Method, device, storage medium and electronic device for processing surgical records

Abstract

The present disclosure relates to a method, device, storage medium and electronic device for processing surgical records, and to the field of information processing technology. The method comprises: extracting target text from the surgical record according to preset rules, and dividing the target text into at least one target text unit, and determining the type corresponding to each target text unit through a pre-trained classification model according to each target text unit, the type being used to indicate whether the corresponding target text unit is related to the target object, and treating the target text unit whose corresponding type is related as a related text unit, and determining the number of target objects included in the related text unit through a pre-trained recognition model according to the related text unit, and determining the total number of target objects included in the surgical record according to the number of target objects included in the related text unit. The present disclosure determines the related text unit through a classification model, and then determines the number through a recognition model, so as to improve the accuracy of number recognition.

Description

Technical Field

[0001] The present disclosure relates to the field of information processing technology, and in particular, to a method, device, storage medium and electronic device for processing surgical records. Background Art

[0002] Surgical records usually contain a lot of valid information, such as the number of surgical incisions, the number of stents, the number of gauze, etc. By analyzing and mining the effective information, it can provide a reference for future evaluation of the surgical situation, prediction of the state trend of the surgical object, and determination of new surgical plans. Therefore, automatically extracting effective information from surgical records can save professionals time reading surgical records. Usually, the named entity recognition method based on sequence labeling can be used to extract effective information. However, sequence labeling of massive surgical records requires a lot of manpower and material resources, the work is complicated, and the accuracy is low. At the same time, the effective information in surgical records often exists in an implicit form, and sequence labeling cannot reflect this implicit form, which further reduces the accuracy. Summary of the invention

[0003] The purpose of the present disclosure is to provide a method, device, storage medium and electronic device for processing surgical records, so as to solve the related problems existing in the prior art.

[0004] In order to achieve the above object, according to a first aspect of an embodiment of the present disclosure, a method for processing a surgical record is provided, the method comprising:

[0005] Extracting target text from the surgical record according to a preset rule, and dividing the target text into at least one target text unit;

[0006] According to each of the target text units, determining the type corresponding to each of the target text units through a pre-trained classification model, wherein the type is used to indicate whether the corresponding target text unit is related to the target object;

[0007] The corresponding target text unit of the relevant type is used as the relevant text unit, and according to the relevant text unit, the number of target objects included in the relevant text unit is determined by a pre-trained recognition model;

[0008] The total number of the target objects included in the surgical record is determined according to the number of the target objects included in the relevant text unit.

[0009] Optionally, extracting the target text from the surgical record according to a preset rule includes:

[0010] Using the text within the specified position range in the surgical record as the target text; and / or,

[0011] The text in the surgical record that matches the specified keyword is used as the target text.

[0012] Optionally, the method further comprises:

[0013] The designated keyword is determined according to the attribute information corresponding to the target object and the surgical area corresponding to the surgical record. The designated keyword includes at least one of the name, value, and location of the target object.

[0014] Optionally, determining the type corresponding to each target text unit by using a pre-trained classification model according to each target text unit includes:

[0015] For each of the target text units, determining a text feature corresponding to the target text unit;

[0016] Inputting the text feature corresponding to the target text unit into a pre-trained first recurrent neural network to obtain a text representation vector output by the first recurrent neural network;

[0017] The type corresponding to the target text unit is determined according to the text representation vector.

[0018] Optionally, determining the number of target objects included in the relevant text unit by using a pre-trained recognition model according to the relevant text unit includes:

[0019] Determining text features corresponding to the relevant text units;

[0020] Inputting the text features corresponding to the relevant text units into a pre-trained second recurrent neural network to obtain a forward text representation vector and a backward text representation vector output by the second recurrent neural network;

[0021] The number of the target objects included in the relevant text unit is determined according to the forward text representation vector and the backward text representation vector.

[0022] Optionally, determining the total number of the target objects included in the surgical record according to the number of the target objects included in the relevant text unit includes:

[0023] The sum of the numbers of the target objects included in all the relevant text units is taken as the total number.

[0024] Optionally, the first recurrent neural network is a long short-term memory network, and the long short-term memory network is trained in the following manner:

[0025] Obtaining a first sample text unit and a real type corresponding to the first sample text unit;

[0026] Determining a text feature corresponding to the first sample text unit;

[0027] Using the text feature corresponding to the first sample text unit as the input of the long short-term memory network, and training the long short-term memory network according to the output of the long short-term memory network and the true type;

[0028] The second recurrent neural network is a bidirectional long short-term memory network, and the bidirectional long short-term memory network is trained in the following manner:

[0029] Acquire a second sample text unit and the newly added number of the target objects in the second sample text unit, where the newly added number is used to represent the number of the target objects newly added in the second sample text unit;

[0030] Determining a text feature corresponding to the second sample text unit;

[0031] The text feature corresponding to the second sample text unit is used as the input of the bidirectional long short-term memory network, and the bidirectional long short-term memory network is trained according to the output of the bidirectional long short-term memory network and the newly added quantity.

