Auxiliary diagnosis method and system based on deep neural network

Abstract

The present invention provides an auxiliary diagnosis method and system based on a deep neural network, which relates to the field of artificial intelligence technology. The method includes obtaining patient information and vectorizing the patient information; if the patient information does not include diagnostic information, inputting the vectorized patient information into a trained deep neural network model, and obtaining a number of auxiliary diagnosis names based on the output of the deep neural network model; if the patient information includes diagnostic information, matching the diagnostic information with the diagnostic names in a preset diagnostic library according to preset matching rules to obtain a number of auxiliary diagnosis names. The present invention utilizes a deep learning algorithm to perform auxiliary diagnosis based on patient information, avoiding the situation where diagnostic recommendations rely on the doctor's input characters and the diagnostic names matched after the input characters are not sorted properly, providing an accurate reference for doctors to make diagnoses and adopt further treatment plans, thereby improving the doctor's work efficiency and the patient's medical experience.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an auxiliary diagnosis system based on deep neural networks. Background Art

[0002] Making a diagnosis is the most important task in the entire medical process. With the continuous improvement of medical information technology, doctors currently mainly enter diagnoses into information systems. However, doctors face a large number of patients seeking medical treatment every day. Clinical diagnosis lacks standard and unified specifications. Manually entering diagnoses greatly reduces doctors' work efficiency and patients' medical experience. On the other hand, since making a diagnosis requires comprehensive consideration of patient information, it places high demands on doctors' professional qualities. Unreasonable diagnoses made by young doctors with little experience will also lead to an increase in medical accidents and medical disputes.

[0003] The development of deep learning technology represented by neural networks has provided certain possibilities for efficient and accurate diagnostic recommendations. Using deep learning to assist doctors in making diagnoses is of great significance in reducing doctors' workload and improving their work efficiency.

[0004] The current auxiliary diagnosis method mainly involves building a database of diagnosis names and matching them based on the doctor's manual input. However, due to incomplete diagnoses in the matching database, the doctor's desired diagnosis cannot be matched, and the full diagnosis must be manually entered. This is undoubtedly time-consuming and labor-intensive for doctors who face a large number of patients every day, and it cannot help doctors obtain more accurate and reasonable diagnoses. Summary of the Invention

[0005] The present invention provides an auxiliary diagnosis system based on deep neural networks, which is used to solve the problem of inaccurate matching results given by existing auxiliary diagnosis methods.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A first aspect of the present invention provides an auxiliary diagnosis method based on a deep neural network, the auxiliary diagnosis method comprising the following steps:

[0008] Acquiring patient information and quantizing the patient information;

[0009] If the patient information does not include diagnosis information, inputting the vectorized patient information into a trained deep neural network model, and obtaining a plurality of auxiliary diagnosis names based on the output of the deep neural network model;

[0010] If the patient information includes diagnosis information, the diagnosis information is matched with the diagnosis names in the preset diagnosis library according to the preset matching rules to obtain a number of auxiliary diagnosis names.

[0011] Furthermore, the auxiliary diagnosis method further comprises, before the above steps, constructing a patient information database D and a diagnosis database D';

[0012] The patient information database includes the patient ID, age, gender, chief complaint and diagnosis name; the diagnosis database includes the diagnosis name and the probability value of the current diagnosis name in each department.

[0013] Furthermore, the patient information includes age, gender and chief complaint.

[0014] Furthermore, the structure of the deep neural network model is specifically as follows:

[0015] The age vector A p Transformed into a fully connected network mlp1 The gender vector S p Transformed into another fully connected network mlp2

[0016] The matrix corresponding to the chief complaint Output after bidirectional LSTM model Vector Y with diagnosis names p Forming the attention matrix Att, d is the dimension of the diagnosis name vector;

[0017] Will Combined with Att, we get Will The predicted diagnostic vector is obtained through a two-layer fully connected network Where W Y and b Y are weights and biases respectively;

[0018] Y P and The focal Loss cross entropy is used as the loss function of the deep neural network model.

[0019] Furthermore, the training of the deep neural network model is specifically as follows:

[0020] The training set is divided into several batches, and each patient in the batch is subjected to the structure of the deep neural network model to obtain the average loss of each batch. Taking minimizing the average loss as the training objective;

[0021] During the training process, set the parameter acc to record the accuracy of the validation set. After training a batch, calculate the current validation set accuracy α. When α is greater than acc, save the model and update acc to α.

