Severe patient sepsis early warning method and system based on AI

Abstract

The invention relates to the technical field of intelligent medical treatment, and discloses an AI-based severe patient sepsis early warning method and system. According to the method, a multi-organ interaction mechanism is deeply analyzed by dynamically constructing an organ-level causal network, and early-stage accurate early warning of sepsis is realized: high-frequency physiological waveforms, asynchronous laboratory indexes and treatment intervention data are fused, and the capture ability of microcirculation failure and autonomic nerve decline is improved; the early warning threshold value is dynamically adjusted based on the treatment coverage degree, and the delay and false alarm defects of a fixed threshold value mechanism are effectively overcome; a pathogen targeted therapy map and an organ function support scheme are automatically generated by a graded triggered clinical action chain, and the clinical decision response time is shortened; according to the method, the risk of missed diagnosis is reduced while the early warning sensitivity is improved, and an earlier and more reliable intervention window is provided for critical patients.

Description

Technical Field

[0001] The present invention relates to the field of smart medical technology, and specifically to an AI-based early warning method and system for sepsis in critically ill patients. Background Art

[0002] Current clinical approaches primarily rely on the modified Early Warning Score (MEWS), Sequential Organ Failure Assessment (SOFA), and rapid SOFA (qSOFA) for sepsis screening, combined with static biomarker testing (such as procalcitonin and lactate) for auxiliary diagnosis. These methods use fixed thresholds to determine organ failure status, manually input vital signs and laboratory data into electronic medical record systems, and perform periodic assessments every 4-6 hours. A few AI models attempt to use LSTM networks to analyze time-series vital sign data, but these models only support early warning based on a single data source and lack the ability to deeply analyze the dynamic interactions between multiple organs.

[0003] Existing technologies have significant problems: it is difficult to effectively integrate multi-source heterogeneous information such as high-frequency physiological waveforms, asynchronous laboratory data, and unstructured text in the ICU environment; it is impossible to quantify the real-time compensatory dependencies between organs, resulting in insufficient sensitivity to early microcirculation and autonomic nervous system dysfunction; at the same time, the fixed risk threshold mechanism does not consider the dynamic impact of therapeutic interventions, which can easily lead to delayed warnings or false alarms. Therefore, an AI-based early warning method and system for sepsis in critically ill patients is proposed. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an AI-based early warning method and system for sepsis in critically ill patients to solve the background technical problems.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] First, an AI-based early warning method for sepsis in critically ill patients, including:

[0007] Step S1: Real-time collection of multi-source asynchronous time series data of ICU patients:

[0008] High-frequency physiological waveforms: original bioelectrical signals and hemodynamic waveforms with a sampling rate ≥ 100 Hz;

[0009] Laboratory indicator series: inflammatory markers, metabolites and immune cell activity indicators;

[0010] Chronological recording of therapeutic interventions: dynamics of vasoactive drug dosage and respiratory support parameter interaction matrix;

[0011] Unstructured text: clinical notes containing descriptions of occult tissue hypoperfusion;

[0012] Step S2, based on the data flow of step S1, dynamically construct an organ-level causal network:

[0013] Taking the infection trigger node as the root, the organ interaction edge weights are dynamically updated through time-varying Granger causality analysis. The organ interaction edge weights represent the compensatory dependence strength between organs in real time.

[0014] Embedded in the organ compensation capacity quantification layer, the microcirculation oscillation attenuation index (MOAI) and blood pressure variation entropy (BVE) are calculated;

[0015] Step S3: Input the organ causal graph output from step S2 and the data stream from step S1 into the multimodal spatiotemporal fusion network:

[0016] High-frequency path: The input is the high-frequency physiological waveform of step S1. The waveform segment features are extracted through a one-dimensional convolutional neural network (1D-CNN), and the temporal convolutional network (TCN) is used to capture long-range dependencies.

[0017] Causal path: Input the organ causal graph from step S2 and use the dynamic graph attention network (DyGAT) to update the organ node state vector;

[0018] Cross-modal alignment module: maps the laboratory indicators in step S1 to the physiological time space through dynamic time warping;

[0019] Step S4: Generate an interpretable warning based on the fusion network output of step S3

[0020] Generate organ dysfunction assessment indicators based on the organ node state vector output by DyGAT;

[0021] Aggregate the evaluation indicators to generate a comprehensive sepsis risk value R sepsis , when R sepsis When >θ(t), a hierarchical clinical action chain is triggered;

[0022] Where θ(t) is the adaptive threshold, which is dynamically calculated based on the patient's baseline status and treatment coverage.

[0023] Preferably, the high-frequency physiological waveform in step S1 specifically includes:

[0024] The original electrocardiogram (ECG) signal that meets the micro-oscillation signal analysis accuracy requirements is used to extract the T-wave micro-oscillation energy;

[0025] The second derivative signal of the photoplethysmography (PPG) that meets the requirements for calculating the perfusion oscillation period is used to calculate the perfusion oscillation period;

[0026] An invasive arterial blood pressure waveform that meets the requirements of blood pressure variation spectrum analysis is used to generate a blood pressure variation spectrum.

[0027] Preferably, in step S2:

[0028] The time-varying Granger causality analysis adopts a sliding window recursive algorithm, and the window length is adaptively adjusted according to the organ compensation state;

[0029] The microcirculation oscillation attenuation index (MOAI) is obtained by integrating the spectrum of the microcirculation perfusion signal and is used to quantify the degree of decline in microvascular regulatory function;

[0030] The calculated blood pressure variation entropy (BVE) is obtained by performing multi-scale entropy analysis on the invasive arterial blood pressure waveform and is used to quantify the state of autonomic nervous system regulation function.

[0031] Preferably, in step S4:

[0032] The organ dysfunction assessment indicators include circulatory system failure index, respiratory system failure index and renal perfusion index;

[0033] The comprehensive sepsis risk value R sepsis The calculation method is:

[0034]

[0035] in is the organ set, β o is the organ weight coefficient, which is determined by the patient's underlying disease, I o is the organ dysfunction index;

[0036] The adjustment formula of the adaptive threshold θ(t) is:

[0037] θ(t)=θ0·φ(C t )

[0038] Where θ0 is the basic risk threshold, C t is the real-time treatment coverage score, φ(·) is the score based on the treatment coverage C t Threshold correction function.

[0039] Preferably, the treatment coverage C t Calculate by the following steps:

[0040] Construct pathogen-antibiotic sensitivity probability matrix M;

[0041] Generate antibiotic coverage vector a based on patient medication records;

[0042] Combined with the suspected pathogen distribution vector p, the coverage score is calculated: C t =p T Ma.

[0043] Preferably, the hierarchical clinical action chain includes:

[0044] First-level action chain: When the comprehensive sepsis risk value R sepsis When >θ(t), a pathogen coverage risk map and fluid resuscitation simulation plan are generated;

[0045] Secondary action chain: When organ dysfunction assessment index I o >η o (η o When the organ index threshold is reached, targeted organ support strategies are pushed.

[0046] Preferably, the pathogen coverage risk map is generated by:

[0047] integrating microbial colonization data with the prevalence of resistant bacteria in wards;

[0048] Calculate the probability P of uncovered pathogens uncover :

[0049]

[0050] Among them S i is the sensitivity of pathogen type i to the current antibiotic regimen; T i is the coverage effectiveness attenuation coefficient based on the duration of medication, and i is the pathogen category index.

[0051] In a second aspect, an AI-based early warning system for sepsis in critically ill patients is provided, for implementing any of the AI-based early warning methods for sepsis in critically ill patients described in the first aspect, comprising:

[0052] Edge data acquisition layer: real-time collection of high-frequency physiological waveforms and treatment intervention data;

[0053] Causal reasoning engine: Dynamically constructs organ-level causal networks and calculates microcirculatory oscillation attenuation index (MOAI) and blood pressure variability entropy (BVE);

[0054] Multimodal fusion module: integrates high-frequency path processor, causal path processor and cross-modal aligner;

[0055] Adaptive early warning center: Generates organ dysfunction assessment indicators and graded clinical action chains, and dynamically adjusts the threshold θ(t).

[0056] Preferably, the edge data acquisition layer includes a signal preprocessing unit for extracting the T wave oscillation energy of the ECG and the perfusion oscillation period of the PPG.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention achieves accurate early warning of sepsis by dynamically constructing an organ-level causal network to deeply analyze the multi-organ interaction mechanism: integrating high-frequency physiological waveforms, asynchronous laboratory indicators and treatment intervention data to improve the ability to capture microcirculatory failure and autonomic nervous system decline; dynamically adjusting the warning threshold based on treatment coverage to effectively overcome the delay and false alarm defects of the fixed threshold mechanism; the hierarchically triggered clinical action chain automatically generates pathogen-targeted treatment maps and organ function support plans, shortening clinical decision-making response time; this method improves warning sensitivity while reducing the risk of missed diagnosis, providing critically ill patients with an earlier and more reliable intervention window.

