A computer-implemented pediatric mechanical ventilation strategy prediction method and system

Abstract

The application discloses a computer-implemented pediatric mechanical ventilation strategy prediction method and system, and relates to the technical field of medical artificial intelligence and clinical auxiliary decision-making.The method reconstructs ventilation physiological characteristics and missing markers according to basic information of a sick child, mechanical ventilation parameter records and laboratory test indexes, and constructs a clinical event flow; the sick child and the ventilation strategy combined with the discretization code are respectively constructed as nodes, and a dynamic interaction event graph is constructed by using concurrent clinical characteristics matched with ventilation parameter adjustment time, time interval and historical backtracking; short-term hidden state and long-term memory vectors are fused through a ventilation interaction state space module, a long-term memory module and a self-adaptive gating mechanism, and a candidate ventilation strategy ranking matrix is output by a scoring network.The application can process multi-source asynchronous and irregular sampling event flows, and assist medical staff in ventilation strategy evaluation.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence and clinical decision support technology, and in particular to a computer-implemented method and system for predicting pediatric mechanical ventilation strategies. Background Technology

[0002] Pediatric respiratory failure is one of the high-risk critical illnesses in pediatric intensive care units. Mechanical ventilation, as a vital life support tool, directly impacts the child's oxygenation, lung mechanics, and subsequent clinical outcomes through its parameter settings. In clinical practice, adjustments to ventilation parameters typically rely on physician experience, bedside monitoring indicators, and guideline recommendations, making it difficult to fully utilize the continuously generated multi-source asynchronous data in electronic medical records. Existing intelligent assistance solutions often employ static features, rule-based models, traditional machine learning, or deep time-series models with a single time scale. While these solutions can provide parameter estimation or knowledge retrieval capabilities in specific scenarios, they typically fail to organize the child, ventilation strategy, and concurrent clinical characteristics into a continuous-time dynamic graph, and struggle to simultaneously characterize frequent ventilation adjustments within a short period and disease progression over a longer time span. Furthermore, pediatric mechanical ventilation data is characterized by limited sample size, irregular event timing, asynchronous laboratory testing and ventilation recording, missing key variables, and cold start in newly admitted children. If relying solely on fixed time windows, static node embeddings, or single memory states, the model may easily overlook the difference between short-term physiological feedback and the cumulative effects of long-term treatment.

[0003] Therefore, how to combine clinical prior knowledge with dynamic modeling to accurately predict pediatric mechanical ventilation strategies has become a technical challenge that urgently needs to be solved. Summary of the Invention

[0004] The main objective of this invention is to provide a computer-implemented method and system for predicting pediatric mechanical ventilation strategies, which aims to combine clinical prior knowledge with dynamic graph modeling to accurately predict pediatric mechanical ventilation strategies.

[0005] This invention proposes a computer-implemented method for predicting pediatric mechanical ventilation strategies, comprising the following steps: Based on the acquired basic information of the child, mechanical ventilation parameter records, and laboratory test indicators, feature reconstruction is performed to obtain reconstructed features characterizing the child's ventilation physiological state and corresponding missing markers, which are then organized into a clinical event flow in chronological order; the child is constructed as a patient node, and multiple core control parameters characterizing the ventilation control strategy in the mechanical ventilation parameter records are discretized and combined to encode a ventilation strategy node; using the ventilation parameter adjustment time as the interaction event timestamp, laboratory test records containing at least one non-missing indicator from the preset laboratory test indicators and whose sampling time is earlier than or equal to the current interaction event timestamp are back-matched as concurrent clinical feature vectors, and a dynamic interaction event graph is constructed using the time interval from the child's last interaction event as the edge weight; using the ventilation interaction state space module, the ventilation interaction state is dynamically adjusted according to the time interval of the current interaction event. The adaptive parameter generation network generates state space parameters, and updates the short-term hidden states of the patient node and the ventilation strategy node through time-interval-based discretized continuous-time states. A long-term memory module maintains the long-term memory vectors of the patient node and the ventilation strategy node, dynamically updating these vectors via bidirectional cross-writing when the current interaction event occurs. An adaptive gating mechanism adaptively fuses the short-term hidden states with the unupdated long-term memory vectors to obtain the final fused representations for each patient node and the ventilation strategy node. The final fused representations of the patient node and each candidate ventilation strategy node are input into a scoring network for correlation measurement, outputting the interaction scores of each candidate ventilation strategy at future time points. A ranking matrix is ​​output based on these interaction scores, where the ranking matrix is ​​only used to assist in clinical parameter evaluation, does not directly generate disease diagnosis conclusions, and does not directly control ventilator execution parameter adjustments.

[0006] Preferably, the process of obtaining the reconstructed features includes: screening pediatric patients under the age of 18 years and removing mechanical ventilation parameter records whose values ​​exceed a preset physiological threshold range or whose missing degree exceeds a preset missing proportion threshold, wherein the preset missing proportion threshold is 50%; converting the tidal volume in the mechanical ventilation parameter records according to the ideal weight in the patient's basic information to obtain the tidal volume per unit weight feature; and constructing the oxygenation index feature characterizing the oxygenation status by using the inhaled oxygen concentration in the mechanical ventilation parameter records and the arterial oxygen partial pressure in the laboratory test indicators.

[0007] Preferably, the laboratory testing indicators include: Arterial oxygen partial pressure Arterial carbon dioxide partial pressure The parameters include blood lactate, white blood cell count, hemoglobin, platelet count, blood glucose, and serum creatinine; wherein, during the feature reconstruction process, the original missing markers of laboratory test indicator variables that have not been sampled or are missing are retained, so that the clinical event stream contains explicit spatially aware features about the missing variable status.

[0008] Preferably, the plurality of core control parameters include four core control parameters, namely: the inhaled oxygen concentration as the first core control parameter, the positive end-expiratory pressure as the second core control parameter, the tidal volume per unit body weight as the third core control parameter, and the ventilation mode as the fourth core control parameter; the ventilation strategy node is obtained by combining and encoding the Cartesian product of the first, second, third, and fourth core control parameters in the discretized value space.

