Premature infant health monitoring and predicting method and system based on deep learning method

Abstract

The invention discloses a premature infant health monitoring and predicting method and system based on a deep learning method. The method comprises the following steps: acquiring basic information data and follow-up visit information data of a premature infant; obtaining a health monitoring result and a health prediction result of the premature infant based on a pre-trained premature infant health monitoring and prediction model according to the basic information data and the follow-up information data; the premature infant health monitoring and prediction model comprises a first input layer, a plurality of types of channels, a first splicing layer, a first full connection layer, a first output layer, a second input layer, a second input layer, an LSTM network, a second splicing layer, a second full connection layer and a second output layer. Through multi-source data integration, multi-channel and LSTM feature extraction, premature infant health monitoring and prediction are realized, and a more comprehensive reference is provided for medical decision.

Description

Technical Field

[0001] The present invention relates to a method and system for monitoring and predicting the health of premature infants based on a deep learning method, and belongs to the technical field of health monitoring and prediction. Background Art

[0002] The clinical diagnosis and treatment technology of "intensive care for premature infants" in NICU in my country has made a series of developments: the clinical application of respiratory therapy technology, surfactant and extracorporeal membrane oxygenation (ECMO) technology has greatly promoted the improvement of the treatment level of respiratory failure in premature infants; the clinical popularization of evidence-based NICU quality improvement concepts has made the treatment technology of critically ill premature infants continue to be standardized. Under this development trend, comprehensive monitoring and prediction of the health status of premature infants has become an urgent need. Traditional health monitoring methods mainly rely on the experience of doctors and manual testing, which has the problems of strong subjectivity, time-consuming, and easy to miss or misdiagnose.

[0003] Advances in science and technology have made it possible to integrate different monitoring technologies. A bedside monitoring device can simultaneously monitor multiple organ functions non-invasively, continuously monitoring heart rate, oxygen saturation, and arterial pressure through a series of non-invasive probes or sensors. Important parameters such as body temperature, respiratory rate, skin perfusion, and blood sugar, as well as biomarkers such as organ function, infection, and metabolism, have been integrated to realize clinical practice of monitoring and changing treatment.

[0004] Artificial intelligence (AI) algorithms have shown great potential in medical work. AI can monitor and interpret various important parameters and physiological and pathological data of premature infants. New deep learning provides a new path for big data analysis and data mining of premature infants' health status. It can discover the patterns hidden in large-scale data in clinical information systems and optimize clinical decision-making. However, there are still a series of challenges for health monitoring and prediction of premature infants, especially for comprehensive prediction and diagnosis based on basic data and follow-up information. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method and system for premature infant health monitoring and prediction based on deep learning methods. Based on a pre-trained premature infant health monitoring and prediction model, premature infant health monitoring and prediction are realized through multi-source data integration, providing a more comprehensive reference for medical decision-making.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] On the one hand, the present invention discloses a method for monitoring and predicting the health of premature infants based on a deep learning method, comprising the following steps:

[0008] Obtain basic information and follow-up information data of premature infants;

[0009] According to the basic information data and the follow-up information data, based on the pre-trained premature infant health monitoring and prediction model, the health monitoring results and health prediction results of the premature infant are obtained;

[0010] The premature infant health monitoring and prediction model includes a first input layer, multiple categories of channels, a first splicing layer, a first fully connected layer, a first output layer, a second input layer, a second input layer, an LSTM network, a second splicing layer, a second fully connected layer, and a second output layer;

[0011] Based on the first input layer, data classification is performed according to the basic information data of the premature infant to obtain system indicator data of multiple categories;

[0012] Based on any category of channels, the extracted features are obtained according to the system indicator data of the corresponding category;

[0013] Based on the first concatenated layer, the extracted features of the channels of multiple categories are fused to obtain a first fused feature;

[0014] Based on the first fully connected layer, extract the first fusion feature to obtain a first fully connected feature;

[0015] Based on the first output layer, and according to the first fully connected feature, obtaining a health monitoring result;

[0016] Based on the second input layer, obtaining time series data according to the follow-up information data;

[0017] Based on the LSTM network, and according to the time series data, LSTM features are obtained;

[0018] Based on the second concatenation layer, the first fully connected feature and the LSTM feature are fused to obtain a second fused feature;

[0019] Based on the second fully connected layer, extract the second fusion feature to obtain a second fully connected feature;

[0020] Based on the second output layer and according to the second fully connected features, a health prediction result is obtained.

[0021] Furthermore, the basic information data of the premature infant includes basic information data, perinatal data, transport data and physiological system indicator data, and the basic information data includes body temperature data, weight data and height data;

[0022] The follow-up information data include premature infant growth curve and physical index data.

[0023] Furthermore, the categories of the system indicator data include respiratory system indicator data, circulatory system indicator data, nervous system indicator data, digestive system indicator data, blood system indicator data, endocrine and metabolic indicator data, infection status indicator data, ROP screening indicator data, hearing screening indicator data, critical illness screening indicator data, and nutrition-related indicator data.