[0059] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a closed-loop architecture diagram of the four-step sepsis early warning system of the present invention;

[0061] Figure 2 This is a flow chart of high-frequency physiological signal processing according to the present invention;

[0062] Figure 3 Dynamically constructing a graph for the organ causal network of the present invention;

[0063] Figure 4 This is a diagram of the multimodal fusion network structure of the present invention;

[0064] Figure 5 A flow chart for adaptive early warning generation of the present invention;

[0065] Figure 6 This is a diagram of the edge-cloud deployment architecture of the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this technical field without making any creative efforts shall fall within the scope of protection of the present invention.

[0067] See also Figure 1-6The present invention provides an AI-based early warning method and system for sepsis in critically ill patients. This system implements closed-loop monitoring through four core steps: Step S1 (multi-source data acquisition) → Step S2 (organ causal network construction) → Step S3 (multimodal fusion) → Step S4 (adaptive warning generation), forming a complete technical chain from microcirculatory signal capture to graded clinical intervention. The specific implementation method is detailed below:

[0068] Step S1: Multi-source asynchronous time series data acquisition

[0069] This step forms the data foundation of the entire sepsis early warning system, aiming to build a complete, real-time portrait of the patient's condition. The core challenge lies in efficiently and accurately capturing the subtle physiological signals that reflect early organ dysfunction (especially microcirculatory failure) and resolving the timescale asynchrony between different data types (high-frequency waveforms, low-frequency discrete indicators, and intervention records). High-quality data is a prerequisite for subsequent organ causal reasoning and accurate early warning. The specific implementation process is as follows:

[0070] 1. High-frequency physiological waveform acquisition

[0071] In the early stages of microcirculatory failure, microcirculatory failure often manifests as subtle changes in millivolt-level bioelectrical or hemodynamic signals, which are easily overwhelmed by noise. Therefore, it is necessary to capture these signals using a sampling frequency far higher than that of conventional clinical monitoring, and design a specific signal processing process to extract key features.

[0072] ECG signal processing flow:

[0073] Acquisition: Acquire raw electrocardiogram (ECG) signals at a sampling rate of at least 200 Hz. This sampling rate should be >100 Hz to ensure that T-wave micro-oscillations are effectively captured.

[0074] Feature extraction: Calculate T wave micro-oscillation energy E TW :

[0075]

[0076] Among them E TW is the T wave micro-oscillation energy (unit: mV 2 ), the lower the value, the more severe the microcirculatory disorder; WT(k) is the kth coefficient after Daubechies 4 wavelet transform, which is used to extract the millivolt oscillation signal; N is the number of single T-band sampling points (determined by heart rate); k is the serial index of the wavelet coefficient (dimensionless, integer);

[0077] T wave microoscillations (TWA) are small alternating changes in T wave morphology and amplitude on the electrocardiogram. They are a sign of unstable myocardial repolarization and are closely related to the risk of myocardial ischemia, microcirculation disorders and malignant arrhythmias. TWBy quantifying the energy of T-wave microoscillations, it reflects the stability of myocardial electrical activity. A decrease in energy indicates insufficient myocardial microcirculatory perfusion or electrophysiological disturbances.

[0078] T wave micro-oscillation energy E TW A decrease in the value is a sensitive indicator of myocardial microcirculation disorder and electrical instability, and can serve as an early warning signal of myocardial involvement in sepsis or septic shock.

[0079] Photoplethysmography (PPG) processing flow:

[0080] Perfusion oscillation reflects microvascular regulatory function, and its effective frequency range has been clinically verified to be 0.04-0.15 Hz (corresponding to a period of 6-25 seconds);

[0081] 1) Acquisition: Acquire a 125 Hz photoplethysmography (PPG) signal. To enhance sensitivity to low-frequency features reflecting microvascular regulation, the second-order derivative of the PPG signal (PPG″(t)) is analyzed.

[0082] 2) Feature extraction: Calculating the perfusion oscillation period

[0083]

[0084] Where T perf The normal value is 6-8 seconds for the perfusion oscillation period (unit: seconds). A value > 15 seconds indicates microcirculatory failure. c is the energy peak frequency in the 0.04-0.15 Hz band; PPG″(t) is the second-order derivative of PPG (enhanced low-frequency characteristics); is the Fourier transform operator; f is the frequency variable.

[0085] The microvascular system has spontaneous, rhythmic contraction and relaxation activity, known as blood flow motion. This activity manifests as oscillations in the low-frequency range of the PPG signal (0.01-0.5Hz). Different physiological mechanisms dominate the oscillations in different frequency bands. perf By identifying the inverse of the dominant oscillation frequency, we quantify the characteristic timescale of microvascular regulatory activity. A prolonged period indicates a decline in microvascular regulatory function (especially endothelial and neurogenic regulation), a key manifestation of microcirculatory dysfunction in sepsis.

[0086] T perf Prolongation indicates the decline of microvascular regulatory function. The decline of vascular regulatory function in the early stage of sepsis is manifested as f c The shift toward low frequency has early warning value for sepsis-related microcirculatory disorders.

[0087] Application examples:

[0088] Patient background: A patient was admitted to the ICU for severe pneumonia on the third day of treatment. His vital signs were stable, but clinical suspicion was that he had occult tissue hypoperfusion.

[0089] ECG processing:

[0090] The patient's current heart rate is measured to be 75 beats per minute (bpm).

[0091] Calculate the number of T-band sampling points N: With a sampling rate of 200 Hz and a heart rate of 75 bpm (1.25 beats per second), the duration of a single heartbeat is 1 / 1.25 = 0.8 seconds. The T-band accounts for approximately 20% of the heartbeat cycle, so its duration is approximately 0.16 seconds. The number of sampling points N = 200 Hz * 0.16 seconds = 32 points.

[0092] Perform db4 wavelet transform on the extracted T band and calculate the coefficient modulus square sum:

[0093] Comparison and conclusion: Normal reference value is usually E TW >0.3mV 2 , patient E TW =0.18mV 2 If the value is lower than normal, it indicates the presence of myocardial microcirculation disorders and electrical instability, and one should be alert to myocardial suppression caused by sepsis.

[0094] PPG treatment:

[0095] The second-order derivative PPG″(t) was calculated for the acquired 125 Hz PPG signal.

[0096] Perform Fourier transform on PPG″(t) and analyze the [0.04, 0.15] Hz frequency band.

[0097] Identify the frequency f corresponding to the maximum spectrum energy peak in the frequency band c =0.05Hz.

[0098] Calculate the perfusion oscillation period T perf =1 / f c =1 / 0.05=20 seconds.

[0099] Comparison and conclusion: The normal perfusion oscillation period ranges from about 6 to 8 seconds. perf = 20 seconds extension, indicating a serious decline in microvascular endothelial-derived regulatory function, which is a clear signal of microcirculatory failure.

[0100] Comprehensive conclusion: Combined with ECG E TW Significantly reduced T and PPG perfThe patient's life expectancy was significantly prolonged. When the vital signs were still in the "stable" stage, he showed clear and synchronous myocardial and systemic microcirculatory dysfunction, which strongly suggested that the pathophysiological process of sepsis was ongoing and that a more in-depth evaluation and early warning process needed to be initiated immediately.

[0101] 2. Laboratory indicator sequence processing

[0102] Laboratory indicators (such as IL-6 and lactate) are key indicators of inflammatory and metabolic status, but their sampling intervals are long (e.g., every 2, 4, or even longer), making them significantly out of sync with high-frequency physiological signals and real-time intervention recordings. This step aims to convert these discrete, asynchronous laboratory data into continuous time series aligned with physiological signals, providing a unified time base for subsequent multimodal fusion.

[0103] Implementation process:

[0104] 1) Data Acquisition: Obtain discrete laboratory test results and their corresponding timestamps from the hospital's laboratory information system (LIS). For example, interleukin-6 (IL-6) is tested every 2 hours, and lactate is tested every 4 hours.

[0105] 2) Time Series Alignment - Cubic Spline Interpolation: This algorithm uses the cubic spline interpolation algorithm. This algorithm uses known discrete data points (timestamps, test values) to construct a smooth cubic polynomial curve passing through all data points. Based on this curve, an indicator estimate can be calculated at any desired time point (for example, every minute).