[0032] According to a second aspect of an embodiment of the present disclosure, a device for processing a surgical record is provided, the device comprising:

[0033] A processing module, used for extracting target text from the surgical record according to a preset rule, and dividing the target text into at least one target text unit;

[0034] A classification module, used to determine the type corresponding to each target text unit through a pre-trained classification model according to each target text unit, wherein the type is used to indicate whether the corresponding target text unit is related to the target object;

[0035] A recognition module, used for taking the corresponding target text unit of the relevant type as the relevant text unit, and determining the number of target objects included in the relevant text unit through a pre-trained recognition model according to the relevant text unit;

[0036] A determination module is used to determine the total number of the target objects included in the surgical record according to the number of the target objects included in the relevant text unit.

[0037] Optionally, the processing module is used to:

[0038] Using the text within the specified position range in the surgical record as the target text; and / or,

[0039] The text in the surgical record that matches the specified keyword is used as the target text.

[0040] Optionally, the device further comprises:

[0041] The keyword determination module is used to determine the specified keyword according to the attribute information corresponding to the target object and the surgical area corresponding to the surgical record. The specified keyword includes at least one of the name, value, and location of the target object.

[0042] Optionally, the classification module includes:

[0043] A first determination submodule is used to determine, for each target text unit, a text feature corresponding to the target text unit;

[0044] A first processing submodule, configured to input a text feature corresponding to the target text unit into a pre-trained first recurrent neural network to obtain a text representation vector output by the first recurrent neural network;

[0045] The classification submodule is used to determine the type corresponding to the target text according to the text representation vector.

[0046] Optionally, the identification module includes:

[0047] A second determination submodule is used to determine the text features corresponding to the relevant text units;

[0048] A second processing submodule, configured to input text features corresponding to the relevant text units into a pre-trained second recurrent neural network to obtain a forward text representation vector and a backward text representation vector output by the second recurrent neural network;

[0049] The identification submodule is used to determine the number of the target objects included in the relevant text unit according to the forward text representation vector and the backward text representation vector.

[0050] Optionally, the determining module is used to:

[0051] The sum of the numbers of the target objects included in all the relevant text units is taken as the total number.

[0052] Optionally, the first recurrent neural network is a long short-term memory network, and the long short-term memory network is trained in the following manner:

[0053] Obtaining a first sample text unit and a real type corresponding to the first sample text unit;

[0054] Determining a text feature corresponding to the first sample text unit;

[0055] Using the text feature corresponding to the first sample text unit as the input of the long short-term memory network, and training the long short-term memory network according to the output of the long short-term memory network and the true type;

[0056] The second recurrent neural network is a bidirectional long short-term memory network, and the bidirectional long short-term memory network is trained in the following manner:

[0057] Acquire a second sample text unit and the newly added number of the target objects in the second sample text unit, where the newly added number is used to represent the number of the target objects newly added in the second sample text unit;

[0058] Determining a text feature corresponding to the second sample text unit;

[0059] The text feature corresponding to the second sample text unit is used as the input of the bidirectional long short-term memory network, and the bidirectional long short-term memory network is trained according to the output of the bidirectional long short-term memory network and the newly added quantity.

[0060] According to a third aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in the first aspect of the embodiment of the present disclosure are implemented.

[0061] According to a fourth aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0062] a memory having a computer program stored thereon;

[0063] A processor is used to execute the computer program in the memory to implement the steps of the method described in the first aspect of the embodiment of the present disclosure.

[0064] Through the above technical solution, the present disclosure first extracts the target text from the surgical record according to the preset rules and divides it into at least one target text unit. Then, the type corresponding to each target text unit is determined through a pre-trained classification model, so that the corresponding type is the relevant target text unit as the relevant text unit. Then, the number of target objects included in the relevant text unit is determined through a pre-trained recognition model, and finally, the total number of target objects included in the surgical record is determined according to the number of target objects included in the relevant text unit. The present disclosure screens out the relevant text units related to the target object in the target text unit through the classification model, and then identifies the number of target objects included in the relevant text unit through the recognition model, thereby determining the total number in the surgical record, which can improve the accuracy of quantity recognition.

[0065] Other features and advantages of the present disclosure will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present disclosure and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present disclosure but do not constitute a limitation of the present disclosure. In the accompanying drawings:

[0067] Figure 1 is a flow chart of a method for processing surgical records according to an exemplary embodiment;

[0068] Figure 2 is a flow chart of another method for processing surgical records according to an exemplary embodiment;

[0069] Figure 3 is a flow chart of another method for processing surgical records according to an exemplary embodiment;

[0070] Figure 4 is a flow chart of another method for processing surgical records according to an exemplary embodiment;

[0071] Figure 5 is a schematic diagram of a classification model according to an exemplary embodiment;

[0072] Figure 6 is a flow chart of another method for processing surgical records according to an exemplary embodiment;

[0073] Figure 7 is a schematic diagram of a recognition model according to an exemplary embodiment;

[0074] Figure 8 is a block diagram of a device for processing surgical records according to an exemplary embodiment;

[0075] Fig. 9 is a block diagram of another surgical record processing device according to an exemplary embodiment;

[0076] Fig.10 is a block diagram of another surgical record processing device according to an exemplary embodiment;

[0077] Fig.11 is a block diagram of another surgical record processing device according to an exemplary embodiment;

[0078] Fig.12 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0079] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0080] Figure 1 is a flowchart of a method for processing surgical records according to an exemplary embodiment. Figure 1 As shown, the method may include the following steps:

[0081] Step 101: extract target text from the surgical record according to preset rules, and divide the target text into at least one target text unit.