[0009] Preferably, each ventilation parameter adjustment event in the dynamic interaction event graph is parameterized as an interaction event edge 5-tuple:

[0010] in, Indicates the current patient node. This indicates the current ventilation strategy node. This indicates the timestamp of the interaction event at the moment the current ventilation parameter adjustment was performed. This indicates the time interval between the current ventilation parameter adjustment and the child's last interaction event; the concurrent clinical feature vector Constructed in the following way: using the current interaction event timestamp Based on the historical timeline, the search is only performed backtracking, and the single laboratory test record that is closest to the current time, contains at least one non-missing indicator among the preset laboratory test indicators, and whose sampling time is earlier than or equal to the current interaction event timestamp is used as the concurrent context feature attribute of the current interaction event edge.

[0011] Preferably, before updating the short-term hidden states of the patient node and the ventilation strategy node, the method further includes: generating a short-term event representation based on the historical local interaction event stream sequence in the dynamic interaction event graph, and inputting the short-term event representation into a temporal context encoder for context compression to obtain a local historical context vector that incorporates irregular time step changes.

[0012] Preferably, in the ventilation interaction state space module, the dynamic process of performing zero-order preserved discretization on the state space parameters and performing state recursion is as follows: The time interval of the current interaction event is... An adaptive parameter generation network is input, which dynamically outputs the continuous parameter matrix of the state space at the current continuous time step; the zero-order hold operator is used to perform a time-space parameter transformation on the continuous parameter matrix of the state space. The exponential discretization sampling transformation is used to obtain a discretized state space parameter matrix; based on the discretized state space parameter matrix, the previously updated hidden state is linearly mapped and recursively mapped to the currently generated local historical context vector to complete the short-term hidden state update of the patient node and the ventilation strategy node.

[0013] Preferably, the long-term memory module maintains a first long-term memory vector for the patient node and a second long-term memory vector for the ventilation strategy node. When the current interaction event occurs, the bidirectional dynamic update includes: reading the current first long-term memory vector corresponding to the patient node and the current second long-term memory vector corresponding to the ventilation strategy node; after calculating the final fusion representations corresponding to the patient node and the ventilation strategy node using an adaptive gating mechanism, the current ventilation strategy feature information is cross-written into the first long-term memory vector corresponding to the patient node to accumulate the disease trajectory, and the current physiological state information of the child is simultaneously cross-written into the second long-term memory vector corresponding to the ventilation strategy node to continuously update the long-term response mode of the ventilation strategy corresponding to the current ventilation strategy node under different pathological states.

[0014] Preferably, the fusion calculation process of the adaptive gating mechanism follows the following mathematical logic: [The short-term hidden states are then...] Compared with the long-term memory vector before the update Dense splicing is performed, and the data is input into a multilayer perceptron; the data is then processed using the external connection of the last layer of the multilayer perceptron. The activation function performs a nonlinear mapping and outputs the corresponding dimension-wise gating coefficients. ; Utilizing the aforementioned dimension-by-dimensional gating coefficients The following convex combination formula is applied to the short-term hidden state and long-term memory vectors to calculate and output the final fused representations corresponding to the patient node and the ventilation strategy node, respectively. : ,in, Represents a node At any moment The final fusion indicates that Represents a node At any moment The dimension-wise gating coefficients, Represents a node At any moment The short-term hidden state. Represents a node The long-term memory vector prior to the current interaction event. Indicates the node index. Indicates the current moment of the interactive event.

[0015] Preferably, the training process of the scoring network and the adaptive parameter generation network is driven by a binary cross-entropy loss function containing a hybrid three-dimensional negative sampling mechanism. The hybrid three-dimensional negative sampling mechanism includes: random negative sampling for establishing the basic discriminative boundary between real ventilation strategies and irrelevant strategies; historical negative sampling for extracting difficult negative samples from ventilation parameter combinations that the current patient has used in the historical disease timeline but not at the current interaction event time, for distinguishing outdated strategies that were historically feasible but are not applicable at present; and inductive negative sampling for constructing negative samples from ventilation parameter combinations that have not appeared in the global historical strategy space during the training phase, for improving the model's inductive generalization ability in the cold start scenario of newly admitted patients.

[0016] This application also discloses a dual-timescale continuous-time dynamic graph mechanical ventilation strategy prediction system, comprising: a feature reconstruction module, used to reconstruct features based on the acquired basic information of the child, mechanical ventilation parameter records, and laboratory test indicators, to obtain reconstructed features characterizing the child's ventilation physiological state and corresponding missing markers, and organize them into a clinical event flow in chronological order; a dynamic graph construction module, used to construct the child as a patient node, and to discretize and combine multiple core control parameters characterizing the ventilation control strategy in the mechanical ventilation parameter records to construct a ventilation strategy node; using the ventilation parameter adjustment time as the interaction event timestamp, backtracking to match laboratory test records containing at least one non-missing indicator from the preset laboratory test indicators and whose sampling time is earlier than or equal to the current interaction event timestamp as concurrent clinical feature vectors, and using the time interval from the child's last interaction event as edge weights to construct a dynamic interaction event graph; and a short-term dynamic modeling module, used to dynamically adjust the adaptive response based on the time interval of the current interaction event using the ventilation interaction state space module. The system generates state space parameters based on the parameter generation network and updates the short-term hidden states of the patient node and the ventilation strategy node through time-interval-based discretized continuous-time states. A long-term memory modeling module maintains long-term memory vectors for the patient node and the ventilation strategy node, dynamically updating these vectors via bidirectional cross-writing when the current interaction event occurs. A gating fusion module adaptively fuses the short-term hidden states with the previously updated long-term memory vectors using an adaptive gating mechanism to obtain the final fused representations for each patient node and the ventilation strategy node. A prediction output module inputs the final fused representations of the patient node and each candidate ventilation strategy node into a scoring network for correlation measurement, outputting the interaction scores for each candidate ventilation strategy at future time points and outputting a ranking matrix based on the interaction scores. A model training module uses a binary cross-entropy loss function with a hybrid three-dimensional negative sampling mechanism to drive the training of the scoring network and the adaptive parameter generation network.