[0106] 3) Generate continuous sequence: Generate a continuous laboratory index time series L with a 1-minute time interval through interpolation lab (t).

[0107] It solves the problem of time asynchrony between laboratory indicators and high-frequency physiological waveforms and treatment records, allowing key indicators such as inflammation and metabolism to be correlated, analyzed and fused with physiological signals reflecting organ function on the same time scale (1-minute resolution).

[0108] 3. Integration of treatment intervention records

[0109] Therapeutic interventions (e.g., vasoactive drugs, ventilator parameters) directly affect the patient's physiological state and organ function. This step aims to structure and quantify these interventions so that their impact on the patient's state can be accurately assessed in subsequent causal analysis and early warning decision-making.

[0110] Implementation process:

[0111] Vasoactive drugs:

[0112] Data structuring: Record the dose (eg, μg / kg / min) of each vasoactive drug (eg, norepinephrine, epinephrine, vasopressin) administered and its precise start, adjustment, and stop times.

[0113] Constructing the dose-time matrix: Create a matrix D vaso The rows of the matrix represent different vasoactive drugs, and the columns represent time (e.g., per minute). The value of the matrix element represents the equivalent dose of the corresponding drug at the corresponding time point (which may need to be standardized, such as norepinephrine equivalents).

[0114] Respiratory parameters:

[0115] Key parameter extraction: Extract core parameters related to oxygenation and ventilation from ventilator records, especially inspired oxygen concentration (FiO2) and positive end-expiratory pressure (PEEP).

[0116] Generate interaction matrix: Create a vector V resp (t)=[FiO2(t),PEEP(t)] T , where FiO2(t) and PEEP(t) represent the set or measured values ​​at time t. T Represents the transpose (usually treated as a column vector).

[0117] Translating complex therapeutic interventions into quantitative, time-synchronized mathematical representations (D vaso (t),V resp This enables the system to consider the perturbations of treatment on organ interactions in organ causal network analysis and accurately assess the coverage and effect of current treatment on patient status (such as oxygenation and circulatory support) in comprehensive risk calculation and threshold adjustment, avoiding misjudging treatment responses as spontaneous improvement of the condition.

[0118] The core of this step lies in addressing the challenge of capturing weak microcirculatory signals through high-frequency sampling and analysis of specific frequency bands. Furthermore, through time series interpolation and intervention quantification matrices, asynchronous laboratory data and treatment records are effectively integrated. Clinical validation has demonstrated a reduced error rate in T-wave microoscillation energy detection compared to traditional methods, providing a high-quality, multi-dimensional, time-aligned input data foundation for subsequent organ causal inference.

[0119] Step S2: Dynamic construction of organ-level causal network

[0120] This step is the core intelligent engine of the early warning system. Traditional scoring systems (such as SOFA) mainly focus on static or cumulative damage to organ function, while the essence of sepsis is the complex and dynamic interaction and compensation / decompensation process between organs caused by systemic inflammatory response. This step aims to quantify the intensity and direction of causal influence between organs in real time (that is, "which organ is affecting which organ, and how big the influence is"), and embed specially designed indicators to quantify the organ's own compensatory reserve capacity (MOAI, BVE). This dynamic causal network can capture the critical points of organ interaction imbalance and compensatory failure earlier and more accurately, and is the key to early warning. The specific implementation process is as follows:

[0121] 1. Time-varying Granger causality analysis process

[0122] Based on the multi-source synchronized data streams provided in step S1, the dynamic causal relationships between different organ systems (nodes) are quantitatively analyzed. Specifically, does the historical data for organ i significantly improve the prediction of the future state of organ j? If so, organ i is considered to have "caused" the changes in organ j in a Granger causal sense, and the strength of this influence is quantified by the degree of prediction improvement.

[0123] Implementation steps:

[0124] 1) Organ node definition: Initialize the organ node set Each organ node is represented by its associated feature vector from step S1 (e.g., a loop node may contain E TW 、T perf and mean arterial pressure MAP; respiratory node includes FiO2, PEEP and oxygenation index; renal node includes urine volume and creatinine interpolation sequence);

[0125] 2) Serum IL-6 concentration was used as the infection trigger root node;

[0126] 3) Dynamic weight calculation: Using a sliding window recursive algorithm, for any two different organ nodes i and j, at time point t, calculate the causal influence weight w of organ i on organ j ij (t):

[0127]

[0128] where w ij (t) is the causal influence weight of organ i on organ j at time t, is the predicted value based on the historical data of organ i, X j (k) is the actual observed value of organ j, T w is the sliding window length (initial 120 minutes);

[0129] w ij(t) quantitatively represents the contribution (or interference) of organ i's historical information to the prediction of organ j's current state within the most recent time window. A larger value indicates a stronger causal influence of organ i on organ j.

[0130] 4) Adaptive window mechanism: initial window length T w = 120 minutes. The system continuously monitors the organ compensation capacity index (see below MOAI). When the microcirculation oscillation attenuation index MOAI>0.3 (indicating that the microvascular regulatory function begins to decline significantly and the organ enters a vulnerable state), the system automatically increases the window length T w Shortened to 60 minutes. At stages when organ compensatory capacity declines and the condition may change rapidly, shortening the analysis window can more quickly capture the latest dynamic changes in the causal relationships between organs and increase the system's sensitivity to impending decompensation events.

[0131] 2. Quantification of organ compensatory capacity

[0132] Analyzing inter-organ interactions alone is insufficient; direct assessment of the functional reserve and compensatory potential of individual organs is also necessary. This workflow calculates two key metrics: the microcirculatory oscillation attenuation index (MOAI), which quantifies microvascular autoregulatory capacity; and blood pressure variability entropy (BVE), which quantifies the autonomic nervous system's ability to regulate cardiovascular function. A decline in these two metrics is a key indicator of the progression of sepsis to shock.

[0133] 1) Calculation of MOAI (Microcirculatory Oscillation Attenuation Index)

[0134] Purpose of the steps: The decline of microvascular regulatory function is manifested as the attenuation of low-frequency oscillation energy. The degree of decline of microvascular regulatory function is quantified by frequency band energy ratio:

[0135]

[0136] Energy effective Energy in the effective compensatory frequency band of 0.04-0.15 Hz (reflecting physiological oscillations); total It is the energy of the entire frequency band from 0.01 to 0.5 Hz (excluding motion noise > 0.5 Hz).

[0137] The healthy microvasculature exhibits active blood flow oscillations (Flowmotion) within a specific frequency band (0.04-0.15 Hz, primarily reflecting endothelial and neural regulation). In sepsis / shock, this intrinsic oscillatory activity attenuates or becomes disrupted. The MOAI quantifies this attenuation by calculating the ratio of the energy in the effective compensatory oscillation frequency band (0.04-0.15 Hz) to the energy in the full frequency band (0.01-0.5 Hz). An elevated MOAI value indicates a relative decrease in effective compensatory oscillation energy, indicating a decline in microvascular regulatory function.

[0138] MOAI is a key indicator for assessing microcirculatory compensatory reserve. A MOAI > 0.35 is set as the threshold, indicating a significant decline in microvascular regulatory function and an important signal for the progression of sepsis and increased risk of shock.

[0139] 2) BVE calculation (blood pressure variation entropy)

[0140] Purpose of the steps: To evaluate the ability of the autonomic nervous system to regulate vascular tone;

[0141] The autonomic nervous system function status is assessed by the complexity of blood pressure fluctuations:

[0142]

[0143] Where τ is the scale factor (1, 2, 3 corresponds to short / medium / long range regulation), A τ (r) is the number of matching templates under the τ scale, B τ (r) is the number of basic templates, r is the similarity tolerance (usually 0.2 times the standard deviation), BVE is the blood pressure variation entropy, and BVE is the blood pressure variation entropy (dimensionless, scalar);

[0144] A healthy autonomic nervous system generates complex, multiscale blood pressure fluctuations. This complexity can be quantified using multiscale entropy (MSE). In sepsis, autonomic function is impaired, and blood pressure variability decreases and tends to be simpler (lower entropy). Blood pressure variation entropy (BVE), based on multiscale entropy analysis, quantifies the complexity and regularity of the arterial blood pressure (ABP) waveform at different time scales. Reduced BVE values ​​indicate failure of autonomic regulation.

[0145] BVE is a key indicator for assessing cardiovascular autonomic compensatory reserve. A BVE threshold of <0.8 indicates a significant decline in autonomic nervous system function, which is associated with sepsis-related sympathetic-adrenal system activation / failure and increases the risk of hemodynamic collapse.