[0024] Furthermore, the channel includes a matrix conversion unit, a plurality of convolutional layers and pooling layers connected in series;

[0025] For any category of channels, based on the matrix conversion unit, the system indicator data of the corresponding category is converted into a data matrix of a preset size;

[0026] Based on the convolution layer, applying a convolution operation to the data matrix or channel extraction feature to extract local features;

[0027] Based on the pooling layer, a pooling operation is performed on the local features to obtain channel extraction features;

[0028] The expression of the preset size is n×n, and satisfies n=ceiling(sqrt(m)), where ceiling represents a function rounded to infinity; sqrt represents a square root function; and m represents the amount of data in the channel with the largest amount of data.

[0029] Furthermore, based on the first concatenated layer, the extracted features of the channels of multiple categories are fused by weighted summation to obtain a first fused feature;

[0030] The expression of the weighted summation is as follows:

[0031] F 1 =ω 1 ·f 1 +ω 2 ·f 2 +…+ω n ·f n +…+ω N ·f N

[0032] In the formula, F 1 represents the first fusion feature; ω n represents the weight of the nth channel; f n Represents the extracted features of the nth channel, n = 1, 2, ..., N, and N represents the total number of channels.

[0033] Furthermore, the LSTM network includes a plurality of LSTM units connected in series;

[0034] According to the hidden state value output by the previous LSTM unit and the current input data, the forget gate, update gate and memory cell candidate values ​​of the current LSTM unit are calculated;

[0035] Calculate the current memory cell value according to the forget gate, the update gate, the memory cell candidate value and the memory cell value of the previous LSTM unit;

[0036] Calculate the hidden state value of the current LSTM unit output according to the hidden state value output by the previous LSTM unit, the current input data and the memory cell value;

[0037] The hidden state value output by the last LSTM unit is used as the output value of the LSTM network.

[0038] Furthermore, the expression of the forget gate is as follows:

[0039]

[0040] In the formula, represents the forget gate; σ represents the Sigmoid activation function; ω f represents the first weight matrix; a t-1 represents the hidden state value output by the t-1th LSTM unit; x t represents the input data of the tth LSTM unit; b f represents the first deviation vector;

[0041] The expression of the update gate is as follows:

[0042]

[0043] In the formula, represents the update gate; ω u represents the second weight matrix; b u represents the second deviation vector;

[0044] The expression of the candidate value of the memory cell is as follows:

[0045] c′ t =tanh(ω c [a t-1 ,x t ]+b c )

[0046] c′ t represents the candidate value of the memory cell; tanh represents the tanh function; ω c represents the third weight matrix; b c represents the third deviation vector;

[0047] The expression of the memory cell value is as follows:

[0048]

[0049] In the formula, c t Represents the memory cell value of the t-th LSTM unit; represents the forget gate; c t-1 Represents the memory cell value of the t-1th LSTM unit; represents the update gate; c′ t represents the candidate value of memory cell;

[0050] The expression of the hidden state value is as follows:

[0051] a t =(ω o [a t-1 ,x t ]+b o )×tanh(c t )

[0052] In the formula, a t Represents the hidden state value output by the tth LSTM unit; ω o represents the fourth weight matrix; b o represents the fourth deviation vector.

[0053] Furthermore, the training method of the premature infant health monitoring and prediction model is as follows:

[0054] Obtain historical basic information data and follow-up information data of premature infants as training sample data, and use the corresponding diagnosis results as training labels;

[0055] According to the training sample data and training labels, the pre-constructed premature infant health monitoring and prediction model is trained based on the preset L2 regularization and cross entropy comprehensive loss function until the preset training termination condition is met to obtain a trained premature infant health monitoring and prediction model.

[0056] Furthermore, the expression of the comprehensive loss function is as follows:

[0057]

[0058]

[0059] Where, L reg represents the comprehensive loss function; m represents the number of training samples; λ represents the regularization coefficient; k represents the number of sample categories; ω k,j Represents the value of the jth element in the kth class sample;

[0060] L exp represents the cross entropy loss function; t ijIndicates whether the true label of the i-th sample is j, if it is j, it is 1, otherwise it is 0; y ij It represents the probability that the i-th sample belongs to j.