[0082] For example, surgical records are usually recorded in text form, and the target text can be extracted from the surgical records according to preset rules. The target text can be understood as text that may be related to the target object. The target object can be, for example, a surgical incision, a stent, a gauze, etc., and the present disclosure does not specifically limit this. The preset rule can be, for example, the beginning part (or the middle part, the ending part) of the surgical record, or the nth to mth sentences in the surgical record, or the part of the surgical record that includes specified keywords. After obtaining the target text, the target text can be divided into at least one target text unit. The target text can be divided according to specified symbols (for example: period, semicolon, carriage return, etc.) to obtain at least one target text unit. For example, if the specified symbol is a period, then each sentence in the target text is a target text unit.

[0083] Step 102, according to each target text unit, the type corresponding to each target text unit is determined by a pre-trained classification model, where the type is used to indicate whether the corresponding target text unit is related to the target object.

[0084] For example, for each target text unit, the target text unit can be input into a pre-trained classification model, and the classification model can output the type corresponding to the target text unit. Among them, the classification model is obtained by training a large number of training samples collected in advance. The classification model can first determine the corresponding text features according to the target text unit, and the text features can characterize the target text unit. After that, the classification model can classify according to the text features corresponding to the target text unit to determine the type of the target text unit. The type is used to indicate whether the corresponding target text unit is related to the target object, that is, it includes two types: related and irrelevant. The corresponding type is related, which can be understood as the target object is described in the target text unit, and the corresponding type is irrelevant, which can be understood as the target object is not described in the target text unit. Then, the classification model can be understood as determining the matching degree of the text features corresponding to the target text unit with each of the two types, that is, the higher the matching degree of the text features with the type, the higher the probability value of the target text unit being the type, and the lower the matching degree, the lower the probability value of the target text unit being the type. The classification model can determine the type with the highest corresponding matching degree as the type of the target text unit.

[0085] Step 103: taking the corresponding target text unit of the relevant type as the relevant text unit, and determining the number of target objects included in the relevant text unit through the pre-trained recognition model according to the relevant text unit.

[0086] Step 104: Determine the total number of target objects included in the surgical record according to the number of target objects included in the relevant text unit.

[0087] For example, after determining the type corresponding to each target text unit, the target text unit of the corresponding type can be used as the relevant text unit, and the relevant text unit can be one or more, and the present disclosure does not specifically limit this. The relevant text unit can be input into a pre-trained recognition model, and the recognition model can output the number of target objects included in the relevant text unit. Wherein, the recognition model can also be obtained by training a large number of training samples collected in advance (the training samples of the training recognition model and the training samples of the training classification model can be the same or different). The recognition model can first determine the corresponding text features according to the relevant text unit, and the text features can characterize the relevant text unit. Afterwards, the recognition model can be identified according to the text features corresponding to the relevant text unit to determine the number of target objects included in the relevant text unit. The recognition model can be understood as determining the matching degree of the text features corresponding to the relevant text unit and each quantity type in a plurality of pre-specified quantity types, that is, the higher the matching degree of the text features corresponding to the relevant text unit and a certain quantity type, the higher the probability value of the number of target objects included in the relevant text unit is the quantity type, and the lower the matching degree, the lower the probability value of the number of target objects included in the relevant text unit is the quantity type. The recognition model can determine the number type with the highest corresponding matching degree as the number of target objects included in the relevant text unit, or determine the number type with a matching degree that meets a preset condition (for example, a matching degree greater than a preset threshold) as the number of target objects included in the relevant text unit. A variety of pre-specified number types can be determined according to the type of target object and the surgical area corresponding to the surgical record. For example, if the target object is a surgical incision and the surgical area corresponding to the surgical record is the thorax, then the number type can be 0, 1, 2, or 3. Finally, the total number of target objects included in the surgical record can be determined according to the number of target objects included in the relevant text unit. For example, if there is one relevant text unit, the number of target objects included in the relevant text unit can be taken as the total number, and if there are multiple relevant text units, the sum of the number of target objects included in multiple relevant text units can be taken as the total number. In this way, the classification model can screen out the relevant text units that describe the target object from multiple target text units, and then use the recognition model to identify the number of target objects included in the relevant text unit, thereby determining the total number in the surgical record, which can effectively improve the accuracy of quantity recognition.

