Item storage and automatic classification method for disinfection supply center based on intelligent algorithm

Abstract

This application belongs to the technical field of item classification in disinfection supply centers, and relates to an item storage and automatic classification method for disinfection supply centers based on intelligent algorithms; the present invention first utilizes image recognition technology and RFID tag technology to quickly identify items, and identify images and departments; then combines electronic medical record information to analyze relevant information of the current case to determine whether there is an infection case; then constructs a risk classification model to further analyze the name, material type and contact type of the item, and integrates features with department and infection case information to achieve accurate classification of items; then, by matching the disinfection standard database, determines the appropriate disinfection process for each type of item, solving the problem of mismatched disinfection processes; finally, stores different types of items in different storage areas to facilitate subsequent disinfection and distribution, thereby improving the overall work efficiency, accuracy and safety of the disinfection supply center.

Description

Technical Field

[0001] The present application belongs to the technical field of item classification in a disinfection supply center, and more specifically, relates to an item storage and automatic classification method for a disinfection supply center based on an intelligent algorithm. Background Art

[0002] The disinfection and supply center is a vital department in a medical institution. It is mainly responsible for processing and supplying medical supplies needed by various departments of the hospital. These items need to undergo a strict cleaning, disinfection, packaging and sterilization process before use to ensure that they will not cause cross infection to patients during use.

[0003] The current steps for the Sterilization Supply Center to process recycled items are as follows:

[0004] Recycling: Used instruments and items should be promptly placed in a sealed waste bin to prevent drying, and the name and quantity of the bag should be recorded on the lid of the bin. Supply room staff should visit clinical departments twice a day to recycle dirty instruments and items, and hand over and record the collection with clinical department staff. After recycling, the recycling staff will hand over the number of items to the cleaning staff, and count and check whether the items in the bag are complete. After each recycling, the recycling bin should be cleaned and disinfected, and stored in a dry place. Used disposable items and medical waste must not be recycled to the disinfection supply center for reprocessing.

[0005] Classification: Classify according to the different materials, properties and contamination conditions of the equipment;

[0006] Cleaning and disinfection: Different instruments and items require different cleaning and disinfection methods. The basic cleaning and disinfection process is as follows:

[0007] Pre-wash (tap water) - cleaning (manual + enzyme) - rinsing (tap water) - disinfection (moist heat disinfection) - final rinse (deionized water or distilled water) - drying - instrument inspection and maintenance - packaging - sterilization - storage - distribution;

[0008] However, currently, classification is done manually based on the understanding of the items, which will result in the following situations:

[0009] Poor classification accuracy: Since the classification of items is based on manual understanding, it is easily affected by personal factors, resulting in inaccurate classification results. Infected items may be mixed with non-infected items, increasing the risk of cross-infection;

[0010] Mismatched disinfection procedures: Due to the inability to accurately identify the infection cases of items, the selection of disinfection procedures is not accurate enough, which may cause insufficient disinfection of items. Summary of the Invention

[0011] The present invention provides an item storage and automatic classification method for a disinfection supply center based on an intelligent algorithm, which aims to solve the technical problems of poor classification accuracy and mismatch of disinfection processes in the classification methods currently used in disinfection supply centers.

[0012] The method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm includes the following steps:

[0013] Step 1: Using image recognition technology, cameras installed at each entrance and sorting area of the disinfection supply center are used to capture images of items. RFID tag technology is then used to identify the department to which the items are referred.

[0014] Step 2: Based on the electronic medical record information and the identified department, obtain relevant information about the current case in the corresponding department. Based on the relevant information of the current case, extract whether there is an infection case within the predetermined time, obtain a label of whether there is an infection case in the corresponding department, and if there is an infection case, extract the name of the infection case;

[0015] Step 3: The collected images, the department, and the labels indicating whether there are any infection cases in the corresponding department are used as input into the constructed risk classification model. The collected images are processed based on the risk classification model to obtain the name, material type, and contact type of the object with the human body. The results obtained from the image processing are then fused with the labels of the department and whether there are any infection cases in the corresponding department to classify the objects and obtain the risk classification results.

[0016] Step 4: Based on the name of the infected case, the material type of the item, and the risk classification results, the results are matched with the pre-built disinfection standard database to determine the disinfection process corresponding to the corresponding risk classification results;

[0017] Step 5: According to the classification results, separate the items with the same risk classification results into the corresponding storage areas and prepare to execute the corresponding disinfection process.

[0018] The present invention first uses image recognition technology and RFID tag technology to quickly identify objects, images and departments; then combines electronic medical record information to analyze relevant information of the current case to determine whether there is an infected case, providing an important decision-making basis for the subsequent disinfection process; then constructs a risk classification model to further analyze the name, material type and contact type of the object, and integrates features with department and infection case information to achieve accurate classification of objects, ensure the separation of infected and non-infected objects, and reduce the risk of cross-infection; then, by matching the disinfection standard database, the appropriate disinfection process is determined for each type of object, solving the problem of mismatched disinfection processes and ensuring the disinfection effect. Finally, different types of objects are stored in different storage areas, so that objects of the same risk classification can be stored in an orderly manner, facilitating subsequent disinfection and distribution, and overall improving the work efficiency, accuracy and safety of the disinfection supply center.