[0022] After training based on the loop constraint mechanism, the accuracy parameter acc is saved.

[0023] Furthermore, the cycle constraint mechanism is specifically as follows:

[0024] Set the update cycle steps and the maximum number of stall steps;

[0025] When the accuracy parameter is updated, the update cycle step number is assigned a value of 0, and when the accuracy parameter is not updated, the update cycle step number is increased by 1;

[0026] When the value of the update loop step number exceeds the maximum stagnation step number, the loop is exited and the model training ends.

[0027] Furthermore, the matching rules are specifically as follows:

[0028] Match the diagnostic information with the diagnostic name in the diagnostic database by first field to obtain the matching set s1;

[0029] Match the diagnostic information with the diagnostic names in the diagnostic database by non-first field to obtain a matching set s2;

[0030] Perform probability analysis on the matching sets s1 and s2 respectively to obtain the corresponding sorted lists l1 and l2;

[0031] The diagnosis names in the sorted lists l1 and l2 are sorted in their respective lists according to the probability values, and the diagnosis names in the two sequence lists are not cross-sorted. The values ​​are taken from the sorting in sequence based on the preset number of auxiliary diagnosis names.

[0032] Furthermore, the probability analysis is specifically as follows:

[0033] The department where the patient is located is obtained, and based on the probability value corresponding to the department in the diagnosis library, all diagnosis names in the matching set are sorted based on the size of the probability value to obtain a sorted list.

[0034] A second aspect of the present invention provides an auxiliary diagnosis system based on a deep neural network, the auxiliary diagnosis system comprising:

[0035] An information collection unit, configured to obtain patient information and quantize the patient information;

[0036] a first auxiliary diagnosis unit, which, when the patient information does not include diagnosis information, inputs the vectorized patient information into a trained deep neural network model, and obtains a plurality of auxiliary diagnosis names based on outputs of the deep neural network model;

[0037] The second auxiliary diagnosis unit, when the patient information includes diagnosis information, matches the diagnosis information with the diagnosis names in the preset diagnosis name library according to a preset matching rule to obtain a plurality of auxiliary diagnosis names.

[0038] A third aspect of the present invention provides a computer storage medium, wherein the computer storage medium stores computer instructions. When the computer instructions are executed on the auxiliary diagnosis system, the auxiliary diagnosis system executes the steps of the auxiliary diagnosis method.

[0039] The auxiliary diagnosis system of the second aspect of the present invention can implement the method of the first aspect and the various implementations of the first aspect, and achieve the same effect.

[0040] The effects provided in the summary of the invention are only the effects of the embodiments, not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects:

[0041] The present invention performs algorithm analysis based on whether there is diagnostic information in the patient information, obtains several auxiliary diagnosis names, and uses a deep learning algorithm to perform auxiliary diagnosis based on the patient information, avoiding the situation where the diagnosis recommendation depends on the doctor's input characters and the unreasonable sorting of the diagnostic names matched after the input characters, providing accurate reference for the doctor to make a diagnosis and adopt further treatment plans, thereby improving the doctor's work efficiency and the patient's medical experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0043] Figure 1 is a schematic flow chart of an embodiment of the method of the present invention;

[0044] Figure 2 is a schematic diagram of the structure of a neural network model in an embodiment of the method of the present invention;

[0045] Figure 3 It is a structural diagram of an embodiment of the system of the present invention. DETAILED DESCRIPTION

[0046] In order to clearly illustrate the technical features of this solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The disclosure below provides many different embodiments or examples for realizing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplicity and clarity and does not itself indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the accompanying drawings are not necessarily drawn to scale. The present invention omits descriptions of well-known components and processing technologies and processes to avoid unnecessary limitations on the present invention.

[0047] like Figure 1 As shown, an auxiliary diagnosis method based on a deep neural network provided by an embodiment of the present invention includes the following steps:

[0048] S1, obtaining patient information, quantizing the patient information, and determining whether the patient information contains diagnostic information;

[0049] S2, if the patient information does not include diagnosis information, inputting the vectorized patient information into a trained deep neural network model, and obtaining a plurality of auxiliary diagnosis names based on the output of the deep neural network model;

[0050] S3. If the patient information includes diagnosis information, the diagnosis information is matched with the diagnosis names in the preset diagnosis library according to the preset matching rules to obtain several auxiliary diagnosis names.