[0088] In summary, the present disclosure first extracts the target text from the surgical record according to preset rules and divides it into at least one target text unit. Then, the type corresponding to each target text unit is determined through a pre-trained classification model, so that the corresponding type is the relevant target text unit as the relevant text unit. Then, the number of target objects included in the relevant text unit is determined through a pre-trained recognition model, and finally, the total number of target objects included in the surgical record is determined according to the number of target objects included in the relevant text unit. The present disclosure uses a classification model to filter out relevant text units related to the target object in the target text unit, and then uses a recognition model to identify the number of target objects included in the relevant text unit, thereby determining the total number in the surgical record, which can improve the accuracy of quantity recognition.

[0089] Figure 2 is a flowchart of another method for processing surgical records according to an exemplary embodiment. Figure 2 As shown, the implementation of step 101 may include:

[0090] Step 1011, taking the text within the specified position range in the surgical record as the target text. And / or,

[0091] Step 1012: Use the text in the surgical record that matches the specified keyword as the target text.

[0092] For example, the text within the specified position range in the surgical record can be used as the target text. The specified position range can be, for example, the beginning part, the middle part, or the end part. The specified position range can also be the sequence number range of sentences, such as the nth to mth sentences in the surgical record, or the first i sentences in the surgical record, or the last j sentences in the surgical record. Furthermore, the text in the surgical record that matches the specified keywords can also be used as the target text. For example, the text in the surgical record that includes the specified keywords can be used as the target text. Taking the target object as the surgical incision and the surgical area corresponding to the surgical record as the chest cavity as an example, the specified position range can be, for example, the beginning part (or the first 6 sentences), the specified keyword can be "incision", and the specified keyword can also be a word that conforms to the structure of numerical value + intercostal space, such as "8th intercostal space", etc.

[0093] Figure 3 is a flowchart of another method for processing surgical records according to an exemplary embodiment. Figure 3 As shown, the method may also include:

[0094] Step 105, determining a designated keyword according to the attribute information corresponding to the target object and the surgical area corresponding to the surgical record, wherein the designated keyword includes at least one of the name, value, and location of the target object.

[0095] For example, the attribute information corresponding to the target object and the surgical area corresponding to the surgical record can be obtained first, wherein the attribute information may include the name, type, value range, etc. of the target object, and the surgical area may be, for example, the thoracic cavity, abdominal cavity, head, etc. Afterwards, the designated keyword can be determined based on the attribute information and the surgical area. Specifically, the name of the target object can be determined based on the attribute information. For example, if the target object is a surgical incision, then "incision" can be used as the designated keyword. For another example, if the target object is a stent, then "stent" can be used as the designated keyword. The positioning part and the corresponding value can also be determined based on the surgical area. For example, if the surgical area is the thoracic cavity, the positioning part can be determined to be the intercostal space, and the corresponding words that conform to the structure of number + intercostal space can be used as designated keywords, such as: "8th intercostal space", "7th intercostal space", etc. For another example, if the surgical area is the abdominal cavity, the positioning part can be determined to be the umbilicus, and the corresponding words that conform to the structure of umbilicus + direction can be used as designated keywords, such as: "left side of the umbilicus", "above the umbilicus", "below the umbilicus", etc.

[0096] Figure 4 is a flowchart of another method for processing surgical records according to an exemplary embodiment. Figure 4 As shown, the implementation of step 102 may include:

[0097] Step 1021: for each target text unit, determine the text feature corresponding to the target text unit.

[0098] Step 1022: input the text feature corresponding to the target text unit into a pre-trained first recurrent neural network to obtain a text representation vector output by the first recurrent neural network.

[0099] Step 1023: Determine the type corresponding to the target text unit according to the text representation vector.

[0100] For example, the type corresponding to the target text unit can be related or irrelevant, so the classification model can be understood as a model for binary classification. Specifically, the classification model may include: an embedding layer, a first recurrent neural network, and an output layer, such as Figure 5As shown. For each target text unit, the target text unit can be input into the embedding layer to obtain the text features (i.e., Embedding) corresponding to the target text unit output by the embedding layer. The embedding layer can be understood as using One-hot to map each word in the target text unit to obtain the corresponding text features, and the text features can characterize each word in the corresponding target text unit. Then the text features corresponding to the target text unit are input into the first recurrent neural network trained in advance to obtain the text representation vector output by the first recurrent neural network. The text representation vector can be understood as the text features corresponding to the target text unit encoded by the first recurrent neural network, and the first recurrent neural network can learn the association between each word in the target text unit, so the text representation vector can characterize the target text unit as a whole. The first recurrent neural network can be any type of RNN (English: Recurrent Neural Network, Chinese: Recurrent Neural Network), for example, it can be an LSTM (English: Long short-term memory, Chinese: Long short-term memory) network, it can also be a BiLSTM (English: Bi-directional LongShort-Term Memory, Chinese: Bidirectional Long Short-Term Memory) network, or it can be other types of RNN, and the present disclosure does not specifically limit this. Finally, the text representation vector can be input into the output layer to obtain the type of the target text unit output by the output layer. The activation function of the output layer can be, for example, a Sigmoid function, that is, the output layer can determine the matching degree of the text representation vector with each of the two types. If the text representation vector has a higher matching degree with the relevant type, then the type corresponding to the target text unit can be determined to be relevant. If the text representation vector has a higher matching degree with the irrelevant type, then the type corresponding to the target text unit can be determined to be irrelevant.