[0019] Preferably, the step of extracting whether there is a case infection within a predetermined time based on the relevant information of the current case is as follows:

[0020] Timestamp acquisition: obtain the timestamp used to identify the department mentioned on the item based on RFID tag technology;

[0021] Determine time: take the day to which the identified timestamp belongs as the predetermined time;

[0022] Extract whether there are infected cases: Extract the information of all cases on that day and determine whether there are infected cases among all cases on that day. If yes, mark it as 1; if not, mark it as 0.

[0023] Preferably, the specific structure of the risk classification model is as follows:

[0024] Input layer: Input the object images captured by the camera, the department, and the infection risk labels extracted from the case information within the department;

[0025] The image processing module includes an object name recognition unit, a material type recognition unit, and a contact type recognition unit;

[0026] The object name recognition unit recognizes the object image through the CNN network and outputs the name of the object;

[0027] The material type recognition unit extracts the material information of the object in the object image based on the CNN network;

[0028] The contact type identification unit infers the contact type of the object based on the name and material type of the object in combination with predefined rules;

[0029] Feature fusion layer: Through weighted feature fusion, the object name, material type, contact type, department information, and infection case label are weighted and spliced to obtain fused features;

[0030] Risk classification module: Classify items based on fusion features and multi-layer perceptrons to obtain risk values, and then divide items into corresponding risk levels based on the risk values.

[0031] Preferably, the item name recognition unit uses ResNet50 as the basic structure, and combines the attention mechanism and multi-scale feature fusion to extract the item name. The structure is as follows:

[0032] Input layer: input the image obtained by the camera;

[0033] Convolutional layer: A pre-trained ResNet50 network model is used for feature extraction. The ResNet50 network model includes multiple convolutional layers and multiple residual modules. Feature extraction is performed based on multiple convolutional layers and multiple residual modules to obtain a high-dimensional feature representation of the image.

[0034] Attention mechanism module: A Squeeze-and-Excitation module is added after the last convolutional layer of the ResNet50 network model. After the input of the last convolutional layer enters the Squeeze-and-Excitation module, global average pooling is first performed. The feature map of each channel is converted into a scalar value through average pooling. Then, a channel weight coefficient is generated based on the fully connected layer. Finally, each channel of the feature map is multiplied by the corresponding weight coefficient to strengthen the focus on key information.

[0035] Fully connected layer: performs a fully connected operation on the features output by the attention mechanism module and finally outputs the name of the recognized item;

[0036] Output layer: Based on a Softmax classification layer, it is used to classify different item names.

[0037] Preferably, the CNN network in the material type recognition unit includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fully connected layer and an output layer;

[0038] The first convolutional layer receives data from the input layer, uses a 3×3 convolution kernel, 64 output channels, a stride of 1, the same padding, and a ReLU activation function; the first convolutional layer is connected to the first maximum pooling layer, and the pooling size of the first maximum pooling layer is 2×2 and the stride is 2;

[0039] The second convolutional layer receives the output of the first maximum pooling layer, uses a 3×3 convolution kernel, 128 output channels, a stride of 1, the same padding, and a ReLU activation function; the second convolutional layer is connected to the second maximum pooling layer, and the pooling size of the second maximum pooling layer is 2×2 and the stride is 2;

[0040] The third convolutional layer receives the output of the second maximum pooling layer, uses a 3×3 convolution kernel, an output channel of 256, a stride of 1, padding of the same, and an activation function of ReLU; wherein the third convolutional layer is connected to the third maximum pooling layer, and the pooling size of the third maximum pooling layer is 2×2 and the stride is 2;

[0041] Fully connected layer: First, the feature map output by the third pooling layer is flattened. The flattened feature map is then input into the fully connected layer for processing and outputting the predicted value of the category. The fully connected layer consists of 128 neurons, uses the ReLU activation function, and sets a dropout rate of 0.5 to prevent overfitting.

[0042] Output layer: The softmax function is used to convert the output of the fully connected layer into a probability distribution, and the category with the highest probability is used as the prediction result of the model.

[0043] Preferably, the specific steps of the feature fusion layer are as follows:

[0044] Assign a learnable weight to each input feature, and then multiply the weight by the corresponding feature to obtain the weighted feature;

[0045] Each weighted feature is concatenated to form a new feature vector, and the new feature vector is input as a fusion feature into the multi-layer perceptron for risk classification.

[0046] Preferably, the multilayer perceptron includes an input layer, a first hidden layer, a second hidden layer, and an output layer;

[0047] The input layer is used to receive fusion features;

[0048] The first hidden layer is used to process the fusion features input in the input layer, wherein the first hidden layer contains n neurons and adopts the ReLU activation function;

[0049] The second hidden layer is used to process the data output by the first hidden layer, wherein the second hidden layer comprises m neurons, where m is greater than n, and adopts a ReLU activation function;

[0050] The output layer consists of one neuron, which outputs the risk value of the item. The output value is processed by the Sigmoid activation function and limited to the range of [0, 1].

