An artificial intelligence auxiliary-based pediatric patient extubation timing prediction method and system

Abstract

This invention discloses an artificial intelligence-assisted method and system for predicting the timing of extubation in pediatric patients, belonging to the field of clinical medical technology. It includes sample information collection: collecting large-sample clinical data of respiratory waveform signals, facial expression video streams, and vital sign data of the child to obtain sample information; and sample information processing: preprocessing the sample information and establishing a repository to store the preprocessed sample information to obtain a database. This invention uses a set state setting method combined with a model generation method, and integrates high-precision ventilator data, child facial expression video data, and vital sign data to predict the timing of extubation, thereby replacing the current subjective judgment method. This provides specific computer-aided decision support for sequential mechanical ventilation treatment during the resuscitation and extubation phase. Simultaneously, different timing prediction models are trained according to different child characteristics to improve the accuracy of computer-aided decision support.

Description

Technical Field

[0001] This invention relates to the field of clinical medical technology, specifically to a method and system for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance. Background Technology

[0002] When a child develops a primary disease that affects respiratory function, intubation is often required for treatment. Once the primary disease is controlled and respiratory function and airway protection are restored, the benefits of continuing to rely on intubation or artificial airway support decrease, while exposure to intubation-related risks (such as airway mucosal damage, increased risk of infection / pneumonia, vocal cord or tracheal complications, airway spasm and secretion retention, and complications related to long-term ventilation) gradually increase.

[0003] Extubation can reduce these complications and promote the recovery of spontaneous breathing and improved mobility in children. Therefore, it should be performed only after evidence-based assessment that the child is safe for extubation, and close monitoring should be maintained after extubation to ensure airway patency. The timing of postoperative extubation has always been a challenge for anesthesiologists. Current expert consensus on extubation criteria includes: recovery of consciousness, ability to open eyes and follow instructions; recovery of airway protective reflexes, such as swallowing and coughing; and regular spontaneous breathing with adequate minute ventilation. It can be seen that these criteria are largely based on the subjective judgment of physicians, and there is still no scientific and objective decision-making standard for children.

[0004] Unlike adults, children cannot cooperate with commands. Muscle relaxation recovery cannot be assessed by checking muscle strength (e.g., raising both hands or gripping the doctor's fingers tightly), nor can their consciousness be determined by obeying commands to open their eyes. Because it's difficult to determine the appropriate time for extubation, medical staff may increase the use of sedatives and muscle relaxants to reduce patient discomfort, potentially leading to serious consequences such as respiratory depression. Relying on the doctor's personal experience and subjective judgment to determine extubation timing inevitably increases clinical insecurity. Therefore, optimizing extubation strategies is crucial for reducing the incidence and mortality of perioperative anesthetic complications in children, improving clinical outcomes, reducing hospital stays, and lowering medical costs. This invention is proposed based on this principle. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-assisted method and system for predicting the timing of extubation in pediatric patients, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance, the method comprising:

[0007] Sample information collection: Collect large-sample clinical data, including respiratory waveform signals, facial expression video streams, and vital sign data of the children, to obtain sample information;

[0008] Sample information processing: Based on the sample information, preprocessing is performed and a repository is established to store the preprocessed sample information to obtain a database;

[0009] Respiratory status setting: Based on the database, the respiratory status of the child under different characteristics is set through the status setting method to obtain the target characteristics and respiratory status. The respiratory status includes the respiratory non-recovery period, the spontaneous breathing recovery period, and the spontaneous breathing period.

[0010] Feature extraction: Based on the target features, the temporal relationship under different respiratory states is explored through the association exploration method to obtain the association information under the target features;

[0011] Training information generation: Based on the association information and the database, the corresponding information is extracted from the information in the database using the information generation method to obtain the training information;

[0012] Base model generation: Based on training information, a timing prediction model under the target features is generated using model generation methods;

[0013] Timing prediction analysis: The timing prediction model based on target features is used to predict the timing of extubation in pediatric patients.