[0101] Figure 6 is a flowchart of another method for processing surgical records according to an exemplary embodiment. Figure 6 As shown, the implementation of step 103 may include:

[0102] Step 1031, determining text features corresponding to relevant text units.

[0103] Step 1032: Input text features corresponding to the relevant text units into a pre-trained second recurrent neural network to obtain a forward text representation vector and a backward text representation vector output by the second recurrent neural network.

[0104] Step 1033: Determine the number of target objects included in the relevant text unit according to the forward text representation vector and the backward text representation vector.

[0105] For example, the number of target objects included in the relevant text unit can be multiple numbers within a certain range, so the recognition model can be understood as a model for multi-classification. Specifically, the recognition model may include: an embedding layer, a second recurrent neural network, and an output layer, such as Figure 7 As shown. The relevant text unit can be input into the embedding layer to obtain the text features corresponding to the relevant text unit output by the embedding layer. The embedding layer can be understood as using One-hot to map each word in the relevant text unit to obtain the corresponding text features, and the text features can represent each word in the relevant text unit. Then the text features corresponding to the relevant text unit are input into the pre-trained second recurrent neural network to obtain the forward text representation vector and the backward text representation vector output by the second recurrent neural network. The second recurrent neural network can be understood as two recurrent neural networks in opposite directions, one for forward encoding the text features corresponding to the relevant text unit to obtain the forward text representation vector, and the other for backward encoding the text features corresponding to the relevant text unit to obtain the backward text representation vector. The second recurrent neural network can learn the association between each word in the relevant text unit and the previous word, and can also learn the association between each word in the relevant text unit and the subsequent word. Therefore, the forward text representation vector can represent the relevant text unit from the front to the back, and the backward text representation vector can represent the relevant text unit from the back to the front. The second recurrent neural network can be, for example, a BiLSTM network, or other types of recurrent neural networks that can extract forward text representation vectors and backward text representation vectors, which are not specifically limited in the present disclosure. Finally, the forward text representation vector and the backward text representation vector can be spliced ​​(i.e., Contact), and the splicing result can be input into the output layer to obtain the number of target objects included in the relevant text unit output by the output layer. The activation function of the output layer can be, for example, a Softmax function, that is, the output layer can determine the matching degree of the splicing result with a variety of quantities, and use the number with the highest matching degree as the number of target objects included in the relevant text unit. For example, the Softmax function can determine the matching degree of the splicing result with 0, 1, 2, and 3. If the matching degree with 2 is the highest, then it can be determined that the number of target objects included in the relevant text unit is 2.

[0106] In one implementation, step 104 may include:

[0107] The sum of the numbers of target objects included in all relevant text units is taken as the total number.

[0108] For example, when there are multiple related text units, the total number may be the sum of the numbers of target objects included in all the related text units, that is, Among them, Num represents the total number, Numx Represents the number of target objects included in the xth related text unit, and there are SUM related text units in total.

[0109] In one application scenario, the first recurrent neural network is a long short-term memory network, which is trained in the following way:

[0110] Step 1) Obtain a first sample text unit and a real type corresponding to the first sample text unit.

[0111] Step 2) Determine the text feature corresponding to the first sample text unit.

[0112] Step 3) Using the text features corresponding to the first sample text unit as the input of the long short-term memory network, and training the long short-term memory network according to the output of the long short-term memory network and the true type.

[0113] For example, when training the long short-term memory network, it is necessary to first obtain training samples, for example, any text unit can be extracted from the historical surgical records as the first sample unit, and whether the first sample text unit is related to the target object is used as the true type corresponding to the first sample text unit. After that, the text features corresponding to the first sample text unit can be determined by using the embedding layer, and then the text features corresponding to the first sample text unit are used as the input of the long short-term memory network, and the long short-term memory network is trained according to the output of the long short-term memory network and the true type. When the first sample text unit is input, the output of the long short-term memory network can match the true type. For example, the loss function (such as cross entropy loss) can be determined according to the output of the long short-term memory network and the true type, and the back propagation algorithm is used to correct the neuron parameters in the long short-term memory network with the goal of reducing the loss function. The neuron parameters can be, for example, the weight (English: Weight) and bias (English: Bias) of the neuron. Repeat the above steps until the loss function meets the preset conditions, such as the loss function is less than the preset loss threshold, so as to achieve the purpose of training the long short-term memory network.

[0114] The second recurrent neural network is a bidirectional long short-term memory. The bidirectional long short-term memory network is trained in the following way:

[0115] Step 4) obtaining a second sample text unit and the newly added number of target objects in the second sample text unit, where the newly added number is used to represent the number of newly added target objects in the second sample text unit.

[0116] Step 5) Determine the text feature corresponding to the second sample text unit.