[0061] On the other hand, the present invention discloses a premature infant health monitoring and prediction system based on a deep learning method, which is applicable to the above-mentioned premature infant health monitoring and prediction method based on a deep learning method, including:

[0062] A data acquisition module is used to obtain basic information data and follow-up information data of premature infants;

[0063] A premature infant health monitoring and prediction model, used to obtain the premature infant health monitoring results and health prediction results based on the basic information data and follow-up information data;

[0064] Wherein, the premature infant health monitoring and prediction model includes: for,

[0065] The first input layer is used to classify data according to the basic information data of the premature infant to obtain system indicator data of multiple categories;

[0066] Multiple categories of channels are used to obtain extracted features according to system indicator data of corresponding categories;

[0067] The first concatenation layer is used to fuse the extracted features of channels of multiple categories to obtain a first fused feature;

[0068] A first fully connected layer is used to extract the first fusion feature to obtain a first fully connected feature;

[0069] A first output layer, used to obtain a health monitoring result according to the first fully connected feature;

[0070] The second input layer is used to obtain time series data according to the follow-up information data;

[0071] An LSTM network is used to obtain LSTM features according to the time series data;

[0072] The second concatenation layer is used to fuse the first fully connected feature and the LSTM feature to obtain a second fused feature;

[0073] A second fully connected layer is used to extract the second fusion feature to obtain a second fully connected feature;

[0074] The second output layer is used to obtain a health prediction result according to the second fully connected feature.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] The method and system for monitoring and predicting the health of premature infants based on deep learning methods of the present invention, firstly, effectively integrates the basic information data of premature infants based on the pre-trained health monitoring and prediction model of premature infants, and improves the comprehensive analysis ability of the model on the overall health status of premature infants;

[0077] Secondly, the introduction of follow-up information and LSTM network can successfully capture the temporal information of the physiological characteristics and potential diseases of premature infants, and improve the model's prediction accuracy of temporal events; at the same time, the model can predict the development and potential diseases of premature infants based on their follow-up information, providing a more accurate and comprehensive reference for medical decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 is a schematic diagram of a premature infant health monitoring and prediction model provided in an embodiment;

[0079] Figure 2 is a schematic diagram of a premature infant health monitoring module provided in an embodiment;

[0080] Figure 3 Schematic diagram of an LSTM unit provided in an embodiment. DETAILED DESCRIPTION

[0081] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0082] Example 1.

[0083] This embodiment 1 provides a method for monitoring and predicting the health of premature infants based on a deep learning method, comprising the following steps:

[0084] Obtain basic information and follow-up information data of premature infants;

[0085] According to the basic information data and follow-up information data, based on the pre-trained premature infant health monitoring and prediction model, the health monitoring results and health prediction results of premature infants are obtained;

[0086] Among them, Figure 1 As shown, the premature infant health monitoring and prediction model includes a first input layer, multiple categories of channels, a first splicing layer, a first fully connected layer, a first output layer, a second input layer, a second input layer, an LSTM network, a second splicing layer, a second fully connected layer, and a second output layer;

[0087] Based on the first input layer, data classification is performed according to the basic information data of premature infants to obtain system indicator data of multiple categories;

[0088] Based on any category of channels, the extracted features are obtained according to the system indicator data of the corresponding category;

[0089] Based on the first concatenation layer, the extracted features of the channels of multiple categories are fused to obtain the first fused features;

[0090] Based on the first fully connected layer, extract the first fusion feature to obtain the first fully connected feature;

[0091] Based on the first output layer, a health monitoring result is obtained according to the first fully connected feature;

[0092] Based on the second input layer, time series data is obtained according to the follow-up information data;

[0093] Based on the LSTM network, LSTM features are obtained according to time series data;

[0094] Based on the second concatenation layer, the first fully connected feature and the LSTM feature are fused to obtain the second fused feature;

[0095] Based on the second fully connected layer, extract the second fusion feature to obtain the second fully connected feature;

[0096] Based on the second output layer, the health prediction result is obtained according to the second fully connected feature.

[0097] The technical concept of the present invention is as follows: first, based on the pre-trained premature infant health monitoring and prediction model, the basic information data of premature infants is effectively integrated, and the comprehensive analysis ability of the model on the overall health status of premature infants is improved;

[0098] Secondly, the introduction of follow-up information and LSTM network can successfully capture the temporal information of the physiological characteristics and potential diseases of premature infants, and improve the model's prediction accuracy of temporal events; at the same time, the model can predict the development and potential diseases of premature infants based on their follow-up information, providing a more accurate and comprehensive reference for medical decision-making.

[0099] This embodiment is mainly divided into two modules, namely, a premature infant health monitoring module and a premature infant health prediction module. The specific principles are as follows:

[0100] 1. Premature baby health monitoring module.

[0101] like Figure 2 As shown, the premature infant health monitoring module includes a first input layer, multiple categories of channels, a first splicing layer, a first fully connected layer and a first output layer in the premature infant health monitoring and prediction model.

[0102] The work performed by the premature infant health monitoring module can be regarded as a classification task, such as Figure 2As shown in the figure, the module divides the basic information data of premature infants into multiple categories, each of which contains a series of related indicators. At the same time, the module contains multiple channels, corresponding to each type of data in a channel. On this basis, after multiple convolutional layers and pooling layers, the data of each channel is extracted for features, and then the first splicing layer is introduced to perform feature fusion by weighted summation. After that, the first fully connected layer is introduced to learn higher-level features, and finally the classification task result output is completed in the first output layer.