[0051] Preferably, step 4 comprises the following steps:

[0052] A predefined risk level disinfection mapping table is queried based on the material type and risk classification results of the item. The risk level disinfection mapping table contains disinfection processes corresponding to different disinfection levels, where the disinfection process includes the name of the disinfection process and the disinfection method. The risk level disinfection mapping table is constructed for each material type. When querying, the risk level disinfection mapping table corresponding to the material type is first obtained based on the material type, and then the corresponding disinfection process in the table is queried based on the risk classification result.

[0053] Finally, the disinfection process of the query is modified based on the name of the infected case.

[0054] Preferably, the specific steps of the disinfection process for correcting the query based on the name of the infection case are as follows:

[0055] Based on the infection case name, the infection case disinfection mapping table is searched to determine the specific disinfection method required for the infection case; wherein the infection case mapping table contains the mapping relationship between the infection case, the step, and the disinfection method;

[0056] Based on the steps and disinfection methods in the table, determine whether the corresponding steps in the disinfection process corresponding to the risk classification result query table use the same disinfection method. If the same disinfection method is used, the disinfection process corresponding to the risk classification query table is used as the output result; if different, the disinfection method involved in the infection case mapping table is replaced with the disinfection method corresponding to the steps involved in the risk classification query table, and the disinfection process after the replacement of the disinfection method is output.

[0057] The beneficial effects of the present invention include:

[0058] The present invention first uses image recognition technology and RFID tag technology to quickly identify objects, images and departments; then combines electronic medical record information to analyze relevant information of the current case to determine whether there is an infected case, providing an important decision-making basis for the subsequent disinfection process; then constructs a risk classification model to further analyze the name, material type and contact type of the object, and integrates features with department and infection case information to achieve accurate classification of objects, ensure the separation of infected and non-infected objects, and reduce the risk of cross-infection; then, by matching the disinfection standard database, the appropriate disinfection process is determined for each type of object, solving the problem of mismatched disinfection processes and ensuring the disinfection effect. Finally, different types of objects are stored in different storage areas, so that objects of the same risk classification can be stored in an orderly manner, facilitating subsequent disinfection and distribution, and overall improving the work efficiency, accuracy and safety of the disinfection supply center. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0060] Figure 1 This is a flowchart of the overall steps provided by an embodiment of the present invention.

[0061] Figure 2 This is a simplified structural diagram of the risk classification model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to make the technical problems, technical solutions and beneficial effects to be solved by this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0063] See also Figure 1 As shown, the method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm includes the following steps:

[0064] Step 1: Using image recognition technology, cameras installed at each entrance and sorting area of the disinfection supply center are used to capture images of items. RFID tag technology is then used to identify the department to which the items are referred.

[0065] As mentioned in the background technology, used instruments and item packages should be promptly placed in a waste bin for sealed storage. At this time, we can use RFID tag technology to identify the corresponding department. We only need to set the corresponding department label on one side of the waste bin, and then scan the code with a scanner to obtain the corresponding department name; as for image recognition technology, we can install a camera above the sorting center. When it is necessary to open the lid to sort the items in the waste bin, we use the camera to collect images of the inside of the waste bin.

[0066] Step 2: Based on the electronic medical record information and the identified department, obtain relevant information about the current case in the corresponding department. Based on the relevant information of the current case, extract whether there is an infection case within the predetermined time, obtain a label of whether there is an infection case in the corresponding department, and if there is an infection case, extract the name of the infection case;

[0067] The steps of extracting whether there is a case infection within a predetermined time based on the relevant information of the current case are as follows:

[0068] Timestamp acquisition: obtain the timestamp used to identify the department mentioned on the item based on RFID tag technology;

[0069] Determine time: take the day to which the identified timestamp belongs as the predetermined time;

[0070] Extract whether there are infected cases: Extract the information of all cases on that day and determine whether there are infected cases among all cases on that day. If yes, mark it as 1; if not, mark it as 0.

[0071] As mentioned in the background technology, medical supplies need to be recycled twice a day in the corresponding department. By obtaining the timestamp when the code is scanned, the case status of the department on the date of the timestamp is obtained. If there is an infected case, it is marked and the name of the infected case is obtained. Then, the infected case is taken into account.

[0072] Step 3: The collected images, the department, and the labels indicating whether there are any infection cases in the corresponding department are used as input into the constructed risk classification model. The collected images are processed based on the risk classification model to obtain the name, material type, and contact type of the object with the human body. The results obtained from the image processing are then fused with the labels of the department and whether there are any infection cases in the corresponding department to classify the objects and obtain the risk classification results.