[0117] Step 6) Using the text features corresponding to the second sample text unit as the input of the bidirectional long short-term memory network, and training the bidirectional long short-term memory network according to the output and the newly added quantity of the bidirectional long short-term memory network.

[0118] For example, when training a bidirectional long short-term memory network, it is also necessary to obtain training samples first. For example, any text unit can be extracted from the historical surgical records as the second sample unit (the second sample text unit can be the same as the first sample text unit or different), and the number of newly added target objects in the second sample text unit is used as the newly added number. When a sentence describes a target object, there may be two situations, one describing a newly added target object, and the other describing a target object that has been described in the previous sentence. In order to avoid repeated counting of target objects, the number of newly added target objects (i.e., the newly added number) can be used as a reference.

[0119] Afterwards, the embedding layer can be used to determine the text features corresponding to the second sample text unit, and then the text features corresponding to the second sample text unit can be used as the input of the bidirectional long short-term memory network, and the bidirectional long short-term memory network can be trained according to the output of the bidirectional long short-term memory network and the newly added quantity. When the second sample text unit is input, the output of the bidirectional long short-term memory network can match the newly added quantity. For example, the loss function (such as cross entropy loss) can be determined according to the output of the bidirectional long short-term memory network and the newly added quantity, and the back propagation algorithm can be used to correct the neuron parameters in the bidirectional long short-term memory network with the goal of reducing the loss function. Repeat the above steps until the loss function meets the preset conditions, such as the loss function is less than the preset loss threshold, so as to achieve the purpose of training the bidirectional long short-term memory network.

[0120] When training the above-mentioned long short-term memory network and bidirectional long short-term memory network, it is only necessary to mark the first sample text unit whether it is related to the target object, and mark the number of newly added target objects for the second sample text unit. Taking the target object as a surgical incision as an example, the first sample text unit is: "After successful anesthesia, the patient first takes a supine position... A small incision is made in the right 8th intercostal axillary line to insert a thoracoscope. Small incisions are made in the right 5th intercostal axillary line and the 7th intercostal axillary line to insert operating instruments... A No. 28 chest drainage tube is placed in the midline of the 8th intercostal space. Count the instruments, no dressings...", which describes the surgical incision, then the true type corresponding to the first sample text unit is related, that is, the first sample text unit is marked as related. For another example, the second sample text unit is: "After successful double-lumen endotracheal intubation and general anesthesia, the patient lies on the left side and undergoes routine disinfection and draping. A small incision is made at the midaxillary line of the 7th intercostal space on the right side for endoscopic exploration... Small incisions of approximately 5 cm and 2 cm in length are made at the anterior and posterior axillary lines of the 4th intercostal space, respectively. Operating instruments are inserted... The incision at the 4th intercostal space is extended, and the chest wall tissue is cut layer by layer...", among which three new surgical incisions are made (including one incision made at the midaxillary line of the 7th intercostal space and two incisions made at the 4th intercostal space), so the number of new target objects in the second sample text unit is 3, and the second text unit is marked as 3.

[0121] Compared with the sequence labeling method in the prior art of labeling each word in a sentence in turn, the information and workload required to label whether it is related to the target object and the number of new target objects are greatly reduced, and the labeling difficulty is also reduced. The operability and accuracy of labeling can be effectively improved, so that the trained long short-term memory network can accurately judge the type corresponding to the target text unit, and the bidirectional long short-term memory network can accurately identify the number of target objects included in the relevant text unit, thereby improving the accuracy of quantity recognition.

[0122] In summary, the present disclosure first extracts the target text from the surgical record according to preset rules and divides it into at least one target text unit. Then, the type corresponding to each target text unit is determined through a pre-trained classification model, so that the corresponding type is the relevant target text unit as the relevant text unit. Then, the number of target objects included in the relevant text unit is determined through a pre-trained recognition model, and finally, the total number of target objects included in the surgical record is determined according to the number of target objects included in the relevant text unit. The present disclosure uses a classification model to filter out relevant text units related to the target object in the target text unit, and then uses a recognition model to identify the number of target objects included in the relevant text unit, thereby determining the total number in the surgical record, which can improve the accuracy of quantity recognition.

[0123] Figure 8is a block diagram of a surgical record processing device according to an exemplary embodiment. Figure 8 As shown, the device 200 includes:

[0124] The processing module 201 is used to extract the target text from the surgical record according to a preset rule, and divide the target text into at least one target text unit.

[0125] The classification module 202 is used to determine the type corresponding to each target text unit through a pre-trained classification model according to each target text unit, and the type is used to indicate whether the corresponding target text unit is related to the target object.

[0126] The recognition module 203 is used to take the corresponding target text unit of the relevant type as the relevant text unit, and determine the number of target objects included in the relevant text unit through the pre-trained recognition model according to the relevant text unit.

[0127] The determination module 204 is used to determine the total number of target objects included in the surgical record according to the number of target objects included in the relevant text unit.

[0128] In one implementation, the processing module 201 may be used to:

[0129] The text in the specified position range in the surgical record is used as the target text. And / or,

[0130] The text in the surgical record that matches the specified keyword is used as the target text.