[0146] 3) Application Examples

[0147] Patient background: A patient developed early signs of septic shock (such as mild lactate elevation and the need for small doses of vasoactive drugs to maintain blood pressure) after surgery due to intra-abdominal infection.

[0148] Causal Network Update:

[0149] The system calculates MOAI in real time = 0.41 (> 0.3 adaptive threshold):

[0150] Triggering the adaptive mechanism: Sliding the causal analysis window length T w Reduced from an initial 120 minutes to 60 minutes to increase sensitivity to rapid changes;

[0151] Recalculating inter-organ causal weights within a shorter 60-minute window;

[0152] Find the causal weight w of kidney (i) → loop (j) ij It rose to 0.48 from 0.22 calculated in the previous 120-minute window.

[0153] After the window is shortened, the system can more sensitively capture the recent and strong negative impact (increased weight) of changes in renal function (which may manifest as decreased urine volume and an upward trend in creatinine) on the circulatory system (blood pressure, microcirculation), suggesting the possible existence of "kidney-heart interaction imbalance" (such as acute kidney injury leading to volume overload and accumulation of metabolic products, thereby aggravating cardiac function suppression and vascular paralysis).

[0154] Compensation index calculation:

[0155] MOAI = 0.41 (>0.35 threshold), confirming a significant decline in microvascular regulatory function;

[0156] BVE=0.73 (<0.8 threshold), confirming that cardiovascular autonomic nervous system regulation function declined significantly.

[0157] This patient not only showed a synchronous decline in microcirculatory (MOAI) and autonomic nervous system (BVE) functions (self-compensation deficiency), but the dynamic causal network also revealed that the negative impact of the kidney on the circulatory system has recently increased significantly (inter-organ interaction imbalance). This pattern of "self-compensation failure + negative enhancement of key organ interactions" is a high-risk warning signal that septic shock is about to enter the decompensated stage. The system should significantly increase the comprehensive sepsis risk value R accordingly. sepsis , and may trigger an early warning.

[0158] The core innovation of this step is to quantify organ interactions and organ compensation capabilities in real time for the first time. ij (t) and dual compensation index (MOAI / BVE), the system can provide insights into the imbalanced relationship between organs and the depletion of their own reserves earlier than traditional methods.

[0159] Step S3: Multimodal spatiotemporal fusion network

[0160] Step S1 provides multi-source heterogeneous data, and step S2 outputs the organ causal network and compensatory indicators. The core challenge of this step is how to effectively fuse these information with different temporal and spatial characteristics (temporal local / long-range characteristics of high-frequency waveforms, topological relationships of organ causal networks, and delayed asynchrony of laboratory indicators) to form a comprehensive and accurate judgment of organ dysfunction. This network adopts a three-way parallel processing architecture to process different types of data respectively, and solves the problem of time inconsistency through cross-modal alignment, and finally outputs a fused feature vector for use in early warning decision-making. The specific architecture is as follows:

[0161] 1. High-frequency path processing flow

[0162] From the high-frequency physiological waveform segment provided in step S1 (e.g., 10-second ECG, PPG, ABP), local transient features (e.g., ST-segment depression, abnormal pulse waveform) and long-range dependent features within the segment (e.g., heart rate variability trend, respiration-pulse coupling) are simultaneously extracted;

[0163] Input: high-frequency physiological waveform segments (ECG, PPG, ABP) of fixed length (e.g., 10 seconds) from step S1;

[0164] Feature extraction:

[0165] 1D-CNN (one-dimensional convolutional neural network):

[0166] Purpose: To capture the local morphological features of a waveform (such as the rising edge, falling edge, amplitude, and duration of a specific waveform);

[0167] Implementation: Using a smaller convolution kernel (size = 5), a kernel size of 5 covers 0.025 seconds (25ms) at a 200Hz sampling rate and 0.04 seconds (40ms) at a 125Hz sampling rate, which can effectively capture local details such as individual heart waves, T waves, and pulse wave characteristic peaks.

[0168] Output: local feature vector F local .

[0169] TCN (Temporal Convolutional Network):

[0170] Purpose: Capturing long-range temporal dependencies within waveform segments (10 seconds). Traditional RNNs / LSTMs can be limited by vanishing gradients and training speed. TCNs utilize dilated convolutions to expand the receptive field and achieve greater efficiency.

[0171] Implementation: Use the empty factor sequence [1,2,4]. This means:

[0172] First convolution layer: dilation factor = 1, receptive field covering 5 points (~25ms@200Hz);

[0173] Second convolution layer: dilation factor = 2, receptive field coverage of approximately 1 + 2 * (5-1) = 9 points (~45ms);

[0174] The third convolution layer has a dilation factor of 4 and a receptive field covering approximately 1 + 4 * (9 - 1) = 33 points (~165ms).

[0175] By superimposing multiple layers, the entire 10-second segment (2000 points @ 200Hz) can be covered.

[0176] Output: long-range dependency feature vector F long .

[0177] Output fusion: The local features F extracted by 1D-CNN are local And the long-range feature F extracted by TCN long Concatenate together to form the final output feature vector of the high-frequency path: F HF =F local ; F long ];

[0178] F HF It comprehensively represents the key information of the original physiological waveform in terms of microscopic morphology and macroscopic trends, and is the basis for judging the immediate organ function status (such as myocardial ischemia and hypoventilation).

[0179] 2. Causal Path Processing

[0180] The organ-level causal network (graph structure) dynamically generated in step S2 is converted into a feature representation that can be learned by the neural network. The key is to use the graph structure information (node ​​organs, edge weights w ij (t)) to update and enrich the state vector of each organ node.

[0181] Graph structure construction: organs (such as circulation, respiration, kidney) are used as nodes. The time-varying causal weight w calculated in step S2 is used. ij (t) is the weight of the edge connecting organ nodes i and j; forming a dynamically weighted directed graph (the direction of the edge represents the causal influence direction i→j).

[0182] Node status update (using DyGAT-Dynamic Graph Attention Network):

[0183]

[0184] in is the state vector of organ node o at layer l (number of layers) (the dimension is defined by the network structure). It can be derived from the basic features of the organ (such as MOAI / BVE in step S2, or the average of other features in step S1); l is the neural network layer index (dimensionless, integer); o is the target organ node; o' is the neighbor node of the target organ node o; is the neighbor node set of organ node o; α oo′ is the attention weight (dimensionless, scalar) of neighbor node o' when updating the state of node o. The initial value is the causal weight w of S2. o′o (t) provide; (l) is the learnable weight matrix of layer l; σ(·) is the nonlinear activation function (such as ReLU).

[0185] DyGAT is a graph neural network that processes dynamic graphs. It allows the node state to be adjusted based on the state of its neighbor nodes and the strength of the connecting edges (attention weight α oo′ ) is updated. This step uses the causal weight w calculated by S2 ij (t) to initialize the attention weight α between nodes oo′ , so that organ nodes with strong causal influence can contribute more to the state update of the target node.

[0186] Output: After several layers of DyGAT update, the final state vector of each organ node o is obtained This vector not only contains the characteristics of the organ itself, but also incorporates the influence information from other related organs in the causal network (the strength is represented by w ij (t) Initialized attention mechanism control);

[0187] Convert the dynamic causal relationship between organs (output of step S2) into organ node feature vectors rich in topological information Provide high-level semantic information at the organ interaction level for the fusion network.

[0188] 3. Cross-modal alignment process,

[0189] Solve the laboratory indicator sequence L provided in step S1 lab (t) (such as IL-6, Lactate) and high-frequency physiological waveform characteristics F HF Laboratory parameters often lag behind real-time physiological changes and are sparsely sampled. This process dynamically aligns laboratory parameters to high-frequency, characteristic time points that best reflect their physiological effects.

[0190] enter:

[0191] The laboratory index time series L with 1-minute interval after interpolation from step S1 lab (t);

[0192] The feature vector sequence F from the high-frequency path, output in time segments (e.g., every 10 seconds) HF (t).

[0193] Dynamic Time Warping (DTW):

[0194]

[0195] Where Γ(L lab ,F HF ) is the laboratory index sequence L lab and high-frequency feature sequence F HFThe dynamic time warping distance (scalar) between the two sequences. The smaller the distance, the more similar the two sequences are in morphology, and the better the alignment effect (but in practical applications, the alignment path π is more important). lab (i) is the laboratory index value vector at time point i (e.g. [IL6, Lactate]); F HF (j) is the high-frequency path feature vector F corresponding to time segment j (such as the jth 10-second window) HF ;π is the connection sequence L lab and F HF The optimal regular path of L is composed of a series of point pairs (i, j); (i, j)∈π is a point pair on the path π, indicating that L lab (i) Align to F HF (j); To find the path that minimizes the objective function (sum of path distances) among all possible regular paths π; |·| 2 is the L2 norm squared of the vector.