[0073] See also Figure 2 As shown, the specific structure of the risk classification model is as follows:

[0074] Input layer: Inputs the object images captured by the camera, and the infection risk labels extracted from the department and the case information within the department. The input layer consists of two input units: one for the object images and one for the infection risk labels extracted from the department and the case information within the department. The object images enter the image processing module, and the infection risk labels extracted from the department and the case information within the department enter the feature fusion layer.

[0075] The image processing module includes an object name recognition unit, a material type recognition unit, and a contact type recognition unit;

[0076] The object name recognition unit recognizes the object image through the CNN network and outputs the name of the object;

[0077] The object name recognition unit uses ResNet50 as the basic structure, where the pre-trained weights are trained on datasets such as ImageNet, and combines the attention mechanism and multi-scale feature fusion to extract object names. The structure is as follows:

[0078] Input layer: Input the image H×W×C acquired by the camera, where H represents the height of the image; W represents the width of the image; and C represents the number of channels. Before inputting the image into the input layer, the image needs to be normalized to ensure that the value of each pixel is within a reasonable range, for example, by mapping the pixel value of the image from 0-255 to the range of 0-1. The input image is scaled or cropped to meet the size requirements of the input layer (for example, 224×224).

[0079] Convolutional layer: A pre-trained ResNet50 network model is used for feature extraction. The ResNet50 network model includes multiple convolutional layers and multiple residual modules. Feature extraction is performed based on multiple convolutional layers and multiple residual modules to obtain a high-dimensional feature representation of the image. The image is passed through multiple convolutional layers for feature extraction. The convolutional layers extract local information of the image (such as low-level features such as edges, textures, and colors). The residual modules use skip connections to avoid small gradients and ensure that features can be effectively propagated in deeper networks. The residual modules output a deep feature map that contains high-level information from the original image (such as object shape and texture). Example:

[0080] Suppose we have an initial input image I0 (size is 224×224×3), and after a series of convolutional layers and residual modules, we get a deep feature map.

[0081] The first residual module (Block1):

[0082] Input feature map size: 224×224×64;

[0083] Convolution operation: extract low-level features such as edges and textures;

[0084] Skip connection: add the input feature map to the convolution output;

[0085] Output feature map: 224×224×64;

[0086] Second residual module (Block2):

[0087] Input feature map: 224×224×64;

[0088] Convolution operation: further extract more complex patterns, local object features, etc.

[0089] Output feature map: 224×224×64;

[0090] Continue similar processing until the 49th residual module (Block49);

[0091] The last convolutional layer:

[0092] The last convolutional layer of ResNet50 outputs a deep feature map of size 7×7×2048; the feature map contains high-level information about the image, such as object shape, texture, and position. This feature map will serve as the input to the attention mechanism module.

[0093] Attention mechanism module: A Squeeze-and-Excitation module is added after the last convolutional layer of the ResNet50 network model. After the input of the last convolutional layer enters the Squeeze-and-Excitation module, global average pooling is first performed. The feature map of each channel is converted into a scalar value through average pooling. Then, a channel weight coefficient is generated based on the fully connected layer. Finally, each channel of the feature map is multiplied by the corresponding weight coefficient to strengthen the focus on key information.

[0094] Exemplary:

[0095] Global Average Pooling: Global Average Pooling is performed on each channel, compressing each 7×7 feature map into a single scalar (average). The output is a 2048-dimensional vector representing the global features of each channel:

[0096]

[0097] Where: x i,j,c represents the eigenvalue of channel c at position (i, j) in the feature map; z c represents the global average pooling value of channel c;

[0098] The fully connected layer generates channel weights: The 2048-dimensional vector obtained by global average pooling is fed into a fully connected layer. This layer generates the weight coefficients for each channel through an activation function. Specifically, the SE module generates the channel weight coefficients through two fully connected layers (FC1 and FC2):

[0099] w c =σ(W2·δ(W1·z));

[0100] Where: W1 and W2 represent the weight matrices of the fully connected layer; σ represents the Sigmoid activation function; δ represents the ReLU activation function;

[0101] Weighted feature map: By multiplying the feature map of each channel by the corresponding weight coefficient w c , achieving the enhancement of important features. The weighted feature map can highlight the key information that needs to be paid attention to in object name recognition.

[0102] Output: The weighted feature map is still of size 7×7×2048, but the features of each channel have been weighted according to the attention mechanism;

[0103] Fully connected layer: Performs a fully connected operation on the features output by the attention mechanism module and finally outputs the name of the identified item. The specific processing steps are as follows:

[0104] Flatten: Flatten the 7×7×2048 feature map into a one-dimensional vector, with a size of 7×7×2048 = 100352 dimensions. The flattened feature vector is input to the fully connected layer, which maps the flattened feature vector (100352 dimensions) to the category space, generating a vector containing category probabilities. The fully connected layer uses the learned weight matrix to achieve input-output mapping. Each output neuron is the weighted sum of the input vector and a set of weight matrices, plus a bias term. The output is an N-dimensional vector representing the predicted score for each category.