[0017] The above technical solution has the following advantages: The method provided by this invention obtains reconstructed features and missing markers characterizing the ventilatory physiological state of children through multi-source data feature reconstruction, and organizes them into a clinical event flow, enabling the model to explicitly perceive the missing states and sampling frequencies of variables. By discretizing and combining multiple core control parameters to construct ventilation strategy nodes, and using time intervals as edge weights to construct a dynamic interactive event graph, the method depicts the high-frequency asynchronous clinical evolution process of pediatric intensive care, reducing the risk of future information leakage through unidirectional backtracking. Utilizing a ventilation interaction state space module and performing state recursion based on zero-order preserved discretization, the method can adjust the amplitude of state evolution according to the length of time intervals, achieving a response to short-term local fluctuations. Combined with the bidirectional cross-writing mechanism of the long-term memory module, the heterogeneous disease course of patients is coupled with the general pattern of ventilation strategies in the long term, alleviating the cold start problem caused by limited sample size or discontinuous strategies. The final fusion representation is obtained by executing the convex combination formula through an adaptive gating mechanism, achieving an adaptive balance between the immediate physiological state and the long-term disease baseline. The scoring network outputs a ranking matrix with recommended ranking and confidence prompts, providing auxiliary decision support for medical staff without directly replacing doctors' diagnosis and treatment decisions or directly controlling ventilator equipment. Attached Figure Description

[0018] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings, wherein: Figure 1 The overall flowchart provided for embodiments of the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the construction of a knowledge graph-enhanced continuous-time event graph provided in an embodiment of the present invention.

[0020] Figure 3 This is a structural diagram of a dual-timescale model provided in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the fusion of short-term state and long-term memory gating provided in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are used to illustrate the technical implementation of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] Combination Figure 1Example 1 provides a computer-implemented method for predicting pediatric mechanical ventilation strategies. This method is mainly used in intensive care settings, especially in pediatric intensive care units, to assist in predicting and supporting mechanical ventilation strategies for children with respiratory failure. In actual clinical practice, pediatric mechanical ventilation data suffers from technical challenges such as high-frequency adjustments, multi-source asynchronous processing, irregular sampling, and severe data gaps. This example constructs a knowledge graph-enhanced continuous-time event graph and combines it with a dual-timescale deep learning architecture to achieve joint modeling of short-term ventilation parameter fluctuations and long-term disease progression. The knowledge graph-enhanced continuous-time event graph refers to a continuous-time dynamic graph composed of patient nodes, ventilation strategy nodes, ventilation parameter adjustment event edges, time interval edge weights, and concurrent clinical feature attributes. The knowledge graph enhancement refers to explicitly connecting patient nodes, ventilation strategy nodes, ventilation parameter adjustment event edges, time interval edge weights, and concurrent clinical feature attributes in a graph structure, and allowing this graph structure relationship to participate in subsequent state recursion, gating fusion, and score prediction. The ventilation interaction state space module refers to a computational module used to generate discretized state space parameters and recursively deduce short-term hidden states based on the time interval of interaction events. The adaptive parameter generation network refers to a multilayer perceptron network that outputs a continuous parameter matrix of the state space based on time intervals. The hybrid three-dimensional negative sampling mechanism refers to a training sample construction mechanism composed of random negative sampling, historical negative sampling, and inductive negative sampling.

[0024] Combination Figure 1 The method in this embodiment first performs a feature reconstruction step. The system acquires basic information about the child, mechanical ventilation parameter records, and laboratory test indicators from the intensive care unit data source in real time or in batches. The basic information includes the child's actual age, gender, actual height, actual weight, and ideal weight, among other basic demographic characteristics. Mechanical ventilation parameter records are obtained from the real-time monitoring data stream of the ventilator or ventilation record forms in the electronic medical record system, specifically including core ventilation variables such as inhaled oxygen concentration, positive end-expiratory pressure, tidal volume, and the currently implemented ventilation mode. Laboratory test indicators include blood gas analysis and routine biochemical test results, specifically covering… Value, arterial oxygen partial pressure Arterial carbon dioxide partial pressure Key physiological variables such as blood lactate, white blood cell count, hemoglobin, platelet count, blood glucose, and serum creatinine are collected. After acquiring the above multi-source asynchronous data, the system first performs pediatric sample screening, extracting samples from pediatric patients under 18 years of age and removing abnormal entries with values ​​exceeding a preset physiological threshold, or mechanical ventilation parameter records with a missing percentage exceeding a preset missing percentage threshold of 50%. For the cleaned data, the system uses the ideal weight in the child's basic information to convert the tidal volume in the mechanical ventilation parameter records. The ideal weight can be calculated using a preset ideal weight formula. We obtained, among which, Indicates ideal weight. This represents the 50th percentile body mass index corresponding to age and gender. Height is represented by the tidal volume. The tidal volume per unit body weight is calculated using the formula of dividing tidal volume by ideal body weight. This characteristic reflects the ventilation intensity in lung-protective ventilation strategies, reducing the impact of individual body size differences on volume assessment. Simultaneously, the system uses the inhaled oxygen concentration recorded from mechanical ventilation parameters and the arterial oxygen partial pressure from laboratory test indicators to construct an oxygenation index characteristic representing oxygenation status. To enable subsequent models to perceive the asynchronous nature of testing in the intensive care environment, for laboratory test indicator variables that have not been sampled or are missing, the system explicitly retains their original missing markers during feature reconstruction, thus enabling the clinical event flow to include explicit spatially aware features regarding variable missing status. All reconstructed features and their corresponding missing markers are organized according to their actual chronological order to generate a unified clinical event flow.

[0025] Combination Figure 2 After constructing the clinical event flow, the system executes a node construction step, transforming heterogeneous clinical entities into topological representations within a graph network. Specifically, the system constructs the child as a patient node and discretizes and encodes multiple core control parameters characterizing the ventilation control strategy from the mechanical ventilation parameter records to construct a ventilation strategy node. These multiple core control parameters include four core control parameters: inhaled oxygen concentration (as the first core control parameter), positive end-expiratory pressure (PEEP) as the second core control parameter, tidal volume per unit body weight as the third core control parameter, and ventilation mode as the fourth core control parameter. Within a preset discretized value space, the system encodes the Cartesian product of the first, second, third, and fourth core control parameters in the discretized value space to obtain the ventilation strategy node. This encoding method preserves the overall synergistic effect of the clinical parameter combination and provides candidate prediction targets for the dynamic graph model.

