Artificial intelligence-based respiratory failure treatment regimen recommendation method and system

Abstract

The embodiment of the application provides a respiratory failure treatment scheme recommendation method and system based on artificial intelligence, which obtains first multi-modal medical collection data of a target object, the first multi-modal medical collection data comprising text modal data, image modal data, the text modal data comprising physiological characteristic data, basic disease data and blood gas analysis data; performs data preprocessing operation on the first multi-modal medical collection data to obtain second multi-modal medical collection data, the data preprocessing comprising data cleaning, normalization processing and time window alignment processing. A medical feature vector set of the target object is extracted from the second multi-modal medical collection data. Classification processing of the target object is performed according to the medical feature vector set to determine an object treatment category label of the target object. According to the object treatment category label, a recommended treatment scheme of the target object is determined, thereby improving the efficiency and accuracy of determining the respiratory failure treatment scheme.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a respiratory failure treatment scheme recommendation method and system based on artificial intelligence. BACKGROUND

[0002] Respiratory failure is a clinical syndrome that seriously threatens human health, and its morbidity and mortality have been high. Timely and accurate formulation of appropriate treatment schemes for patients with respiratory failure is crucial for improving patient prognosis and reducing mortality.

[0003] In the prior art, the determination of respiratory failure treatment schemes mainly relies on the clinical experience of doctors and traditional medical judgment methods. Doctors usually preliminarily assess the patient's condition according to single type of physiological indicator data of the patient, such as heart rate, blood pressure, etc., combined with their own professional knowledge and past treatment experience. However, this evaluation method based on single physiological indicator and subjective experience has obvious limitations. Single physiological indicator can only reflect one aspect of the patient's condition and cannot fully and accurately present the overall condition of the patient. Moreover, the experience level and judgment standard of different doctors are different, which may lead to inconsistency and inaccuracy in the formulation of treatment schemes.

[0004] In addition, for some complex respiratory failure cases, it is difficult to capture subtle changes in the condition and potential influencing factors by relying on single physiological indicators. For example, the patient may have multiple underlying diseases, which may interact with each other and then have a complex effect on the treatment of respiratory failure. However, traditional methods often fail to consider these multiple factors comprehensively, thus easily leading to suboptimal selection of treatment schemes and delay of treatment opportunity for patients.

[0005] Moreover, most existing methods use simple classification rules, which are often based on limited clinical experience and cannot well adapt to the individual needs of different patients. For some patients with special conditions or complex conditions, these simple classification rules may not accurately classify the patients into appropriate treatment categories, thus failing to provide accurate treatment scheme recommendations.

[0006] That is, the existing technology has problems such as insufficient data utilization, inaccurate evaluation method, and lack of individualization in the formulation of respiratory failure treatment schemes. SUMMARY

[0007] The present application provides a respiratory failure treatment scheme recommendation method and system based on artificial intelligence to improve the efficiency and accuracy of determining respiratory failure treatment schemes.

[0008] In a first aspect, the present application provides a respiratory failure treatment scheme recommendation method based on artificial intelligence, comprising:

[0009] obtaining first multi-modal medical acquisition data of a target object, the first multi-modal medical acquisition data comprising text modal data, image modal data, the text modal data comprising physiological feature data, basic disease data, blood gas analysis data;

[0010] performing data preprocessing operation on the first multi-modal medical acquisition data to obtain second multi-modal medical acquisition data, the data preprocessing comprising data cleaning, normalization processing, and time window alignment processing;

[0011] extracting a medical feature vector set of the target object from the second multi-modal medical acquisition data;

[0012] performing classification processing of the target object according to the medical feature vector set to determine an object treatment category label of the target object;

[0013] determining a recommended treatment scheme of the target object according to the object treatment category label.

[0014] Optionally, the classification processing of the target object according to the medical feature vector set to determine the object treatment category label of the target object comprises:

[0015] performing clustering analysis according to the medical feature vector set to obtain a clustering analysis result of the target object, and determining the object treatment category label of the target object according to the clustering analysis result and a mapping relationship between the clustering analysis result and the object treatment category label;

[0016] or,

[0017] inputting the medical feature vector set into a pre-trained decision tree model to obtain the object treatment category label of the target object.

[0018] Optionally, the clustering analysis according to the medical feature vector set to obtain the clustering analysis result of the target object comprises:

[0019] determining an initial clustering center according to a medical feature vector included in the medical feature vector set;

[0020] obtaining a medical weight coefficient corresponding to each medical feature vector based on a treatment risk quantitative index of the medical feature vector;

[0021] determine distances between each of the medical feature vectors and the initial clustering center according to each of the medical feature vectors and a medical weight coefficient corresponding to each of the medical feature vectors, and generate an initial clustering result based on the distances between each of the medical feature vectors and the initial clustering center;

[0022] iteratively update the initial clustering center according to the initial clustering result, a medical feature vector of the target object, and a medical weight coefficient until a preset termination condition is met, to obtain a clustering analysis result of the target object.

[0023] Optionally, the clustering analysis based on the set of medical feature vectors to obtain the clustering analysis result of the target object comprises:

[0024] construct an initial state of hierarchical clustering based on distribution characteristics of each medical feature vector in the set of medical feature vectors, each of the medical feature vectors in the initial state being an independent initial cluster;

[0025] obtain a medical weight coefficient corresponding to each of the medical feature vectors based on a treatment risk quantitative index of the medical feature vector;

[0026] calculate a weighted distance between each of the independent initial clusters according to each of the medical feature vectors and the medical weight coefficient corresponding to each of the medical feature vectors, to generate a clustering distance matrix;

[0027] iteratively merge independent initial clusters with a weighted distance less than or equal to a preset distance threshold based on each of the weighted distances in the clustering distance matrix, until a clustering number of the independent initial clusters reaches a preset number, to obtain the clustering analysis result of the target object.

[0028] Optionally, before the classification processing of the target object based on the set of medical feature vectors to determine an object treatment category label of the target object, the method further comprises a training method of the decision tree model, specifically comprising:

[0029] construct an initial state of hierarchical clustering based on distribution characteristics of each sample medical feature vector in a set of sample medical feature vectors of sample objects, each of the sample medical feature vectors in the initial state being an independent initial cluster;

[0030] obtain a medical weight coefficient corresponding to each of the sample medical feature vectors based on a treatment risk quantitative index of the sample medical feature vector;

[0031] calculate a weighted distance between each of the independent initial clusters according to each of the sample medical feature vectors and the medical weight coefficient corresponding to each of the sample medical feature vectors, to generate a sample clustering distance matrix;

[0032] merge, according to the weighted distances in the sample clustering distance matrix, independent initial clusters with a weighted distance less than or equal to a preset distance threshold to a preset number of clusters, to obtain an initial classification result of the sample object;

[0033] obtain a constraint condition of an initial decision tree model, the constraint condition including a Gini coefficient constraint and a medical feature vector threshold constraint;

[0034] train the initial decision tree model according to the initial classification result and the constraint condition, to generate the decision tree model.

[0035] Optionally, the training of the initial decision tree model according to the initial classification result and the constraint condition to generate the decision tree model includes:

[0036] obtain a classification label corresponding to each sample medical feature vector in the initial classification result, and the Gini coefficient constraint and the medical feature vector threshold constraint in the constraint condition;

[0037] based on the classification label, traverse each medical feature vector in the sample medical feature vector set, calculate the Gini coefficient of the medical feature vector under different split thresholds, and select the split feature with the smallest Gini coefficient and the corresponding split threshold as a candidate split node;

[0038] based on the medical feature vector threshold constraint, perform normalization verification on the split threshold in the candidate split node, if the split threshold is within the numerical range of the normalized medical feature vector, the candidate split node is retained, otherwise the candidate split node is removed;

[0039] based on the retained candidate split node, generate a split priority queue in ascending order of Gini coefficient, select the candidate split node with the smallest Gini coefficient from the split priority queue as a current node split condition, and divide the sample medical feature vector set into two subsets;

[0040] recursively traverse the two subsets, repeat the steps of calculating the Gini coefficient, selecting the split feature and the split threshold, verifying the split threshold, and generating the split priority queue, until a preset stopping condition is met, the preset stopping condition including that the Gini coefficient of the subset is zero, the split threshold of all medical feature vectors does not meet the medical feature vector threshold constraint, or the maximum recursion depth is reached;

[0041] generate a tree structure including classification rules and split thresholds according to all nodes after recursive splitting, as the decision tree model.

[0042] Optionally, after the classification processing of the target object is performed according to the medical feature vector set, and the object treatment category label of the target object is determined, the method further comprises:

[0043] According to the preset batch processing configuration parameter, the illness monitoring quantitative data of the target object is acquired, and the illness monitoring quantitative data comprises vital sign quantitative data, blood gas analysis quantitative data, and lesion trend quantitative data;

[0044] The progress weight corresponding to the vital sign quantitative data, the blood gas analysis quantitative data, and the lesion trend quantitative data is acquired respectively;

[0045] According to the vital sign quantitative data, the blood gas analysis quantitative data, the lesion trend quantitative data, and the progress weight, the illness progress quantitative indicator of the target object is determined based on an attention mechanism;

[0046] Based on the illness progress quantitative indicator, the current object treatment category label of the target object, and the hierarchical conversion threshold corresponding to the illness progress quantitative indicator, it is determined whether the object treatment category label of the target object needs to be adjusted;

[0047] If the object treatment category label of the target object needs to be adjusted, the recommended treatment scheme of the target object is determined based on the adjusted object treatment category label.