[0014] Furthermore, the process of preprocessing sample information and establishing a repository to store the preprocessed sample information to obtain a database is as follows: Information preprocessing is performed based on the sample information to obtain preprocessed sample information, which includes respiratory information, facial expression information, and vital sign information. A correlation is established between the respiratory information, facial expression information, and vital sign information. Based on the correlation, the respiratory information, facial expression information, and vital sign information of a single child are extracted to obtain subsample information. Basic information of the child corresponding to the subsample information is obtained. Basic information includes extubation information and child characteristics, including the child's age, gender, and weight. A correlation is established between the basic information and the subsample information. A database is established to store the subsample information, basic information, and correlation.

[0015] Furthermore, the process of pre-setting the child's respiratory status is as follows: setting a basic status, which includes the non-recovery respiratory status, the positive recovery respiratory status, and the recovery respiratory status; splitting the basic information to obtain extubation information and child characteristics; selecting the child characteristics to obtain target features; extracting subsample information with the same target features from the database to obtain relevant information; filling the non-recovery respiratory status, the positive recovery respiratory status, and the recovery respiratory status with relevant information based on the extubation information to obtain the non-recovery respiratory period, the spontaneous breathing recovery period, and the spontaneous breathing period; and integrating the non-recovery respiratory period, the spontaneous breathing recovery period, and the spontaneous breathing period to obtain the respiratory status of the target features.

[0016] Furthermore, the association exploration method includes: extracting subsample information corresponding to the target feature from the database to obtain selection information; determining the position of the selection information in the respiratory state to obtain the selection position; determining the time of the selection position to obtain the selection time; extracting the extubation time from the extubation information corresponding to the selection information to obtain the target time; obtaining the difference between the selection time and the target time to obtain the predicted time; and integrating the predicted time and the selection information to obtain association information.

[0017] Furthermore, the information generation method includes: splitting related information to obtain selected information and prediction time; determining whether there is a regular change between different selected information and prediction time to obtain a judgment result; when the judgment result indicates that there is a regular change between different selected information and prediction time, extracting the regular change to derive information from the selected information and prediction time to obtain derived information and derived time; integrating the selected information, prediction time, derived information, and derived time to obtain training information; when the judgment result indicates that there is no regular change between different selected information and prediction time, the selected information and prediction time are used as training information.

[0018] Furthermore, the model generation method includes:

[0019] Data acquisition and processing: The training information is split to obtain the selected information and the prediction time. The correspondence between the selected information and the prediction time is established. The selected information, the prediction time and the correspondence are integrated to obtain the training data. The selected information is the imported feature and the prediction time is the exported feature.

[0020] Model selection and training: Select a deep learning model as the model base, and import the imported and exported features into the model base for the model base to learn and obtain the initial model;

[0021] Model adjustment and optimization: Preset validation information and validation time, import the validation information into the initial model to obtain the result time, extract the difference between the result time and the validation time to obtain the difference information, and adjust and optimize the initial model based on the difference information to obtain the timing prediction model;

[0022] Model attribution determination: Based on training information, target features are determined, and the timing prediction model is the timing prediction model under the target features.

[0023] Furthermore, the timing prediction method includes: obtaining information about pediatric patients requiring extubation timing prediction to obtain information to be predicted, the information to be predicted including feature information and basic information, the information items of the basic information corresponding to the information items of the sample information, the information items of the feature information corresponding to the information items of the child's features, extracting a timing prediction model under the target feature corresponding to the feature information to obtain a selection model, importing the basic information into the selection model to obtain the prediction time, extracting the respiratory state under the target feature corresponding to the feature information to obtain a selection state, and performing information matching between the basic information and several states in the selection state to obtain a target state, the target state being one of the following: respiratory non-recovery period, spontaneous breathing recovery period, and spontaneous breathing period.

[0024] An AI-assisted pediatric patient extubation timing prediction system is provided, which uses the aforementioned AI-assisted pediatric patient extubation timing prediction method.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] This AI-assisted method and system for predicting the timing of extubation in pediatric patients uses a state setting method combined with a model generation method to introduce multimodal data intelligent analysis. It combines high-precision ventilator data, facial expression video data of the child, and vital sign data to predict the timing of extubation, thus replacing the current subjective judgment method. This provides specific computer-aided decision support for sequential mechanical ventilation treatment during the resuscitation and extubation phase. At the same time, different timing prediction models are trained according to different child characteristics to improve the accuracy of computer-aided decision support.