[0026] Combination Figure 2The system constructs a dynamic interactive event graph based on the aforementioned nodes and clinical event flow. Using the time of ventilation parameter adjustment as the timestamp of the interactive event, the system backtracks historically to match laboratory test records containing at least one non-missing indicator from the preset laboratory test indicators, and whose sampling time is earlier than or equal to the current interactive event timestamp, as concurrent clinical feature vectors. The system then uses the time interval from the child's last interactive event as the edge weight to construct the dynamic interactive event graph. In this way, each ventilation parameter adjustment event in the dynamic interactive event graph is parameterized as an interactive event edge quintuple, as shown below:

[0027] in, Indicates the current patient node. This indicates the current ventilation strategy node. This indicates the timestamp of the interaction event at the moment the current ventilation parameter adjustment was performed. This indicates the time interval between the current ventilation parameter adjustment and the child's last interaction event. This represents the concurrent clinical feature vector. The concurrent clinical feature vector... Constructed in the following way: using the current interaction event timestamp Based on this, the system searches backward along the historical timeline, matching the single laboratory test record closest to the current moment that contains at least one non-missing indicator from the preset laboratory test indicators and whose sampling time is earlier than or equal to the timestamp of the current interaction event as the concurrent context feature attribute of the current interaction event edge. This one-way backtracking data matching mechanism can reduce the risk of future information leakage and ensure that the model's predictions at any historical moment are based on the clinical information known at that time. By organizing patients, ventilation strategies, irregular time intervals, and concurrent clinical features in a dynamic interaction event graph, the system can characterize the high-frequency, asynchronous clinical evolution of pediatric intensive care.

[0028] Combination Figure 3After constructing the dynamic interaction event graph, the system enters the computation and inference phase of the dual-timescale model. First, the system uses the ventilation interaction state space module to dynamically adjust the adaptive parameter generation network's state space parameters based on the time interval of the current interaction event. Then, it updates the short-term hidden states of the patient node and the ventilation strategy node through time-interval-based discretized continuous-time state recursion. Before updating the short-term hidden states of the patient node and the ventilation strategy node, the system also includes generating short-term event representations based on the historical local interaction event stream sequence in the dynamic interaction event graph. These short-term event representations are then input into a temporal context encoder for context compression, resulting in a local historical context vector that incorporates irregular time-step changes. In the ventilation interaction state space module, the dynamic process of performing zero-order preserved discretization on the state space parameters and performing state recursion is as follows: the system sets the time interval of the current interaction event... The system takes an adaptive parameter generator as input and dynamically outputs a continuous parameter matrix in the state space for the current consecutive time step. and input mapping matrix The state-space continuous parameter matrix is ​​subjected to a time interval using the zero-order preserve operator. The exponential discretization sampling transformation yields the discretized state-space parameter matrix, which is calculated as follows: ;in, Indicates time The discrete state transition matrix, Indicates time Discrete input mapping matrix, Represents a continuous state transition matrix. Represents a continuous input mapping matrix. Represents the identity matrix. This represents matrix exponentiation. Based on the discretized state-space parameter matrix, the system performs a linear mapping recursion between the previously updated hidden state and the currently generated local historical context vector to complete the short-term hidden state update of the patient node and the ventilation strategy node. The recursive form is as follows: ;in, Indicates ventilation strategy node At any moment The updated short-term hidden state. Indicates ventilation strategy node In the short-term hidden state prior to the current interactive event, This represents the local history context vector output by the temporal context encoder. For the patient node... The system adopts isomorphic state recursion. ;in, Represents patient node At any moment The updated short-term hidden state. Represents patient node In the short-term hidden state prior to the current interactive event, This represents a local historical context vector generated from the stream of local interaction events related to the patient's history.

[0029] Combination Figure 3 To simultaneously consider the long-term disease progression, the system utilizes a long-term memory module to maintain long-term memory vectors for both the patient node and the ventilation strategy node. When a current interaction event occurs, these long-term memory vectors are dynamically updated via bidirectional cross-writing. The long-term memory module maintains a first long-term memory vector for the patient node and a second long-term memory vector for the ventilation strategy node. The bidirectional dynamic update process during the current interaction event is as follows: the system first reads the current first long-term memory vector corresponding to the patient node and the current second long-term memory vector corresponding to the ventilation strategy node. After calculating the final fused representations for each of the patient node and the ventilation strategy node using an adaptive gating mechanism, the system cross-writes the current ventilation strategy feature information into the first long-term memory vector corresponding to the patient node to accumulate the disease progression trajectory, and simultaneously cross-writes the current physiological state information of the child into the second long-term memory vector corresponding to the ventilation strategy node to continuously update the long-term response pattern of the ventilation strategy under different pathological states. The specific writing function can be represented as follows: as well as ;in, and These represent the updated long-term memory vectors of the patient node and the ventilation strategy node, respectively. and These represent the long-term memory vectors of the patient node and the ventilation strategy node prior to the current interaction event, respectively. and Indicates the memory retention coefficient. and This indicates writing to the mapping function. and This represents the final fusion representation. and This represents the embedding of patient node features and the embedding of ventilation strategy features. This represents the concurrent clinical feature vector. This bidirectional cross-writing mechanism couples the heterogeneous disease course of patients with the common patterns of ventilation strategies over a long period, which can alleviate the cold start problem caused by limited sample size or discontinuous strategies.

[0030] Combination Figure 4After acquiring the short-term hidden state representing short-term fluctuations and the long-term memory vector representing long-term trends, the system uses an adaptive gating mechanism to adaptively fuse the short-term hidden state with the long-term memory vector before the update, obtaining the final fused representations corresponding to the patient node and the ventilation strategy node, respectively. The fusion calculation process of the adaptive gating mechanism follows the following mathematical logic: the system first fused the short-term hidden state with the long-term memory vector before the update, obtaining the final fused representations corresponding to the patient node and the ventilation strategy node. Compared with the long-term memory vector before the update Dense splicing is performed, and the data is input into a multilayer perceptron; then, the data from the last external layer of the multilayer perceptron is utilized. The activation function performs a nonlinear mapping and outputs the corresponding dimension-wise gating coefficients. Finally, using the aforementioned dimension-by-dimensional gating coefficients... The final fused representation of the output node is calculated by performing a convex combination formula on the short-term hidden state and long-term memory vector. The specific calculation formula is shown below:

[0031] in, Represents a node At any moment The final fusion indicates that Represents a node At any moment The dimension-wise gating coefficients, Represents a node At any moment The short-term hidden state. Represents a node The long-term memory vector prior to the current interaction event. Indicates the node index. This indicates the current moment of the interaction event. Through this dimension-wise gating fusion mechanism, the model can adaptively adjust the weights of short-term and long-term information according to the child's current clinical stage. During acute changes, the gating coefficient allows the model to rely more on short-term dynamics, while when the condition is relatively stable or the historical trajectory is long, the gating coefficient allows the model to make more use of long-term memory.