[0051] In one implementation of this embodiment, before the above step S1 , the step further includes constructing a patient information database D and a diagnosis database D′.

[0052] The patient information database includes the patient ID, age, gender, chief complaint and diagnosis name; the diagnosis database includes the diagnosis name and the probability value of the current diagnosis name in each department.

[0053] The process of constructing the patient information database D is as follows: extract cases corresponding to diagnoses from the database, select diagnoses with a number of cases between 100 and 500, so that the medical records corresponding to the diagnoses in D are evenly distributed, and construct the information in these selected medical records into the patient information database D, where each data is in the format of the following Table 1:

[0054]

[0055] Table 1

[0056] The construction process of the diagnosis database D' is as follows: extract all diagnosis sets {d} in the patient information database D, calculate the corresponding probability of {d} in each hospital department (for example, the probability of diagnosis d1 in a certain department is Where n1 is the number of times d1 appears in the department, and s is the total number of diagnoses in the department), the diagnosis library D' is obtained, as shown in Table 2 below:

[0057] Diagnosis name Department of Orthopedics Obstetrics and Gynecology Cardiology Pediatrics hypertension 0.0015 0.0012 0.4 0.00001 rash 0.0001 0.0011 0.0002 0.0019 Femoral head necrosis 0.014 0.000001 0.000001 0.000011

[0058] Table 2

[0059] In addition, a symptom and sign database D was constructed symp , including symptoms and signs and their corresponding aliases, as shown in Table 3 below:

[0060] Symptom name Corresponding alias stomach ache Abdominal pain, stomach pain, stomach pain

[0061] Table 3

[0062] In step S1, the patient information includes age, gender, chief complaint and diagnosis information, and is represented by vectors respectively.

[0063] Age is divided into infants (0-5 years old), teenagers (6-14 years old), young people (15-39 years old), middle-aged people (40-59 years old) and elderly people (>=60 years old). Each age group is one-hot encoded. For example, infants are coded as [1,0,0,0,0], and young people are coded as [0,0,1,0,0], which is recorded as A p .

[0064] Onehot encoding is performed on gender. The onehot encodings of male and female are [1,0] and [0,1] respectively, denoted as S p .

[0065] For the chief complaint, we first encode 4750 common medical characters (including Chinese characters, punctuation marks, and special symbols) and express them in an 80-dimensional vector. The i-th character is recorded as The main complaint can be expressed as a matrix r represents the row of the matrix.

[0066] For diagnostic information, the neural network model M recommends 200 common diagnoses (assuming there are 200 diagnoses in the patient information database D). These 200 common diagnostic names are encoded. The encoding rule is to one-hot encode all the diagnoses contained in patient p and then add them together to obtain the diagnosis vector, which is recorded as Y. p .

[0067] In step S2, the construction of the deep neural network model M includes:

[0068] (i) Divide the training set, validation set and test set:

[0069] In the patient information database D, the number of patients corresponding to each diagnosis to be predicted is divided into a training set, a validation set, and a test set in a ratio of 65%, 15%, and 20%. The training set corresponding to each diagnosis name is combined as the training set of the neural network model M. The same applies to the validation set and test set.

[0070] (ii) Figure 2 As shown, the structure of the deep neural network model is constructed:

[0071] The age vector A p Transformed into a three-layer fully connected network mlp1 The activation function between layers is tanh; the gender vector S p Transformed into another three-layer fully connected network mlp2 The activation function between layers is tanh.

[0072] The matrix corresponding to the chief complaint Output after bidirectional LSTM model Vector Y with diagnosis names p Forming the attention matrix Att, d is the dimension of the diagnosis name vector. The LSTM model mainly consists of a forget gate, an input gate, and an output gate. Specifically, the output f of the forget gate at time t is t =σ(W f ·[h t-1 ,x t ]+b f ), where x t is the word vector of the t-th character, h t-1 is the output of the hidden state at time t-1, W f and b f They are the matrix parameters and bias parameters of the linear transformation respectively; the input gate is mainly determined by two matrices i t =σ(W i ·[h t-1 ,x t ]+b i )and where i t is to decide which information needs to be updated, W i and b i is the corresponding linear transformation weight matrix and bias parameter, and Is the alternative content for updating. The long-term memory at time t is The short-term memory output by the final output gate at time t is h t =o t tanh(C t ), where o t=σ(W O ·[h t-1 ,x t ]+b O ) and W O and b O are the corresponding weights and biases. Since the output at time t is related to the context, a bidirectional LSTM is used, and the semantic output of the t-th character is where [h t ,h t ′] is the new vector that combines the last dimension of the two vectors. The final complaint matrix The output after bidirectional LSTM is