[0103] 1.1 Based on the first input layer, data classification is performed according to the basic information data of premature infants to obtain system indicator data of multiple categories.

[0104] The basic information data of premature infants include basic information data, perinatal data, transfer data and physiological system indicator data. The basic information data include body temperature data, weight data and height data; the physiological system indicator data include body temperature, blood sugar, routine 24-hour body fluid intake and output, blood biochemistry (electrolytes, blood sugar, creatinine, bilirubin, blood calcium); the first input layer obtains various basic information data related to premature infants, classifies them and distributes them to different channels for the next step of processing.

[0105] The data of premature infants can be classified according to the main systems of the human body, including respiratory system indicator data, circulatory system indicator data, nervous system indicator data, digestive system indicator data, blood system indicator data, endocrine and metabolic indicator data, infection status indicator data, ROP screening indicator data, hearing screening indicator data, critical illness screening indicator data, and nutrition-related indicator data.

[0106] These indicators represent the health status of premature infants in different aspects and are used to judge the main diseases of each system. For example, in the respiratory system, the changes in respiratory rate, three-recess sign, cyanosis of lips, apnea monitoring, alveolar-arterial oxygen pressure difference, oxygenation index (OI), arterial / alveolar oxygen pressure ratio and other indicators are used to judge whether the patient has symptoms such as premature respiratory distress syndrome, premature bronchial or pulmonary dysplasia, pneumonia, etc.; in the circulatory system, the extremity temperature, impaired consciousness, edema, urine volume, heart rate, electrocardiogram, blood pressure and other indicators are used to judge whether the patient has hypotension, patent ductus arteriosus and other diseases.

[0107] The changes in indicators of each system correspond to the diseases of that system. The independence of indicators and diseases between systems is relatively high. It is necessary to assign each type of indicator data to a separate channel and extract features for each system of premature infants. This classification method can more carefully reflect the health status of premature infants in different physiological systems, help ensure that the multi-channel neural network can process information from different systems in a targeted manner, and provide more comprehensive and accurate input for the model. This process lays the foundation for the integration of multi-channel data in the premature infant health monitoring module.

[0108] 1.2 Based on any category of channels, the extracted features are obtained according to the system indicator data of the corresponding category.

[0109] Each channel consists of a matrix conversion unit, multiple convolutional layers and pooling layers connected in series.

[0110] Firstly, for any category of channels, based on the matrix conversion unit, the system indicator data of the corresponding category is converted into a data matrix of a preset size;

[0111] The expression of the preset size is n×n, and satisfies n=ceiling(sqrt(m)), where ceiling represents a function rounded to infinity; sqrt represents a square root function; and m represents the number of data in the channel with the largest amount of data. The empty spaces in the data matrix are filled with 0.

[0112] Secondly, based on the convolution layer, a convolution operation is applied to the data matrix or channel extraction features, and the local features are extracted by sliding the convolution kernel on the matrix.

[0113] Since the number of indicators related to the main systems of the human body is small and the amount of calculation is also small, a smaller step size can be used to extract features more finely. The padding size is set to 1, the strides is set to 1, the convolution kernel size kernel_size is set to 3, and the number of convolution kernels filter is set to 5; after the convolution operation, the Relu activation function is used to process each element in the matrix. The expression is as follows:

[0114]

[0115] In the formula, i represents the value of a single element in the matrix; i ′ Represents the result of operating the element i using the Relu activation function.

[0116] Furthermore, based on the pooling layer, the local features are pooled, the maximum pooling algorithm is used, a 2×2 filter is used, and a step size of 2 is used for scanning, and the maximum value is selected for output, thereby reducing the data dimension while retaining the local features to obtain the channel extraction features;

[0117] Finally, repeat the above convolution and pooling operations several times to complete the feature extraction process.

[0118] The whole process helps to extract representative features from the raw data, providing more informative input for subsequent neural network learning. Such feature extraction operations provide the model with a deeper and more effective understanding of the physiological information of different systems in premature infants.

[0119] 1.3 Based on the first concatenated layer, the extracted features of channels of multiple categories are fused to obtain the first fused features.

[0120] The features extracted from each channel are fused together by weighted summation, and the features of each channel are assigned a weight according to their importance. Assuming there are N channels, the features of each channel are represented by f 1 ,f 2 ,…,f N , the corresponding weight is ω 1 ,ω 2 ,…,ω N , the expression of weighted summation is as follows:

[0121] F 1 =ω 1 ·f 1 +ω 2 ·f 2 +…+ω n ·f n +…+ω N ·f N

[0122] In the formula, F 1 represents the first fusion feature; ω n represents the weight of the nth channel; f n Represents the extracted features of the nth channel, n = 1, 2, ..., N, and N represents the total number of channels.

