Bayesian network structure containing a state transition probability matrix is ​​constructed. A dynamic assessment framework for vascular tension reserve that can evolve over time is established. A sequential importance resampling particle filtering algorithm is employed. At each uniform time interval, the posterior distribution of the latent variable is recursively updated using newly acquired observation node data to estimate the probability density function of the current vasoactivity-dependent risk. Under uncertainty conditions, the vaso-dependent risk estimate is continuously calibrated. The updated posterior distribution of the latent variable is then calculated. First-order time difference is used to obtain the temporal drift rate of latent variable states. When the temporal drift rate exceeds the drift threshold for two consecutive time phases, it is marked as a critical inflection point in the evolution from compensation to decompensation. This identifies the critical transition moment when hemodynamics slides from compensation to decompensation. The drift threshold is calculated in real time by multiplying the standard deviation of the posterior distribution of the latent variable by a preset adaptive coefficient, so that the inflection point detection threshold is dynamically adjusted with the estimated confidence level. When the temporal drift rate first exceeds the drift threshold, an inflection point warning flag is activated. If the temporal drift rate of the next time phase continues to be higher than the drift threshold, the inflection point is confirmed, so as to avoid inflection point misjudgment caused by transient disturbances, avoid false inflection points triggered by transient disturbances, and improve clinical reliability. In some embodiments, the probability of early metabolic instability and the infection-driven confidence label are used as two observation nodes, and the risk of vasoactivity dependence is defined as a latent variable. A discrete-time temporal Bayesian network structure containing a state transition probability matrix is ​​constructed. Specifically, the latent variable is a continuous variable, representing the current patient's potential dependence on vasoactive drug support. The closer the value is to 1, the more depleted the vascular tone reserve is, and the higher the probability of drug intervention. In the observation nodes, the probability of early metabolic instability is the continuous value output in step 3, and the infection-driven confidence label is the three-state discrete value output in step 4. The state transition probability matrix is ​​parameterized using a Gaussian random walk model, and the standard deviation of the transition noise is fixed at 0.02. The observation likelihood function is modeled according to the data types of the two types of observation nodes: for continuous observations, a Beta distribution is used to describe its likelihood under given conditions. The conditional probability density of the latent variable values ​​is obtained by linearly transforming the two shape parameters of the Beta distribution. Specifically, the shape parameter α is mapped to 1 plus 5 times the latent variable value, and β is mapped to 1 plus 5 times 1 minus the latent variable value. For discrete tri-state observations, the Softmax function is used to map the latent variable values ​​to the conditional probabilities of the three categories. The mapping weight matrix has a dimension of 1 row and 3 columns, and the initial values ​​of each element are set according to clinical priors and remain fixed after the model is deployed. The time step of the temporal Bayesian network is consistent with the unified time grid, that is, a discrete phase every 5 minutes. The initial prior distribution of the latent variable is set as a normal distribution with a mean of 0.15 and a standard deviation of 0.08, and truncated in the interval between 0 and 1 to reflect the clinical prior knowledge that most emergency patients have not yet entered a vasoactive-dependent state in the early stage of admission. In some embodiments, a sequential importance resampling particle filter algorithm is employed. At each uniform time interval, the posterior distribution of the latent variables is recursively updated using newly acquired observation node data to estimate the probability density function of the current vasculogenic risk. The total number of particles in the particle filter is set to 1000. Initially, the particle swarm is randomly sampled from the aforementioned initial prior distribution, and the weights of each particle are uniformly initialized. During each phase of recursion, the state of each particle is predicted based on the state transition probability matrix, i.e., a Gaussian noise perturbation with a mean of zero and a standard deviation of 0.02 is applied to each particle. Then, the likelihood weight of each particle is calculated using the newly arrived observation data, with the weight value being the product of the likelihoods of continuous and discrete observations. After the weight update is completed, a system resampling step is executed. Resampling is triggered when the number of effective particles is less than two-thirds of the total number to suppress weight degradation. The updated latent variables are then processed... The posterior distribution is used to calculate its first-order time difference, which is the posterior mean of the current time phase minus the posterior mean of the previous time phase, divided by the time interval (5 minutes) to obtain the temporal drift rate of the latent variable state. The drift threshold is calculated in real time by multiplying the standard deviation of the latent variable posterior distribution by a preset adaptive coefficient, which is fixed at 2.0. When the temporal drift rate exceeds the drift threshold for two consecutive time phases, it is marked as a critical inflection point moment in the evolution from the compensation period to the decompensation period. A two-stage confirmation mechanism is adopted to avoid false triggering caused by instantaneous noise disturbances: when the temporal drift rate first exceeds the drift threshold, only the inflection point warning flag is activated, and no final inflection point mark is generated. If the temporal drift rate of the next time phase continues to be higher than the drift threshold, the inflection point is confirmed and the inflection point moment is recorded. At the same time, the latent variable mean and drift rate value at that moment are cached for subsequent use. If the drift rate of the next time phase falls back below the threshold, the warning flag is canceled and no inflection point record is generated.