[0105] Output layer: Based on a Softmax classification layer, it is used to classify different item names. The Softmax classification layer (Softmax function) converts the network output into a probability distribution. Each element of this probability distribution is between [0, 1], and the sum of all elements is 1. The output of the Softmax function is the predicted probability of each category. The network will eventually select the category with the largest probability as the final prediction result.

[0106] The loss function of the object name recognition unit uses the cross entropy loss function to calculate the difference between the true label and the probability of the model output; the loss value is used for backpropagation to update the model parameters.

[0107] The material type recognition unit extracts the material information of the object in the object image based on the CNN network;

[0108] The CNN network in the material type recognition unit includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fully connected layer and an output layer;

[0109] The first convolutional layer receives data from the input layer, uses a 3×3 convolution kernel, 64 output channels, a stride of 1, the same padding, and a ReLU activation function; the first convolutional layer is connected to the first maximum pooling layer, and the pooling size of the first maximum pooling layer is 2×2 and the stride is 2;

[0110] The second convolutional layer receives the output of the first maximum pooling layer, uses a 3×3 convolution kernel, 128 output channels, a stride of 1, the same padding, and a ReLU activation function; the second convolutional layer is connected to the second maximum pooling layer, and the pooling size of the second maximum pooling layer is 2×2 and the stride is 2;

[0111] The third convolutional layer receives the output of the second maximum pooling layer, uses a 3×3 convolution kernel, an output channel of 256, a stride of 1, padding of the same, and an activation function of ReLU; wherein the third convolutional layer is connected to the third maximum pooling layer, and the pooling size of the third maximum pooling layer is 2×2 and the stride is 2;

[0112] Fully connected layer: First, the feature map output by the third pooling layer is flattened. The flattened feature map is then input into the fully connected layer for processing and outputting the predicted value of the category. The fully connected layer consists of 128 neurons, uses the ReLU activation function, and sets a dropout rate of 0.5 to prevent overfitting.

[0113] Output layer: The softmax function is used to convert the output of the fully connected layer into a probability distribution, and the category with the highest probability is used as the prediction result of the model.

[0114] The contact type identification unit infers the contact type of the object based on the name and material type of the object in combination with predefined rules;

[0115] For example: contact types include direct contact, invasive contact, epidermal contact, and indirect contact;

[0116] Direct contact is defined as surgical instruments, medical devices or any items that come into direct contact with the patient's skin, mucous membranes or tissues.

[0117] Invasive contact is defined as emphasizing the invasiveness of surgery or medical procedures, which means that surgical instruments or medical tools enter the patient's body and come into contact with internal tissues or organs.

[0118] Epidermal contact is contact with the patient's surface skin;

[0119] Indirect contact refers to indirect contact between an object and a patient or the environment, usually when an object transmits pathogens by contacting the environment or surface.

[0120] For example, the object type we identified is a scalpel, which is generally used to cut skin diseases deep into the subcutaneous tissue, muscle layer or even deeper. Therefore, based on this logic, the scalpel is an invasive contact.

[0121] Another example is bedding, which directly contacts the epidermis, so it is epidermal contact.

[0122] We predefine the above rules to determine the contact type of the corresponding object. Next, we can build a mapping table that contains the mapping between object names and contact types. As mentioned in the above rules, a scalpel corresponds to invasive contact. Given a mapping relationship, we can query the contact relationship of the corresponding object through the mapping table.

[0123] Feature fusion layer: Through weighted feature fusion, the object name, material type, contact type, department information, and infection case label are weighted and spliced to obtain fused features;

[0124] The specific steps of the feature fusion layer are as follows:

[0125] Assign a learnable weight to each input feature, and then multiply the weight by the corresponding feature to obtain the weighted feature;

[0126] Each weighted feature is concatenated to form a new feature vector, and the new feature vector is input as a fusion feature into the multi-layer perceptron for risk classification.

[0127] Based on the above, we can see that the input of our risk classification module includes weighted item name, material type, contact type, department information and infection case label; the level of risk is most correlated with the contact type and infection case label, so a higher initial weight can be assigned, while a lower weight is assigned to the department information, item name and material type; the sum of the weights corresponding to the item name, material type, contact type, department information and infection case label is 1; for example, the initial weights of the contact type and infection case label are both 0.35; the weights of the other three input features are all 0.1.

[0128] Risk classification module: classifies items based on fusion features and multi-layer perceptrons to obtain risk values, and then divides items into corresponding risk levels based on the risk values;

[0129] The multilayer perceptron includes an input layer, a first hidden layer, a second hidden layer and an output layer;

[0130] The input layer is used to receive fusion features;

[0131] The first hidden layer is used to process the fusion features input in the input layer, wherein the first hidden layer contains n neurons and adopts the ReLU activation function;

[0132] The second hidden layer is used to process the data output by the first hidden layer, wherein the second hidden layer comprises m neurons, where m is greater than n, and adopts a ReLU activation function;

[0133] The output layer consists of one neuron, which outputs the risk value of the item. The output value is processed by the Sigmoid activation function and limited to the range of [0, 1].

