A propofol dosing prediction model based on a neural network and a construction method thereof

Abstract

The application provides a propofol administration prediction model based on a neural network and a construction method thereof, and belongs to the technical field of medical anesthesia models. In view of the problems that the existing propofol administration lacks quantitative calculation and does not consider time difference, the propofol administration prediction model construction method comprises the following steps: taking collected patient basic information and propofol medication data as input variables and preprocessing, constructing a neural network model of an interactive constraint mechanism and a weighted multi-task loss function, linearly fusing an induction dose and a prediction dose of propofol administration, inputting the input variables into the neural network model for training, using a SHAP algorithm to interpret the neural network model, inputting the input variables into the neural network model for testing, and obtaining a propofol administration prediction model. The application can generate feature contribution value sorting and visualized results for the prediction results of the administration of each patient, and clearly indicates the positive or negative influence of variables such as weight, age, gender and physiological rhythm on the prediction results.

Description

Technical Field

[0001] This invention relates to the field of medical anesthesia model technology, and in particular to a propofol dosing prediction model based on neural networks and its construction method. Background Technology

[0002] In recent years, the number of people undergoing painless examinations and surgeries worldwide has increased significantly, including but not limited to painless endoscopy, hysteroscopy, and oocyte retrieval. In China, approximately 7 million people undergo painless gastrointestinal endoscopy (GI) annually. The advantages of painless procedures (such as rapid recovery, enhanced comfort, and earlier discharge) have significantly improved patient satisfaction and compliance. Consequently, the precise use of sedatives and anesthetics is crucial for optimizing procedural efficiency and patient turnover. Propofol is the most widely used anesthetic in painless examinations and surgeries. Propofol dosage is influenced by various factors, including age, sex, weight, circadian rhythm, insomnia, anxiety, and smoking. Currently, the impact of circadian rhythm on propofol use is receiving increasing attention.

[0003] Circadian rhythms are intrinsic, modulated oscillations of physiological structure and function with a period of approximately 24 hours. Studies have shown that propofol administration and efficacy are influenced by circadian rhythms. For example, in patients undergoing laparoscopic surgery, surgeries performed between 8:00 and 18:00 showed higher anesthesia depth index, mean arterial pressure (MBP), and heart rate (HR) compared to surgeries performed between 22:00 and 5:00. In another study, patients undergoing elective laparoscopic abdominal surgery between 8:00 and 12:00 required higher doses of propofol than those undergoing surgery between 18:00 and 22:00. These studies, comparing daytime and nighttime surgeries, limit the applicability of propofol. Furthermore, in short-duration gynecological outpatient surgeries (abortion and curettage) performed between 8:30 and 11:30, the dosage of propofol was found to be lower than that used in surgeries performed between 14:00 and 17:00. These existing qualitative studies confirm the temporal variability in propofol administration. However, due to the lack of quantitative calculations, practical guidance for propofol use in clinical practice cannot be provided. Additionally, existing dosing prediction models do not consider the temporal variability in propofol administration, resulting in inaccurate dosages for painless gastrointestinal endoscopy anesthesia at different time points, still requiring manual adjustment of the dosage.

[0004] Therefore, researching and constructing a feasible dosing prediction model for automatically adjusting propofol dosage that takes into account circadian rhythms is expected to accelerate patient turnover and recovery, and reduce adverse reactions in outpatients. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention aims to construct a propofol dosing prediction model that incorporates the specific factor of circadian rhythm, based on the propofol dosage of clinical patients undergoing painless gastroscopy and colonoscopy during the daytime. Furthermore, it proposes a propofol dosing prediction system and method based on neural networks, thereby effectively solving the problem that existing dosing prediction models in the background technology do not consider the temporal variability of propofol dosing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention discloses a method for constructing a propofol dosing prediction model based on a neural network, comprising the following steps: The collected patient basic information and propofol medication data were used as input variables, and the input variables were preprocessed. A neural network model with an interactive constraint mechanism and a weighted multi-task loss function was constructed to linearly fuse the induction dose and predicted dose of propofol administration. The preprocessed input variables were then input into the neural network model for training. The trained neural network model was interpreted using the SHAP interpretability analysis module, and then the neural network model was tested using preprocessed input variables to obtain the propofol dosing prediction model.