[0028] The process of outputting dynamic early warning grading results to guide triage intervention is as follows: Integrating three types of information—the mean of the posterior distribution of the current temporal latent variables, the temporal drift rate, and the infection-driven confidence label—a continuous value estimate of the early risk probability of sepsis is obtained through multivariate logistic regression mapping. Multidimensional pathophysiological signals are condensed into intuitive continuous risk quantification indicators. Based on pre-defined risk probability interval thresholds, the early risk probability of sepsis is mapped to four risk levels: low, medium, high, and extremely high. Each level corresponds to a differentiated emergency monitoring frequency and intervention preparation triggering conditions, providing differentiated treatment priority criteria for clinical decision-making. A dynamic early warning grading result containing risk level, dominant risk driving factors, and inflection point markers is generated and output to the emergency triage decision support system in the form of structured data messages for triage intervention strategy formulation, providing the triage system with directly callable structured decision support information. In some embodiments, the three types of information—the mean of the posterior distribution of the latent variables at the current time, the time-series drift rate, and the infection-driven confidence label—are integrated, and a continuous value estimate of the early risk probability of sepsis is obtained through multivariate logistic regression mapping. Specifically, the input features of the logistic regression model include three dimensions: the mean of the latent variables (reflecting the central estimate of the current vasoactivity-dependent risk), the time-series drift rate (obtained by dividing the difference between the posterior mean of the previous time step and the posterior mean of the current time step by 5 minutes, with positive and negative values ​​indicating an upward or downward trend in risk, respectively), and the infection-driven confidence label (encoded as a three-state discrete value, with 1 for high confidence, -1 for unobserved, and 0 for low confidence). The logistic regression model uses the same methods as steps 2 to 4 during the training phase. The historical dataset of emergency patients was used. Overfitting was prevented by L2 regularization constraints, with the regularization coefficient C set to 1.0. The regression coefficients of each feature were determined by five-fold cross-validation and then fixed, and were not updated during the runtime. Each effective time grid point (5 minutes) was input with the three-dimensional feature vector of the current time phase. After mapping by the Sigmoid function, a continuous probability value was output, which is the early risk probability of sepsis. If the current grid point was marked as an invalid grid point or a low signal-to-noise ratio grid point due to data quality reasons, the logistic regression mapping was not performed on that grid point. The calculation continued after the next effective grid point was restored. When three consecutive grid points were invalid, a data quality alarm was triggered and the most recent effective probability value was used. However, the confidence level was reduced by 5% for each grid point until a new effective grid point was reached and then reset. In some embodiments, the early risk probability of sepsis is mapped to four risk levels—low, medium, high, and very high—based on pre-defined risk probability interval thresholds. Each level corresponds to a differentiated emergency monitoring frequency and intervention preparation triggering conditions. The risk probability interval thresholds are determined by calibrating the Youden index maximization method using training set samples: the low-risk interval is a probability value less than 0.20, corresponding to routine observation and treatment, with a monitoring frequency of reassessing vital signs every hour and performing the full process calculations from steps 1 to 5; the medium-risk interval is between 0.20 and 0.45, corresponding to enhanced monitoring and treatment, with the monitoring frequency increased to reassessing heart rate and blood pressure every 30 minutes and retesting lactate and procalcitonin every 2 hours; the high-risk interval is between 0.45 and 0.75, corresponding to initiating preparations for sepsis bundle therapy, including establishing two or more intravenous access lines and preparing blood cultures. Prepare nutrition and antibiotics, and notify the emergency team to stand by; the extremely high risk range is defined as a probability value greater than or equal to 0.75, in which case fluid resuscitation should be initiated immediately (rapid infusion of 30 mL / kg crystalloid solution) and vasoactive drugs such as norepinephrine should be prepared, and the patient should be prioritized for triage to the emergency area; the generated structured data message includes timestamp, patient identifier, risk level, risk probability value, dominant risk driver (the largest absolute value of the standardized regression coefficient among the three input features of logistic regression; if it is the mean of the latent variable, it is labeled "vasoactivity-dependent risk driver"; if it is the drift rate, it is labeled "decompensated trend driver"; if it is an infection label, it is labeled "infection driver"), inflection point marker and inflection point time, and is pushed to the emergency triage decision support system display screen and mobile terminal in JSON format for triage nurses and emergency physicians to collaboratively develop intervention strategies.