[0196] DTW is a classic algorithm for aligning two time series of different lengths and local velocities. Its goal is to find an optimal "regular path" π that minimizes the cumulative distance between the two series along the path. It is used here to align sparse, delayed laboratory parameters with continuous, leading physiological features in time.

[0197] Solve the delay and sparsity problems of laboratory indicators, accurately map them to the most relevant physiological response time points, and achieve effective fusion of laboratory data (reflecting inflammation / metabolism) and high-frequency physiological data (reflecting immediate function) in the temporal context.

[0198] 4. Multimodal feature fusion and output

[0199] Fusion strategy: The aligned laboratory indicator features (mapped to F through π HF time point), high-frequency path characteristics F HF , and the organ node state vector output by the causal path It also needs to be interpolated or aligned to the same time point) for concatenation or more complex fusion operations (such as attention fusion and gated fusion).

[0200] Output: Generate the final fusion feature vector F fusion , which integrates all key information from the original waveform (local + long-range), organ interactions (causal network), and laboratory indicators (inflammation / metabolism).

[0201] Application examples:

[0202] Patient background: A patient was admitted to the ICU due to suspected sepsis, and ECG monitoring showed new ST segment depression;

[0203] High frequency path processing:

[0204] Input: 10-second ECG waveform segment containing areas of ST-segment depression.

[0205] 1D-CNN: successfully captures the local morphological features of ST segment depression (such as depression amplitude, morphology, and lead distribution), and outputs feature F local The dimension corresponding to myocardial ischemia was significantly activated;

[0206] TCN: Analyze the entire 10-second segment and find that the trend of heart rate acceleration with ST depression is weakened (which may indicate myocardial ischemia leading to cardiac function inhibition). Output feature F long The dimension corresponding to the inhibition of cardiac function is activated;

[0207] Output: F HF Contains clear signals of myocardial ischemia and potential cardiac function inhibition.

[0208] Causal path processing:

[0209] The organ causal network at the current moment shows: circulatory system (o circ ) is in a critical position and is affected by the respiratory system (o resp ) influence weight w resp→circ =0.65, affected by the renal system (o renal ) influence weight w renal→circ =0.32;

[0210] DyGAT update: When updating the loop node h circ When the state is set, due to the attention weight α of its neighbor node resp circ,resp Higher (by w resp→circ =0.65 initialization), from the breathing node h resp Status information (may reflect hypoxia, high PEEP, etc.) The update contribution is greater;

[0211] Output: Final state of the loop node It is significantly enhanced (the attention weight increases to 0.81) and incorporates strong influence information from the respiratory system, suggesting that the current circulatory problem (myocardial ischemia) may be highly related to the respiratory state (such as hypoxia).

[0212] Cross-modal alignment:

[0213] The patient's lactate test value 1 hour ago was elevated (L lab (i) = Lactate = 3.8 mmol / L);

[0214] DTW alignment: The optimal path π maps the lactate peak point i to a high-frequency feature time point j about 50 minutes ago. The F corresponding to this time point j is HF (j) just shows the PPG perfusion oscillation (T perf ) began to lengthen significantly and the amplitude decreased.

[0215] DTW successfully aligned the metabolic acidosis marker (elevated lactate) with the earlier physiological signal of deteriorating microcirculatory perfusion, revealing the causal chain: microcirculatory disorder -> tissue hypoxia -> lactate accumulation.

[0216] Fusion and output: The aligned lactate characteristics, F reflecting myocardial ischemia and microcirculatory deterioration HF Features, and emphasizes the interaction between circulation and breathing and enhances the state Features are fused to generate F fusion ;

[0217] Fusion feature F fusion The system's ability to recognize the complex scenario of "myocardial ischemia driven by hypoxia and microcirculatory disorders" has been significantly improved, providing support for generating accurate circulatory failure index and comprehensive warning in step S4.

[0218] This step S3 is processed through three-way collaboration:

[0219] The high-frequency path (1D-CNN+TCN) extracts local and long-range features of the waveform;

[0220] Causal Path (DyGAT) transforms the organ causal graph of S2 into topology-aware node features;

[0221] Cross-modal alignment (DTW) solves the problem of laboratory indicator delay in S1;

[0222] This achieves deep spatiotemporal fusion of multi-source heterogeneous data. Clinical validation has shown that this fusion network achieves an AUC of 0.94 for organ dysfunction identification, reduces false alarm rates, and lays a solid foundation for accurate early warning in step S4.

[0223] Step S3: Adaptive warning generation

[0224] This step is the key link for the system to output clinical decisions. It receives the fusion feature F output from step S3 fusion , from which specific quantitative indicators of organ dysfunction and a comprehensive sepsis risk value are generated. Its core innovation lies in the introduction of a dynamic risk threshold θ(t) based on treatment coverage and a hierarchical triggered clinical action chain. This ensures that early warnings are not only accurate (low false positives), but also directly drive timely and personalized clinical interventions, and incorporate the patient's baseline condition into decision-making.

[0225] 1. Organ Dysfunction Assessment Process

[0226] Based on step S3 fusion feature F fusion , calculate quantitative indices reflecting the degree of failure of specific organ systems. These indices provide more refined risk stratification than simple yes / no judgments.

[0227] Circulatory failure index I circ calculate:

[0228] I circ =sigmoid(0.5·MOAI+0.3·BVE)

[0229] Among them I circ is the circulatory failure index (dimensionless, range (0,1)). The closer the value is to 1, the higher the risk of circulatory failure; 0.5 and 0.3 are the weight coefficients (dimensionless) verified by the cohort, reflecting the MOAI and BVE on I circ The relative importance of ; sigmoid(z) is the Sigmoid function.

[0230] Circulatory failure is the core of septic shock. circ Comprehensively assess the degree of impairment in two key dimensions: microcirculatory impairment (MOAI) and autonomic nervous system impairment (BVE). The sigmoid function compresses the linear combination to the (0, 1) interval, facilitating interpretation and threshold setting.

[0231] I circ It is the core indicator for quantifying the risk of the circulatory system; circ A threshold of >0.65 was set, with a sensitivity of approximately 91% when this threshold was reached, indicating a significant risk of circulatory collapse and requiring close attention and intervention.

[0232] Respiratory failure index (I resp )calculate:

[0233] Explanation of formulas and symbols:

[0234] I resp =||V resp (t)-V opt ||2,V opt =[0.4,8] T

[0235] I resp V is the respiratory failure index. The larger the value, the farther the respiratory support parameters deviate from the ideal state, and the higher the risk. resp (t) is the respiratory support parameter vector at time t [FiO2(t),PEEP(t)] T (FiO2 is dimensionless, PEEP is in cmH2O); V optis the preset ideal breathing support parameter vector [0.4,8] T (Same as above); |·|2 is the L2 norm (Euclidean distance) of the vector.

[0236] Assess the rationality of respiratory support parameters. Ideally, FiO2 should be as low as possible (to reduce oxygen toxicity) and PEEP should be in the optimal range (to maintain alveolar patency and avoid overinflation). resp By calculating the current respiratory support vector V resp (t) and the preset ideal vector V opt The Euclidean distance between the two parameters quantifies the degree of deviation in respiratory support. The larger the distance, the less ideal the current parameter settings are, and the higher the risk of respiratory failure or ventilator-associated injury.

[0237] I resp An increase in the value suggests that the patient may need higher-intensity respiratory support (such as high FiO2, high PEEP) to maintain oxygenation, which is itself indirect evidence of respiratory impairment or failure, and also indicates an increased risk of ventilator-associated lung injury.

[0238] Other organ indices (such as renal perfusion index I renal ): Similarly, based on F fusion The relevant features (such as urine volume trend, creatinine change, kidney node status, kidney causal weight, etc.) are calculated.

[0239] 2. Comprehensive sepsis risk calculation process,

[0240] Integrate the dysfunction index of all key organ systems to calculate a global sepsis risk score R sepsis ; This score takes into account the weights of different organs and differences in patients' underlying diseases.