[0027] Meanwhile, by setting different characteristics of patients through the set state setting method, and setting their respiratory status according to different characteristics of patients, in the subsequent extubation timing prediction, the respiratory status of the patient can be quickly determined by analyzing the information to be predicted of the patient, so as to assist clinicians in judging the timing of extubation. Through the set timing prediction model, the timing of extubation can be predicted based on the patient's specific breathing, facial video and vital signs information, and the specific extubation timing time can be output to further assist clinicians in performing the extubation operation.

[0028] Meanwhile, by setting the information generation method, it is determined whether there is a pattern between different selection times and prediction times under the same target features. If there is a pattern, information derivation is performed on the selection time and prediction time according to the pattern to complete the generation of derived information and derived time, so as to expand the richness of training information and increase the training sample size for subsequent model training. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0030] Figure 2 This is a schematic diagram of the regularity judgment structure of the present invention;

[0031] Figure 3 This is a schematic diagram of the model generation structure of the present invention;

[0032] Figure 4 This is a schematic diagram of the correlation exploration method of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Extubation in pediatric patients (usually referring to the removal of endotracheal intubation or tracheostomy-related ventilation support) is a comprehensive assessment process. The core is to determine whether the child has the ability to breathe spontaneously and maintain airway patency. This requires assessing stable vital signs, improved respiratory reserve and ventilation / oxygenation indicators, recovery of cough and swallowing reflexes, and sufficient consciousness and muscle strength to maintain airway patency. At the same time, a ventilation test or extubation assessment is usually performed (if possible, observe inspiratory flow rate, work of breathing, and, if necessary, perform an airway leak test). After extubation, close monitoring of respiratory rate, blood oxygen saturation, work of breathing, and signs of stridor / aspiration is necessary. Sedation and analgesia strategies, airway suction, and necessary nebulization / oxygenation support should also be implemented before and after extubation to reduce the risk of reintubation and respiratory failure.

[0035] Example 1:

[0036] like Figures 1-4 As shown, the present invention provides a technical solution: a method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance, the method comprising:

[0037] Sample information collection: Collect large-sample clinical data, including respiratory waveform signals, facial expression video streams, and vital sign data of the children, to obtain sample information;

[0038] It is important to note that respiratory waveform information includes tidal volume, minute ventilation, airway pressure, flow rate, FiO2, etc., while facial expression video streams include subtle facial features and expression changes. These data are used to extract important features such as the child's swallowing reflex, facial movements, and pain response. Vital signs data include key indicators such as heart rate, blood pressure, and blood oxygen saturation. The specific collection process can be based on information from previous extubation procedures of children, and sample information can also be obtained before constructing this method. Specifically, respiratory data acquisition involves connecting a dedicated BicoreII data acquisition hardware based on a programmable microcontroller to the data transmission port of the respiratory platform for parameter acquisition. A unified data exchange protocol is written in Java and data is acquired and transmitted through the open data port of the respiratory platform. Data is written to a MongoDB database at the data aggregation terminal, and interface with a big data system is achieved through a software interface. Data acquisition employed a high-density, equally spaced sampling pattern, recording all parameters of the respiratory platform every 20ms. This density allowed for a near-complete recording and reconstruction of the detailed characteristics of respiratory parameters for each respiratory cycle. Simultaneously, key treatment procedures and time points prior to PACU were collected based on the resuscitation log. Facial video data acquisition utilized a high-resolution medical-grade camera to capture facial video data of the infant throughout the resuscitation phase, highlighting subtle facial features and expression changes. This data was used to extract crucial features such as the infant's swallowing reflex, facial movements, and pain response. The acquired video data will be processed and encoded in real time and stored on a high-speed, high-capacity solid-state drive for subsequent cutting, analysis, and annotation. Each video file will be precisely labeled, including the child's identification information, acquisition timestamp, and other key information, so that the facial video data can be quickly and accurately retrieved and used when needed. Vital signs collection: The child's vital signs data, including key indicators such as heart rate, blood pressure, and blood oxygen saturation, will be monitored in real time by a monitor and recorded automatically by software. These vital signs data will be collected in 5-minute cycles and stored periodically in the surgical anesthesia database for later data analysis and retrospective studies.