[0131] Fig. 9 is a block diagram of another surgical record processing device according to an exemplary embodiment. Fig. 9 As shown, the device 200 may also include:

[0132] The keyword determination module 205 is used to determine the specified keyword according to the attribute information corresponding to the target object and the surgical area corresponding to the surgical record. The specified keyword includes at least one of the name, value, and location of the target object.

[0133] Fig.10 is a block diagram of another surgical record processing device according to an exemplary embodiment. Fig.10 As shown, the classification module 202 may include:

[0134] The first determination submodule 2021 is used to determine, for each target text unit, a text feature corresponding to the target text unit.

[0135] The first processing submodule 2022 is used to input the text feature corresponding to the target text unit into a pre-trained first recurrent neural network to obtain a text representation vector output by the first recurrent neural network.

[0136] The classification submodule 2023 is used to determine the type corresponding to the target text unit according to the text representation vector.

[0137] Fig.11 is a block diagram of another surgical record processing device according to an exemplary embodiment. Fig.11 As shown, the identification module 203 may include:

[0138] The second determining submodule 2031 is used to determine the text features corresponding to the relevant text units.

[0139] The second processing submodule 2032 is used to input the text features corresponding to the relevant text units into the pre-trained second recurrent neural network to obtain the forward text representation vector and the backward text representation vector output by the second recurrent neural network.

[0140] The identification submodule 2033 is used to determine the number of target objects included in the relevant text unit according to the forward text representation vector and the backward text representation vector.

[0141] In another application scenario, the determination module 204 may be used to:

[0142] The sum of the numbers of target objects included in all relevant text units is taken as the total number.

[0143] In one application scenario, the first recurrent neural network is a long short-term memory network, which is trained in the following way:

[0144] Step 1) Obtain a first sample text unit and a real type corresponding to the first sample text unit.

[0145] Step 2) Determine the text feature corresponding to the first sample text unit.

[0146] Step 3) Using the text features corresponding to the first sample text unit as the input of the long short-term memory network, and training the long short-term memory network according to the output of the long short-term memory network and the true type.

[0147] The second recurrent neural network is a bidirectional long short-term memory. The bidirectional long short-term memory network is trained in the following way:

[0148] Step 4) obtaining a second sample text unit and the newly added number of target objects in the second sample text unit, where the newly added number is used to represent the number of newly added target objects in the second sample text unit.

[0149] Step 5) Determine the text feature corresponding to the second sample text unit.

[0150] Step 6) Using the text features corresponding to the second sample text unit as the input of the bidirectional long short-term memory network, and training the bidirectional long short-term memory network according to the output and the newly added quantity of the bidirectional long short-term memory network.

[0151] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0152] In summary, the present disclosure first extracts the target text from the surgical record according to preset rules and divides it into at least one target text unit. Then, the type corresponding to each target text unit is determined through a pre-trained classification model, so that the corresponding type is the relevant target text unit as the relevant text unit. Then, the number of target objects included in the relevant text unit is determined through a pre-trained recognition model, and finally, the total number of target objects included in the surgical record is determined according to the number of target objects included in the relevant text unit. The present disclosure uses a classification model to filter out relevant text units related to the target object in the target text unit, and then uses a recognition model to identify the number of target objects included in the relevant text unit, thereby determining the total number in the surgical record, which can improve the accuracy of quantity recognition.

[0153] Fig.12 FIG. 1 is a block diagram of an electronic device 300 according to an exemplary embodiment. Fig.12 As shown, the electronic device 300 may include: a processor 301 , a memory 302 . The electronic device 300 may also include one or more of a multimedia component 303 , an input / output (I / O) interface 304 , and a communication component 305 .

[0154] The processor 301 is used to control the overall operation of the electronic device 300 to complete all or part of the steps in the above-mentioned surgical record processing method. The memory 302 is used to store various types of data to support the operation of the electronic device 300, and these data may include, for example, instructions for any application or method used to operate on the electronic device 300, and application-related data, such as contact data, sent and received messages, pictures, audio, video, etc. The memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Read-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, magnetic disk or optical disk. The multimedia component 303 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 302 or sent through the communication component 305. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 304 provides an interface between the processor 301 and other interface modules, and the other interface modules may be keyboards, mice, buttons, etc. These buttons may be virtual buttons or physical buttons. The communication component 305 is used for wired or wireless communication between the electronic device 300 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IOT, eMTC, or other 5G, etc., or a combination of one or more of them, is not limited here. Therefore, the corresponding communication component 305 may include: Wi-Fi module, Bluetooth module, NFC module, etc.

[0155] In an exemplary embodiment, the electronic device 300 can be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned surgical record processing method.

[0156] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, and when the program instructions are executed by a processor, the steps of the above-mentioned method for processing surgical records are implemented. For example, the computer-readable storage medium may be the above-mentioned memory 302 including program instructions, and the above-mentioned program instructions may be executed by the processor 301 of the electronic device 300 to complete the above-mentioned method for processing surgical records.