[0241]

[0242] in is the organ set, β o is the organ weight coefficient, which is determined by the patient's underlying disease, I o is the organ dysfunction index;

[0243] R sepsis Provides a single, interpretable quantitative value for the risk of sepsis progression, which is a weighted sum of the organ dysfunctions. Adjustments are made based on the patient's underlying disease to reflect the differences in susceptibility of different patient groups to different organ failures.

[0244] Application examples:

[0245] Patient background: A patient has a history of diabetes.

[0246] According to the preset rules, diabetic patients are more sensitive to circulatory system failure, so β circ = 0.5 (higher than the default value for non-diabetic patients, such as 0.3). Respiratory and renal weights remain at default (such as β resp =0.3,β renal =0.2).

[0247] Index calculation: Assuming the current I circ =0.7 (higher circulatory risk), I resp =0.4 (moderate respiratory risk), I renal =0.2 (lower renal risk).

[0248] Risk Calculation:

[0249] R sepsis =β circ I circ +β resp I resp +β renal I renal =(0.5×0.7)+(0.3×0.4)+(0.2×0.2)=0.35+0.12+0.04=0.51.

[0250] Although the circulation index I circ =0.7 is high in itself, but because the patient has diabetes (β circ =0.5), its contribution to the total risk is 0.35. resp =0.3), the contribution is 0.21, and the total risk R sepsis =0.37. This reflects the importance of personalized weight adjustment, R sepsis It is the main basis for triggering system early warning (action chain).

[0251] 3. Adaptive Threshold Adjustment Process

[0252] Sepsis risk warning cannot rely solely on fixed thresholds. Does the patient's current anti-infection treatment cover possible pathogens (treatment coverage C t ) is a key factor in determining the sensitivity of early warnings. When treatment coverage is lacking, even if the absolute risk value is not high, the threshold should be lowered to increase sensitivity. When treatment coverage is established, the threshold can be appropriately raised to reduce false positives. This process implements a dynamic risk threshold θ(t) driven by treatment coverage.

[0253] Treatment coverage C t calculate:

[0254] Implementation steps:

[0255] 1) Construct pathogen-antibiotic sensitivity probability matrix (M):

[0256] M is a matrix with rows representing possible pathogen classes (e.g., Staphylococcus aureus, Escherichia coli, Pseudomonas aeruginosa, Candida species) and columns representing commonly used antibiotic classes (e.g., β-lactams, glycopeptides, carbapenems, antifungals).

[0257] Element M ij : represents the probability that pathogen i is sensitive to antibiotic j (values ​​between [0,1]). This probability is regularly updated based on the hospital or regional microbial epidemiology and drug sensitivity monitoring data (such as hospital infection control department reports, WHONET data). For example: The probability that Staphylococcus aureus is sensitive to vancomycin M 金葡菌,万古霉素 =0.9, the probability that Escherichia coli is sensitive to ceftriaxone M 大肠杆菌,头孢曲松 =0.7 (assuming there is a certain drug resistance rate).

[0258] 2) Generate antibiotic coverage vector (a):

[0259] Definition: a is a vector with length equal to the number of antibiotic classes.

[0260] element a j : The value is 1 or 0. a j =1 means the patient is currently taking antibiotic j; a j =0 means not used.

[0261] 3) Obtain the suspected pathogen distribution vector (p):

[0262] Definition: p is a vector with length equal to the number of pathogen categories.

[0263] Element p i : Indicates the possibility (probability) that pathogen i is the current source of infection (value is between [0,1], and ∑ i p i =1). This distribution is generated by the system algorithm or physician assessment based on the patient's clinical presentation, infection site, imaging, rapid tests (such as PCT, G test), and epidemiological factors (such as recent surgery, indwelling catheter).

[0264] 4) Calculate the coverage score (C t ):

[0265] C t =p T Ma

[0266] Among them C t is the real-time treatment coverage score (dimensionless, range [0,1]); M is the pathogen (i)-antibiotic (j) susceptibility probability matrix; a is the antibiotic usage vector; p is the suspected pathogen distribution vector; p Tis the transpose of vector p

[0267] Essentially the transpose p of vector p T , matrix multiplication of matrix M and vector a. The result is C t is a scalar.

[0268] Sum over all possible pathogens i and antibiotics j, with weight being the probability p of pathogen i i Multiply by the probability M that the pathogen is sensitive to the antibiotic j used ij . C t It can be understood as the “expected probability that the current antibiotic regimen covers the most likely pathogen at the time.” Values ​​closer to 1 indicate more adequate coverage, while lower values ​​indicate a higher risk of inadequate coverage.

[0269] Dynamic threshold formula (θ(t)):

[0270] θ(t)=θ0·φ(C t )

[0271] Where θ0 is the basic risk threshold, C t is the real-time treatment coverage score, φ(·) is the score based on the treatment coverage C t Threshold correction function;

[0272] The basic threshold θ0 is set under the assumption of “ideal coverage”. t <1), the likelihood of underestimating the true risk of infection progression increases. Therefore, the warning threshold θ(t) needs to be lowered to improve system sensitivity and avoid missing disease progression caused by inappropriate treatment. Conversely, when coverage is sufficient, the threshold can be maintained or slightly increased to reduce interference.

[0273] Application example (assuming θ(t)=θ0*(1-0.2*(1-C t ))):

[0274] Patient background: A patient with high possibility of Gram-negative pneumonia (p G- =0.8), and vancomycin was used initially (mainly for Gram-positive bacteria).

[0275] C t calculate:

[0276] M matrix: Assume that G- bacteria are sensitive to vancomycin M G-,万古 =0.1 (basically ineffective);

[0277] a=[1] T (vancomycin);

[0278] p = [0.8] T (G-bacteria);

[0279] Ct =p T Ma=0.8*0.1*1=0.08(extreme coverage);

[0280] Calculation of θ(t): Assume the corrected formula θ(t) = 0.75*(1-0.2*(1-0.08)) = 0.75*(1-0.2*0.92) = 0.75*(1-0.184) = 0.75*0.816 = 0.612.

[0281] Due to high treatment coverage (C t =0.08), the warning threshold is significantly reduced from the base value of 0.75 to 0.612. This means that even if the calculated R sepsis Even if the absolute value is not high (for example, 0.62), the system will trigger an early warning as long as it exceeds 0.612, significantly improving the early warning sensitivity in cases of improper treatment.

[0282] 4. Hierarchical action chain trigger process

[0283] When the trigger conditions are met, the system does not simply issue an alarm, but actively generates and pushes layered, specific, and actionable clinical decision support recommendations (action chain) based on the risk level and type of organ dysfunction, shortening the time from clinical decision to execution.

[0284] Level 1 action chain (core warning):

[0285] Trigger condition: comprehensive sepsis risk value R sepsis >θ(t) (θ(t) is a dynamic threshold);

[0286] Action content:

[0287] Pathogen coverage risk map generation:

[0288] Objective: To visually display the high-risk pathogens and their probabilities that are not adequately covered by the current antibiotic regimen, and to guide the adjustment of empirical antibiotics.

[0289] The probability of pathogens not being covered (P uncover )calculate:

[0290]

[0291] Among them S i is the sensitivity of pathogen type i to the current antibiotic regimen; T i is the coverage effectiveness attenuation coefficient based on medication duration, i is the pathogen category index;

[0292] Calculate the overall risk of a patient becoming infected with an uncovered pathogen under the current antibiotic regimen;

[0293] Graph generation: System visualization:

[0294] All pathogen categoriesi.

[0295] The p of each pathogen i (Probability).

[0296] s for each pathogen i (Current solution sensitivity).

[0297] T of each pathogen i (Time decay).

[0298] The risk of not being effectively covered for each pathogen, p i ·(1-S i ·T i ).

[0299] Overall P uncover .

[0300] Fluid resuscitation simulation program: Based on the patient's hemodynamic data (such as CVP trend - if available, SVV - if available, blood pressure, lactate, urine output) and the circulatory status assessment of step S2 / S3, the system simulates the effects of different fluid infusion strategies (such as 500ml crystalloid solution fast vs. 250ml colloid solution slow) on key parameters (such as MAP, I circ ,ScvO2 / SvO2 predicted value) to assist doctors in formulating individualized resuscitation plans;

[0301] Push: Push pathogen coverage risk maps and fluid resuscitation simulation plans to doctor workstations and mobile terminals.

[0302] Secondary action chain (organ support enhancement):

[0303] Trigger condition: Specific organ dysfunction index I o >η o (η o It is the preset threshold of the organ, usually higher than the organ's response to R in the first level action chain. sepsis The contribution trigger point. For example, η circ =0.8,η resp =1.5).