[0032] Finally, the system enters the prediction output stage. The system inputs the final fused representation of the patient node and the final fused representation of each candidate ventilation strategy node into the scoring network for correlation measurement, outputting the interaction scores of each candidate ventilation strategy at future time steps, and outputting a ranking matrix based on the interaction scores. The scoring network calculates the affinity or inner product metric between the two, and the output score is further converted into the interaction probability or ranking result of each candidate ventilation strategy at future time steps through an activation function; in one implementation, the scoring network can be represented as... ;in, Represents patient node With ventilation strategy node At any moment Interaction score, This represents the Sigmoid activation function. This represents a multilayer perceptron. This represents the final fusion representation of the patient nodes. This represents the final fused representation of the candidate ventilation strategy nodes. This represents a product of dimensions. The final ranking matrix output by the system not only includes the recommended ranking of candidate ventilation strategies but also specific probability values ​​and confidence indicators, thus providing stable and interpretable auxiliary reference information for healthcare professionals when adjusting clinical parameters. It is important to note that this ranking matrix and probability output are only used to assist healthcare professionals in their judgment; they do not directly replace doctors' diagnostic and treatment decisions, nor do they directly control the ventilator equipment or drive the ventilator to perform parameter adjustments. Therefore, artificial intelligence-based decision support is achieved while ensuring medical safety.

[0033] Combination Figure 1 This embodiment of the invention, based on the first embodiment, further expands on the specific implementation details of data acquisition and feature reconstruction. In actual operation, the feature reconstruction module 101 connects to and extracts basic data from the pediatric intensive care information system. To improve data accuracy, the system first executes clinical sample screening rules. The system sets an age threshold to screen pediatric patients under 18 years of age and automatically filters adult data. Simultaneously, the system initiates a medical rule cleaning procedure to detect anomalies in mechanical ventilation parameter records. For example, when the inhaled oxygen concentration is greater than 100% or less than 21%, or when the positive end-expiratory pressure is greater than 30 cmH2O, the system removes these abnormal records caused by instrument errors or data entry mistakes. In addition, for records with an overall missing percentage exceeding a preset missing percentage threshold, such as when the missing percentage of core control parameters in a child's mechanical ventilation record exceeds 50%, the system removes the entire line of that single ventilation record to ensure the quality of the training dataset.

[0034] After data cleaning is completed, the feature reconstruction module 101 uses the child's actual height and gender from the child's basic information to calculate the ideal weight using a preset formula. The child's ideal weight was calculated, among which, Indicates ideal weight. This represents the 50th percentile body mass index corresponding to age and gender. The ideal weight is used to represent height. Next, the feature reconstruction module 101 performs feature conversion on the raw tidal volume recorded in the mechanical ventilation parameters based on this ideal weight. Specifically, the raw tidal volume is divided by the ideal weight to calculate and construct the tidal volume characteristic per unit weight, with the standard unit being milliliters per kilogram of body weight, i.e., 6 mL / kg to 8 mL / kg. This feature reconstruction can reduce the interference of large weight fluctuations caused by excessive obesity, generalized edema, or severe emaciation in critically ill children on the assessment of ventilation volume, enabling the model to capture the volume limitation indicators in lung-protective ventilation strategies. Simultaneously, for real-time assessment of the child's oxygenation status, the feature reconstruction module 101 links and calls laboratory test indicators to obtain the arterial oxygen partial pressure from the most recent arterial blood gas analysis, and combines it with the current inhaled oxygen concentration to perform a formula calculation. The specific calculation formula is arterial oxygen partial pressure divided by inhaled oxygen concentration, thereby constructing the oxygenation index characteristic representing oxygenation status.

[0035] In clinical pediatric intensive care settings, laboratory tests are often intermittently and irregularly triggered, while the waveforms and parameters of mechanical ventilation are continuously and frequently recorded. This leads to a serious asynchrony between laboratory testing and ventilation recording. To enable the model to proactively perceive and utilize the non-uniform sampling characteristics of this data, the feature reconstruction module 101 retains the original missing markers for laboratory test indicators that have not been sampled or are missing during the feature reconstruction process. For example, if a child does not have a blood test for serum creatinine or glucose at a certain time point, the system does not use traditional mean filling or forward resampling for mechanical falsification. Instead, it directly assigns a specific mask placeholder value and simultaneously constructs a binary missing indicator variable, where a value of 1 indicates that the variable is present at the current time, and a value of 0 indicates that the variable is missing at the current time. Through this design, the generated clinical event stream not only includes the reconstructed features of the child but also the explicit variable missing state spatial awareness features, enabling subsequent time-series models to infer the frequency of examinations and the severity of the child's condition based on the missing pattern.

[0036] Combination Figure 2Based on the above embodiments, Embodiment 3 of the present invention further elaborates on the construction of the dynamic interactive event graph and the discretized combination encoding of the ventilation strategy nodes. In actual operation, the dynamic graph construction module 102 constructs the child as a patient node in the topological network and performs discretized combination encoding on the four core control parameters characterizing the ventilation control strategy in the mechanical ventilation parameter record to construct the ventilation strategy node. These four core control parameters are the inhaled oxygen concentration as the first core control parameter, the positive end-expiratory pressure as the second core control parameter, the tidal volume per unit body weight as the third core control parameter, and the ventilation mode as the fourth core control parameter. Since the control parameters of the ventilator are physically continuously changing, directly using them as continuous values ​​for state prediction would result in a large prediction space and reduce the robustness of clinical decision-making. Therefore, this embodiment divides each parameter into intervals within a preset discretized value space. For example, the inhaled oxygen concentration is divided into several discrete intervals with a step size of 5%, such as 21% to 25% and 26% to 30%; the positive end-expiratory pressure is divided into several discrete intervals with a step size of 2 cmH2O, such as 0 to 2 cmH2O and 3 to 4 cmH2O; the tidal volume per unit body weight is statically divided into discrete intervals with a step size of 1 mL / kg; and the ventilation mode is encoded into discrete classification labels according to the actual clinical type, such as synchronized intermittent mandatory ventilation mode, pressure-controlled ventilation mode, and high-frequency oscillatory ventilation mode. The dynamic graph construction module 102 generates a global strategy space consisting of a finite number of discrete strategy states by performing Cartesian product operations on all possible interval combinations of these four core control parameters in the discretized value space. Each unique Cartesian product combination is encoded as a unique ventilation strategy node.