[0123] This process aims to fuse the features from different channels together by weighted summation. The fused features contain information from different channels and can more comprehensively reflect the overall physiological state of premature infants, so that the system can more comprehensively analyze the correlation between different data sources. This helps to improve the model's ability to comprehensively analyze complex physiological information, thereby more accurately diagnosing or predicting diseases.

[0124] 1.4 Based on the first fully connected layer, extract the first fusion feature to obtain the first fully connected feature.

[0125] Based on the first output layer and the first fully connected feature, the health monitoring results are obtained, and then the classification task results are output to classify various diseases that may exist in premature infants.

[0126] The entire process of the premature infant health monitoring module integrates multi-source data into a comprehensive model, enabling the system to more comprehensively capture the various physiological information of premature infants. Through the previous feature extraction and fusion process, the model has learned the key features in different channels, and the first fully connected layer can further integrate these features, improve the ability to grasp complex information, and finally accurately classify diseases in the first output layer.

[0127] 2. Premature baby health prediction module.

[0128] like Figure 1 As shown, the premature infant health prediction module includes the first input layer, multiple categories of channels, the first splicing layer, the first fully connected layer, the second input layer, the second input layer, the LSTM network, the second splicing layer, the second fully connected layer and the second output layer in the premature infant health monitoring and prediction model.

[0129] The work completed by the premature infant health prediction module can be regarded as a prediction task. Based on the data and training methods used by the premature infant health monitoring module, the premature infant follow-up related information is added, and the LSTM network is used to perform time series modeling.

[0130] 2.1 Based on the basic information data of premature infants, classification, feature extraction and fusion are performed to obtain the first fully connected feature. The principle is consistent with the premature infant health monitoring module in the first part, so it will not be repeated here.

[0131] 2.2 Based on the second input layer, time series data is obtained according to the follow-up information data.

[0132] The follow-up information data of premature infants, including growth curves of premature infants, changes in physical indicators, etc., are integrated into an independent channel in the format of time series data to form one-dimensional time series data.

[0133] 2.2 Based on the LSTM network, LSTM features are obtained according to time series data.

[0134] The LSTM network consists of multiple LSTM units connected in series. The structure of each LSTM unit is as follows: Figure 3 As shown, the specific principles are as follows:

[0135] According to the hidden state value output by the previous LSTM unit and the current input data, the forget gate, update gate and memory cell candidate values ​​of the current LSTM unit are calculated;

[0136] Calculate the current memory cell value based on the forget gate, update gate, memory cell candidate value, and the memory cell value of the previous LSTM unit;

[0137] Calculate the hidden state value of the current LSTM unit output according to the hidden state value output by the previous LSTM unit, the current input data and the memory cell value;

[0138] The hidden state value output by the last LSTM unit is used as the output value of the LSTM network.

[0139] The forget gate determines whether to discard some information. The expression of the forget gate is as follows:

[0140]

[0141] In the formula, represents the forget gate; σ represents the Sigmoid activation function; ω f represents the first weight matrix; a t-1 represents the hidden state value output by the t-1th LSTM unit; x t represents the input data of the tth LSTM unit; b f represents the first deviation vector;

[0142] The update gate determines the data stored in the memory cell. The expression of the update gate is as follows:

[0143]

[0144] In the formula, represents the update gate; ω u represents the second weight matrix; b u represents the second deviation vector;

[0145] The expression of the candidate value of the memory cell is as follows:

[0146] c′ t =tanh(ω c [a t-1 ,x t ]+b c )

[0147] c′ t represents the candidate value of the memory cell; tanh represents the tanh function; ω c represents the third weight matrix; b c represents the third deviation vector;

[0148] The expression for the memory cell value is as follows:

[0149]

[0150] In the formula, c t Represents the memory cell value of the t-th LSTM unit; represents the forget gate; c t-1Represents the memory cell value of the t-1th LSTM unit; represents the update gate; c′ t represents the candidate value of memory cell;

[0151] The expression of the hidden state value is as follows:

[0152] a t =(ω o [a t-1 ,x t ]+b o )×tanh(c t )

[0153] In the formula, a t Represents the hidden state value output by the tth LSTM unit; ω o represents the fourth weight matrix; b o represents the fourth deviation vector.

[0154] In this step, the LSTM network is used to model the time series data. A many-to-one structure is used to input the follow-up information of premature infants in time steps. The input of each time step is a time point of the follow-up information, and the feature is output in the last LSTM unit.

[0155] 2.3 Based on the second concatenation layer, the first fully connected feature and the LSTM feature are fused to obtain the second fused feature.

[0156] The fusion of the first fully connected features and the LSTM features requires comprehensive consideration of the different types of features extracted by the two neural networks to retain their respective characteristics and thus improve the expressive power of the model. The features extracted by the multi-channel neural network from different physiological systems mainly reflect the static physiological state of premature infants before discharge; the LSTM network performs time series modeling on the follow-up information of premature infants, and the output hidden state sequence captures the dependencies in the time series information. This part of the features mainly reflects the growth trend and health changes of premature infants during the follow-up period. The first fully connected feature and the hidden state sequence output by the LSTM are connected together using a connection operation to form a new comprehensive feature vector.