[0029] Example 3, as Figure 1 As shown, based on Embodiments 1 and 2, the present invention also provides a sepsis early risk prediction system based on multimodal medical data, which is used to implement a sepsis early risk prediction method based on multimodal medical data. The system includes a data acquisition and alignment module, a metabolic instability quantification module, an infection-driven discrimination module, a risk evolution early warning module, and a visualization platform. Each module is connected in sequence to form a complete processing link from data acquisition to early warning output. The data acquisition and alignment module synchronously acquires the surface infrared thermal imaging temperature distribution matrix, the continuous ECG lead QRS complex amplitude time-series signal, and the dynamic data of serum lactate concentration and procalcitonin obtained from bedside rapid detection. Using the ECG R wave peak detection time as the global synchronization benchmark, it performs time stamp hard alignment on each data stream, applies median filtering to the thermal imaging matrix to suppress spatial salt-and-pepper noise, performs adaptive wavelet threshold denoising on the ECG signal to eliminate electromyographic interference and baseline drift, and uses cubic spline imputation for missing values ​​in biochemical data. The preprocessed medical data of each modality are arranged according to a 5-minute uniform time grid and combined to construct a standardized multimodal input matrix, providing a time-consistent multi-source input basis for downstream analysis modules. The metabolic instability quantification module calculates the thermal imaging spatial heterogeneity coefficient and the QRS amplitude variation coefficient, multiplies them, and performs quantile normalization based on the historical distribution of normal populations to construct a tissue perfusion-compensation index. Using the time series sequence of the tissue perfusion-compensation index combined with the principal components of thermal imaging and the QRS amplitude change rate as the input feature set, a gradient boosting tree model is used to extract cross-modal spatiotemporal interaction features, outputting a nonlinear representation vector of microcirculatory metabolic instability. Simultaneously, using the tissue perfusion-compensation index, time-domain indicators of heart rate variability, absolute body surface temperature, and patient trauma and surgical history identifiers as inputs, an anomaly score is obtained through an isolated forest model. When the anomaly score is higher than the 95th percentile of the training set, the initial instability probability is attenuated and corrected, with a maximum attenuation coefficient of 0.6. The output is an early metabolic instability probability corrected for interference isolation, capturing microcirculatory dysfunction signals in the early stages of infection while reducing interference from non-infectious factors. The infection-driven discrimination module embeds the corrected early metabolic instability probability as a prior weight into the XGBoost dynamic baseline comparison framework. It constructs a steady-state baseline feature model using historical data from the first effective grid point after the patient's admission to the emergency department within 24 hours. The baseline is continuously updated using a 288-grid sliding window, and the Mahalanobis distance between the current feature vector and the baseline is calculated as a deviation index. Simultaneously, it nests a hard rule decision tree based on the direction of two consecutive changes in the lactate-to-calcitonin ratio. Based on conditions such as the ratio increasing in the same direction and the absolute concentration increasing synchronously or changing in opposite directions, it outputs high-confidence and low-confidence infection-driven judgments respectively. The XGBoost deviation analysis and hard rule output are jointly inferred to generate three types of infection-driven labels and their confidence levels: high-confidence, low-confidence, and unobserved. The confidence level decays by 5% per grid point when there is no new biochemical data, until it is reset after new data arrives, providing a dynamic and quantifiable discrimination basis for the etiology of sepsis. The risk evolution early warning module uses the probability of early metabolic instability and the confidence label of infection drive as observation nodes, defines vasoactivity-dependent risk as a latent variable, constructs a discrete-time Bayesian network, and uses a particle filtering algorithm to recursively update the posterior distribution of the latent variable. The total number of particles is 1000, and the standard deviation of the transfer noise is 0.02. The first-order time difference is calculated on the updated posterior distribution of the latent variable to obtain the temporal drift rate. A two-stage confirmation mechanism is used to determine the inflection point: when the drift rate first exceeds the threshold obtained by multiplying the posterior standard deviation of the latent variable by the adaptive coefficient 2.0, an early warning flag is activated. If the drift rate continues to exceed the threshold in the next time phase, the inflection point is confirmed and recorded. The system records the time of occurrence; it integrates three types of information—latent variable posterior mean, time-series drift rate, and infection label—and maps them to an early risk probability of sepsis through logistic regression. Structured early warning messages are output according to four risk levels: low, medium, high, and very high. This captures the evolutionary inflection point before hemodynamic decompensation occurs, extending the effective intervention window. After receiving the early warning messages, the visualization platform simultaneously displays risk level color blocks, real-time risk probability curves, inflection point markers, and dominant risk drivers in a dashboard format. It also pushes differentiated intervention suggestions to the emergency triage decision support system display screen and mobile terminals, enabling triage nurses and emergency physicians to collaboratively develop intervention strategies.