[0039] Sample information processing: Based on the sample information, preprocessing is performed and a repository is established to store the preprocessed sample information to obtain a database;

[0040] It should be noted that the preprocessing of the sample information uses the first-digit interpolation method to correct any minor defects, duplications, or significant anomalies (such as significant data fluctuations caused by treatment procedures) that may exist within the data.

[0041] Respiratory status setting: Based on the database, the respiratory status of the child under different characteristics is set through the status setting method to obtain the target characteristics and respiratory status. The respiratory status includes the respiratory non-recovery period, the spontaneous breathing recovery period, and the spontaneous breathing period.

[0042] It is important to note that by setting different characteristics of patients through the state setting method, and setting their respiratory status according to different characteristics, it is possible to build a prediction model for the timing of extubation for patients with different characteristics.

[0043] Feature extraction: Based on the target features, the temporal relationship under different respiratory states is explored through the association exploration method to obtain the association information under the target features;

[0044] It is important to note that by exploring the temporal relationship between extubation timing under different respiratory states using the established correlation exploration method, correlation information under the target features can be obtained. This correlation information is then used to facilitate subsequent information generation.

[0045] Training information generation: Based on the association information and the database, the corresponding information is extracted from the information in the database using the information generation method to obtain the training information;

[0046] It is important to note that by using the information generation method set up, the associated information is combined with the database to extract information for model training. The training information is then used to build the timing prediction model in the future.

[0047] Base model generation: Based on training information, a timing prediction model under the target features is generated using model generation methods;

[0048] It is important to note that the timing prediction model under the target features is obtained by training the model based on the training information through the set model generation method.

[0049] Timing prediction analysis: The timing prediction model based on target features is used to predict the timing of extubation in pediatric patients.

[0050] It is important to note that the timing prediction model is used to predict the timing of extubation for pediatric patients based on the timing prediction method. The prediction result is how much time after the current time, such as the current time, 10 minutes later, 20 minutes later, 30 minutes later, and 60 minutes later, when extubation can be performed.

[0051] like Figure 1As shown, the process of preprocessing sample information and establishing a database to store the preprocessed sample information is as follows: Preprocessing is performed on the sample information to obtain preprocessed sample information, which includes respiratory information, facial expression information, and vital sign information. A correlation is established between respiratory information, facial expression information, and vital sign information. Based on this correlation, the respiratory information, facial expression information, and vital sign information of a single child are extracted to obtain subsample information. Basic information of the child corresponding to the subsample information is obtained. Basic information includes extubation information and child characteristics, including the child's age, gender, and weight. A correlation is established between the basic information and the subsample information. A database is established to store the subsample information, basic information, and correlation.

[0052] It is important to note that the preprocessing process involves using front-end numerical interpolation to correct for minor omissions, duplications, and significant anomalies (such as significant data fluctuations caused by treatment procedures) that may exist within the data. The preprocessing process also includes, but is not limited to, data cleaning and noise reduction. The process of establishing correlations between respiratory information, facial expression information, and vital signs information involves associating information obtained from the same child to facilitate subsequent information retrieval. Based on these correlations, the process of extracting respiratory, facial expression, and vital signs information from individual children to obtain subsample information is also included. Since the database contains multiple subsample information, the process of obtaining the basic information of the children corresponding to the subsample information involves identifying the corresponding children based on the subsample information and obtaining extubation information based on when the child was successfully extubated.