[0073] Output of age Gender output Combined with the attention matrix Att of the main complaint, we get Will Through the last two layers of fully connected networks, the predicted diagnosis vector is obtained Where W Y and b Y are weights and biases respectively;

[0074] Y P and The focal Loss cross entropy is used as the loss function of the deep neural network model, that is,

[0075] (iii) Train the parameters of the neural network model:

[0076] The training set is divided into several batches, and the size of each batch is set to 20. All batches pass through the model M once, which is called an epoch. The network is trained using stochastic gradient descent. Each patient in the batch passes through the structure of the deep neural network model to obtain the average loss of each batch. To minimize the average loss L batch As a training goal;

[0077] During the training process, the model is automatically saved, that is, the parameter acc (initial value is set to 0) is set to record the accuracy of the validation set. When a batch of data passes through the neural network model, the loss L is obtained. batch , using the back-propagation algorithm to minimize the loss L batch In this way, the parameters of the model M are updated. After each epoch of training, the accuracy of all data in the validation set is calculated using the model with the current parameters (the calculation method is n1 is the total number of correct diagnosis names predicted by the validation set, and n2 is the total number of sample diagnosis names in the validation set). Compare the values ​​of α and acc. When α is greater than the current acc value, save the model and update the acc value (i.e., let α = acc).

[0078] After training based on the loop constraint mechanism, the accuracy parameter acc is saved.

[0079] The specific loop constraint mechanism is as follows: after training for several epochs, overfitting may occur, that is, the α obtained each time will no longer increase or will continue to oscillate. To address this situation and reduce unnecessary loops to obtain the optimal model, we set three parameters: update_wait_step (initial value 0), maximum wait_step (initial value assumed to be 10), and break_flag (initial value False). The specific operation method is that when the parameters of the model are updated (that is, when α is greater than the current acc), the update stagnation step update_wait_step is reassigned to 0. When the parameters of the model are no longer updated (that is, α is less than the current acc), the update stagnation step update_wait_step value is increased by 1. It is judged whether the update stagnation step update_wait_step is greater than the maximum stagnation step max_wait_step. If not, the exit loop flag break_flag is still False. If so, the exit loop flag break_flag becomes True. After each epoch, break_flag is judged. If break_flag is True, the loop training is completed. If it is False, the next loop is continued.

[0080] After the loop is completed or exits the loop, the parameters saved by the program are used as the parameters of the final model M.

[0081] In step S2, the vectorized patient information is input into a trained deep neural network model, and the specific process of obtaining a plurality of auxiliary diagnosis names based on the output of the deep neural network model is as follows:

[0082] S21, for patient p, convert its condition information into a vector Then input model M, the model output value is a 200-dimensional vector (we assume there are 200 diseases), The first n largest components are recorded, the position numbers of the n largest components are obtained, and these numbers are arranged from large to small according to the size of the component value. Then, the diagnosis name is obtained according to the position number and the number of the diagnosis name. Finally, the diagnosis name is displayed in the order of arrangement, which is the diagnosis recommended by model M for patient p.

[0083] S22: If the doctor has filled in the diagnosis name in these 200 diagnoses (because a patient may have multiple diagnoses), then select the largest n components excluding the filled-in diagnosis from the 200 diagnoses and recommend n diagnoses according to the method of step S21.

[0084] The diagnoses recommended in step S21 and step S22 serve as auxiliary diagnosis results for doctors.

[0085] In step S3, the matching rules are specifically as follows: match the diagnostic information with the diagnostic name in the diagnostic library by the first field to obtain a matching set s1; match the diagnostic information with the diagnostic name in the diagnostic library by the non-first field to obtain a matching set s2; perform probability analysis on the matching sets s1 and s2 respectively to obtain corresponding sorting lists l1 and l2.