[0134] In this embodiment, since the data dimension used is small, the multi-layer perceptron does not use the Dropout layer, as follows:

[0135] The input features are as follows:

[0136] Item name (text feature): If using pre-trained word embeddings (such as Word2Vec or GloVe), each item name will be mapped to a fixed-dimensional vector (32 or 64 dimensions).

[0137] Material type (discrete category): usually one-hot encoding is used. Assuming there are 6 material types, the dimension is 6.

[0138] Contact type (discrete category): If one-hot encoding is used, assuming there are 5 contact types, the dimension is 5.

[0139] Department information (discrete categories): Department information is processed through the embedding layer, assuming the embedding dimension is 6.

[0140] Infection case label (binary variable): This is a scalar with a value of 0 or 1 and a dimension of 1.

[0141] Therefore, after embedding and encoding, the input dimension will be the sum of the following parts:

[0142] Item Name: 64 (assuming 64-dimensional embedding)

[0143] Material Type: 6

[0144] Contact type: 5

[0145] Department Information: 6

[0146] Infection case label: 1

[0147] Therefore, the total dimensionality of the input features is: 64 + 6 + 5 + 6 + 1 = 82. This dimensionality is not particularly high for a neural network. Compared to some other complex tasks, an 82-dimensional input does not overly complicate the network, nor does it require a heavy reliance on dropout to prevent overfitting. Furthermore, weighted feature fusion assigns learnable weights to each input feature, allowing the model to adaptively adjust feature importance based on the training data. Weighted feature fusion makes the contribution of each feature more explicit, thus helping the model process features more efficiently. This approach reduces the model's over-reliance on all features, indirectly preventing overfitting.

[0148] In this example, the embedding layer processes department information, converting it into a dense, low-dimensional vector. This better captures the relationship between department and item risk. Different departments have different infection risks. For example, surgery may have a higher risk of external infection, while departments like the ICU may face a higher risk of cross-infection. This allows the model to automatically adjust item risk assessments based on department information, providing more accurate risk classification.

[0149] Secondly, the contact type reflects how an item comes into contact with the human body or other surfaces, which is related to the path of infectious disease transmission. In medical and public health settings, the type of contact an item has a significant impact on infection risk. Through one-hot encoding, different contact types are converted into discrete features, helping the model understand the different impacts of different types of contact on item risk. The infection case label directly indicates whether the item is associated with an infection case, which is crucial for item risk assessment. As a binary feature (0 or 1), it clearly expresses the infection association of the item, helping the model quickly and accurately identify infected items.

[0150] However, in the prior art, whether a department has an infection case and whether a corresponding disinfection process needs to be carried out is known through manual notification. The present invention obtains whether a corresponding department has an infection case directly through electronic medical records, and predicts the risk type based on the information obtained, so that the final prediction result is fully considered and the technical problem of forgetfulness in the prior art is avoided. Although disinfection monitoring will be carried out after disinfection to prevent items containing pathogens from being shipped out, it will also cause duplication of workload. Therefore, we have effectively avoided the problem of duplication of workload; and the present invention is more intelligent, avoiding the technical problems of poor classification accuracy and mismatch of disinfection processes caused by manual classification.

[0151] Step 4: Based on the name of the infected case, the material type of the item, and the risk classification results, the results are matched with the pre-built disinfection standard database to determine the disinfection process corresponding to the corresponding risk classification results;

[0152] The step 4 comprises the following steps:

[0153] Based on the material type and risk classification results of the item, a predefined risk level disinfection mapping table is queried (see Table 2, which is a risk level disinfection mapping table for stainless steel). The risk level disinfection mapping table contains disinfection processes corresponding to different disinfection levels, where the disinfection process includes the name of the disinfection process and the disinfection method. The risk level disinfection mapping table is constructed for each material type. When querying, the risk level disinfection mapping table corresponding to the material type is first obtained based on the material type, and then the corresponding disinfection process in the table is queried based on the risk classification results.

[0154] Table 2: Risk level disinfection mapping table for stainless steel materials

[0155]

[0156]

[0157]

[0158] Based on Table 2 above, a risk level disinfection mapping table for stainless steel materials is given. We provide corresponding risk level disinfection mapping tables for each material. For example, corresponding risk level disinfection mapping tables are provided for glass, fabric, plastic and other materials respectively. Of course, the above examples are not limitations of the present invention. Regarding the level setting, we can set multiple levels, not just five levels. Specifically, different levels can be established according to different disinfection process standards in actual conditions.

[0159] Finally, the disinfection process of the query is modified based on the name of the infected case.