[0048] Optionally, the progress weight corresponding to the vital sign quantitative data, the blood gas analysis quantitative data, and the lesion trend quantitative data is acquired respectively, comprising:

[0049] From the illness progress history data of a plurality of sample patients pre-constructed, a sample medical feature vector set of each sample patient is acquired, each sample medical feature vector in the sample medical feature vector set comprises vital sign features, blood gas analysis features, and lesion trend features, and has been normalized;

[0050] An illness deterioration index of the sample patient is extracted as a regression target variable, and the illness deterioration index is generated according to the occurrence time and severity quantization of a preset deterioration event;

[0051] The vital sign features, the blood gas analysis features, and the lesion trend features are taken as independent variables, and a linear regression model or a logistic regression model is fitted, and the regression coefficients corresponding to each feature are outputted;

[0052] Based on the absolute value size of the regression coefficient of each feature, the importance indicator of each feature is calculated;

[0053] The importance indicators are normalized so that the sum of the progress weights of each feature is 1, to generate progress weights corresponding to the vital sign quantitative data, the blood gas analysis quantitative data, and the lesion trend quantitative data.

[0054] Optionally, the determination of whether the object treatment category label of the target object needs to be adjusted based on the disease progression quantitative indicator, the current object treatment category label of the target object, and the hierarchical transition threshold corresponding to the disease progression quantitative indicator comprises:

[0055] A preset hierarchical transition threshold range associated with the current object treatment category label is obtained, the preset hierarchical transition threshold range comprising a deterioration threshold upper limit and an improvement threshold lower limit;

[0056] The disease progression quantitative indicator is compared with the deterioration threshold upper limit in a first comparison, and if the disease progression quantitative indicator is greater than or equal to the deterioration threshold upper limit, a category up-regulation instruction is generated;

[0057] The disease progression quantitative indicator is compared with the improvement threshold lower limit in a second comparison, and if the disease progression quantitative indicator is less than or equal to the improvement threshold lower limit, a category down-regulation instruction is generated;

[0058] According to the category up-regulation instruction or the category down-regulation instruction, an adjusted object treatment category label of the target object is determined, and it is verified whether the adjusted object treatment category label is within a preset label hierarchical range;

[0059] If the adjusted object treatment category label is outside the preset label hierarchical range, the current object treatment category label is maintained unchanged; otherwise, the adjusted object treatment category label is taken as an updated object treatment category label.

[0060] In a second aspect, the present application provides a respiratory failure treatment scheme recommendation system based on artificial intelligence, which comprises a machine readable storage medium and a processor, the machine readable storage medium storing machine executable instructions, and the processor, when executing the machine executable instructions, implements the aforementioned respiratory failure treatment scheme recommendation method based on artificial intelligence.

[0061] In a third aspect, the present application provides a computer readable storage medium, which stores computer execution instructions, and the computer execution instructions, when executed by a processor, are used to implement the method according to any one of the first aspect.

[0062] In a fourth aspect, the present application provides a computer program product for implementing the method of any one of the first aspect when executed by a processor.

[0063] The method and system for recommending a respiratory failure treatment scheme based on artificial intelligence provided by the present application overcome the limitations of the prior art relying only on single physiological indicator data by acquiring first multi-modal medical collection data containing text modalities (physiological characteristic data, basic disease data, blood gas analysis data) and image modalities, and can reflect the condition of the target object from multiple dimensions. The data preprocessing operations including data cleaning, normalization processing, and time window alignment processing are performed on the collected data, effectively solving the problems of noise, missing values, inconsistency, and format standard deviation in the medical data in the prior art, and improving the quality and usability of the data. The medical feature vector set is extracted from the preprocessed second multi-modal medical collection data, and the multi-modal data is converted into quantifiable and analyzable feature vectors, which facilitates further processing and mining of potential information in the data. The object treatment category label is determined by classifying the medical feature vector set, which avoids the drawbacks of simple classification rules that cannot meet individual needs, and can more accurately classify the target object. Finally, the recommended treatment scheme is determined according to the object treatment category label, realizing the individualization and precision of the respiratory failure treatment scheme recommendation, and improving the pertinence and effectiveness of the treatment scheme. BRIEF DESCRIPTION OF DRAWINGS

[0064] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and serve to explain the principles of the present application, together with the description.

[0065] Figure 1 A flowchart of a method for recommending a respiratory failure treatment scheme based on artificial intelligence provided by an embodiment of the present application;

[0066] Figure 2 A flowchart of another method for recommending a respiratory failure treatment scheme based on artificial intelligence provided by an embodiment of the present application;

[0067] Figure 3 A flowchart of yet another method for recommending a respiratory failure treatment scheme based on artificial intelligence provided by an embodiment of the present application;

[0068] Figure 4 A structural diagram of a system for recommending a respiratory failure treatment scheme based on artificial intelligence provided by an embodiment of the present application.

[0069] The specific embodiments of the present application have been shown by the above drawings, and will be described in more detail hereinafter. These drawings and the written description are not intended to restrict the scope of the present application concept in any way, but to illustrate the present application concept to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0070] In order to enable persons skilled in the art to better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. According to the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0071] The terms "first", "second", and the like in the specification and claims of the present application and the above drawings are used to distinguish different objects, not to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product, or end including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product, or end.

[0072] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily mutually exclusive or alternative to other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0073] Figure 1 A flowchart of a respiratory failure treatment scheme recommendation method based on artificial intelligence provided by the embodiments of the present application is shown. It should be understood that in other embodiments, the order of some steps of the respiratory failure treatment scheme recommendation method based on artificial intelligence of the present embodiment can be shared according to actual needs, or some steps can be omitted or maintained. As shown in the figure, the method can include the following steps: Figure 1

[0074] Step S110: acquiring first multi-modal medical acquisition data of a target object.

[0075] The first multi-modal medical acquisition data includes text modal data, image modal data, and the text modal data includes physiological feature data, basic disease data, and blood gas analysis data. ​

[0076] In practical application scenarios, the acquisition of the first multi-modal medical data of the target object involves multiple different data sources and acquisition methods. In terms of text modal data acquisition, physiological characteristic data acquisition relies on various professional medical monitoring equipment. For example, heart rate data can be monitored in real time through wearable heart rate monitoring bracelets or electrocardiogram monitors in hospital wards. These devices continuously record the target object's heart rate changes and transmit the data to the corresponding storage system. For blood pressure data, an electronic sphygmomanometer is usually used for measurement, and the measured data is automatically stored or manually entered into the medical information system. Respiratory rate data can be obtained through respiratory sensing bands or camera-based respiratory monitoring technology, and the collected data is integrated into a unified data storage.

[0077] The acquisition of basic disease data mainly comes from the target object's medical history. The hospital's electronic medical record system is an important data source, which records in detail the various diseases the target object has ever suffered from, including the diagnosis time, treatment process, and recovery information. In addition, the doctor's inquiry record is also an important way to obtain basic disease data. During the communication with the target object, the doctor will inquire in detail about the medical history and record these information.

[0078] Blood gas analysis data is obtained by collecting blood samples from the target object and then using a professional blood gas analyzer for detection. When collecting blood samples, appropriate blood collection sites such as arteries are usually selected to obtain accurate blood gas analysis results. After blood collection, the sample is quickly sent to the laboratory for detection by professional personnel using a blood gas analyzer. During the detection process, the instrument analyzes various components in the blood, such as oxygen partial pressure, carbon dioxide partial pressure, and acid-base balance, and outputs the detection results in digital form, which is stored in the hospital's laboratory information system.

[0079] For image modal data acquisition, it mainly involves hospital imaging equipment. Chest X-ray is obtained by taking a chest X-ray of the target object using an X-ray machine. During the shooting process, the target object needs to maintain the correct posture according to the doctor's instructions to ensure that a clear and accurate image is taken. After shooting, the image data of the X-ray film will be transmitted to the Picture Archiving and Communication System (PACS) of the hospital for storage and management.

[0080] CT scan image acquisition is through multi-layer scanning of the target object's chest by a CT scanner. The CT scanner will rotate around the target object, scanning the chest from different angles, and then reconstructing a series of cross-sectional image data obtained by scanning to generate three-dimensional CT scan images. These image data are also stored in the PACS system for subsequent viewing and analysis.

[0081] Step S120: performing data preprocessing on the first multimodal medical collected data to obtain second multimodal medical collected data.

[0082] The data preprocessing includes data cleaning, normalization and time window alignment.

[0083] After the first multimodal medical data is acquired, it is necessary to perform data preprocessing operations on the data because the data may have noise, missing values, outliers, different scales and time stamps, and other problems.