[0157] After connecting the first fully connected feature and the LSTM feature output, the model can more comprehensively understand the health status of premature infants. The static features provide detailed information about the physiological state before discharge, while the time series features capture the dynamic changes of premature infants during follow-up. This fusion strategy helps improve the model's adaptability to the task of predicting premature infant diseases, enabling it to more accurately capture the complex characteristics of the overall physiological state of premature infants.

[0158] 2.4 Based on the second fully connected layer, extract the second fusion feature to obtain the second fully connected feature; based on the second output layer, use the Softmax function to obtain the health prediction result according to the second fully connected feature.

[0159] The expression of the Softmax function is as follows:

[0160]

[0161] In the formula, ω i Represents the output value of the i-th node; ω c Represents the output value of any node c; C represents the number of categories; the Softmax function can be used to convert the output values ​​of multiple categories into a probability distribution with a range of [0,1] and a value of 1, thereby achieving the task of predicting various diseases that may exist in premature infants.

[0162] The entire process can not only fully capture the patient's multi-faceted physiological information, but also accurately predict diseases based on the follow-up information of premature infants, providing reliable and comprehensive support for medical decision-making. This comprehensive process enables the system to achieve more outstanding performance in the task of predicting premature infant diseases.

[0163] 3. Model training.

[0164] The training method of the premature infant health monitoring and prediction model is as follows:

[0165] The historical basic information data and follow-up information data of premature infants are obtained through the hospital diagnosis and treatment system as training sample data, and the corresponding diagnosis results are used as training labels;

[0166] According to the training sample data and training labels, based on the preset L2 regularization and cross entropy comprehensive loss function, the pre-constructed premature infant health monitoring and prediction model is trained until the preset training termination conditions are met to obtain a trained premature infant health monitoring and prediction model.

[0167] In this embodiment, the back propagation algorithm is used to train the model by minimizing the comprehensive loss function and adjusting the weights in the multi-channel neural network, LSTM model and fully connected layer. The comprehensive loss function not only takes into account the cross entropy function, but also further uses the L2 regularization method to prevent the model from overfitting.

[0168] The expression of the comprehensive loss function is as follows:

[0169]

[0170]

[0171] Where, L regrepresents the comprehensive loss function; m represents the number of training samples; λ represents the regularization coefficient; k represents the number of sample categories; ω k,j Represents the value of the jth element in the kth class sample;

[0172] L exp represents the cross entropy loss function; t ij Indicates whether the true label of the i-th sample is j, if it is j, it is 1, otherwise it is 0; y ij It represents the probability that the i-th sample belongs to j.

[0173] In summary, by adopting the technical solution of the present invention, health monitoring and prediction of premature infants based on multi-source data integration is realized. The present invention has the following beneficial effects in solving the problems of data integration and time series prediction in health monitoring of premature infants:

[0174] Comprehensive data utilization: Effective integration of basic data, including perinatal conditions, transfer conditions, and indicators of major physiological systems, improves the model's ability to comprehensively analyze the overall health status of premature infants.

[0175] In terms of multi-source data integration, the premature infant health monitoring module uses a multi-channel neural network structure to achieve effective integration of multi-source data. It mainly includes perinatal conditions, transportation conditions, and basic indicator information of major systems such as the respiratory system, circulatory system, nervous system, and digestive system. Each data source is assigned to a different channel, allowing the model to process information from different systems at the same time.

[0176] Through the convolutional layer and pooling layer, the model can extract the features in each channel and then fuse these features. This helps the system to comprehensively analyze the correlation between different data sources, so that the model can better understand the overall health status of premature infants. By integrating multi-source data into a comprehensive model, the system can more comprehensively capture the multi-faceted physiological information of premature infants, providing a basis for accurate disease diagnosis.

[0177] Capturing temporal information: The introduction of temporal models such as long short-term memory networks successfully captured the temporal information of the physiological characteristics and potential diseases of premature infants, improving the model's prediction accuracy for temporal events.

[0178] The present invention selects models such as the long short-term memory network (LSTM) in deep learning for training. The LSTM network is a recurrent neural network suitable for processing time series data, which can capture the time series dependency of data and has good expression ability for the time series information of different systems of premature infants. Through multiple layers of convolutional layers and LSTM layers, the model can effectively learn spatiotemporal features, thereby improving the model's ability to grasp the physiological changes of premature infants. The parameters of the model are learned through the training set, and are continuously adjusted through the back-propagation algorithm to minimize the loss function to improve the generalization performance of the model.