[0037] Combination Figure 2 After the nodes are defined, the dynamic graph construction module 102 uses the ventilation parameter adjustment time as the interaction event timestamp to construct interaction event edges connecting the patient node and the ventilation strategy node. Each ventilation parameter adjustment event is mathematically and topologically parameterized as a quintuple of interaction event edges, the standard mathematical expression of which is shown below:

[0038] in, Indicates the current patient node. This indicates the current ventilation strategy node. This indicates the timestamp of the interaction event at the moment the current ventilation parameter adjustment was performed. This indicates the time interval between the current ventilation parameter adjustment and the child's last interaction event. This represents the concurrent clinical feature vector. This time interval... This is directly used as the edge weight of the current interaction event edge to explicitly represent the time span between two treatment adjustments. The concurrent clinical feature vector... Dynamic construction is performed using the following time-series backtracking mechanism: Dynamic graph construction module 102 uses the current interaction event timestamp. Based on this, no data retrieval in the future is permitted. Instead, retrieval is only performed back along the historical timeline. The most recent single laboratory test record that passes non-missing indicator verification and is earlier than or equal to the timestamp of the current interaction event is selected as the concurrent context feature attribute of the current interaction event edge. This design aligns with the actual decision-making logic of ICU physicians, who, when adjusting ventilator parameters, can only refer to currently available blood gas and biochemical indicators and cannot predict future test results.

[0039] Combination Figure 3 Based on the above embodiments, Embodiment 4 of the present invention elaborates in detail the dual-timescale deep network computing architecture composed of the short-term dynamic modeling module 103 and the long-term memory modeling module 104. To capture the immediate feedback of frequent parameter adjustments on the physiological state of the child within a short period, the short-term dynamic modeling module 103 is configured with a ventilation interaction state space module. When reading the edge of the current interaction event, the system first generates a short-term event representation based on the historical local interaction event stream sequence in the dynamic interaction event graph, and inputs this short-term event representation into the temporal context encoder for context compression. The temporal context encoder is implemented through a set of one-dimensional temporal convolutional networks, thereby obtaining a local historical context vector that incorporates irregular time step changes. Subsequently, the system sets the time interval of the current interaction event... The input is an adaptive parameter generation network, which consists of a multi-layer fully connected neural network and a non-linear activation function. It can dynamically output the continuous parameter matrix of the state space for the current continuous time step according to the span of the time interval. and input mapping matrix To solve the continuous-time state-space equations in a digital computer, the short-term dynamic modeling module 103 uses a zero-order hold operator to perform a time-space optimization on the continuous parameter matrix of the state space. The exponential discretization sampling transformation is used to calculate the discretized state-space parameter matrix, which is calculated in the form of: ;in, Indicates time The discrete state transition matrix, Indicates time Discrete input mapping matrix, Represents a continuous state transition matrix. Represents a continuous input mapping matrix. This represents the identity matrix. Finally, based on this discretized state-space parameter matrix, the system performs a linear mapping and recursive summation between the previously updated hidden state and the currently generated local historical context vector, thereby completing the update of the short-term hidden state of the ventilation strategy node. The recursive form is as follows: ;in, Indicates ventilation strategy node At any moment The updated short-term hidden state. Indicates ventilation strategy node In the short-term hidden state prior to the current interactive event, This represents the local history context vector output by the temporal context encoder.

[0040] Meanwhile, to document the long-term disease progression trajectory of the children and the long-term response patterns of ventilation strategies, the long-term memory modeling module 104 introduces a long-term memory mechanism. The long-term memory modeling module 104 maintains a first long-term memory vector for each patient node and a second long-term memory vector for each ventilation strategy node in the memory. When a current interaction event occurs, the system dynamically updates these long-term memory vectors using a bidirectional cross-writing method. Specifically, the system first reads the current first long-term memory vector corresponding to the patient node and the current second long-term memory vector corresponding to the ventilation strategy node from the global storage. After the gating fusion module 105 calculates the final fusion representations corresponding to the patient node and the ventilation strategy node, the system cross-writes the current ventilation strategy feature information and immediate physiological feedback into the first long-term memory vector corresponding to the patient node, allowing the patient's memory vector to be continuously updated during treatment. Simultaneously, the system cross-writes the current physiological state information of the child into the second long-term memory vector corresponding to the ventilation strategy node, enabling the ventilation strategy node to accumulate its long-term response patterns under different pathological states. In one feasible writing method, the patient node long-term memory vector and the ventilation strategy node long-term memory vector are respectively calculated according to the formula... and Updated, where the meanings of each symbol are the same as those in the aforementioned long-term memory write function.

[0041] Combination Figure 4 Based on the above embodiments, Embodiment 5 of the present invention further elaborates on the mathematical implementation of the gating fusion module 105 and the loss function and training mechanism of the model training module. The gating fusion module 105 utilizes an adaptive gating mechanism to adaptively fuse the short-term hidden state with the long-term memory vector before update, outputting the final fused representations corresponding to the patient node and the ventilation strategy node, respectively, as input for subsequent score prediction. Its fusion calculation process follows mathematical combinatorial logic: the system converts the short-term hidden state... Compared with the long-term memory vector before the update Dense concatenation is performed to form a high-dimensional joint feature vector, which is then input into a multilayer perceptron. The final layer of the multilayer perceptron is then used to perform [the following process / process]: The activation function performs a nonlinear mapping on the joint feature vector, thereby outputting the corresponding dimension-wise gating coefficients. Subsequently, the system generates the final fused representation using the following convex combination formula:

[0042] in, Represents a node At any moment The final fusion indicates that Represents a node At any moment The dimension-wise gating coefficients, Represents a node At any moment The short-term hidden state. Represents a node The long-term memory vector prior to the current interaction event. Indicates the node index. This represents the current moment of the interaction event. Through this convex combination formula, the system can achieve an adaptive dynamic balance between the high-frequency fluctuations of the immediate physiological state and the stable evolution of the long-term disease baseline.