[0160] The specific steps of the disinfection process for modifying the query based on the name of the infected case are as follows:

[0161] Based on the name of the infected case, the infected case disinfection mapping table is queried to determine the specific disinfection method required for the infected case; the infected case mapping table contains the mapping relationship between the infected case, the step, and the disinfection method, see Table 3;

[0162] Table 3. Infection case mapping table

[0163]

[0164]

[0165] Based on the steps and disinfection methods in the table, determine whether the corresponding steps in the disinfection process corresponding to the risk classification result query table use the same disinfection method. If the same disinfection method is used, the disinfection process corresponding to the risk classification query table is used as the output result; if different, the disinfection method involved in the infection case mapping table is replaced with the disinfection method corresponding to the step involved in the risk classification query table, and the disinfection process after the replacement disinfection method is output;

[0166] For example, referring to Table 2, if we output the highest risk, that is, risk level 1, then the corresponding disinfection process in risk level 1 is output; if the name of the infected case is tuberculosis, then query whether the disinfection and sterilization method output in Table 2 meets the requirements of tuberculosis in Table 3. Based on Table 2, it can be seen that it does not meet the requirements, so it is necessary to replace the disinfection and sterilization method combined with the disease in Table 3 with the disinfection and sterilization method output in Table 2, and finally output the adjusted disinfection process.

[0167] Based on the above, different infection cases (such as tuberculosis, hepatitis, drug-resistant bacteria, etc.) have different requirements for disinfection methods. These requirements are usually closely related to factors such as the infectivity, drug resistance, and transmission route of the pathogen. By adjusting the disinfection process according to the specific type of infection case, the targetedness and effectiveness of the disinfection can be ensured. By adjusting the disinfection process based on the type of infection case, cross-contamination of different pathogens can be effectively avoided. For example, for airborne pathogens such as tuberculosis, measures such as moist heat disinfection and ultraviolet disinfection are required, while for blood-borne diseases such as hepatitis, vinyl gas disinfection may be more suitable. Such refined management can reduce the risk of cross-infection, especially in highly sensitive environments (such as operating rooms, ICUs, etc.).

[0168] Step 5: According to the classification results, separate the items with the same risk classification results into the corresponding storage areas and prepare to execute the corresponding disinfection process.

[0169] The present invention first uses image recognition technology and RFID tag technology to quickly identify objects, images and departments; then combines electronic medical record information to analyze relevant information of the current case to determine whether there is an infected case, providing an important decision-making basis for the subsequent disinfection process; then constructs a risk classification model to further analyze the name, material type and contact type of the object, and integrates features with department and infection case information to achieve accurate classification of objects, ensure the separation of infected and non-infected objects, and reduce the risk of cross-infection; then, by matching the disinfection standard database, the appropriate disinfection process is determined for each type of object, solving the problem of mismatched disinfection processes and ensuring the disinfection effect. Finally, different types of objects are stored in different storage areas, so that objects of the same risk classification can be stored in an orderly manner, facilitating subsequent disinfection and distribution, and overall improving the work efficiency, accuracy and safety of the disinfection supply center.

[0170] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm, characterized in that: The following steps are involved: Step 1: Using image recognition technology, cameras installed at each entrance and sorting area of the disinfection supply center are used to capture images of items. RFID tag technology is then used to identify the department to which the items are referred. Step 2: Based on the electronic medical record information and the identified department, obtain relevant information about the current case in the corresponding department. Based on the relevant information of the current case, extract whether there is an infection case within the predetermined time, obtain a label of whether there is an infection case in the corresponding department, and if there is an infection case, extract the name of the infection case; Step 3: The collected images, the department, and the labels indicating whether there are any infection cases in the corresponding department are used as input into the constructed risk classification model. The collected images are processed based on the risk classification model to obtain the name, material type, and contact type of the object with the human body. The results obtained from the image processing are then fused with the labels of the department and whether there are any infection cases in the corresponding department to classify the objects and obtain the risk classification results. The specific structure of the risk classification model is as follows: Input layer: Input the object images captured by the camera, the department, and the infection risk labels extracted from the case information within the department; The image processing module includes an object name recognition unit, a material type recognition unit, and a contact type recognition unit; The object name recognition unit recognizes the object image through the CNN network and outputs the name of the object; The material type recognition unit extracts the material information of the object in the object image based on the CNN network; The contact type identification unit infers the contact type of the object based on the name and material type of the object in combination with predefined rules; Feature fusion layer: Through weighted feature fusion, the object name, material type, contact type, department information, and infection case label are weighted and spliced to obtain fused features; Risk classification module: classifies items based on fusion features and multi-layer perceptrons to obtain risk values, and then divides items into corresponding risk levels based on the risk values; Step 4: Based on the name of the infected case, the material type of the item, and the risk classification results, the results are matched with the pre-built disinfection standard database to determine the disinfection process corresponding to the corresponding risk classification results; Step 5: According to the classification results, separate the items with the same risk classification results into the corresponding storage areas and prepare to execute the corresponding disinfection process.

2. The method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm according to claim 1, characterized in that: The steps of extracting whether there is a case infection within a predetermined time based on the relevant information of the current case are as follows: Timestamp acquisition: obtain the timestamp used to identify the department mentioned on the item based on RFID tag technology; Determine time: take the day to which the identified timestamp belongs as the predetermined time; Extract whether there are infected cases: Extract the information of all cases on that day and determine whether there are infected cases among all cases on that day. If yes, mark it as 1; if not, mark it as 0.