[0084] Data cleaning is the first step in preprocessing. For physiological characteristic data in text modal data, outliers must be detected and addressed first. For example, heart rate data normally fluctuates within a certain range. If the heart rate value recorded at a given moment deviates significantly from the normal heart rate range for the target subject, it may be an outlier. Outliers can be identified by setting upper and lower thresholds. The upper threshold can be determined based on medical knowledge and the individual condition of the target subject, and the lower threshold is similarly determined. For identified outliers, the following approaches can be used: If the outlier is caused by equipment failure or data entry error, the data record can be deleted directly. If the outlier is caused by a unique physiological condition of the target subject but still has some reference value, interpolation methods, such as linear interpolation, can be used to estimate a reasonable value based on the normal data points before and after the outlier.

[0085] Basic disease data may contain duplicate records or incorrect disease names. This can be achieved by establishing a standard dictionary of disease names, comparing collected disease names against the dictionary, and correcting any non-compliant names. Furthermore, data deduplication algorithms, such as hashing algorithms, can be used to deduplicate basic disease data to ensure data accuracy and uniqueness.

[0086] Cleaning blood gas analysis data also requires attention to outliers. Since each indicator in blood gas analysis data has a normal physiological range, such as blood oxygen partial pressure and carbon dioxide partial pressure, thresholds can be set based on these normal ranges, and data that exceeds these thresholds can be marked and processed. For missing blood gas analysis data, machine learning algorithms, such as regression analysis, can be used to predict missing values ​​based on other relevant data about the target subject, such as physiological characteristics and underlying disease data.

[0087] For image modal data, there may be problems such as image blur, noise interference, etc. Image enhancement techniques can be used to improve the clarity and quality of the image. For example, using histogram equalization method to enhance the contrast of the image, by adjusting the gray scale distribution of the image, so that the details in the image are more clear and visible. For noise in the image, filtering algorithms can be used to remove, such as Gaussian filtering, which smoothes the noise in the image by weighted average of each pixel point in the image, while preserving the edge information of the image.

[0088] Normalization processing is to unify the data of different features to the same scale range, so as to facilitate subsequent analysis and calculation. For physiological feature data in text modal data, such as heart rate, blood pressure and other continuous data, the min-max normalization method can be used. The specific calculation process of this method is: for each data point, first subtract the minimum value of the feature, then divide by the difference between the maximum and minimum values of the feature. In this way, the data can be mapped to the interval [0, 1]. For example, for heart rate data, let the heart rate value at a certain time be H, the minimum value of the feature be H_min, and the maximum value be H_max, then the normalized heart rate value H_norm=(H-H_min) / (H_max-H_min).

[0089] For basic disease data, since it is discrete data, it can be converted into numerical data by using one-hot encoding. Specifically, a binary vector is created for each disease, and the length of the vector is equal to the number of disease types. If the target object has a certain disease, the corresponding vector position is 1, otherwise it is 0. Then normalize the one-hot encoded data, you can use the simple mean-standard deviation normalization method, that is, first calculate the mean and standard deviation of each feature, then subtract the mean from each data point, and then divide by the standard deviation.

[0090] For blood gas analysis data, you can also choose the appropriate normalization method according to its data type. If it is continuous data, you can use min-max normalization or mean-standard deviation normalization; if it is discrete data, you can use one-hot encoding and then normalize.

[0091] The time window alignment processing is to ensure the consistency of data of different modalities in time. Since the physiological feature data, blood gas analysis data, etc. may be collected at different time points, and the collection time of image modal data may also be different from the text modal data, time window alignment processing is needed. A fixed time window can be set, for example, in hours. For each time window, the physiological feature data, blood gas analysis data, etc. collected in the time window are integrated. If some data is missing in a certain time window, interpolation or missing value filling method can be used for processing. For image modal data, it needs to be associated with the corresponding time window. If there is no corresponding image data in a certain time window, the image data of the previous and next time windows can be considered for interpolation or the last collected image data can be used for substitution.

[0092] Step S130: extracting a medical feature vector set of the target object from the second multi-modal medical acquisition data.

[0093] After completing the data preprocessing, the second multi-modal medical acquisition data is obtained, and then the medical feature vector set of the target object needs to be extracted from these data.

[0094] For feature extraction of text modal data, for physiological feature data, the heart rate value, blood pressure, respiratory rate, etc. after normalization processing can be combined into a vector. For example, assuming that the physiological feature data includes heart rate value H norm, systolic blood pressure SBP norm, diastolic blood pressure DBP norm and respiratory rate RR norm, the physiological feature vector can be:

[0095] P = [H norm, SBP norm, DBP norm, RR norm].

[0096] For basic disease data, after one-hot encoding and normalization processing, the binary vector corresponding to each disease is combined into a longer vector. Assuming that there are n kinds of diseases, the basic disease vector B = [b_1, b_2,..., b_n] obtained after processing, where b_i represents the normalized binary value corresponding to the i-th disease.

[0097] For blood gas analysis data, the normalized partial pressure of oxygen, partial pressure of carbon dioxide, pH, etc. indicators are combined into a vector. Let the partial pressure of oxygen be PO2 norm, the partial pressure of carbon dioxide be PCO2 norm, and the pH be pH norm, then the blood gas analysis vector G = [PO2 norm, PCO2 norm, pH norm].

[0098] For feature extraction of image modality data, a convolutional neural network (CNN) can be used for processing. First, the preprocessed image data is input into the CNN. The CNN is usually composed of convolutional layers, pooling layers, and fully connected layers. In the convolutional layer, multiple different convolution kernels are used to slide on the image and perform convolution operations to extract local features of the image. Each convolution kernel generates a feature map, and multiple convolution kernels generate multiple feature maps. For example, assuming there are m convolution kernels, each with a size of k x k, after the convolutional layer, m feature maps with a size of W x H (W and H are the width and height of the feature map) are obtained.

[0099] The role of the pooling layer is to downsample the feature maps output by the convolutional layer, reducing the dimension of the data while preserving important feature information. Common pooling methods include max pooling and average pooling. Taking max pooling as an example, the maximum value in each pooling window is selected as the output of the window, resulting in a smaller size feature map.

[0100] After multiple processing by the convolutional layer and the pooling layer, the feature map of the last layer is unfolded into a one-dimensional vector, which is then input into the fully connected layer. The fully connected layer performs linear transformation and non-linear activation on the input vector, and finally outputs a fixed-length image feature vector I.

[0101] Finally, the physiological feature vector P extracted from the text modality data, the basic disease vector B, the blood gas analysis vector G, and the image feature vector I extracted from the image modality data are spliced to form a medical feature vector set F = [P, B, G, I] of the target object. During splicing, the dimensions of the vectors need to be matched to ensure the consistency and usability of the medical feature vector set.

[0102] Step S140: performing classification processing on the target object according to the medical feature vector set to determine the object treatment category label of the target object.

[0103] Specifically, the classification processing of the target object according to the medical feature vector set to determine the object treatment category label of the target object includes two ways.

[0104] Method one: performing clustering analysis according to the medical feature vector set to obtain a clustering analysis result of the target object, and determining the object treatment category label of the target object according to the clustering analysis result and a mapping relationship between the clustering analysis result and the object treatment category label.

[0105] One possible implementation, taking K-Means clustering as an example, includes the following steps:

[0106] Step S1411: Determine initial cluster centers according to the medical feature vectors included in the medical feature vector set.

[0107] When performing clustering analysis, determining initial cluster centers is a key first step. One commonly used method is random selection, that is, randomly selecting several medical feature vectors from the medical feature vector set as initial cluster centers. In specific implementation, a unique number can be assigned to each medical feature vector, and then a random number generator is used to generate a specified number of random numbers, and according to these numbers, the corresponding medical feature vectors are selected from the medical feature vector set as initial cluster centers.

[0108] Another more optimized method is the K-means++ algorithm. The specific steps of this algorithm are as follows: First, randomly select a medical feature vector as the first initial cluster center. Then, calculate the distance of other medical feature vectors to the initial cluster center, here the Euclidean distance can be used for calculation. For each medical feature vector x, its Euclidean distance d(x, c_1)=√((x_1-c_1_1)^2+(x_2-c_1_2)^2+...+(x_n-c_1_n)^2) to the initial cluster center c_1 is calculated, where x_i and c_1_i represent the values of the i-th dimension of the medical feature vector x and the initial cluster center c_1 respectively.

[0109] Next, select the medical feature vector farthest from the first initial cluster center as the second initial cluster center. Then, for the remaining medical feature vectors, calculate their shortest distances to all selected initial cluster centers. For example, for medical feature vector x, calculate d_min(x)=min(d(x, c_1), d(x, c_2)). Then, according to these shortest distances, select the medical feature vector with the largest distance as the next initial cluster center. Repeat this process until the required number of initial cluster centers is selected.

[0110] Step S1412: Obtain medical weight coefficients corresponding to each medical feature vector based on the treatment risk quantification indicators of the medical feature vectors.