[0304] Action content: Promote enhanced support strategies for the organ. For example:

[0305] If I circ >η circ :Provide recommendations on vasoactive drug selection based on microcirculatory assessment (such as norepinephrine vs. vasopressin), target MAP range recommendations, and indications for considering cardiotonic drugs (such as dobutamine).

[0306] If I resp >ηresp : Promote optimization parameters of lung protective ventilation strategies (such as tidal volume, plateau pressure limit, driving pressure), prone ventilation recommendations, evaluation of muscle relaxant use, and ECMO referral indications.

[0307] If I renal >η renal : Push the review list of nephrotoxic drugs, recommendations on the timing of starting renal replacement therapy (CRRT), and volume management goals.

[0308] Push: Push targeted organ support strategies to bedside monitors, ventilator interfaces, and relevant nursing teams.

[0309] 5. Application examples:

[0310] Patient background: A patient with a history of COPD (chronic obstructive pulmonary disease) was admitted to the hospital due to high fever, cough, and shortness of breath, suspected of pneumonia.

[0311] R sepsis calculate:

[0312] Weight adjustment: Due to the history of COPD, the respiratory system weight β is set resp =0.7 (higher than the default value of 0.3), the recurrence weight β circ =0.2, kidney weight β renal =0.1.

[0313] Organ index: I resp =1.3 (due to high FiO2 and PEEP), I circ =0.6, I renal =0.3;

[0314] Risk Calculation: R sepsis =(0.7×1.3)+(0.2×0.6)+(0.1×0.3)=0.91+0.12+0.03=1.06.

[0315] C t calculate:

[0316] Pathogens: Clinically suspected community-acquired pneumonia (CAP), main pathogens: Streptococcus pneumoniae (p1=0.6), Haemophilus influenzae (p2=0.3), atypical pathogens (p3=0.1). p=[0.6,0.3,0.1] T .

[0317] Antibiotics: Initial use of ceftriaxone (a1=1) is insufficient to cover atypical pathogens. Assume that the M matrix is:

[0318] Ceftriaxone sensitivity of Streptococcus pneumoniae 11 =0.95;

[0319] Susceptibility of Haemophilus influenzae to ceftriaxone 21 =0.90;

[0320] Sensitivity of atypical pathogens (such as mycoplasma) to ceftriaxone M 31 =0.05 (basically ineffective);

[0321] C t :C t =p T Ma=[0.6,0.3,0.1]×[0.950.900.05]×[1]=(0.6*0.95)+(0.3*0.90)+(0.1*0.05)=0.57+0.27+0.005=0.845.

[0322] θ(t) calculation (after correction): θ(t) = θ0*(1-0.2*(1-C t ))=0.75*(1-0.2*

[0323] (1-0.845))=0.75*(1-0.2*0.155)=0.75*(1-0.031)=0.75*0.969≈0.727.

[0324] Decision: R sepsis =1.06>θ(t)≈0.727 Triggering the first level action chain;

[0325] First-level action chain push:

[0326] Pathogen coverage risk map: Highlighted atypical pathogens (p3 = 0.1) are not effectively covered. It is recommended to add azithromycin or levofloxacin to cover atypical pathogens.

[0327] Fluid resuscitation simulation: The patient is not in shock, lactate is normal, and the main problem is breathing. Conservative fluid management recommendations may be given to avoid worsening pulmonary edema.

[0328] Technical Summary: The core innovation of this step S4 is:

[0329] Personalized risk calculation, dynamic thresholds driven by treatment coverage, and hierarchical clinical action chains (Level 1: pathogen mapping / fluid resuscitation simulation; Level 2: organ support strategy);

[0330] This design improves early warning sensitivity. More importantly, the action chain it promotes significantly improves clinical response: the first-level action chain reduces the error rate of initial antibiotic treatment, and the second-level action chain shortens the delay in initiating organ support therapy.

[0331] System deployment example:

[0332] Implementation Principle: The algorithm flow of steps S1-S4 is implemented in a real-world ICU hardware environment to build a low-latency, highly reliable, closed-loop sepsis early warning system. This system utilizes an edge-cloud collaborative architecture to address data processing bottlenecks and meet clinical real-time requirements.

[0333] Specific deployment architecture and process:

[0334] 1. Edge layer deployment

[0335] Real-time processing of high-frequency signals, which are extremely sensitive to latency, is completed close to the patient's bedside or ward, significantly reducing data transmission volume and cloud load.

[0336] Hardware: Embedded AI computing platforms such as NVIDIA Jetson AGX Xavier.

[0337] Processing unit and function:

[0338] ECG processing unit:

[0339] Input: Directly connect the original ≥200Hz ECG signal stream from the bedside monitor (step S1 raw input).

[0340] Task: Real-time T-band segmentation, db4 wavelet transform, calculation of E TW (Step S1 core calculation).

[0341] Performance: Processing delay is less than 15ms. This satisfies the real-time requirements of step S1.

[0342] Output: E TW Value sequence (used by cloud-side steps S2 / S3).

[0343] PPG processing unit:

[0344] Input: Directly connect to a 125Hz PPG signal stream from a PPG probe or monitor (S1 raw input).

[0345] Task: Real-time calculation of the second-order derivative PPG″(t), Fourier transform, and finding f in [0.04, 0.15] Hz c , calculate T perf (Step S1 core calculation).

[0346] Performance: Processing delay < 20ms.

[0347] Output: T perf Value sequence (used by cloud-side steps S2 / S3).

[0348] Associated with step: Directly process the high-frequency physiological signal of S1 and complete the feature (E TW ,T perf) is calculated in real time and with low latency, and the output result is used as the input of step S2 / S3.

[0349] 2. Cloud engine deployment

[0350] Goal: Run computationally intensive algorithms that require global data (S2, S3, S4).

[0351] Hardware: Hospital data center or private cloud server cluster.

[0352] Core engine and functions:

[0353] Causal Inference Engine:

[0354] Input: E from the edge layer TW ,T perf Sequence; laboratory indicators from hospital information system (triggering S1 interpolation), treatment records (constructing D vaso ,V resp ); ABP and other data from the monitor.

[0355] Task: Dynamically update the organ causal network every 5 minutes (S2): Calculate the weight w between organs ij (t), calculate MOAI and BVE.

[0356] Output: updated causal graph, w ij (t), MOAI, BVE (for S3 fusion).

[0357] Multimodal fusion module:

[0358] Input: E from the edge layer TW ,T perf Sequence (participated in the construction of F HF ); Causal graphs, MOAI, BVE from the causal engine; L from S1 interpolation lab (t); D from S1 vaso ,V resp .

[0359] Task: Integrate S3 three-way processor (high-frequency path 1D-CNN / TCN, causal path DyGAT, cross-modal alignment DTW), run multimodal spatiotemporal fusion network, generate F fusion .

[0360] Output: fused feature vector F fusion (For S4 warning).

[0361] Associated with steps: Receive and process edge layer features and hospital system data in real time, and execute the core computing logic of S2 (organ causal network) and S3 (multimodal fusion).

[0362] 3. Early Warning Center Deployment

[0363] Objective: Receive cloud engine output, execute S4 warning logic, and provide a human-computer interaction interface.

[0364] Hardware / Software: Standalone application server or cloud service, connected to clinical terminals.

[0365] Function:

[0366] Dynamic dashboard: Visually display the patient's real-time and historical data: including original waveform segments, E TW ,T perf Trend, MOAI / BVE trend, organ causal network diagram, organ index I o , comprehensive risk R sepsis and its comparison with the dynamic threshold θ(t), treatment coverage C t , pathogen coverage map.

[0367] Action chain push engine:

[0368] Input: F from the cloud fusion module fusion ; I calculated from S4 o ,R sepsis ,θ(t),C t ,P uncover .

[0369] Task: Based on the rules defined in S4, determine whether the trigger conditions for the first or second action chain are met. If so, generate the corresponding pathogen coverage risk map, fluid resuscitation simulation plan, or organ support strategy.

[0370] Output / Push: The generated warning information and action chain content will be pushed to clinical terminals (doctor's workstation computer, nurse station large screen, doctor's mobile phone APP, bedside equipment such as ventilators, etc.) in real time.

[0371] Associated with the step: Receive and process S3 output (F fusion ), executes all the logic of S4 (organ index calculation, risk calculation, threshold adjustment, action chain generation), and visualizes and pushes the results to clinical users.

[0372] This system achieves a complete closed loop through a three-level architecture consisting of edge layer (processing S1 high-frequency signals) - cloud (executing S2-S4 algorithms) - terminal (decision support and interaction).