[0086] The diagnosis names in the sorted lists l1 and l2 are sorted in order according to the probability values ​​in their respective lists, and the diagnosis names in the two sequence lists are not cross-sorted. The values ​​are taken from the sorting list in order based on the number of preset auxiliary diagnosis names. Specifically: if the total number of diagnosis names |l1|+|l2| is less than the required number of recommended diagnosis names n, all diagnoses in l1 and l2 are displayed in the order of l1 first and l2 last. All diagnosis names in l1 are displayed in order according to the order in l1, and all diagnosis names in l2 are displayed in order according to the order in l2. If the total number of diagnosis names |l1|+|l2| is greater than the required number of recommended diagnosis names n, we divide it into two cases: (1) |l1| is greater than or equal to n, and the first n diagnosis names in l1 are displayed; (2) |l1| is less than n, and all diagnosis names in l1 are displayed in order, and the first n-|l1| diagnosis names in l2 are displayed after l1 in the order in l2.

[0087] The probability analysis is specifically as follows:

[0088] Get the department where the patient is located, and based on the probability value corresponding to the department in the diagnosis library, sort all the diagnosis names in the matching set based on the probability value to obtain a sorted list. Specifically: find the probability value of all the diagnosis names in s1 under the department where the patient is located according to the matched diagnosis name collection s1, and sort all the diagnosis names in s1 from large to small according to the probability value of each diagnosis name to obtain a list l1, and remove diagnoses with a probability of 0 from l1. If l1 contains diagnoses that have been filled in by the doctor, they are also removed from l1. Similarly, the set s2 obtained in step (301) is obtained in the same way as s1 to obtain the probability list l2.

[0089] like Figure 3As shown, the present invention also provides an auxiliary diagnosis system based on a deep neural network, and the auxiliary diagnosis system includes an information acquisition unit, a first auxiliary diagnosis unit and a second auxiliary diagnosis unit.

[0090] The information acquisition unit is used to obtain patient information and vectorize the patient information; when the patient information does not include diagnostic information, the first auxiliary diagnosis unit inputs the vectorized patient information into a trained deep neural network model, and obtains a number of auxiliary diagnosis names based on the output of the deep neural network model; when the patient information includes diagnostic information, the second auxiliary diagnosis unit matches the diagnostic information with the diagnostic names in a preset diagnostic name library according to preset matching rules to obtain a number of auxiliary diagnosis names.

[0091] The present invention also provides a computer storage medium, in which computer instructions are stored. When the computer instructions are run on the auxiliary diagnosis system, the auxiliary diagnosis system executes the steps of the auxiliary diagnosis method.

[0092] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. An auxiliary diagnosis method based on deep neural network, characterized by: The auxiliary diagnosis method comprises the following steps: Acquiring patient information and quantizing the patient information; If the patient information does not include diagnosis information, inputting the vectorized patient information into a trained deep neural network model, and obtaining a plurality of auxiliary diagnosis names based on the output of the deep neural network model; If the patient information includes diagnosis information, the diagnosis information is matched with the diagnosis names in the preset diagnosis library according to the preset matching rules to obtain a number of auxiliary diagnosis names; The auxiliary diagnosis method further comprises, before the above steps, constructing a patient information database D and a diagnosis database D'; The patient information database includes patient ID, age, gender, chief complaint and diagnosis name; the diagnosis database includes diagnosis name and the probability value of the current diagnosis name in each department; The process of constructing the patient information database D is as follows: extract cases corresponding to diagnoses from the database, select diagnoses with a number of cases between 100 and 500, so that the medical records corresponding to the diagnoses in D are evenly distributed, and construct the information in these selected medical records into the patient information database D, where each data entry includes the patient ID, age, gender, chief complaint, and diagnosis name; The construction process of the diagnosis database D' is as follows: extract all diagnosis sets {d} in the patient information database D, calculate the probability of {d} corresponding to each hospital department, and obtain the diagnosis database D', where the diagnosis probability under a certain department is Where n1 is the number of times the diagnosis appears in the department, and s is the total number of diagnoses in the department; Constructing a symptom and sign database symp , including symptoms and signs and their corresponding aliases; The matching rules are specifically as follows: Match the diagnostic information with the diagnostic name in the diagnostic database by first field to obtain the matching set s1; Match the diagnostic information with the diagnostic names in the diagnostic database by non-first field to obtain a matching set s2; Perform probability analysis on the matching sets s1 and s2 respectively to obtain the corresponding sorted lists l1 and l2; The diagnosis names in the sorted lists l1 and l2 are sorted in order according to the probability values ​​in their respective lists, and the diagnosis names in the two sequence lists are not cross-sorted, and values ​​are taken from the sorting in order based on the number of preset auxiliary diagnosis names; Specifically, if the total number of diagnostic names |l1|+|l2| is less than the required number of recommended diagnostic names n, all diagnoses in l1 and l2 will be displayed in the order of l1 first and l2 last. All diagnostic names in l1 will be displayed in the order of l1, and all diagnostic names in l2 will be displayed in the order of l2. If the total number of diagnostic names |l1|+|l2| is greater than the required number of recommended diagnostic names n, we divide it into two cases: (1) |l1| is greater than or equal to n, the first n diagnostic names in l1 will be displayed; (2) |l1| is less than n, all diagnostic names in l1 will be displayed in order, and the first n-|l1| diagnostic names in l2 will be displayed after l1 in the order of l2. The probability analysis is specifically as follows: Obtain the department where the patient is located, and based on the probability value corresponding to the department in the diagnosis library, sort all the diagnosis names in the matching set based on the probability value to obtain a sorted list; Specifically: find the probability values ​​of all diagnosis names in s1 under the department where the patient is located according to the matched diagnosis name collection s1, sort all diagnosis names in s1 from large to small according to the probability value of each diagnosis name to obtain list l1, and remove diagnoses with a probability of 0 from l1. If l1 contains diagnoses that have been filled in by the doctor, they are also removed from l1; use the same method to obtain probability list l2 from set s2 according to the method of s1.