[0053] like Figure 1 As shown, the process of presetting the respiratory status of the child is as follows: a basic status is set, which includes the non-recovery respiratory status, the positive recovery respiratory status, and the recovery respiratory status. The basic information is split to obtain extubation information and child characteristics. The child characteristics are selected to obtain target features. Subsample information with the same target features is extracted from the database to obtain relevant information. Based on the extubation information, the relevant information is filled into the non-recovery respiratory status, the positive recovery respiratory status, and the recovery respiratory status to obtain the non-recovery respiratory period, the spontaneous breathing recovery period, and the spontaneous breathing period. The non-recovery respiratory period, the spontaneous breathing recovery period, and the spontaneous breathing period are integrated to obtain the respiratory status of the target features.

[0054] It is important to note that the process of selecting patient characteristics to obtain target features can be randomized, but it is necessary to select all patient characteristics as target features. The process of extracting sub-sample information with the same target features from the database to obtain relevant information involves extracting sub-sample information with the same target features, converting all sub-sample information into relevant information, and adding the respiratory information range, video information range, and vital sign range composed of relevant information to the corresponding base state to obtain the respiratory state. The respiratory state is a set of states with specific range information. Specific ranges for the non-recovery breathing period, spontaneous breathing recovery period, and spontaneous breathing period are as follows: Non-recovery breathing period: Respiratory waveform: Inspiratory trigger delay >150ms in pressure support mode, end-expiratory carbon dioxide waveform variability (CV) >30%, flow-time curve shows sawtooth asynchronous wave (AsynchronyIndex ≥25%); Facial expression: Eyelid closure >80% (infrared thermal imaging analysis), facial... Muscle EMG activity <10uV, no pain expression (NPIS score ≤2); Vital signs: respiratory rate <10 breaths / min (newborn) or <15 breaths / min (child), diaphragmatic electrical activity Edi <3uV, oxygenation index (OI) >8; Spontaneous breathing recovery period: respiratory waveform: spontaneous breathing trigger success rate >70%, shallow and rapid breathing index (RSBI) <8 (child) / <5 (newborn), synchronicity of chest and abdominal movements (phase difference <15°), facial expression: periodicity Frequent frowning (frequency 0.1-0.3Hz), nasal flaring index (NFLI) 2-4, brief appearance of distress expression (NPIS score 3-5); Vital signs: Respiratory rate: neonates 40-60 breaths / min, children 25-35 breaths / min, tidal volume 6-8 ml / kg (variability <20%), heart rate variability (HRV) LF / HF ratio 1.5-2.5; Respiratory waveform: continuous for 1 hour without ventilator triggering, end-expiratory lung impedance change rate <10%.

[0055] Airway closure index (ACI) <0.15; Facial expression: relaxed face (EMG 15-25uV), spontaneous smile (frequency >1 time / 5 minutes), NPIS score ≤1; Vital signs: oxygen saturation >95% (FiO2 ≤0.4), end-tidal carbon dioxide 35-45 mmHg (normal waveform), absence of paradoxical chest and abdominal movements.

[0056] like Figure 1 and Figure 4 As shown, the association exploration method includes: extracting subsample information corresponding to the target feature from the database to obtain selection information, determining the position of the selection information in the respiratory state to obtain the selection position, determining the time of the selection position to obtain the selection time, extracting the extubation time from the extubation information corresponding to the selection information to obtain the target time, obtaining the difference between the selection time and the target time to obtain the predicted time, and integrating the predicted time and selection information to obtain association information.

[0057] It is important to note that the process of determining the location of the selected information within the respiratory state to obtain the selection location involves determining the selection location based on the time the selected information reflects within a specific respiratory state. The process of determining the selection time within the selection location involves determining the generation time of the selected information. The process of extracting the extubation time from the extubation information corresponding to the selected information to obtain the target time involves extracting the extubation time from the selected information. By analyzing the interval between the selection time and the target time, the predicted time for extubation can be obtained. Figure 4 As shown, Figure 4 The selected time corresponds to the generation time of the selected information, and the separator points represent the time separation nodes for different breathing states.