[0111] In order to more accurately perform clustering analysis, the treatment risk quantification index of the medical feature vector needs to be considered. The evaluation of the treatment risk quantification index needs to consider multiple factors, such as the severity of the target object's condition, the difficulty of treatment, the possibility of complications, and the like. The treatment risk of each medical feature vector can be scored by a group of experienced doctors according to the information contained in the medical feature vector through expert evaluation. For example, a higher score is given to a medical feature vector with a more severe condition, greater treatment difficulty, and a high possibility of complications; a lower score is given to a medical feature vector with a less severe condition, less treatment difficulty, and a low possibility of complications.

[0112] Then, the medical weight coefficient corresponding to each medical feature vector is calculated according to the treatment risk quantification index. The specific calculation method is to divide the treatment risk quantification index of each medical feature vector by the sum of the treatment risk quantification indexes of all medical feature vectors. Let the treatment risk quantification index of the medical feature vector x_i be r_i, and the sum of the treatment risk quantification indexes of all medical feature vectors be R = ∑r_j (j from 1 to N, N is the total number of medical feature vectors), then the medical weight coefficient w_i corresponding to the medical feature vector x_i is r_i / R.

[0113] Step S1413: Determine the distance between each medical feature vector and the initial clustering center according to each medical feature vector and the medical weight coefficient corresponding to each medical feature vector, and generate an initial clustering result based on the distance between each medical feature vector and the initial clustering center.

[0114] After the initial clustering center and the medical weight coefficient are determined, the distance between each medical feature vector and the initial clustering center needs to be calculated. Here, the weighted Euclidean distance method is used for calculation. For a medical feature vector x and an initial clustering center c, the weighted Euclidean distance d_w(x, c) = √(w_1*(x_1-c_1)^2+w_2*(x_2-c_2)^2+...+w_n*(x_n-c_n)^2), where w_i is the medical weight coefficient corresponding to the i-th dimension of the medical feature vector x, x_i and c_i are the values of the i-th dimension of the medical feature vector x and the initial clustering center c, respectively.

[0115] According to the calculated weighted Euclidean distance, each medical feature vector is assigned to the cluster in which the initial clustering center with the closest distance is located. For example, for a medical feature vector x, if d_w(x, c_1) < d_w(x, c_2) <... < d_w(x, c_k) (k is the number of initial clustering centers), then the medical feature vector x is assigned to the cluster in which the initial clustering center c_1 is located. In this way, an initial clustering result is generated.

[0116] Step S1414: iteratively updating the initial clustering center according to the initial clustering result, the medical feature vector of the target object and the medical weight coefficient, until a preset termination condition is met, to obtain a clustering analysis result of the target object.

[0117] According to the initial clustering result, the initial clustering center needs to be iteratively updated. For each cluster, the weighted average value of all medical feature vectors in the cluster is calculated, and the weighted average value is taken as the new clustering center. The specific calculation method is as follows: assuming that there are m medical feature vectors x_1, x_2, …, x_m in a certain cluster, and the corresponding medical weight coefficients are w_1, w_2, …, w_m respectively, then the value of the i-th dimension of the new clustering center c_new is c_new_i=(∑(w_j*x_j_i)) / (∑w_j) (j from 1 to m).

[0118] Then, the weighted Euclidean distance of each medical feature vector and the new clustering center is recalculated, and the medical feature vectors are redistributed to different clusters according to the distance, to obtain a new clustering result. The above iterative updating process is repeated until a preset termination condition is met. The preset termination condition can include that the clustering center no longer changes, that is, the difference between the new clustering center and the clustering center of the last time in each dimension is less than a very small threshold value; or the maximum number of iterations is reached to avoid the algorithm from falling into an infinite loop. When the preset termination condition is met, the clustering analysis result of the target object is obtained.

[0119] Another possible implementation, taking hierarchical clustering as an example, the clustering analysis according to the medical feature vector set to obtain the clustering analysis result of the target object includes the following steps:

[0120] Step S1421: based on the distribution characteristics of each medical feature vector in the medical feature vector set, an initial state of hierarchical clustering is constructed, and each medical feature vector in the initial state is an independent initial cluster.

[0121] In this embodiment, when hierarchical clustering method is used for clustering analysis, an initial state of hierarchical clustering needs to be constructed first. According to the distribution characteristics of each medical feature vector in the medical feature vector set, each medical feature vector is regarded as an independent initial cluster. In this way, there are as many independent initial clusters as the number of medical feature vectors in the initial state.

[0122] Step S1422: based on the treatment risk quantification index of the medical feature vector, the medical weight coefficient corresponding to each medical feature vector is obtained.

[0123] This step can refer to the aforementioned step S1412, which will not be described here.

[0124] Step S1423: Calculate the weighted distance between each of the independent initial clusters according to each of the medical feature vectors and the medical weight coefficients corresponding to each of the medical feature vectors, and generate a cluster distance matrix;

[0125] After the medical weight coefficients are determined, the weighted distance between each of the independent initial clusters needs to be calculated. For two independent initial clusters C_i and C_j, where C_i contains medical feature vectors x_i1, x_i2, …, x_im and C_j contains medical feature vectors x_j1, x_j2, …, x_jn, when calculating the weighted distance between them, first calculate the weighted Euclidean distance between each medical feature vector in C_i and each medical feature vector in C_j. For medical feature vectors x_ik and x_jl, the weighted Euclidean distance is equal to the square root of the sum of the square of the first dimension value of the medical weight coefficient multiplied by the square of the difference between the first dimension value of x_ik and the first dimension value of x_jl, the square of the second dimension value of the medical weight coefficient multiplied by the square of the difference between the second dimension value of x_ik and the second dimension value of x_jl, and so on, up to the square of the p-th dimension value of the medical weight coefficient multiplied by the square of the difference between the p-th dimension value of x_ik and the p-th dimension value of x_jl.

[0126] Then, the average of these weighted Euclidean distances is taken as the weighted distance between the two independent initial clusters. The specific calculation formula is: let two independent initial clusters C_i and C_j, the weighted distance between them is equal to (the sum of the weighted Euclidean distances between each medical feature vector in C_i and each medical feature vector in C_j) divided by (the number of medical feature vectors in C_i multiplied by the number of medical feature vectors in C_j).

[0127] After calculating the weighted distance between all independent initial clusters, a cluster distance matrix is generated. This matrix is a symmetric matrix, with rows and columns corresponding to each independent initial cluster, and the elements in the matrix represent the weighted distance between the two independent initial clusters corresponding to the two independent initial clusters.

[0128] Step S1424: According to each of the weighted distances in the cluster distance matrix, iteratively merge the independent initial clusters whose weighted distances are less than or equal to a preset distance threshold until the number of clusters of the independent initial clusters reaches a preset number, and obtain the clustering analysis result of the target object.

[0129] According to the weighted distances in the cluster distance matrix, perform an iterative merging operation. First, find two independent initial clusters in the cluster distance matrix whose weighted distances are less than or equal to a preset distance threshold. The preset distance threshold is a pre-set parameter used to control the conditions of clustering merging. When two clusters C_i and C_j that meet the conditions are found, they are merged into a new cluster C_new.

[0130] The merged new cluster C_new contains all medical feature vectors in the original two clusters. Next, the weighted distance between the new cluster C_new and other unmerged clusters needs to be recalculated. The calculation method is similar to the calculation of the weighted distance between the two independent initial clusters, that is, the weighted Euclidean distance between each medical feature vector in C_new and each medical feature vector in other clusters is calculated, and then the average value is taken.

[0131] Update the cluster distance matrix by deleting the rows and columns corresponding to the original clusters C_i and C_j, and adding the rows and columns corresponding to the new cluster C_new. Repeat the above iterative merging process to find clusters that meet the conditions for merging and update the cluster distance matrix.

[0132] Stop the iteration until the number of independent initial clusters reaches the preset number. The preset number is a parameter set according to actual needs, which determines the number of final clusters. At this time, the clustering analysis result of the target object is obtained.

[0133] Finally, according to the clustering analysis result and the pre-established mapping relationship between the clustering analysis result and the object treatment category label of the object, the object treatment category label of the target object is determined. This mapping relationship can be established through historical data and expert experience, for example, a certain specific cluster is corresponding to a certain specific treatment category label. If the medical feature vector of the target object is divided into a certain cluster, the corresponding object treatment category label can be determined according to the mapping relationship.

[0134] Method two: input the medical feature vector set into a pre-trained decision tree model to obtain the object treatment category label of the target object.

[0135] Before inputting the medical feature vector set into the decision tree model, the decision tree model needs to be trained. Figure 2 Another flowchart of a respiratory failure treatment scheme recommendation method based on artificial intelligence provided by the embodiments of the present application is provided. As shown in the figure, before inputting the medical feature vector set into the decision tree model, the method can further include: Figure 2

[0136] Step S210: based on the distribution characteristics of each sample medical feature vector in the sample medical feature vector set of the sample object, an initial state of hierarchical clustering is constructed.

[0137] Among them, each sample medical feature vector in the initial state is an independent initial cluster.