[0179] Prediction of follow-up information: The model can predict the development and potential diseases of premature infants based on their follow-up information, providing a more accurate reference for medical decision-making.

[0180] The present invention uses a multi-channel neural network structure similar to the premature infant health monitoring module. Follow-up information includes growth curves and basic physical indicators, etc., which are also assigned to corresponding channels.

[0181] Through the convolutional layer and the fully connected layer, the model can automatically learn the patterns in the follow-up information. This enables the system to predict the development and potential diseases of premature infants based on their follow-up information. By introducing this specialized module, the system has increased its attention to the growth and development of premature infants after discharge, especially the monitoring and prediction of "NICU premature infants", providing a more comprehensive reference for medical decision-making.

[0182] Example 2.

[0183] This embodiment 2 provides a premature infant health monitoring and prediction system based on a deep learning method, which is applicable to the premature infant health monitoring and prediction method based on a deep learning method in embodiment 1, including:

[0184] A data acquisition module is used to obtain basic information data and follow-up information data of premature infants;

[0185] The premature infant health monitoring and prediction model is used to obtain the premature infant health monitoring results and health prediction results based on the basic information data and follow-up information data;

[0186] Among them, the premature infant health monitoring and prediction model includes:

[0187] The first input layer is used to classify data based on the basic information data of premature infants to obtain system indicator data of multiple categories;

[0188] Multiple categories of channels are used to obtain extracted features according to system indicator data of corresponding categories;

[0189] The first concatenation layer is used to fuse the extracted features of channels of multiple categories to obtain a first fused feature;

[0190] The first fully connected layer is used to extract the first fusion feature and obtain the first fully connected feature;

[0191] A first output layer, used for obtaining a health monitoring result according to the first fully connected feature;

[0192] The second input layer is used to obtain time series data based on the follow-up information data;

[0193] LSTM network, used to obtain LSTM features based on time series data;

[0194] The second concatenation layer is used to fuse the first fully connected feature and the LSTM feature to obtain the second fused feature;

[0195] The second fully connected layer is used to extract the second fusion feature and obtain the second fully connected feature;

[0196] The second output layer is used to obtain a health prediction result according to the second fully connected feature.

[0197] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0198] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0199] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0200] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0201] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for monitoring and predicting the health of premature infants based on deep learning methods, characterized in that: The steps include: Obtain basic information and follow-up information data of premature infants; According to the basic information data and the follow-up information data, based on the pre-trained premature infant health monitoring and prediction model, the health monitoring results and health prediction results of the premature infant are obtained; The premature infant health monitoring and prediction model includes a first input layer, multiple categories of channels, a first splicing layer, a first fully connected layer, a first output layer, a second input layer, a second input layer, an LSTM network, a second splicing layer, a second fully connected layer, and a second output layer; Based on the first input layer, data classification is performed according to the basic information data of the premature infant to obtain system indicator data of multiple categories; Based on any category of channels, the extracted features are obtained according to the system indicator data of the corresponding category; Based on the first concatenated layer, the extracted features of the channels of multiple categories are fused to obtain a first fused feature; Based on the first fully connected layer, extract the first fusion feature to obtain a first fully connected feature; Based on the first output layer, and according to the first fully connected feature, obtaining a health monitoring result; Based on the second input layer, obtaining time series data according to the follow-up information data; Based on the LSTM network, and according to the time series data, LSTM features are obtained; Based on the second concatenation layer, the first fully connected feature and the LSTM feature are fused to obtain a second fused feature; Based on the second fully connected layer, extract the second fusion feature to obtain a second fully connected feature; Based on the second output layer and according to the second fully connected features, a health prediction result is obtained.

2. The method for monitoring and predicting premature infant health based on deep learning method according to claim 1, characterized in that: The basic information data of premature infants include basic information data, perinatal data, transfer data and physiological system indicator data. The basic information data include body temperature data, weight data and height data; the follow-up information data include premature infant growth curve and physical indicator data.

3. The method for monitoring and predicting premature infant health based on deep learning method according to claim 1, characterized in that: The categories of the system indicator data include respiratory system indicator data, circulatory system indicator data, nervous system indicator data, digestive system indicator data, blood system indicator data, endocrine and metabolic indicator data, infection status indicator data, ROP screening indicator data, hearing screening indicator data, critical illness screening indicator data, and nutrition-related indicator data.

4. The method for monitoring and predicting premature infant health based on deep learning method according to claim 1, characterized in that: The channel includes a matrix conversion unit, a plurality of convolutional layers and pooling layers connected in series; For any category of channels, based on the matrix conversion unit, the system indicator data of the corresponding category is converted into a data matrix of a preset size; Based on the convolution layer, applying a convolution operation to the data matrix or channel extraction feature to extract local features; Based on the pooling layer, a pooling operation is performed on the local features to obtain channel extraction features; The expression of the preset size is n×n, and satisfies n=ceiling(sqrt(m)), where ceiling represents a function rounded to infinity; sqrt represents the square root function; m represents the amount of data in the channel with the largest amount of data.