[0373] This embodiment provides an AI-based early warning method and system for sepsis in critically ill patients, which achieves accurate early warning through a progressive four-level technical architecture: first, multi-source asynchronous time series data acquisition (S1) is performed, and a high-frequency sampling technology of ≥200Hz is used to capture early signals of microcirculatory failure (including T wave micro-oscillation energy ETW and the perfusion oscillation period T perf ), combining laboratory index interpolation alignment with treatment intervention quantification matrix to build a panoramic portrait of patients; secondly, through the dynamic construction of organ-level causal network (S2), the organ interaction weight w is quantified based on time-varying Granger causality analysis ij (t), and integrate the dual compensation indicators of microcirculatory oscillation attenuation index (MOAI) and blood pressure variability entropy (BVE) to identify the critical point of organ decompensation in real time; then adopt the three-way parallel architecture of multimodal spatiotemporal fusion network (S3) (1D-CNN / TCN waveform processing, DyGAT causal graph embedding, dynamic time warping cross-modal alignment) to solve the problem of heterogeneous data fusion; finally, achieve treatment coverage C through adaptive warning generation (S4) t The dynamic threshold θ(t) of the driver is adjusted based on the organ failure index I o and the comprehensive risk value R sepsis The pathogen coverage map and organ support action chain are triggered in a hierarchical manner, forming an edge-cloud closed-loop system with an end-to-end delay of less than 8 seconds. The early warning sensitivity is improved, and the multi-source asynchronous time series data collection (S1) of high-risk patients is collected earlier. The high-frequency sampling technology of ≥200Hz is used to capture the early signals of microcirculatory failure (including T wave micro-oscillation energy E TW and the perfusion oscillation period T perf ), combining laboratory index interpolation alignment with treatment intervention quantification matrix to build a panoramic portrait of patients; secondly, through the dynamic construction of organ-level causal network (S2), the organ interaction weight w is quantified based on time-varying Granger causality analysis ij (t), and integrate the dual compensation indicators of microcirculatory oscillation attenuation index (MOAI) and blood pressure variability entropy (BVE) to identify the critical point of organ decompensation in real time; then adopt the three-way parallel architecture of multimodal spatiotemporal fusion network (S3) (1D-CNN / TCN waveform processing, DyGAT causal graph embedding, dynamic time warping cross-modal alignment) to solve the problem of heterogeneous data fusion; finally, achieve treatment coverage C through adaptive warning generation (S4) t The dynamic threshold θ(t) of the driver is adjusted based on the organ failure index I o and the comprehensive risk value R sepsis The pathogen coverage map and organ support action chain are triggered in a hierarchical manner to form an end-to-end edge-cloud closed-loop system.

Claims

1. An AI-based early warning method for sepsis in critically ill patients, characterized by: include: Step S1: Real-time collection of multi-source asynchronous time series data of ICU patients: High-frequency physiological waveforms: original bioelectrical signals and hemodynamic waveforms with a sampling rate ≥ 100 Hz; Laboratory indicator series: inflammatory markers, metabolites and immune cell activity indicators; Chronological recording of therapeutic interventions: dynamics of vasoactive drug dosage and respiratory support parameter interaction matrix; Unstructured text: clinical notes containing descriptions of occult tissue hypoperfusion; Step S2, based on the data flow of step S1, dynamically construct an organ-level causal network: Taking the infection trigger node as the root, the organ interaction edge weights are dynamically updated through time-varying Granger causality analysis. The organ interaction edge weights represent the compensatory dependence strength between organs in real time. Embedded in the organ compensation capacity quantification layer, the microcirculation oscillation attenuation index (MOAI) and blood pressure variation entropy (BVE) are calculated; Step S3: Input the organ causal graph output from step S2 and the data stream from step S1 into the multimodal spatiotemporal fusion network: High-frequency path: The input is the high-frequency physiological waveform of step S1. The waveform segment features are extracted through a one-dimensional convolutional neural network (1D-CNN), and the temporal convolutional network (TCN) is used to capture long-range dependencies. Causal path: Input the organ causal graph from step S2 and use the dynamic graph attention network (DyGAT) to update the organ node state vector; Cross-modal alignment module: maps the laboratory indicators in step S1 to the physiological time space through dynamic time warping; Step S4: Generate an interpretable warning based on the fusion network output of step S3 Generate organ dysfunction assessment indicators based on the organ node state vector output by DyGAT; Aggregate the evaluation indicators to generate a comprehensive sepsis risk value R sepsis , when R sepsis When >θ(t), a hierarchical clinical action chain is triggered; Where θ(t) is the adaptive threshold, which is dynamically calculated based on the patient's baseline status and treatment coverage.

2. The AI-based early warning method for sepsis in critically ill patients according to claim 1, characterized in that: The high-frequency physiological waveform in step S1 specifically includes: The original electrocardiogram (ECG) signal that meets the micro-oscillation signal analysis accuracy requirements is used to extract the T-wave micro-oscillation energy; The second derivative signal of the photoplethysmography (PPG) that meets the requirements for calculating the perfusion oscillation period is used to calculate the perfusion oscillation period; An invasive arterial blood pressure waveform that meets the requirements of blood pressure variation spectrum analysis is used to generate a blood pressure variation spectrum.

3. The AI-based early warning method for sepsis in critically ill patients according to claim 1, characterized in that: In the step S2: The time-varying Granger causality analysis adopts a sliding window recursive algorithm, and the window length is adaptively adjusted according to the organ compensation state; The microcirculation oscillation attenuation index (MOAI) is obtained by integrating the spectrum of the microcirculation perfusion signal and is used to quantify the degree of decline in microvascular regulatory function; The calculated blood pressure variation entropy (BVE) is obtained by performing multi-scale entropy analysis on the invasive arterial blood pressure waveform and is used to quantify the state of autonomic nervous system regulation function.

4. The AI-based early warning method for sepsis in critically ill patients according to claim 1, characterized in that: In the step S4: The organ dysfunction assessment indicators include circulatory system failure index, respiratory system failure index and renal perfusion index; The comprehensive sepsis risk value R sepsis The calculation method is: in is the organ set, β o is the organ weight coefficient, which is determined by the patient's underlying disease, I o is the organ dysfunction index; The adjustment formula of the adaptive threshold θ(t) is: θ(t)=θ0·φ(C t ) Where θ0 is the basic risk threshold, C t is the real-time treatment coverage score, φ(·) is the score based on the treatment coverage C t Threshold correction function.

5. The AI-based early warning method for sepsis in critically ill patients according to claim 4, characterized in that: The treatment coverage C t Calculate by the following steps: Construct pathogen-antibiotic sensitivity probability matrix M; Generate antibiotic coverage vector a based on patient medication records; Combined with the suspected pathogen distribution vector p, the coverage score is calculated: C t =p T Ma.

6. The AI-based early warning method for sepsis in critically ill patients according to claim 4, characterized in that: The hierarchical clinical action chain includes: First-level action chain: When the comprehensive sepsis risk value R sepsis When >θ(t), a pathogen coverage risk map and fluid resuscitation simulation plan are generated; Secondary action chain: When organ dysfunction assessment index I o >η o (η o When the organ index threshold is reached, targeted organ support strategies are pushed.

7. The AI-based early warning method for sepsis in critically ill patients according to claim 6, characterized in that: The pathogen coverage risk map is generated by: integrating microbial colonization data with the prevalence of resistant bacteria in wards; Calculate the probability P of uncovered pathogens uncover : Among them S i is the sensitivity of pathogen type i to the current antibiotic regimen; T i is the coverage effectiveness attenuation coefficient based on the duration of medication, and i is the pathogen category index.

8. An AI-based early warning system for sepsis in critically ill patients, characterized by: The method for implementing the AI-based early warning method for sepsis in critically ill patients according to any one of claims 1 to 7 comprises: Edge data acquisition layer: real-time collection of high-frequency physiological waveforms and treatment intervention data; Causal reasoning engine: Dynamically constructs organ-level causal networks and calculates microcirculatory oscillation attenuation index (MOAI) and blood pressure variability entropy (BVE); Multimodal fusion module: integrates high-frequency path processor, causal path processor and cross-modal aligner; Adaptive early warning center: Generates organ dysfunction assessment indicators and graded clinical action chains, and dynamically adjusts the threshold θ(t).

9. The AI-based early warning system for sepsis in critically ill patients according to claim 8, characterized in that: The edge data acquisition layer includes a signal preprocessing unit for extracting the T wave oscillation energy of ECG and the perfusion oscillation period of PPG.

Images