[0138] ​To train the decision tree model, a set of sample medical feature vectors of sample objects is first used. Similar to the construction process of the medical feature vector of the target object, the sample medical feature vector of the sample object is also obtained from the multi-modal medical acquisition data of the sample object after preprocessing and feature extraction.

[0139] According to the distribution characteristics of each sample medical feature vector in the set of sample medical feature vectors, each sample medical feature vector is regarded as an independent initial cluster. In this way, in the initial state of hierarchical clustering, the number of independent initial clusters is equal to the number of sample medical feature vectors.

[0140] Step S220: Based on the treatment risk quantization indicators of the sample medical feature vectors, obtain the medical weight coefficients corresponding to each sample medical feature vector.

[0141] Similarly, in order to consider the treatment risk of the sample medical feature vector, the medical weight coefficient corresponding to each sample medical feature vector needs to be obtained. The treatment risk of each sample medical feature vector can be scored by a professional doctor according to the information contained in the sample medical feature vector through expert evaluation. These information includes physiological characteristics, underlying diseases, blood gas analysis results and imaging features of the sample object.

[0142] Then, the treatment risk quantization indicator of each sample medical feature vector is divided by the sum of the treatment risk quantization indicators of all sample medical feature vectors to obtain the corresponding medical weight coefficient. Let the treatment risk quantization indicator of the sample medical feature vector y_i be s_i, and the sum of the treatment risk quantization indicators of all sample medical feature vectors be S, which is equal to the sum of the treatment risk quantization indicators of each sample medical feature vector (i from 1 to the total number of sample medical feature vectors), then the medical weight coefficient v_i corresponding to the sample medical feature vector y_i is equal to s_i divided by S.

[0143] Step S230: According to each sample medical feature vector and the medical weight coefficient corresponding to each sample medical feature vector, calculate the weighted distance between the independent initial clusters to generate a sample cluster distance matrix.

[0144] After determining the medical weight coefficient, the weighted distance between each independent initial cluster is calculated. For two independent initial clusters D_i and D_j, where D_i contains sample medical feature vectors y_i1, y_i2, …, y_im, and D_j contains sample medical feature vectors y_j1, y_j2, …, y_jn, the weighted Euclidean distance between each sample medical feature vector in D_i and each sample medical feature vector in D_j is first calculated.

[0145] The weighted Euclidean distance of sample medical feature vectors y_ik and y_jl is equal to the square root of the sum of the square of the first dimension value of the medical weight coefficient multiplied by the square of the difference between the first dimension value of y_ik and the first dimension value of y_jl, the square of the second dimension value of the medical weight coefficient multiplied by the square of the difference between the second dimension value of y_ik and the second dimension value of y_jl, and so on to the square of the p-th dimension value of the medical weight coefficient multiplied by the square of the difference between the p-th dimension value of y_ik and the p-th dimension value of y_jl.

[0146] Then, the average of these weighted Euclidean distances is taken as the weighted distance between the two independent initial clusters, which is calculated by adding the weighted Euclidean distance between each sample medical feature vector in D_i and each sample medical feature vector in D_j, and then dividing by the product of the number of sample medical feature vectors in D_i and the number of sample medical feature vectors in D_j.

[0147] After calculating the weighted distance between all independent initial clusters, a sample cluster distance matrix is generated. The rows and columns of the sample cluster distance matrix correspond to each independent initial cluster, and the elements in the sample cluster distance matrix represent the weighted distance between the corresponding two independent initial clusters.

[0148] Step S240: According to the weighted distances in the sample cluster distance matrix, iteratively merge the independent initial clusters with a weighted distance less than or equal to a preset distance threshold until the number of clusters reaches a preset number, to obtain an initial classification result of the sample object.

[0149] According to the weighted distances in the sample cluster distance matrix, perform an iterative merging operation. In the sample cluster distance matrix, find two independent initial clusters with a weighted distance less than or equal to a preset distance threshold, and merge them into a new cluster.

[0150] The new cluster after merging contains all sample medical feature vectors in the original two clusters. Recalculate the weighted distance between the new cluster and other unmerged clusters, and update the sample cluster distance matrix.

[0151] Repeat the above iterative merging process to find clusters that meet the conditions for merging and update the sample cluster distance matrix. Stop iteration when the number of independent initial clusters reaches a preset number, and obtain an initial classification result of the sample object.

[0152] Step S250: Obtain the constraint conditions of the initial decision tree model, including Gini coefficient constraints and medical feature vector threshold constraints.

[0153] To train a suitable decision tree model, constraints of an initial decision tree model need to be obtained. A Gini coefficient constraint is used to measure the purity of a decision tree node, i.e., the consistency of samples in the node. The smaller the Gini coefficient, the more consistent the classes of samples in the node.

[0154] A medical feature vector threshold constraint is used to limit the value range of a medical feature vector used when splitting a decision tree. For example, for some medical feature vectors, there may be a reasonable upper and lower limit range for their values. When splitting a decision tree, the split threshold needs to be within this range.

[0155] Step S260: According to each of the initial classification results and the constraints, training the initial decision tree model to generate the decision tree model.

[0156] Specifically, this step can include the following sub-steps:

[0157] Step S261: Obtaining the classification label corresponding to each sample medical feature vector in the initial classification result, and the Gini coefficient constraint and medical feature vector threshold constraint in the constraint condition.

[0158] When training an initial decision tree model, first, the classification label corresponding to each sample medical feature vector in the initial classification result needs to be obtained. These classification labels represent the class to which the sample object belongs, such as different treatment classes.

[0159] At the same time, the Gini coefficient constraint and the medical feature vector threshold constraint in the constraint condition are obtained. The Gini coefficient constraint is a pre-set threshold used to control the purity of a decision tree node; the medical feature vector threshold constraint specifies the value range of a medical feature vector when splitting a decision tree. This medical feature vector threshold constraint can be determined according to a large amount of clinical data, for example, when the target object's PaO2<60 mmHg, the decision tree model can directly determine the target object as a specific object treatment class label (such as a high-risk label).

[0160] Step S262: Based on the classification label, traversing each medical feature vector in the set of sample medical feature vectors, calculating the Gini coefficient of the medical feature vector under different split thresholds, and selecting the split feature with the smallest Gini coefficient and the corresponding split threshold as the candidate split node.

[0161] Based on the classification label, each medical feature vector in the set of sample medical feature vectors is traversed. For each medical feature vector, different split thresholds are tried. For each split threshold, the set of sample medical feature vectors is divided into two subsets, and the Gini coefficient is calculated according to the sample distribution of the two subsets.

[0162] The calculation method of the Gini coefficient is as follows: assuming that a node contains N samples, and the number of samples belonging to category c is n c, the Gini coefficient of the node is equal to 1 minus the sum of the squares of the proportions of the samples of each category c in the total number of samples. For each split threshold, the Gini coefficients of the two subsets after splitting are calculated, and then a weighted average is taken according to the number of samples in the two subsets to obtain the Gini coefficient under the split threshold.

[0163] The split feature and the corresponding split threshold with the smallest Gini coefficient are screened out as the candidate split node. This process is to find the split mode that can make the purity of the subsets after splitting the highest.

[0164] Step S263: According to the medical feature vector threshold constraint, the split threshold in the candidate split node is normalized and checked. If the split threshold is in the numerical range of the normalized medical feature vector, the candidate split node is retained, otherwise the candidate split node is removed.

[0165] According to the medical feature vector threshold constraint, the split threshold in the candidate split node is normalized and checked. In the data preprocessing stage, the medical feature vector is normalized so that its value range is within a certain interval. The split threshold in the candidate split node is compared with the numerical range of the normalized medical feature vector.

[0166] If the split threshold is within the numerical range, the candidate split node is retained because the split threshold is reasonable; otherwise, the candidate split node is removed to ensure that the split threshold used when the decision tree is split is effective.

[0167] Step S264: Based on the retained candidate split nodes, a split priority queue is generated in ascending order of Gini coefficient, and the candidate split node with the smallest Gini coefficient is selected from the split priority queue as the current node split condition, and the sample medical feature vector set is divided into two subsets.

[0168] Based on the retained candidate split nodes, a split priority queue is generated in ascending order of Gini coefficient. The candidate split nodes with smaller Gini coefficients are arranged at the front of the queue, indicating that these split nodes can make the purity of the subsets after splitting higher.

[0169] The candidate split node with the smallest Gini coefficient is selected from the queue as the split condition of the current node. According to the split condition, the sample medical feature vector set is divided into two subsets. For example, if the split condition is that a certain medical feature vector is greater than a certain split threshold, the samples in the sample medical feature vector set whose medical feature vector is greater than the split threshold are divided into one subset, and the samples whose medical feature vector is less than or equal to the split threshold are divided into another subset.

[0170] Step S265: recursively traversing the two subsets, repeating the steps of calculating the Gini coefficient, screening the split feature and split threshold, verifying the split threshold, and generating the split priority queue, until a preset stopping condition is met, the preset stopping condition including that the Gini coefficient of the subset is zero, that all medical feature vectors do not satisfy the medical feature vector threshold constraint, or that the maximum recursion depth is reached.