2. The auxiliary diagnosis method based on deep neural network according to claim 1, characterized in that: The structure of the deep neural network model is specifically as follows: The age vector A p Transformed into a three-layer fully connected network mlp1 The activation function between layers is tanh; the gender vector S p Transformed into another three-layer fully connected network mlp2 The activation function between layers is tanh; The matrix corresponding to the chief complaint Output after bidirectional LSTM model Vector Y with diagnosis names p Forming the attention matrix Att, d is the dimension of the diagnosis name vector; for the chief complaint, first encode the 4750 characters commonly used in medicine, and express these characters with an 80-dimensional vector, with the i-th character being recorded as The main complaint can be expressed as a matrix r represents the row of the matrix; Will Combined with Att, we get Will The predicted diagnostic vector is obtained through a two-layer fully connected network Where W Y and b Y are weights and biases respectively; Y P and The focal Loss cross entropy is used as the loss function of the deep neural network model.

3. The auxiliary diagnosis method based on deep neural network according to claim 1, characterized in that: The training of the deep neural network model is specifically as follows: The training set is divided into several batches, and each patient in the batch is subjected to the structure of the deep neural network model to obtain the average loss of each batch. Taking minimizing the average loss as the training objective; During the training process, set the parameter acc to record the accuracy of the validation set. After training a batch, calculate the current validation set accuracy α. When α is greater than acc, save the model and update acc to α. After training based on the loop constraint mechanism, the accuracy parameter acc is saved.

4. The auxiliary diagnosis method based on deep neural network according to claim 3 is characterized in that: The cycle constraint mechanism is specifically as follows: Set the update cycle steps and the maximum number of stall steps; When the accuracy parameter is updated, the update cycle step number is assigned a value of 0, and when the accuracy parameter is not updated, the update cycle step number is increased by 1; When the value of the update loop step number exceeds the maximum stagnation step number, the loop is exited and the model training ends.

5. The auxiliary diagnosis system based on deep neural network is characterized by: The system is used to implement the method according to claim 1, and the auxiliary diagnosis system includes: An information collection unit, configured to obtain patient information and quantize the patient information; a first auxiliary diagnosis unit, which, when the patient information does not include diagnosis information, inputs the vectorized patient information into a trained deep neural network model, and obtains a plurality of auxiliary diagnosis names based on outputs of the deep neural network model; The second auxiliary diagnosis unit, when the patient information includes diagnosis information, matches the diagnosis information with the diagnosis names in the preset diagnosis name library according to a preset matching rule to obtain a plurality of auxiliary diagnosis names.

6. A computer storage medium storing computer instructions, wherein: When the computer instructions are executed on the auxiliary diagnosis system according to claim 5 , the auxiliary diagnosis system is caused to execute the steps of the auxiliary diagnosis method according to any one of claims 1 to 4 .

Images