[0030] Example 4: This invention also provides an early warning system for emergency sepsis based on tissue perfusion-compensation index and isolated forest correction. In this example, suspected patients admitted to the emergency department of a tertiary hospital from January 2024 to December 2025 were selected as the research subjects, and a total of 1860 valid samples were finally included. Among them, 820 patients were clinically diagnosed with sepsis (positive group), and 1040 patients were clearly diagnosed with non-infectious systemic inflammatory response syndrome (negative group). The ratio of positive to negative samples was approximately 1:1.27, which is basically consistent with the 1:1.2 control ratio in the above examples. Multimodal data acquisition was initiated immediately upon admission to the emergency department for all patients: surface infrared thermal imaging used a 320×240 pixel detection array, a frame rate of 1 frame / second, and a temperature resolution better than 0.05℃; the ECG signal sampling rate was 500Hz, with the QRS complex in lead II selected as the main analysis channel; rapid bedside detection of serum lactate concentration and procalcitonin dynamic data was performed, with blood sampling intervals not exceeding 2 hours; the peak detection time of the ECG R wave was used as the global synchronization benchmark, and time stamp hard alignment was implemented for each data stream, constructing a standardized multimodal input matrix based on a 5-minute uniform time grid. First, for each effective time grid point, the ratio of the spatial standard deviation to the mean of the surface infrared thermal imaging temperature distribution matrix is ​​calculated to obtain the thermal imaging heterogeneity coefficient. At the same time, the QRS group amplitude variation coefficient within the 5-minute sliding time window is calculated. After multiplying the two, the quantile normalization method based on the distribution of 2000 age-matched healthy subjects is used to map it to the 0 to 1 interval to construct the tissue perfusion-compensation index. Test data shows: The mean tissue perfusion-compensation index of patients in the positive group within the first 6 hours after admission to the emergency department was 0.62±0.14, while that in the negative group was 0.31±0.11, and the difference between the two groups was statistically significant (P<0.001). Further, the first 5 principal components of the tissue perfusion-compensation index time series combined with the dimensionality reduction of the thermal imaging temperature matrix and the QRS amplitude change rate were used as the input feature set. Cross-modal spatiotemporal interaction features were extracted through a gradient boosting tree model, and a 128-dimensional nonlinear quantitative characterization vector of early microcirculation metabolic instability was output. The gradient boosting tree model was set with 100 trees, a maximum depth of 5 layers, a learning rate of 0.1, a feature subsampling ratio of 0.8, and five-fold cross-validation. Secondly, using tissue perfusion-compensation index, heart rate variability time-domain index, absolute body surface temperature and patient trauma and surgical history identifiers as input dimensions, an isolated forest model is trained on a historical sample set, containing 100 isolated trees, with each subtree sampling 256 samples. The contamination rate parameter is automatically set based on the actual proportion of non-infectious stress events in the training set, and the preset abnormal score threshold is the 95th percentile of the abnormal scores in the training set. The test results show that: In the 1040 negative group patients, the isolated forest model successfully identified 763 cases with significant non-infectious stress interference (including trauma, acute heart failure, heatstroke, etc.), with an accuracy rate of 73.4%. In the 820 positive group patients, only 89 cases were misclassified as having significant non-infectious stress interference, with a misclassification rate of 10.9%. After interference isolation correction, the mean probability of early metabolic instability was 0.71±0.16 in the positive group and 0.24±0.13 in the negative group, and the discrimination between the two groups was significantly improved compared with before correction (AUC increased from 0.78 to 0.86). Third, the corrected probability of early metabolic instability is embedded as a prior weight into the XGBoost dynamic baseline comparison framework. This framework maintains a steady-state baseline feature model for each patient based on their own historical data. The baseline is continuously updated using a sliding window of 288 grid points (corresponding to 24 hours). The XGBoost model is set with 50 trees, a maximum depth of 4 layers, a learning rate of 0.05, and a binary log loss objective function. At the same time, a hard rule decision tree based on the direction of change of the ratio of serum lactate to procalcitonin is nested: if the ratio shows an increasing trend in the same direction and the absolute concentrations of both increase synchronously, the output is a high-confidence infection-driven state; if the ratio changes in opposite directions or procalcitonin remains below 0.5 ng / mL, the output is a low-confidence infection-driven state; otherwise, the output is a state to be observed. Test data shows that: Among the 820 positive patients, the proportion of high-confidence infection-driven labels was 68.3%, the proportion under observation was 22.1%, and the proportion of low-confidence labels was 9.6%. Among the 1040 negative patients, the proportion of low-confidence infection-driven labels was 71.5%, the proportion under observation was 21.2%, and the proportion of high-confidence labels was 7.3%. The Kappa consistency coefficient between the infection-driven confidence labels generated by the joint inference and the final clinical diagnosis was 0.74, indicating that the labeling system has a good ability to distinguish etiologies. Fourth, using the probability of early metabolic instability and the confidence label of infection drive as observation nodes, and defining vasoactivity-dependent risk as a latent variable, a discrete-time temporal Bayesian network containing a state transition probability matrix is ​​constructed. A sequential importance resampling particle filtering algorithm is adopted, with a total of 1000 particles and a transition noise standard deviation of 0.02. The posterior distribution of the latent variable is recursively updated every 5 minutes based on new observation data. The first-order temporal difference is calculated on the updated posterior distribution of the latent variable to obtain the temporal drift rate. The drift threshold is calculated in real time by multiplying the standard deviation of the posterior distribution of the latent variable by an adaptive coefficient of 2.0. A two-stage confirmation mechanism is adopted: when the drift rate first exceeds the threshold, an inflection point warning flag is activated; if the drift rate continues to be higher than the threshold in the next time phase, the inflection point is confirmed. The test results show that: Among the 820 positive patients, 617 cases successfully captured the critical inflection point of the evolution from compensated to decompensated within 12 hours of admission to the emergency department, with an inflection point capture rate of 75.2%. The average lead time from inflection point confirmation to the actual occurrence of hemodynamic decompensation in clinical practice was 4.7 ± 1.8 hours. Among the 1040 negative patients, only 63 cases experienced false inflection point reports, with a false report rate of 6.1%. Fifth, by integrating three types of information—the mean of the posterior distribution of the current latent variables, the temporal drift rate, and the infection-driven confidence label—a continuous value of the early risk probability of sepsis is obtained through multivariate logistic regression mapping; the risk probability interval thresholds are calibrated according to the Youden index maximization method: low risk (<0.20), medium risk (0.20-0.45), high risk (0.45-0.75), and very high risk (≥0.75). The test set validation results show: The four-level risk stratification system had an overall predictive accuracy of 87.6% for the early risk of sepsis, a sensitivity of 85.3%, a specificity of 89.4%, a positive predictive value of 86.1%, and a negative predictive value of 88.8%. Among patients in the very high-risk group, 92.7% progressed to septic shock or required vasoactive drug support within 24 hours, while only 3.2% in the low-risk group experienced disease progression. In summary, this embodiment verifies the effectiveness of the sepsis early risk prediction method based on multimodal medical data in real emergency scenarios. It constructs a tissue perfusion-compensation index by fusing infrared thermographic temperature heterogeneity distribution and ECG QRS amplitude variation, uses isolated forests to screen for non-infectious stress interference, and introduces XGBoost dynamic baseline comparison and hard-rule decision trees to generate infection-driven confidence labels. Finally, it captures the inflection point of the evolution from compensated to decompensated stages using a temporal Bayesian network, achieving dynamic quantitative assessment and four-level early warning classification of early sepsis risk. This method demonstrates good performance in terms of inflection point capture lead time, risk classification accuracy, and consistency in infection-driven discrimination, providing quantitative decision support for emergency triage.