[0171] The two subsets obtained by recursive traversal are recursively traversed, and the above steps of calculating the Gini coefficient, screening the split feature and split threshold, verifying the split threshold, and generating the split priority queue are repeated for each subset.

[0172] When the preset stopping condition is met, the recursion is stopped. The preset stopping condition includes that the Gini coefficient of the subset is zero, indicating that the samples in the subset belong to the same class and do not need to be split; that all medical feature vectors do not satisfy the medical feature vector threshold constraint, i.e., no suitable split method is found; or that the maximum recursion depth is reached, avoiding overfitting caused by too deep decision trees.

[0173] Step S266: generating a tree structure containing classification rules and split thresholds from all nodes after recursive splitting, as the decision tree model.

[0174] When the recursive process is completed, a tree structure is generated from all nodes after recursive splitting. Each node contains a split feature, a split threshold, and a classification rule. The classification rule is determined according to the child nodes of the node, for example, if a node is split into two child nodes corresponding to different classes, the classification rule of the node can be determined according to the value of the sample in the split feature.

[0175] These nodes are organized in a tree structure to form a decision tree model containing classification rules and split thresholds. This decision tree model can be used to classify the medical feature vector set of the target object and determine its object treatment class label.

[0176] Step S150: determining the recommended treatment plan for the target object according to the object treatment class label.

[0177] After determining the object treatment class label of the target object, the recommended treatment plan for the target object is determined according to the pre-established mapping relationship between the object treatment class label and the recommended treatment plan. This mapping relationship can be established through expert experience, clinical research, and historical data.

[0178] For example, for a certain object treatment category label, it can correspond to a certain treatment plan, including drug treatment, respiratory support treatment, etc. The drug treatment plan can select the appropriate drug type, dose and use frequency according to the condition and physical condition of the target object. The respiratory support treatment plan can include non-invasive ventilation, invasive ventilation and other different ways.

[0179] For example, the object treatment category label includes low risk, medium risk and high risk. The recommended treatment plan corresponding to each object treatment category label can be, for example: matching a relatively mild recommended treatment plan such as non-invasive ventilation for a low-risk target object; matching a medium-intensity recommended treatment plan such as invasive mechanical ventilation for a medium-risk target object; and matching a high-intensity recommended treatment plan such as extracorporeal membrane oxygenation (ECMO) for a high-risk target object.

[0180] In one possible implementation, after the classification processing of the target object according to the medical feature vector set is performed and the object treatment category label of the target object is determined, the condition of the target object can be monitored, and the object treatment category label of the target object can be dynamically adjusted according to the result of the condition monitoring, so as to match a more suitable recommended treatment plan for the target object. Figure 3 Another flowchart of a respiratory failure treatment plan recommendation method based on artificial intelligence provided by an embodiment of the present application is shown in FIG. 6. As shown in FIG. 6, the method can further include: Figure 3

[0181] Step S310: obtaining condition monitoring quantitative data of the target object according to a preset batch processing configuration parameter.

[0182] The condition monitoring quantitative data includes vital sign quantitative data, blood gas analysis quantitative data, and lesion trend quantitative data.

[0183] The preset batch processing configuration parameter is a parameter preset in advance, which is used to control the acquisition process of the condition monitoring quantitative data. According to these parameters, the condition monitoring quantitative data of the target object can be obtained from the related data source.

[0184] The vital sign quantitative data can be obtained from real-time monitoring equipment, such as heart rate, blood pressure, respiratory rate, etc. These data reflect the basic physical function state of the target object. The blood gas analysis quantitative data is obtained by detecting the blood sample of the target object, including partial pressure of oxygen, partial pressure of carbon dioxide, acid-base value, etc., which can directly reflect the respiratory function and acid-base balance state of the target object.

[0185] The lesion trend quantitative data can be obtained by analyzing the results of multiple image examinations of the target object. For example, by comparing the chest CT scan images at different times, the size, number, density and other changes of the lung lesions are analyzed to quantify the development trend of the lesions.​

[0186] Step S320: Obtain the progression weight corresponding to the vital sign quantification data, the blood gas analysis quantification data, and the lesion trend quantification data, respectively.

[0187] Specifically, the specific process of this step is as follows:

[0188] Step S321: Obtain a sample medical feature vector set of each sample patient from the pre-constructed disease progression history data of a plurality of sample patients, each sample medical feature vector in the sample medical feature vector set comprising vital sign features, blood gas analysis features, and lesion trend features, and having been normalized.

[0189] The pre-constructed disease progression history data of a plurality of sample patients contains a large amount of disease information of sample patients at different times. From these data, a sample medical feature vector set of each sample patient is extracted. Each sample medical feature vector comprises vital sign features, blood gas analysis features, and lesion trend features.

[0190] During the extraction process, these features are normalized to eliminate the scale difference between different features. Normalization can use methods such as min-max normalization or mean-standard deviation normalization to ensure that the value range of all features is within a reasonable interval.

[0191] Step S322: Extract the disease deterioration index of the sample patient as the regression target variable, the disease deterioration index being generated according to the occurrence time and severity quantification of a pre-set deterioration event.

[0192] The disease deterioration index is extracted from the disease progression history data of the sample patient as the regression target variable. The pre-set deterioration event can include exacerbation of respiratory failure, the need for invasive ventilation, etc. According to the occurrence time and severity of these deterioration events, the disease deterioration index is quantified.

[0193] For example, different weights can be assigned to different deterioration events, and a weighted score is calculated according to the occurrence time and severity of the deterioration event as the disease deterioration index.

[0194] Step S323: Fit a linear regression model or a logistic regression model with the vital sign features, the blood gas analysis features, and the lesion trend features as independent variables, and output the regression coefficients corresponding to each feature.

[0195] The normalized vital sign features, blood gas analysis features, and lesion trend features are used as independent variables, and the disease deterioration index is used as the dependent variable to fit a linear regression model or a logistic regression model.

[0196] The general form of a linear regression model is that the dependent variable equals a constant term plus the sum of each independent variable multiplied by the corresponding regression coefficient. A logistic regression model is used to handle binary classification problems, such as determining whether a disease condition is worsening. By fitting the model, the regression coefficient corresponding to each feature can be obtained.

[0197] Step S324: Based on the absolute value of the regression coefficient of each feature, the importance indicator of each feature is calculated.

[0198] Based on the absolute value of the regression coefficient of each feature, the importance indicator of each feature is calculated. The larger the absolute value of the regression coefficient, the greater the influence of the feature on the disease condition worsening index, and the higher the importance indicator.

[0199] Suppose the regression coefficients of the vital sign features, blood gas analysis features, and lesion trend features are and respectively, then their importance indicators are equal to the absolute values of their respective regression coefficients.

[0200] Step S325: The importance indicators are normalized so that the sum of the progression weights of each feature is 1, generating progression weights corresponding to the vital sign quantization data, the blood gas analysis quantization data, and the lesion trend quantization data.

[0201] The importance indicators of each feature are normalized to convert them into progression weights. The specific method is to divide the importance indicator of each feature by the sum of the importance indicators of all features.

[0202] Suppose the progression weights corresponding to the vital sign quantization data, the blood gas analysis quantization data, and the lesion trend quantization data are and respectively, then the progression weight corresponding to the vital sign quantization data is equal to the importance indicator of the vital sign feature divided by (the importance indicator of the vital sign feature plus the importance indicator of the blood gas analysis feature plus the importance indicator of the lesion trend feature), the progression weight corresponding to the blood gas analysis quantization data is equal to the importance indicator of the blood gas analysis feature divided by (the importance indicator of the vital sign feature plus the importance indicator of the blood gas analysis feature plus the importance indicator of the lesion trend feature), and the progression weight corresponding to the lesion trend quantization data is equal to the importance indicator of the lesion trend feature divided by (the importance indicator of the vital sign feature plus the importance indicator of the blood gas analysis feature plus the importance indicator of the lesion trend feature). After such calculation, the progression weights corresponding to the vital sign quantization data, the blood gas analysis quantization data, and the lesion trend quantization data are generated, and the sum of these weights is 1, which can reasonably reflect the importance of each part of the data in assessing the progression of the disease condition.

[0203] Step S330: determining the illness progression quantification index of the target object based on the vital sign quantification data, the blood gas analysis quantification data, the lesion trend quantification data, and the progression weight according to an attention mechanism.

[0204] The attention mechanism can make the model pay more attention to the data part that has a greater impact on the illness progression. When determining the illness progression quantification index of the target object, the relationship between each quantification data and the corresponding progression weight is first considered.

[0205] For the vital sign quantification data, the data of each dimension is associated with the vital sign progression weight. For example, the vital sign quantification data includes multiple dimensions such as heart rate, blood pressure, and respiratory rate. The data of each dimension is multiplied by the vital sign progression weight. Assuming that the vital sign quantification data is a vector containing multiple dimension values, each dimension value in the vector is multiplied by the vital sign progression weight to obtain the vital sign quantification data vector after the weight action.