5. The method for monitoring and predicting premature infant health based on deep learning method according to claim 1, characterized in that: Based on the first concatenated layer, the extracted features of the channels of multiple categories are fused by weighted summation to obtain a first fused feature; The expression of the weighted summation is as follows: F1=ω1·f1+ω2·f2+…+ω n ·f n +…+oh N ·f N In the formula, F1 represents the first fusion feature; ω n represents the weight of the nth channel; f n Represents the extracted features of the nth channel, n = 1, 2, ..., N, and N represents the total number of channels.

6. The method for monitoring and predicting premature infant health based on deep learning method according to claim 1, characterized in that: The LSTM network includes a plurality of LSTM units connected in series; According to the hidden state value output by the previous LSTM unit and the current input data, the forget gate, update gate and memory cell candidate values ​​of the current LSTM unit are calculated; Calculate the current memory cell value according to the forget gate, the update gate, the memory cell candidate value and the memory cell value of the previous LSTM unit; Calculate the hidden state value of the current LSTM unit output according to the hidden state value output by the previous LSTM unit, the current input data and the memory cell value; The hidden state value output by the last LSTM unit is used as the output value of the LSTM network.

7. The method for monitoring and predicting premature infant health based on deep learning method according to claim 6, characterized in that: The expression of the forget gate is as follows: In the formula, represents the forget gate; σ represents the Sigmoid activation function; ω f represents the first weight matrix; a t-1 represents the hidden state value output by the t-1th LSTM unit; x t represents the input data of the tth LSTM unit; b f represents the first deviation vector; The expression of the update gate is as follows: In the formula, represents the update gate; ω u represents the second weight matrix; b u represents the second deviation vector; The expression of the candidate value of the memory cell is as follows: c′ t =tanh(ω c [a t-1 ,x t ]+b c ) c′ t represents the candidate value of the memory cell; tanh represents the tanh function; ω c represents the third weight matrix; b c represents the third deviation vector; The expression of the memory cell value is as follows: In the formula, c t Represents the memory cell value of the t-th LSTM unit; represents the forget gate; c t-1 Represents the memory cell value of the t-1th LSTM unit; represents the update gate; c′ t represents the candidate value of memory cell; The expression of the hidden state value is as follows: a t =(ω o [a t-1 ,x t ]+b o )×tanh(c t ) In the formula, a t Represents the hidden state value output by the tth LSTM unit; ω o represents the fourth weight matrix; b o represents the fourth deviation vector.

8. The method for monitoring and predicting premature infant health based on deep learning method according to claim 1, characterized in that: The training method of the premature infant health monitoring and prediction model is as follows: Obtain historical basic information data and follow-up information data of premature infants as training sample data, and use the corresponding diagnosis results as training labels; According to the training sample data and training labels, the pre-constructed premature infant health monitoring and prediction model is trained based on the preset L2 regularization and cross entropy comprehensive loss function until the preset training termination condition is met to obtain a trained premature infant health monitoring and prediction model.

9. The method for monitoring and predicting premature infant health based on deep learning method according to claim 8, characterized in that: The expression of the comprehensive loss function is as follows: Where, L reg represents the comprehensive loss function; m represents the number of training samples; λ represents the regularization coefficient; k represents the number of sample categories; ω k,j Represents the value of the jth element in the kth class sample; L exp represents the cross entropy loss function; t ij Indicates whether the true label of the i-th sample is j, if it is j, it is 1, otherwise it is 0; y ij It represents the probability that the i-th sample belongs to j.

10. A premature infant health monitoring and prediction system based on a deep learning method, applicable to the premature infant health monitoring and prediction method based on a deep learning method according to any one of claims 1 to 9, characterized in that: include: A data acquisition module is used to obtain basic information data and follow-up information data of premature infants; A premature infant health monitoring and prediction model, used to obtain the premature infant health monitoring results and health prediction results based on the basic information data and follow-up information data; Wherein, the premature infant health monitoring and prediction model includes: for, The first input layer is used to classify data according to the basic information data of the premature infant to obtain system indicator data of multiple categories; Multiple categories of channels are used to obtain extracted features according to system indicator data of corresponding categories; The first concatenation layer is used to fuse the extracted features of channels of multiple categories to obtain a first fused feature; A first fully connected layer is used to extract the first fusion feature to obtain a first fully connected feature; A first output layer, used to obtain a health monitoring result according to the first fully connected feature; The second input layer is used to obtain time series data according to the follow-up information data; An LSTM network is used to obtain LSTM features according to the time series data; The second concatenation layer is used to fuse the first fully connected feature and the LSTM feature to obtain a second fused feature; A second fully connected layer is used to extract the second fusion feature to obtain a second fully connected feature; The second output layer is used to obtain a health prediction result according to the second fully connected feature.

Images