[0058] like Figure 1 and Figure 2 As shown, the information generation method includes: splitting related information to obtain selected information and prediction time; judging whether there is a regular change between different selected information and prediction time to obtain a judgment result; when the judgment result indicates that there is a regular change between different selected information and prediction time, extracting the regular change to derive information from the selected information and prediction time to obtain derived information and derived time; integrating the selected information, prediction time, derived information, and derived time to obtain training information; when the judgment result indicates that there is no regular change between different selected information and prediction time, the selected information and prediction time are used as training information.

[0059] It is important to note that the process of determining whether there is a regular change in the selected information and prediction time can be approached by grouping individual selected information and prediction time into groups. For each group, the "selected information parameter—corresponding prediction time—result indicator (such as error / accuracy)" can be compiled into uniform data. Then, a graph (e.g., plotting indicator curves for each group according to prediction time, or creating a heatmap to show the differences between groups in different prediction time periods) can be used to visually determine whether the differences between groups systematically increase / converge or trend with prediction time and selected parameters. Further, statistical modeling can be used to formally test this "regular change." For example, the selected information parameter, prediction time, and their interaction terms can be included in the model, with a focus on testing the significance of the interaction term "group × prediction time" (and, if necessary, "group × selected information parameter × prediction time"). If the interaction term is significant and the effect size indicates that the differences between groups change with prediction time and selected parameters, then a regular change can be considered to exist between different groups. When a regular change exists, the selected information and prediction time within each group can be further enriched with this regular change to enrich the sample information. Figure 2 As shown, the associated information includes multiple different selection information and the prediction time corresponding to the selection information.

[0060] like Figure 1 and Figure 3 As shown, the model generation methods include:

[0061] Data acquisition and processing: The training information is split to obtain the selected information and the prediction time. The correspondence between the selected information and the prediction time is established. The selected information, the prediction time and the correspondence are integrated to obtain the training data. The selected information is the imported feature and the prediction time is the exported feature.

[0062] It is important to note that by establishing a correspondence between the selected information and the predicted time, it is easy to extract the corresponding selected information and predicted time as sub-training data. The selected information includes respiratory information, facial feature information, and vital sign information, while the derived feature is the specific time.

[0063] Model selection and training: Select a deep learning model as the model base, and import the imported and exported features into the model base for the model base to learn and obtain the initial model;

[0064] It should be noted that the model base can be a support vector machine or a recurrent neural network, etc., and the initial model is obtained by sequentially importing the sub-training data into the model base for training.

[0065] Model adjustment and optimization: Preset validation information and validation time, import the validation information into the initial model to obtain the result time, extract the difference between the result time and the validation time to obtain the difference information, and adjust and optimize the initial model based on the difference information to obtain the timing prediction model;

[0066] It is important to note that the preset verification information and verification time must conform to the target features. That is, the verification information and verification time are obtained by acquiring respiratory information, facial feature information, vital sign information, and the difference between the information acquisition time and the successful extubation time under the target features. The verification information is imported into the initial model to complete the model training. By comparing the two results, the training of the initial model and its parameters are optimized and adjusted according to the differences in the results to obtain the timing prediction model.

[0067] Model attribution determination: Based on training information, target features are determined, and the timing prediction model is the timing prediction model under the target features.

[0068] It is important to note that a relationship needs to be established between the target features and the timing prediction model in order to find and select the appropriate timing prediction model in subsequent use.

[0069] like Figure 1As shown, the timing prediction method includes: obtaining information about pediatric patients requiring extubation timing prediction to obtain information to be predicted, which includes feature information and basic information. The information items of the basic information correspond to the information items of the sample information, and the information items of the feature information correspond to the information items of the child's features. The timing prediction model under the target feature corresponding to the feature information is extracted to obtain the selection model. The basic information is imported into the selection model to obtain the prediction time. The respiratory state under the target feature corresponding to the feature information is extracted to obtain the selection state. The basic information is matched with several states in the selection state to obtain the target state. The target state is one of the following: the respiratory non-recovery period, the spontaneous breathing recovery period, and the spontaneous breathing period.

[0070] It is important to note that the process of obtaining information on pediatric patients who need to predict the timing of extubation is the information acquisition process. Specifically, it involves obtaining information on children who need to predict the timing of extubation, matching the basic information with several selected states to obtain the target state, and quickly matching the child's respiratory status through information matching to facilitate real-time judgment by clinicians.