[0206] Similarly, for the blood gas analysis quantification data, it is also a vector containing multiple dimension values. Each dimension value in the vector is multiplied by the blood gas analysis progression weight to obtain the blood gas analysis quantification data vector after the weight action.

[0207] For the lesion trend quantification data, it is also a vector containing multiple dimension values. Each dimension value in the vector is multiplied by the lesion trend progression weight to obtain the lesion trend quantification data vector after the weight action.

[0208] Then, the data after the weight processing is integrated based on the attention mechanism. The attention mechanism can dynamically adjust the attention degree according to the importance of different data to the illness progression in different situations. An attention function is used to calculate the attention coefficient of each data dimension. For each dimension value in the vital sign quantification data vector after the weight action, the corresponding attention coefficient is calculated using the attention function; for each dimension value in the blood gas analysis quantification data vector and the lesion trend quantification data vector after the weight action, the corresponding attention coefficient is also calculated using the attention function.

[0209] Finally, the data adjusted by the attention coefficient is summarized to obtain the illness progression quantification index. Specifically, the dimension values of the vital sign quantification data after the weight and attention coefficient action are added to obtain the summary value of the vital sign quantification data; the dimension values of the blood gas analysis quantification data after the weight and attention coefficient action are added to obtain the summary value of the blood gas analysis quantification data; the dimension values of the lesion trend quantification data after the weight and attention coefficient action are added to obtain the summary value of the lesion trend quantification data. Adding the three summary values together, the illness progression quantification index of the target object is obtained.

[0210] Step S340: Based on the disease progression quantification indicator, the current subject treatment category label of the target subject, and the level transition threshold corresponding to the disease progression quantification indicator, it is determined whether the subject treatment category label of the target subject needs to be adjusted.

[0211] Specifically, the specific process of this step can be shown as follows:

[0212] Step S341: A preset level transition threshold range associated with the current subject treatment category label is obtained, and the preset level transition threshold range includes a deterioration threshold upper limit and an improvement threshold lower limit.

[0213] Different subject treatment category labels correspond to different preset level transition threshold ranges. These threshold ranges are pre-set according to a large amount of clinical data and expert experience. A mapping relationship table can be used to store the correspondence between the subject treatment category label and the preset level transition threshold range. According to the current subject treatment category label of the target subject, the deterioration threshold upper limit and the improvement threshold lower limit associated with the current subject treatment category label are found from the mapping relationship table.

[0214] Step S342: The disease progression quantification indicator is compared with the deterioration threshold upper limit in a first comparison, and if the disease progression quantification indicator is greater than or equal to the deterioration threshold upper limit, a category up instruction is generated.

[0215] The calculated disease progression quantification indicator and the deterioration threshold upper limit are compared. If the disease progression quantification indicator is greater than or equal to the deterioration threshold upper limit, it means that the disease of the target subject may be in a state of deterioration, and more aggressive treatment measures need to be taken. At this time, a category up instruction can be generated, which indicates that the subject treatment category label of the target subject is to be adjusted to a higher level of treatment category.

[0216] Step S343: The disease progression quantification indicator is compared with the improvement threshold lower limit in a second comparison, and if the disease progression quantification indicator is less than or equal to the improvement threshold lower limit, a category down instruction is generated.

[0217] Then, the disease progression quantification indicator and the improvement threshold lower limit are compared. If the disease progression quantification indicator is less than or equal to the improvement threshold lower limit, it means that the disease of the target subject has improved significantly, and the original aggressive treatment measures may not be needed. In this case, a category down instruction can be generated, which indicates that the subject treatment category label of the target subject is to be adjusted to a lower level of treatment category.

[0218] Step S344: According to the category up-regulation instruction or the category down-regulation instruction, the adjusted object treatment category label of the target object is determined, and it is verified whether the adjusted object treatment category label is within the preset label level range.

[0219] According to the generated category up-regulation instruction or category down-regulation instruction, the object treatment category label of the target object is adjusted. If it is a category up-regulation instruction, the object treatment category label is adjusted to a higher level of treatment category; if it is a category down-regulation instruction, the object treatment category label is adjusted to a lower level of treatment category.

[0220] After adjustment, it is necessary to verify whether the adjusted object treatment category label is within the preset label level range. The preset label level range specifies the reasonable value range of the object treatment category label, which can avoid unreasonable label adjustment. The preset label level range can be stored in a preset label level list, and it is checked whether the adjusted object treatment category label is in the list.

[0221] Step S345: If the adjusted object treatment category label exceeds the preset label level range, the current object treatment category label is maintained; otherwise, the adjusted object treatment category label is taken as the updated object treatment category label.

[0222] If the adjusted object treatment category label exceeds the preset label level range, it indicates that the adjustment is unreasonable, and the current object treatment category label is maintained to ensure the stability and rationality of the treatment plan. If the adjusted object treatment category label is within the preset label level range, it is taken as the updated object treatment category label for subsequent treatment plan recommendation.

[0223] Step S350: If the object treatment category label of the target object needs to be adjusted, the recommended treatment plan of the target object is determined based on the adjusted object treatment category label.

[0224] If it is determined through the previous steps that the object treatment category label of the target object needs to be adjusted, the recommended treatment plan of the target object is re-determined according to the adjusted object treatment category label. Similarly, according to the pre-established mapping relationship between the object treatment category label and the recommended treatment plan, the recommended treatment plan corresponding to the adjusted object treatment category label is found from the mapping relationship.

[0225] For example, if the adjusted subject treatment category label corresponds to a higher level of treatment regimen, it can increase the dosage of medication used, replace more effective drugs, or adopt more aggressive respiratory support treatment methods. If the adjusted subject treatment category label corresponds to a lower level of treatment regimen, it can reduce the use of drugs, reduce the intensity of respiratory support, etc. In this way, the recommended treatment regimen is dynamically adjusted according to the progress of the target subject's condition, thereby improving the effectiveness and pertinence of treatment.

[0226] Throughout the process, when data collection is involved, privacy-sensitive data such as patient personal information, detailed medical record information, etc. that may be involved can be protected by various privacy protection and leakage prevention techniques. In the data collection stage, the data can be encrypted using symmetric encryption algorithms or asymmetric encryption algorithms to convert the data into ciphertext, and only authorized personnel can decrypt and view it. In terms of data storage, a secure storage system is used, and strict access permission control is set, so that only personnel with corresponding permissions can access the data. At the same time, regular backups are made to prevent data loss or damage. During data transmission, a secure network protocol such as SSL / TLS protocol is used to ensure the security of the data during transmission. Through these technical means, the privacy-sensitive data of patients can be effectively protected to prevent data leakage and misuse.

[0227] The above embodiments describe in detail the various steps and implementation details of the artificial intelligence-based respiratory failure treatment regimen recommendation method, enabling those skilled in the art to implement the solution of the invention based on these descriptions. Starting from obtaining multi-modal medical acquisition data of the target subject, through data preprocessing, feature extraction, classification processing, etc. to determine the subject treatment category label, and then dynamically adjusting the label according to the progress of the condition and determining the recommended treatment regimen, the entire process forms a complete closed loop, which can provide more accurate and personalized treatment regimen recommendations for respiratory failure patients.

[0228] Figure 4 A structural diagram of an artificial intelligence-based respiratory failure treatment regimen recommendation system 100 provided by the embodiments of the present application is shown in FIG. 1. As shown in FIG. 1, a processor 120 can be used in the artificial intelligence-based respiratory failure treatment regimen recommendation system 100 and used to execute the functions in the present application. Figure 4

[0229] The artificial intelligence-based respiratory failure treatment regimen recommendation system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the artificial intelligence-based respiratory failure treatment regimen recommendation method of the present application. Although the present application only shows one server, for the sake of convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.​

[0230] For example, the artificial intelligence-based respiratory failure treatment regimen recommendation system 100 can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. The artificial intelligence-based respiratory failure treatment regimen recommendation system 100 can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof, for example. The method of the present application can be implemented according to these program instructions. The artificial intelligence-based respiratory failure treatment regimen recommendation system 100 also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0231] For ease of illustration, only one processor is described in the artificial intelligence-based respiratory failure treatment regimen recommendation system 100. However, it should be noted that the artificial intelligence-based respiratory failure treatment regimen recommendation system 100 in the present application can also include multiple processors, so the steps performed by one processor described in the present application can also be jointly performed or individually performed by multiple processors. For example, if the processor of the artificial intelligence-based respiratory failure treatment regimen recommendation system 100 performs steps A and B, it should be understood that steps A and B can also be jointly performed by two different processors or individually performed in one processor. For example, a first processor performs step A, a second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0232] The embodiment of the present application discloses a computer readable storage medium storing a computer program for electronic data exchange, wherein the computer program causes a computer to perform the steps in the artificial intelligence-based respiratory failure treatment regimen recommendation method described in the foregoing embodiments.

[0233] The embodiment of the present application discloses a computer program product including a non-transitory computer readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps in the artificial intelligence-based respiratory failure treatment regimen recommendation method described in the foregoing embodiments.