3. The method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm according to claim 1, characterized in that: The item name recognition unit uses ResNet50 as the basic structure and combines the attention mechanism and multi-scale feature fusion to extract the item name. The structure is as follows: Input layer: input the image obtained by the camera; Convolutional layer: A pre-trained ResNet50 network model is used for feature extraction. The ResNet50 network model includes multiple convolutional layers and multiple residual modules. Feature extraction is performed based on multiple convolutional layers and multiple residual modules to obtain a high-dimensional feature representation of the image. Attention mechanism module: A Squeeze-and-Excitation module is added after the last convolutional layer of the ResNet50 network model. After the input of the last convolutional layer enters the Squeeze-and-Excitation module, global average pooling is first performed. The feature map of each channel is converted into a scalar value through average pooling. Then, a channel weight coefficient is generated based on the fully connected layer. Finally, each channel of the feature map is multiplied by the corresponding weight coefficient to strengthen the focus on key information. Fully connected layer: performs a fully connected operation on the features output by the attention mechanism module and finally outputs the name of the recognized item; Output layer: Based on a Softmax classification layer, it is used to classify different item names.

4. The method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm according to claim 1, characterized in that: The CNN network in the material type recognition unit includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, a fully connected layer and an output layer; The first convolutional layer receives data from the input layer, uses a 3×3 convolution kernel, 64 output channels, a stride of 1, the same padding, and a ReLU activation function; the first convolutional layer is connected to the first maximum pooling layer, and the pooling size of the first maximum pooling layer is 2×2 and the stride is 2; The second convolutional layer receives the output of the first maximum pooling layer, uses a 3×3 convolution kernel, 128 output channels, a stride of 1, the same padding, and a ReLU activation function; the second convolutional layer is connected to the second maximum pooling layer, and the pooling size of the second maximum pooling layer is 2×2 and the stride is 2; The third convolutional layer receives the output of the second maximum pooling layer, uses a 3×3 convolution kernel, an output channel of 256, a stride of 1, padding of the same, and an activation function of ReLU; wherein the third convolutional layer is connected to the third maximum pooling layer, and the pooling size of the third maximum pooling layer is 2×2 and the stride is 2; Fully connected layer: First, the feature map output by the third pooling layer is flattened. The flattened feature map is then input into the fully connected layer for processing and outputting the predicted value of the category. The fully connected layer consists of 128 neurons, uses the ReLU activation function, and sets a dropout rate of 0.5 to prevent overfitting. Output layer: The softmax function is used to convert the output of the fully connected layer into a probability distribution, and the category with the highest probability is used as the prediction result of the model.

5. The method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm according to claim 1, characterized in that: The specific steps of the feature fusion layer are as follows: Assign a learnable weight to each input feature, and then multiply the weight by the corresponding feature to obtain the weighted feature; Each weighted feature is concatenated to form a new feature vector, and the new feature vector is input as a fusion feature into the multi-layer perceptron for risk classification.

6. The method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm according to claim 1, characterized in that: The multilayer perceptron includes an input layer, a first hidden layer, a second hidden layer and an output layer; The input layer is used to receive fusion features; The first hidden layer is used to process the fusion features input in the input layer, wherein the first hidden layer contains n neurons and adopts the ReLU activation function; The second hidden layer is used to process the data output by the first hidden layer, wherein the second hidden layer comprises m neurons, where m is greater than n, and adopts a ReLU activation function; The output layer consists of one neuron, which outputs the risk value of the item. The output value is processed by the Sigmoid activation function and limited to the range of [0, 1].

7. The method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm according to claim 1, characterized in that: The step 4 comprises the following steps: A predefined risk level disinfection mapping table is queried based on the material type and risk classification results of the item. The risk level disinfection mapping table contains disinfection processes corresponding to different disinfection levels, where the disinfection process includes the name of the disinfection process and the disinfection method. The risk level disinfection mapping table is constructed for each material type. When querying, the risk level disinfection mapping table corresponding to the material type is first obtained based on the material type, and then the corresponding disinfection process in the table is queried based on the risk classification result. Finally, the disinfection process of the query is modified based on the name of the infected case.

8. The method for storing and automatically classifying items in a disinfection supply center based on an intelligent algorithm according to claim 7, characterized in that: The specific steps of the disinfection process for modifying the query based on the name of the infected case are as follows: Based on the name of the infected case, the infected case disinfection mapping table is searched to determine the specific disinfection method required for the infected case; The infection case mapping table contains the mapping relationship between infection cases, steps, and disinfection methods; Based on the steps and disinfection methods in the table, determine whether the corresponding steps in the disinfection process corresponding to the risk classification result query table use the same disinfection method. If the same disinfection method is used, the disinfection process corresponding to the risk classification query table is used as the output result; if different, the disinfection method involved in the infection case mapping table is replaced with the disinfection method corresponding to the steps involved in the risk classification query table, and the disinfection process after the replacement of the disinfection method is output.

Images