[0071] Example 2:

[0072] The timing of extubation for children is predicted by analyzing their respiratory waveform signals, facial expression video streams, and vital sign data separately.

[0073] Analysis of respiratory parameters: Preprocessing the data using a bandpass filter effectively removes high-frequency noise and low-frequency drift, making the main characteristics of the signal clearer. After filtering, methods such as moving average or median filtering are used to further reduce signal fluctuations. Moving average smooths the signal and eliminates the influence of occasional extreme values ​​by averaging local data points. Median filtering reduces noise by taking the median of data within a local window, and is particularly suitable for removing spike noise. These methods extract representative features such as the signal's mean, standard deviation, peak value, and volatility. After feature extraction, deep learning algorithms are used for data analysis.

[0074] Video image analysis: The video file was read and frames extracted using Python's OpenCV library. The Haarcascade frontal face detector provided by OpenCV was used to detect faces from the images. The extracted facial images underwent preprocessing, including grayscale conversion, histogram equalization, mean filtering, and smoothing, to reduce the impact of lighting conditions and noise, thereby improving the quality of subsequent feature extraction. Features were extracted from the preprocessed images using deep learning techniques. Initial feature extraction was performed using a VGG-based network, or a hybrid model such as the Facial paindetection vision transformer (FPD-ViT) was employed to better capture global information and local details. These features were then integrated for a more comprehensive understanding of facial expression dynamics. Finally, the extracted features were fed into a classifier, such as a Transformer or a hybrid CNN-LSTM model, to identify the different breathing states represented by facial expressions.

[0075] Analysis of vital signs: Filters are used to remove noise and interference from vital sign data (such as heart rate, blood pressure, and blood oxygen saturation). Representative physiological features are extracted from the processed data. After feature extraction, machine learning or deep learning algorithms are applied for data analysis. Models are built using algorithms such as Support Vector Machines (SVM). SVM can identify heart rate, blood pressure, and blood oxygen saturation characteristics under different physiological states by inputting the extracted features into a classifier, thereby predicting whether the patient has regained spontaneous breathing.

[0076] Multimodal Feature Extraction: The Mixture of Experts (MoE) mechanism improves prediction performance and computational efficiency by combining multiple expert models. First, multiple independent expert models are constructed based on different modalities of data, each expert specializing in a specific type of data. A gating network is introduced to dynamically select the most relevant expert based on the features of the input data. The selected expert uses a recurrent neural network (RNN) to extract modal features, which are then processed by a multilayer perceptron (MLP) to generate a comprehensive output. This mechanism effectively combines the strengths of various experts, making MoE perform exceptionally well in handling complex tasks, especially in large-scale data processing. The features of each modality are weighted against the confidence scores of the evaluated experts to obtain the prediction target: the probability of extubation at the current time, 10 minutes, 20 minutes, 30 minutes, and 60 minutes later, as well as auxiliary output indicators such as facial expression classification and current breathing state classification.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance, the method comprising: Sample information collection: Collect large-sample clinical data, including respiratory waveform signals, facial expression video streams, and vital sign data of the children, to obtain sample information; The method is characterized in that it further includes: Sample information processing: Based on the sample information, preprocessing is performed and a repository is established to store the preprocessed sample information to obtain a database; Respiratory status setting: Based on the database, the respiratory status of the child under different characteristics is set through the status setting method to obtain the target characteristics and respiratory status. The respiratory status includes the respiratory non-recovery period, the spontaneous breathing recovery period, and the spontaneous breathing period. Feature extraction: Based on the target features, the temporal relationship under different respiratory states is explored through the association exploration method to obtain the association information under the target features; Training information generation: Based on the association information and the database, the corresponding information is extracted from the information in the database using the information generation method to obtain the training information; Base model generation: Based on training information, a timing prediction model under the target features is generated using model generation methods; Timing prediction analysis: The timing prediction model based on target features is used to predict the timing of extubation in pediatric patients.