[0043] To enhance the discriminative and generalization performance of the prediction system, the model training module employs a hybrid three-dimensional negative sampling mechanism, using a binary cross-entropy loss function to drive the joint training of the scoring network and the adaptive parameter generation network. This embodiment utilizes a hybrid three-dimensional negative sampling mechanism comprising three sampling dimensions. The first dimension is random negative sampling, where the system randomly extracts strategy combinations from the global ventilation strategy space that have not occurred at the current moment to construct negative samples, establishing a basic discriminative boundary between real ventilation strategies and irrelevant strategies. The second dimension is historical negative sampling, where the system retrieves difficult negative samples from ventilation parameter combinations that the current patient previously used but have ceased or not used at the current interaction timeline. This sampling dimension enables the model to learn and distinguish outdated strategies that were historically feasible but are currently inapplicable. The third dimension is inductive negative sampling, where the system constructs negative samples from ventilation parameter combinations not present in the global historical strategy space during the training phase, improving the model's inductive generalization ability in cold-start scenarios for newly admitted patients. The sampling ratio of the three types of negative samples can be set to 1:1:1, or it can be proportionally expanded according to the number of positive samples in the training set. The binary cross-entropy loss function can be expressed as follows: ;in, Indicates training loss, This represents the training sample set consisting of positive and negative samples. This represents summing over the samples in the training sample set. Representing patient nodes, Indicates the ventilation strategy node, Indicates the moment of the interaction event. This represents the sample label, with a value of 1 for positive samples and a value of 0 for negative samples. This represents the interaction score output by the rating network. Represents a logarithmic function.

[0044] The modules and units involved in the various embodiments of this invention can be implemented in software or hardware, for example, using a combination of electronic components such as application-specific integrated circuits (ASICs) or field-programmable gate arrays (FPGAs). In practical applications, the above functions can be assigned to different functional modules according to specific needs, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above.

Claims

1. A computer-implemented method for predicting pediatric mechanical ventilation strategies, characterized in that, Includes the following steps: Based on the acquired basic information of the child, mechanical ventilation parameter records, and laboratory test indicators, feature reconstruction was performed to obtain reconstructed features characterizing the child's ventilation physiological state and corresponding missing markers, which were then organized into a clinical event flow in chronological order. The child was constructed as a patient node, and multiple core control parameters characterizing the ventilation control strategy in the mechanical ventilation parameter records were discretized and combined to encode them into a ventilation strategy node. Using the ventilation parameter adjustment time as the timestamp of the interaction event, laboratory test records containing at least one non-missing indicator from the preset laboratory test indicators and whose sampling time is earlier than or equal to the timestamp of the current interaction event are matched back in the historical direction as concurrent clinical feature vectors. A dynamic interaction event graph is constructed using the time interval from the child's last interaction event as the edge weight. The state space parameters generated by the adaptive parameter generation network are dynamically adjusted according to the time interval of the current interaction event using the ventilation interaction state space module. The short-term hidden states of the patient node and the ventilation strategy node are updated recursively through discrete continuous-time states based on time intervals. The long-term memory vector of the patient node and the ventilation strategy node is maintained using the long-term memory module. When the current interaction event occurs, the long-term memory vector is dynamically updated through bidirectional cross-writing. The short-term hidden states are adaptively fused with the long-term memory vector before the current interaction event using an adaptive gating mechanism to obtain the final fused representations of the patient node and the ventilation strategy node respectively. The final fusion representation of the patient node and the final fusion representation of each candidate ventilation strategy node are input into a scoring network for correlation measurement, outputting the interaction score of each candidate ventilation strategy at a future time, and outputting a ranking matrix based on the interaction score. This ranking matrix is ​​only used to assist in clinical parameter assessment, does not directly generate disease diagnosis conclusions, and does not directly control ventilator execution parameter adjustments. Before updating the short-term hidden states of the patient node and the ventilation strategy node, the method further includes: generating a short-term event representation based on the historical local interaction event stream sequence in the dynamic interaction event graph, and inputting the short-term event representation into a temporal context encoder for context compression to obtain a local historical context vector that incorporates irregular time step changes. In the ventilation interaction state space module, the dynamic process of performing zero-order preserved discretization processing on the state space parameters and performing state recursion is as follows: the time interval of the current interaction event is... An adaptive parameter generation network is input, dynamically outputting the continuous parameter matrix of the state space at the current continuous time step; the zero-order hold operator is used to perform a time-space parameter transformation on the continuous parameter matrix of the state space. The exponential discretization sampling transformation is used to obtain a discretized state space parameter matrix; based on the discretized state space parameter matrix, the previously updated hidden state is linearly mapped and recursively mapped to the currently generated local history context vector to complete the short-term hidden state update of the patient node and the ventilation strategy node. The long-term memory module maintains a first long-term memory vector for the patient node and a second long-term memory vector for the ventilation strategy node. When the current interaction event occurs, the bidirectional dynamic update includes: reading the current first long-term memory vector corresponding to the patient node and the current second long-term memory vector corresponding to the ventilation strategy node; after calculating the final fusion representations corresponding to the patient node and the ventilation strategy node using an adaptive gating mechanism, the current ventilation strategy feature information is cross-written into the first long-term memory vector corresponding to the patient node to accumulate the disease trajectory, and the current physiological state information of the child is simultaneously cross-written into the second long-term memory vector corresponding to the ventilation strategy node to continuously update the long-term response mode of the ventilation strategy corresponding to the current ventilation strategy node under different pathological states.