[0234] The apparatus embodiments described above are only illustrative, wherein the modules described as separate components can or can not be physically separated, and the components displayed as modules can or can not be physical modules, i.e., can be located in one place or distributed on multiple network modules. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0235] Through the specific description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be realized by means of software and necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, including Read-Only Memory (ROM), Random Access Memory (RAM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), One-time Programmable Read-Only Memory (OTPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Compact Disc Read-Only Memory (CD-ROM) or other optical disk storage, magnetic disk storage, magnetic tape storage, or any other medium capable of storing data and computer readable.

[0236] Finally, it should be noted that: the above disclosed is only the preferred embodiment of the present application, only for the purpose of describing the technical solutions of the present application, not for its limitation; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand; it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for recommending treatment options for respiratory failure based on artificial intelligence, characterized in that: include: Acquire first multimodal medical collected data of the target object, where the first multimodal medical collected data includes text modality data and image modality data, and the text modality data includes physiological characteristic data, basic disease data, and blood gas analysis data; Performing data preprocessing on the first multimodal medical collected data to obtain second multimodal medical collected data, wherein the data preprocessing includes data cleaning, normalization, and time window alignment; extracting a medical feature vector set of the target object from the second multimodal medical acquisition data; Performing classification processing on the target object according to the medical feature vector set to determine the object treatment category label of the target object includes (1) or (2): (1) determining an initial cluster center based on the medical feature vectors included in the medical feature vector set; obtaining a medical weight coefficient corresponding to each medical feature vector based on the treatment risk quantification index of the medical feature vector; determining a distance between each medical feature vector and the initial cluster center based on each medical feature vector and the medical weight coefficient corresponding to each medical feature vector, and generating an initial clustering result based on the distance between each medical feature vector and the initial cluster center; iteratively updating the initial cluster center based on the initial clustering result, the medical feature vector of the target object, and the medical weight coefficient until a preset termination condition is met, obtaining a cluster analysis result of the target object, and determining an object treatment category label of the target object based on the cluster analysis result and the mapping relationship between the cluster analysis result and the object treatment category label; or, (2) Based on the distribution characteristics of each sample medical feature vector in the sample medical feature vector set of the sample object, construct an initial state of hierarchical clustering, wherein each sample medical feature vector in the initial state is an independent initial cluster; based on the treatment risk quantification index of the sample medical feature vector, obtain the medical weight coefficient corresponding to each sample medical feature vector; according to each sample medical feature vector and the medical weight coefficient corresponding to each sample medical feature vector, calculate the weighted distance between the independent initial clusters to generate a sample cluster distance matrix; according to each weighted distance in the sample cluster distance matrix, iteratively merge the independent initial clusters whose weighted distance is less than or equal to the preset distance threshold until the number of clusters reaches the preset number, and obtain the initial classification result of the sample object; Obtaining constraints of an initial decision tree model, the constraints including a Gini coefficient constraint and a medical feature vector threshold constraint; training the initial decision tree model based on the initial classification results and the constraints to generate a decision tree model; inputting the medical feature vector set into the decision tree model to obtain a treatment category label for the target subject; Acquire the target subject's disease monitoring quantitative data according to preset batch configuration parameters, wherein the disease monitoring quantitative data includes vital sign quantitative data, blood gas analysis quantitative data, and pathological trend quantitative data; Obtaining progression weights corresponding to the vital sign quantitative data, the blood gas analysis quantitative data, and the pathological trend quantitative data respectively; Determining a quantitative indicator of the target subject's disease progression based on an attention mechanism according to the quantitative vital sign data, the quantitative blood gas analysis data, the quantitative lesion trend data, and the progression weight; determining whether it is necessary to adjust the subject treatment category label of the target subject based on the disease progression quantification indicator, the current subject treatment category label of the target subject, and a level conversion threshold corresponding to the disease progression quantification indicator; If the subject treatment category label of the target subject needs to be adjusted, a recommended treatment regimen for the target subject is determined based on the adjusted subject treatment category label.

2. The method for recommending a treatment plan for respiratory failure based on artificial intelligence according to claim 1, characterized in that: The performing cluster analysis based on the medical feature vector set to obtain the cluster analysis result of the target object includes: Based on the distribution characteristics of each medical feature vector in the medical feature vector set, constructing an initial state of hierarchical clustering, wherein each medical feature vector in the initial state is an independent initial cluster; Based on the treatment risk quantification index of the medical feature vector, obtaining the medical weight coefficient corresponding to each medical feature vector; Calculating the weighted distances between the independent initial clusters based on the medical feature vectors and the medical weight coefficients corresponding to the medical feature vectors to generate a cluster distance matrix; According to each weighted distance in the cluster distance matrix, independent initial clusters whose weighted distance is less than or equal to a preset distance threshold are iteratively merged until the number of clusters of the independent initial clusters reaches a preset number, thereby obtaining a cluster analysis result of the target object.

3. The method for recommending a treatment plan for respiratory failure based on artificial intelligence according to claim 1, characterized in that: The step of training the initial decision tree model according to the initial classification results and the constraint conditions to generate the decision tree model includes: Obtaining the classification label corresponding to each sample medical feature vector in the initial classification result, as well as the Gini coefficient constraint and the medical feature vector threshold constraint in the constraint conditions; Based on the classification label, traverse each medical feature vector in the sample medical feature vector set, calculate the Gini coefficient of the medical feature vector under different splitting thresholds, and select the splitting feature with the smallest Gini coefficient and the corresponding splitting threshold as a candidate splitting node; According to the medical feature vector threshold constraint, a normalization check is performed on the splitting threshold in the candidate splitting node; if the splitting threshold is within the numerical range of the normalized medical feature vector, the candidate splitting node is retained; otherwise, the candidate splitting node is eliminated; Based on the retained candidate split nodes, sorting them in ascending order of the Gini coefficient to generate a split priority queue, selecting the candidate split node with the smallest Gini coefficient from the split priority queue as the current node split condition, and dividing the sample medical feature vector set into two subsets; Recursively traversing the two subsets, repeatedly calculating the Gini coefficient, screening the splitting features and the splitting threshold, verifying the splitting threshold, and generating the splitting priority queue until a preset stopping condition is met, wherein the preset stopping condition includes that the Gini coefficient of the subset is zero, the splitting thresholds of all medical feature vectors do not meet the medical feature vector threshold constraint, or the maximum recursion depth is reached; A tree structure including classification rules and splitting thresholds is generated according to all nodes after recursive splitting as the decision tree model.

4. The method for recommending a treatment plan for respiratory failure based on artificial intelligence according to claim 1, characterized in that: The obtaining of the progress weights corresponding to the vital sign quantitative data, the blood gas analysis quantitative data, and the pathological trend quantitative data respectively includes: Obtaining a sample medical feature vector set of each sample patient from pre-constructed disease progression history data of multiple sample patients, wherein each sample medical feature vector in the sample medical feature vector set includes vital sign features, blood gas analysis features, and pathological trend features, and has been normalized; Extracting the disease exacerbation index of the sample patient as a regression target variable, wherein the disease exacerbation index is quantitatively generated according to the occurrence time and severity of a preset exacerbation event; Using the vital sign characteristics, the blood gas analysis characteristics, and the pathological trend characteristics as independent variables, fitting a linear regression model or a logistic regression model, and outputting the regression coefficient corresponding to each characteristic; Calculate the importance index of each feature based on the absolute value of the regression coefficient of each feature; The importance index is normalized so that the sum of the progress weights of the features is 1, and progress weights corresponding to the vital sign quantitative data, the blood gas analysis quantitative data, and the pathological trend quantitative data are generated.

5. The method for recommending a treatment plan for respiratory failure based on artificial intelligence according to claim 1, characterized in that: The determining whether it is necessary to adjust the subject treatment category label of the target subject based on the disease progression quantification indicator, the current subject treatment category label of the target subject, and the level conversion threshold corresponding to the disease progression quantification indicator includes: Obtaining a preset level conversion threshold range associated with the current subject treatment category label, wherein the preset level conversion threshold range includes an upper deterioration threshold limit and a lower improvement threshold limit; performing a first comparison between the disease progression quantification indicator and the upper deterioration threshold, and generating a category upward adjustment instruction if the disease progression quantification indicator is greater than or equal to the upper deterioration threshold; Performing a second comparison between the disease progression quantification index and the improvement threshold lower limit, and generating a category downgrade instruction if the disease progression quantification index is less than or equal to the improvement threshold lower limit; determining an adjusted subject treatment category label for the target subject according to the category-up instruction or the category-down instruction, and verifying whether the adjusted subject treatment category label is within a preset label hierarchy range; If the adjusted object treatment category label exceeds the preset label level range, the current object treatment category label is maintained unchanged; otherwise, the adjusted object treatment category label is used as the updated object treatment category label.

6. An artificial intelligence-based respiratory failure treatment plan recommendation system, characterized in that: The invention comprises a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by a computer, the method for recommending a respiratory failure treatment plan based on artificial intelligence according to any one of claims 1 to 5 is implemented.

Images