[0031] Example 5: This invention also provides a test of a sepsis decompensation early warning system based on particle filter recursion and two-stage inflection point confirmation. This example is based on an early sepsis risk prediction system, constructing a complete processing link including a data acquisition and alignment module, a metabolic instability quantification module, an infection-driven discrimination module, a risk evolution early warning module, and a visualization platform. A prospective validation test was conducted in the emergency department of a teaching hospital. The test period was from March 2025 to September 2025, and a total of 920 suspected patients with infection in the emergency department were included, of which 398 were finally diagnosed with sepsis and 522 were patients with non-infectious systemic inflammatory response syndrome. The system hardware environment is as follows: the infrared thermal imaging equipment uses a 320×240 pixel detection array with a frame rate of 1 frame / second; the electrocardiogram monitor has a sampling rate of 500Hz; the bedside biochemical analyzer supports rapid detection of serum lactate and procalcitonin. All data is transmitted to the data processing server in real time through the hospital intranet and a standardized multimodal input matrix is ​​constructed according to a 5-minute uniform time grid. First, the metabolic instability quantification module calculates the thermal imaging spatial heterogeneity coefficient and the QRS amplitude variation coefficient, multiplies the two and performs quantile normalization based on the historical distribution of normal populations to construct the tissue perfusion-compensation index. Test data shows: The mean tissue perfusion-compensation index of patients in the sepsis group was 0.58±0.13 in the first 3 hours after admission to the emergency department, while that in the non-infection group was 0.29±0.10, with a significant difference between the two groups (P<0.001). The 128-dimensional nonlinear characterization vector of microcirculatory metabolic instability extracted by the gradient boosting tree model achieved an AUC of 0.84, a sensitivity of 81.2%, and a specificity of 83.7% in the five-fold cross-validation. Simultaneously, using tissue perfusion-compensation index, time-domain index of heart rate variability, absolute value of body surface temperature, and patient trauma and surgical history markers as inputs, anomaly scores were obtained through an isolated forest model. When the anomaly score was higher than the 95th percentile of the training set, the initial instability probability was attenuated and corrected, with a maximum attenuation coefficient of 0.6. After correction, the early metabolic instability probability was 0.69±0.15 in the sepsis group and 0.22±0.12 in the non-infection group, and the AUC increased to 0.87. Secondly, the infection-driven discrimination module embeds the corrected early metabolic instability probability as a prior weight into the XGBoost dynamic baseline comparison framework. It constructs a steady-state baseline feature model using historical data from the first effective grid point after the patient's admission to the emergency department within 24 hours. The baseline is continuously updated using a 288-grid sliding window, and the Mahalanobis distance between the current feature vector and the baseline is calculated as a deviation index. At the same time, it nests a hard rule decision tree based on the direction of change of the lactate-to-calcitonin ratio in two consecutive tests. The test results show that: Among 398 sepsis patients, 65.8% had high-confidence infection-driven labels, 24.6% were pending observation, and 9.6% had low-confidence labels. Among 522 non-infected patients, 69.4% had low-confidence infection-driven labels, 23.8% were pending observation, and 6.8% had high-confidence labels. The Kappa consistency coefficient between the jointly inferred infection-driven confidence labels and the final clinical diagnosis was 0.72. The confidence level decreased by 5% per grid point when no new biochemical data was available, and was reset upon the arrival of new data, effectively avoiding inflated label confidence levels. Third, the risk evolution early warning module uses the probability of early metabolic instability and the confidence label of infection drive as observation nodes, defines vasoactivity-dependent risk as a latent variable, and constructs a discrete-time temporal Bayesian network. The particle filtering algorithm is used to recursively update the posterior distribution of the latent variable, with a total of 1000 particles and a transfer noise standard deviation of 0.02. The first-order time difference is calculated on the updated posterior distribution of the latent variable to obtain the temporal drift rate. A two-stage confirmation mechanism is used to determine the inflection point: when the drift rate first exceeds the threshold obtained by multiplying the posterior standard deviation of the latent variable by the adaptive coefficient 2.0, the early warning flag is activated. If the threshold is continuously exceeded in the next time phase, the inflection point is confirmed and the time is recorded. Test results show that: Among 398 patients with sepsis, the critical inflection point of the evolution from compensated to decompensated phase was successfully captured in 301 cases within 12 hours of admission to the emergency department, with an inflection point capture rate of 75.6%. The average lead time from inflection point confirmation to the actual occurrence of hemodynamic decompensation in clinical practice was 5.1 ± 2.0 hours. Among 522 non-infected patients, only 34 cases had false inflection point reports, with a false report rate of 6.5%. The two-stage confirmation mechanism effectively avoided false inflection points caused by transient noise disturbances, thus improving clinical reliability. Fourth, by integrating three types of information—latent variable posterior mean, time-series drift rate, and infection label—and mapping them to the early risk probability of sepsis through logistic regression, a structured early warning message is output according to four risk levels: low, medium, high, and very high. After receiving the early warning message, the visualization platform simultaneously displays the risk level color block, real-time risk probability curve, inflection point marker, and dominant risk driving factors in the form of a dashboard, and pushes differentiated intervention suggestions to the emergency triage decision support system display screen and mobile terminal. The test set validation results show: The four-level risk stratification system demonstrated an overall predictive accuracy of 86.9% for early sepsis risk, with a sensitivity of 84.7%, a specificity of 88.6%, a positive predictive value of 85.4%, and a negative predictive value of 88.1%. In the extremely high-risk group, 91.3% of patients progressed to septic shock or required vasoactive drug support within 24 hours, while only 2.9% of patients in the low-risk group experienced disease progression. The average message push latency of the visualization platform was 1.2 ± 0.4 seconds, meeting the real-time triage needs of the emergency department. In summary, this embodiment, through prospective clinical validation testing, demonstrates the feasibility and effectiveness of a sepsis early risk prediction system based on multimodal medical data in a real emergency environment. The system achieves temporal consistency of multi-source data through a data acquisition alignment module, early capture of microcirculatory metabolic instability and correction of non-infectious stress interference through a metabolic instability quantification module, dynamic and quantifiable identification of infectious etiologies through an infection-driven discrimination module, and early capture of the inflection point from compensated to decompensated stages and four-level risk grading warning through a risk evolution early warning module. Test results show that the system meets the expected design goals in terms of inflection point capture lead time, risk grading accuracy, infection-driven discrimination consistency, and real-time response performance. It can provide structured and readily available decision support information for emergency triage and has good clinical application prospects.