2. The computer-implemented pediatric mechanical ventilation strategy prediction method according to claim 1, characterized in that, The process of obtaining the reconstructed features includes: screening pediatric patients under the age of 18 years and removing mechanical ventilation parameter records whose values ​​exceed a preset physiological threshold range or whose missing percentage exceeds a preset missing percentage threshold, wherein the preset missing percentage threshold is 50%; converting the tidal volume in the mechanical ventilation parameter records according to the ideal weight in the patient's basic information to obtain the tidal volume per unit weight feature; and constructing the oxygenation index feature characterizing the oxygenation status by using the inhaled oxygen concentration in the mechanical ventilation parameter records and the arterial oxygen partial pressure in the laboratory test indicators.

3. The computer-implemented pediatric mechanical ventilation strategy prediction method according to claim 1, characterized in that, The laboratory testing indicators include: Arterial oxygen partial pressure Arterial carbon dioxide partial pressure The parameters include blood lactate, white blood cell count, hemoglobin, platelet count, blood glucose, and serum creatinine; wherein, during the feature reconstruction process, the original missing markers of laboratory test indicator variables that have not been sampled or are missing are retained, so that the clinical event stream contains explicit spatially aware features about the missing variable status.

4. The computer-implemented pediatric mechanical ventilation strategy prediction method according to claim 1, characterized in that, The multiple core control parameters include four core control parameters: the inhaled oxygen concentration as the first core control parameter, the positive end-expiratory pressure as the second core control parameter, the tidal volume per unit body weight as the third core control parameter, and the ventilation mode as the fourth core control parameter; the ventilation strategy node is obtained by combining and encoding the Cartesian product of the first, second, third, and fourth core control parameters in the discretized value space.

5. The computer-implemented pediatric mechanical ventilation strategy prediction method according to claim 1, characterized in that, Each ventilation parameter adjustment event in the dynamic interaction event graph is parameterized as an interaction event edge 5-tuple: in, Indicates the current patient node. This indicates the current ventilation strategy node. This indicates the timestamp of the interaction event at the moment the current ventilation parameter adjustment was performed. This indicates the time interval between the current ventilation parameter adjustment and the child's last interaction event; the concurrent clinical feature vector Constructed in the following way: using the current interaction event timestamp Based on the historical timeline, the search is only performed backtracking, and the single laboratory test record that is closest to the current time, contains at least one non-missing indicator among the preset laboratory test indicators, and whose sampling time is earlier than or equal to the current interaction event timestamp is used as the concurrent context feature attribute of the current interaction event edge.

6. The computer-implemented pediatric mechanical ventilation strategy prediction method according to claim 1, characterized in that, The fusion calculation process of the adaptive gating mechanism follows the following mathematical logic: the short-term hidden state Compared with the long-term memory vector before the update Dense splicing is performed, and the data is input into a multilayer perceptron; the data is then processed using the external connection of the last layer of the multilayer perceptron. The activation function performs a nonlinear mapping and outputs the corresponding dimension-wise gating coefficients. ; Using the aforementioned dimension-by-dimensional gating coefficient The following convex combination formula is applied to the short-term hidden state and long-term memory vectors to calculate and output the final fused representations corresponding to the patient node and the ventilation strategy node, respectively. : ,in, Represents a node At any moment The final fusion indicates that Represents a node At any moment The dimension-wise gating coefficients, Represents a node At any moment The short-term hidden state. Represents a node The long-term memory vector prior to the current interaction event. Indicates the node index. Indicates the current moment of the interactive event.

7. The computer-implemented pediatric mechanical ventilation strategy prediction method according to claim 1, characterized in that, The training process of the scoring network and the adaptive parameter generation network is driven by a binary cross-entropy loss function that includes a hybrid three-dimensional negative sampling mechanism. The hybrid three-dimensional negative sampling mechanism includes: random negative sampling to establish the basic discriminative boundary between real ventilation strategies and irrelevant strategies; historical negative sampling to extract difficult negative samples from ventilation parameter combinations that the current patient has used in the historical disease timeline but not at the current interaction time, to distinguish outdated strategies that were historically feasible but are not applicable now; and inductive negative sampling to construct negative samples from ventilation parameter combinations that have not appeared in the global historical strategy space during the training phase, to improve the model's inductive generalization ability in cold start scenarios for newly admitted patients.

8. A dual-timescale continuous-time dynamic graph mechanical ventilation strategy prediction system, employing the computer-implemented pediatric mechanical ventilation strategy prediction method as described in any one of claims 1 to 7, characterized in that, include: The feature reconstruction module is used to reconstruct features based on the acquired basic information of the child, mechanical ventilation parameter records, and laboratory test indicators, to obtain reconstructed features that characterize the child's ventilation physiological state and corresponding missing markers, and organize them into a clinical event flow in chronological order. The dynamic graph construction module is used to construct the child as a patient node and to discretize and combine multiple core control parameters that characterize the ventilation control strategy in the mechanical ventilation parameter record to construct a ventilation strategy node. Using the ventilation parameter adjustment time as the timestamp of the interaction event, the system backtracks to match laboratory test records that contain at least one non-missing indicator from the preset laboratory test indicators and whose sampling time is earlier than or equal to the timestamp of the current interaction event as concurrent clinical feature vectors. The system also constructs a dynamic interaction event graph using the time interval from the child's last interaction event as the edge weight. The short-term dynamic modeling module is used to dynamically adjust the state space parameters generated by the adaptive parameter generation network based on the time interval of the current interaction event using the ventilation interaction state space module. The system also updates the short-term hidden states of the patient node and the ventilation strategy node by recursively updating the discretized continuous-time state based on the time interval. The long-term memory modeling module is used to maintain the long-term memory vectors of the patient node and the ventilation strategy node using the long-term memory module, and dynamically update the long-term memory vectors through bidirectional cross-writing when the current interaction event occurs. The gating fusion module is used to adaptively fuse the short-term hidden state with the long-term memory vector before the update using an adaptive gating mechanism to obtain the final fused representations corresponding to the patient node and the ventilation strategy node, respectively. The prediction output module is used to input the final fusion representation of the patient node and the final fusion representation of each candidate ventilation strategy node into the scoring network for correlation measurement, output the interaction score of each candidate ventilation strategy at a future time, and output a ranking matrix based on the interaction score. The model training module is used to drive the training of the scoring network and the adaptive parameter generation network using a binary cross-entropy loss function that includes a hybrid three-dimensional negative sampling mechanism.

Images