[0157] In another exemplary embodiment, a computer program product is also provided. The computer program product includes a computer program executable by a programmable device, and the computer program has a code portion for executing the above-mentioned surgical record processing method when executed by the programmable device.

[0158] The preferred embodiments of the present disclosure are described in detail above in conjunction with the accompanying drawings; however, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept of the present disclosure, a variety of simple modifications can be made to the technical solution of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.

[0159] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present disclosure will not further describe various possible combinations.

[0160] In addition, various embodiments of the present disclosure may be arbitrarily combined, and as long as they do not violate the concept of the present disclosure, they should also be regarded as the contents disclosed by the present disclosure.

Claims

1. A method for processing surgical records, characterized in that: The method comprises: Extracting target text from the surgical record according to a preset rule, and dividing the target text into at least one target text unit; According to each of the target text units, a type corresponding to each of the target text units is determined by a pre-trained classification model, wherein the type is used to indicate whether the corresponding target text unit is related to the target object; the classification model is used to determine the matching degree between the text feature corresponding to the target text unit and each of two types, wherein the two types include related and unrelated; The corresponding target text unit of the relevant type is used as the relevant text unit, and according to the relevant text unit, the number of target objects included in the relevant text unit is determined by a pre-trained recognition model; the recognition model is used to determine the matching degree of the text feature corresponding to the relevant text unit with each quantity type of a plurality of pre-specified quantity types, and the plurality of pre-specified quantity types are determined according to the type of the target object and the surgical area corresponding to the surgical record; The total number of the target objects included in the surgical record is determined according to the number of the target objects included in the relevant text unit.

2. The method according to claim 1, characterized in that The step of extracting target text from the surgical record according to preset rules includes: Using the text within the specified position range in the surgical record as the target text; and / or, The text in the surgical record that matches the specified keyword is used as the target text.

3. The method according to claim 2, characterized in that The method further comprises: The designated keyword is determined according to the attribute information corresponding to the target object and the surgical area corresponding to the surgical record. The designated keyword includes at least one of the name, value, and location of the target object.

4. The method according to claim 1, characterized in that: Determining the type corresponding to each target text unit by using a pre-trained classification model according to each target text unit includes: For each of the target text units, determining a text feature corresponding to the target text unit; Inputting the text feature corresponding to the target text unit into a pre-trained first recurrent neural network to obtain a text representation vector output by the first recurrent neural network; The type corresponding to the target text unit is determined according to the text representation vector.

5. The method according to claim 1, characterized in that The step of determining the number of target objects included in the relevant text unit by using a pre-trained recognition model according to the relevant text unit includes: Determining text features corresponding to the relevant text units; Inputting the text features corresponding to the relevant text units into a pre-trained second recurrent neural network to obtain a forward text representation vector and a backward text representation vector output by the second recurrent neural network; The number of the target objects included in the relevant text unit is determined according to the forward text representation vector and the backward text representation vector.

6. The method according to claim 1, characterized in that Determining the total number of the target objects included in the surgical record according to the number of the target objects included in the relevant text unit includes: The sum of the numbers of the target objects included in all the relevant text units is taken as the total number.

7. The method according to claim 1, characterized in that The classification model includes a first recurrent neural network, which is a long short-term memory network. The long short-term memory network is trained in the following way: Obtaining a first sample text unit and a real type corresponding to the first sample text unit; Determining a text feature corresponding to the first sample text unit; Using the text feature corresponding to the first sample text unit as the input of the long short-term memory network, and training the long short-term memory network according to the output of the long short-term memory network and the true type; The recognition model includes a second recurrent neural network, which is a bidirectional long short-term memory network. The bidirectional long short-term memory network is trained in the following way: Acquire a second sample text unit and the newly added number of the target objects in the second sample text unit, where the newly added number is used to represent the number of the target objects newly added in the second sample text unit; Determining a text feature corresponding to the second sample text unit; The text feature corresponding to the second sample text unit is used as the input of the bidirectional long short-term memory network, and the bidirectional long short-term memory network is trained according to the output of the bidirectional long short-term memory network and the newly added quantity.

8. A surgical record processing device, characterized in that: The device comprises: A processing module, used for extracting target text from the surgical record according to a preset rule, and dividing the target text into at least one target text unit; A classification module, used to determine the type corresponding to each target text unit through a pre-trained classification model according to each target text unit, wherein the type is used to indicate whether the corresponding target text unit is related to the target object; the classification model is used to determine the matching degree between the text feature corresponding to the target text unit and each of the two types, wherein the two types include related and unrelated; A recognition module, used for taking a corresponding target text unit of a related type as a related text unit, and determining the number of target objects included in the related text unit through a pre-trained recognition model according to the related text unit; the recognition model is used for determining a matching degree between a text feature corresponding to the related text unit and each of a plurality of pre-specified quantity types, wherein the plurality of pre-specified quantity types are determined according to the type of the target object and the surgical area corresponding to the surgical record; A determination module is used to determine the total number of the target objects included in the surgical record according to the number of the target objects included in the relevant text unit.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

10. An electronic device, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the method according to any one of claims 1 to 7.

Images