[0032] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting the early risk of sepsis based on multimodal medical data, characterized in that, Includes the following steps: Step 1: Collect the patient's multimodal medical data, including infrared thermal imaging, electrocardiogram signals, serum lactate and procalcitonin dynamic data, and construct a standardized multimodal input matrix after timestamp alignment and noise filtering; Step 2: Using infrared thermal imaging data and electrocardiogram signal data from the standardized multimodal input matrix as input, and based on the coupled analysis of thermal imaging heterogeneity distribution and QRS amplitude variation, a tissue perfusion-compensation index is constructed, and a gradient boosting tree is combined to extract spatiotemporal features to quantify the early nonlinear characterization of microcirculatory metabolic instability. Step 3: The tissue perfusion-compensation index, along with relevant data from the standardized multimodal input matrix, is used as input. An isolated forest is introduced to screen out non-infectious stress interference terms, including trauma and fever, and the corrected probability of early metabolic instability is output. Step 4: The probability of early metabolic instability is embedded as a prior weight into the XGBoost dynamic baseline comparison framework, coupled with a decision tree of nested "yes / no" infectious etiology discrimination rules, and infection-driven confidence labels are generated based on the direction of change of the ratio of serum lactate to procalcitonin in two consecutive times. Step 5: Using the probability of early metabolic instability and the confidence label of infection drive as observation nodes, construct a temporal Bayesian network, dynamically update the latent variable posterior distribution of vasoactivity-dependent risk through recursive reasoning, calculate the first-order time difference of the updated latent variable posterior distribution, obtain the temporal drift rate of the latent variable posterior distribution, and capture the evolution inflection point from compensation to decompensation. Step 6: Based on the posterior distribution of latent variables and their temporal drift rate, analyze the early risk probability of sepsis and map it to four risk levels: low, medium, high, and extremely high. Output dynamic early warning classification results to guide triage and intervention.