2. The method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance according to claim 1, characterized in that: The process of preprocessing sample information and establishing a database to store the preprocessed sample information is as follows: Preprocessing is performed on the sample information to obtain preprocessed sample information, which includes respiratory information, facial expression information, and vital sign information. Correlation relationships are established between respiratory information, facial expression information, and vital sign information. Based on these relationships, the respiratory information, facial expression information, and vital sign information of individual children are extracted to obtain subsample information. Basic information of the children corresponding to the subsample information is obtained, including extubation information and child characteristics, such as age, gender, and weight. Correlation relationships are established between the basic information and the subsample information. A database is then established to store the subsample information, basic information, and correlation relationships.

3. The method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance according to claim 2, characterized in that: The process of pre-setting the respiratory status of the child is as follows: Set a basic status, which includes the non-recovery respiratory status, the positive recovery respiratory status, and the recovery respiratory status; break down the basic information to obtain extubation information and child characteristics; select the child characteristics to obtain target features; extract subsample information with the same target features from the database to obtain relevant information; fill the non-recovery respiratory status, the positive recovery respiratory status, and the recovery respiratory status with the relevant information based on the extubation information to obtain the non-recovery respiratory period, the spontaneous breathing recovery period, and the spontaneous breathing period; integrate the non-recovery respiratory period, the spontaneous breathing recovery period, and the spontaneous breathing period to obtain the respiratory status of the target features.

4. The method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance according to claim 1, characterized in that: The association exploration method includes: extracting subsample information corresponding to the target feature from the database to obtain selection information; determining the position of the selection information in the respiratory state to obtain the selection position; determining the time of the selection position to obtain the selection time; extracting the extubation time from the extubation information corresponding to the selection information to obtain the target time; obtaining the difference between the selection time and the target time to obtain the predicted time; and integrating the predicted time and the selection information to obtain association information.

5. The method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance according to claim 1, characterized in that: The information generation method includes: splitting related information to obtain selected information and prediction time; judging whether there is a regular change between different selected information and prediction time to obtain a judgment result; when the judgment result indicates that there is a regular change between different selected information and prediction time, extracting the regular change to derive information from the selected information and prediction time to obtain derived information and derived time; integrating the selected information, prediction time, derived information, and derived time to obtain training information; when the judgment result indicates that there is no regular change between different selected information and prediction time, the selected information and prediction time are used as training information.

6. The method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance according to claim 1, characterized in that: The model generation method includes: Data acquisition and processing: The training information is split to obtain the selected information and the prediction time. The correspondence between the selected information and the prediction time is established. The selected information, the prediction time and the correspondence are integrated to obtain the training data. The selected information is the imported feature and the prediction time is the exported feature. Model selection and training: Select a deep learning model as the model base, and import the imported and exported features into the model base for the model base to learn and obtain the initial model; Model adjustment and optimization: Preset validation information and validation time, import the validation information into the initial model to obtain the result time, extract the difference between the result time and the validation time to obtain the difference information, and adjust and optimize the initial model based on the difference information to obtain the timing prediction model; Model attribution determination: Based on training information, target features are determined, and the timing prediction model is the timing prediction model under the target features.

7. The method for predicting the timing of extubation in pediatric patients based on artificial intelligence assistance according to claim 1, characterized in that: The timing prediction method includes: obtaining information about pediatric patients requiring extubation timing prediction to obtain information to be predicted, which includes feature information and basic information. The information items of the basic information correspond to the information items of the sample information, and the information items of the feature information correspond to the information items of the child's features. The timing prediction model under the target feature corresponding to the feature information is extracted to obtain the selection model. The basic information is imported into the selection model to obtain the prediction time. The respiratory state under the target feature corresponding to the feature information is extracted to obtain the selection state. The basic information is matched with several states in the selection state to obtain the target state. The target state is one of the following: respiratory incomplete recovery period, spontaneous breathing recovery period, and spontaneous breathing period.

8. A pediatric patient extubation timing prediction system based on artificial intelligence assistance, characterized in that: The method for predicting the timing of extubation in pediatric patients based on any one of claims 1-7 was used.

Images