2. The method for early risk prediction of sepsis based on multimodal medical data according to claim 1, characterized in that, In step 1, the process of constructing the standardized multimodal input matrix is ​​as follows: Simultaneously acquire the patient's surface infrared thermal imaging temperature distribution matrix and continuous electrocardiogram QRS complex amplitude timing signal, and obtain dynamic data of serum lactate concentration and procalcitonin obtained from bedside rapid detection; Using the peak detection time of the ECG R wave as the global synchronization benchmark, timestamp hard alignment is performed on each data stream, and median filtering is applied to the surface infrared thermal imaging temperature distribution matrix to suppress spatial salt-and-pepper noise. Adaptive wavelet threshold denoising is performed on the continuous ECG lead QRS complex amplitude time series signal. Missing values ​​in the dynamic data of serum lactate concentration and procalcitonin were imputed using cubic spline interpolation. The preprocessed medical data of each modality were arranged according to a uniform time grid and combined to construct a standardized multimodal input matrix.

3. The method for early risk prediction of sepsis based on multimodal medical data according to claim 1, characterized in that, In step 2, the process of constructing the tissue perfusion-compensation index and extracting spatiotemporal features using a combined gradient boosting tree is as follows: The ratio of the spatial standard deviation to the mean of the surface infrared thermal imaging temperature distribution matrix is ​​calculated to obtain the thermal imaging heterogeneity coefficient. At the same time, the coefficient of variation of the QRS amplitude of the continuous electrocardiogram QRS complex amplitude time sequence signal within the sliding time window is calculated to quantify the amplitude of stroke volume compensation driven by myocardial electrical activity. The tissue perfusion-compensation index was constructed by multiplying the thermal imaging heterogeneity coefficient with the QRS amplitude variation coefficient and performing normalization mapping to characterize the dynamic shift in the coupling relationship between peripheral microcirculation blood flow distribution and sympathetic tension. Using the tissue perfusion-compensation index time series combined with the principal components of thermal imaging and the amplitude change rate as input feature set, cross-modal spatiotemporal interaction features are extracted through a gradient boosting tree model, and an early nonlinear quantitative characterization vector of microcirculatory metabolic instability is output.

4. The method for early risk prediction of sepsis based on multimodal medical data according to claim 1, characterized in that, In step 3, the process of outputting the corrected probability of early metabolic instability is as follows: Using tissue perfusion-compensation index, time-domain index of heart rate variability, absolute value of body surface temperature, and patient trauma and surgical history markers as input dimensions, an isolated forest model was trained on a historical sample set to construct an isolation discriminant hyperplane for non-infectious stress events; The current patient's input dimension features are input into the trained isolated forest model to obtain anomaly scores and determine whether they belong to the category of non-infectious stress interference items. Based on the preset anomaly scoring threshold, the early nonlinear quantization representation vector is weighted and mapped and inversely probabilistically corrected. After removing the mixed offset, it is transformed into an early metabolic instability probability value after interference isolation correction.

5. The method for early risk prediction of sepsis based on multimodal medical data according to claim 4, characterized in that, When the abnormal score of the isolated forest model is higher than the preset threshold, it is determined that the current patient has a significant non-infectious stress interference item, and the confidence weight of the probability of early metabolic instability is reduced; when the abnormal score is lower than the preset threshold, the original probability output is retained.

6. The method for early risk prediction of sepsis based on multimodal medical data according to claim 1, characterized in that, In step 4, the process of generating infection-driven credibility tags is as follows: The corrected probability of early metabolic instability is used as a priori weight and embedded into the XGBoost dynamic baseline contrast framework, which maintains a steady-state baseline feature model for each patient based on their own historical data. In the XGBoost decision path, a hard rule decision tree containing "yes / no" infectious etiology is coupled and nested. Based on the change direction and consistency of the ratio of serum lactate to procalcitonin obtained from two consecutive tests, it is determined whether the current metabolic abnormality tends to be infection-driven. Based on the results of the ratio change direction determination and the XGBoost deviation analysis, dynamically updated infection drive confidence labels are generated to distinguish between three types of states: high confidence infection drive, low confidence infection drive, and unobserved state.

7. The method for early risk prediction of sepsis based on multimodal medical data according to claim 6, characterized in that, If the ratio of lactate to procalcitonin shows an increasing trend in the same direction for two consecutive times and the absolute concentrations of both increase synchronously, then the decision tree output is a highly reliable infection driver. If the ratio changes in opposite directions or procalcitonin remains below the preset threshold, the decision tree output is low-confidence infection-driven, incorporating biochemical dynamics into the infection etiology discrimination logic.

8. The method for early risk prediction of sepsis based on multimodal medical data according to claim 1, characterized in that, In step 5, the process of capturing the inflection point of the evolution from the compensatory period to decompensation is as follows: Using the probability of early metabolic instability and the confidence label of infection drive as two observation nodes, and defining the risk of vasoactivity dependence as a latent variable, a discrete-time temporal Bayesian network structure containing a state transition probability matrix is ​​constructed. The sequential importance resampling particle filter algorithm is adopted. At each time interval, the latent variable posterior distribution is recursively updated based on the newly acquired observation node data to estimate the probability density function of the current vasculogenic risk. The first-order time difference of the updated latent variable posterior distribution is calculated to obtain the temporal drift rate of the latent variable state. When the temporal drift rate exceeds the drift threshold for two consecutive time phases, it is marked as the critical inflection point moment of the evolution from the compensatory period to the decompensatory period. The drift threshold is calculated in real time by multiplying the standard deviation of the posterior distribution of the latent variable by a preset adaptive coefficient. When the time drift rate first exceeds the drift threshold, an inflection point warning flag is activated. If the time drift rate of the next time phase continues to be higher than the drift threshold, the inflection point is confirmed.

9. A method for early risk prediction of sepsis based on multimodal medical data according to claim 8, characterized in that, In step 6, the process of outputting dynamic early warning and grading results to guide triage and intervention is as follows: By integrating three types of information—the mean of the posterior distribution of the current latent variables, the time-series drift rate, and the infection-driven confidence label—a continuous value estimate of the early risk probability of sepsis is obtained through multivariate logistic regression mapping. Based on the pre-defined risk probability interval thresholds, the early risk probability of sepsis is mapped to four risk levels: low, medium, high, and very high. Each level corresponds to a different frequency of emergency monitoring and triggering conditions for intervention preparation. The system generates dynamic early warning classification results that include risk level, dominant risk drivers, and inflection point indicators. These results are then output to the emergency triage decision support system in the form of structured data messages for the formulation of triage intervention strategies.

10. A sepsis early risk prediction system based on multimodal medical data, used to implement the sepsis early risk prediction method based on multimodal medical data as described in any one of claims 1-9, characterized in that, The visualization platform includes the following modules for communication connections: The data acquisition and alignment module is used to simultaneously acquire multimodal medical data, including surface infrared thermal imaging, electrocardiogram signals, and bedside biochemical test data. It performs hard timestamp alignment with the peak time of the electrocardiogram R wave as the global benchmark. After median filtering, wavelet denoising, and spline interpolation, a standardized multimodal input matrix is ​​constructed according to a 5-minute uniform grid. The metabolic instability quantification module is used to calculate the thermal imaging spatial heterogeneity coefficient and QRS amplitude variation coefficient and fuse them into a tissue perfusion-compensation index. The spatiotemporal interaction features are extracted through gradient boosting tree and the nonlinear characterization vector of microcirculation metabolic instability is output. At the same time, the isolated forest is used to screen out non-infectious stress interference and correct the output early metabolic instability probability, so as to achieve the initial separation of infectious and non-infectious factors. The infection-driven discrimination module is used to embed the corrected early metabolic instability probability as a prior weight into the XGBoost dynamic baseline comparison framework, calculate the feature deviation, and nest a hard rule decision tree based on the change direction of the lactate to procalcitonin ratio to jointly infer and generate three types of infection-driven labels and their confidence levels: high confidence, low confidence, and unobserved. The risk evolution early warning module is used to construct a temporal Bayesian network with instability probability and infection label as observation nodes, and recursively update the vasoactivity-dependent risk latent variable through particle filtering. It captures the decompensated evolution inflection point by combining the latent variable drift rate and the two-stage inflection point confirmation mechanism. It integrates three types of information and maps them to the early risk probability of sepsis through logistic regression. It outputs structured early warning messages to the triage system according to four risk levels.

Images