[0007] As a further preferred embodiment of the above technical solution, the neural network model includes an input layer, a shared feature extraction layer, a branching layer, and an output layer, wherein: The input layer is used to input preprocessed input variables; The shared feature extraction layer is used to extract weighted features from the preprocessed input variables. The shared feature extraction layer includes a fully connected layer, a batch normalization layer, an activation function, and an attention mechanism, which are used to extract input variables and calculate the importance weights of each variable. The branch layer includes an induced dose branch layer and a maintenance dose branch layer. Both the induced dose branch layer and the maintenance dose branch layer contain multi-layer fully connected networks. The weighted features are input into the induced dose branch layer and the maintenance dose branch layer, respectively, for induced dose prediction and maintenance dose prediction. The output layer is used to output the predicted value of the induced amount per unit body weight and the predicted value of the maintenance amount per unit body weight per unit time, respectively.

[0008] A further optimized scheme is as follows: an interaction constraint mechanism is introduced between the induced quantity branch layer and the maintenance quantity branch layer. The interaction constraint mechanism adopts a Cross-Stitch structure to linearly fuse the weighted feature tasks input to the induced quantity branch layer and the maintenance quantity branch layer, and to determine the synergistic and restrictive relationship between the induced quantity and the maintenance quantity. The Cross-Stitch structure is expressed as follows: , ,

[0009] Where: parameter matrix Given a 2×2 trainable weight matrix, f induce and f maintain f′ represents the features of the induced and maintenance branches before the interaction, respectively. induce and f′ maintain These represent the features after being processed by the interaction constraint mechanism; α represents the proportion of the induced quantity branch features that retain their own features during interaction updates, 1-α represents the proportion obtained from the maintenance quantity branch features, β represents the proportion of the maintenance quantity branch features that retain their own features during interaction updates, and 1-β represents the proportion obtained from the induced quantity branch features.

[0010] A further optimized scheme is as follows: a weighted multi-task loss function is introduced between the induced dose branch layer and the maintenance dose branch layer to linearly fuse the induced dose and maintenance dose characteristics, further determining the synergistic and restrictive relationship between the induced dose and the maintenance dose. The expression of the weighted multi-task loss function is as follows:

[0011] In the formula: and These represent the actual and predicted values ​​of the induced amount per unit body weight, respectively. and These represent the actual and predicted values ​​of the maintenance duration per unit body weight per unit time, respectively. Indicates mean square error; These are the loss weighting coefficients for different sub-tasks; This is a cross-task regular expression.

[0012] Secondly, this invention discloses a propofol dosing prediction system based on a neural network, wherein a propofol dosing prediction model based on a neural network is mounted on the prediction system, and the prediction system includes: The data collection and processing module collects patients' basic information and propofol medication data as input variables, and preprocesses the input variables. The data input module is used to input the preprocessed input variables into the propofol dosing prediction model; The solution module uses the propofol dosing prediction model to solve and linearly fuse the input variables, and performs induction dose prediction and maintenance dose prediction respectively. The output module outputs the predicted values ​​of induced weight per unit body weight and maintenance weight per unit time.

[0013] A further preferred option is that the prediction system also includes a SHAP interpretability analysis module, which is used to calculate the marginal contribution of the input variables in each prediction result.

[0014] Thirdly, this invention discloses a propofol dosing prediction method based on a neural network. This prediction method is implemented based on the aforementioned propofol dosing prediction system based on a neural network, and includes the following steps: Basic information and propofol usage data of individual patients were obtained as input variables; Input variables are fed into the propofol dosing prediction system, and standardized or normalized preprocessing, feature extraction, branch prediction and interaction constraint calculation are performed. Finally, the predicted values ​​of induced dose per unit weight and maintenance dose per unit weight per unit time for individual patients are output. The marginal contribution ranking and visualization results of the input variables to the predicted values ​​of induced and sustained weight per unit of body weight per unit time were calculated using the SHAP algorithm.

[0015] Furthermore, the present invention also discloses a computer-readable storage medium storing a computer program or instructions, the computer program or instructions being used to cause a processor to execute the steps of the aforementioned propofol dosing prediction method based on a neural network.

[0016] Furthermore, the present invention also discloses a computer device, the computer device including a memory and a processor, the memory storing a computer program or instructions, the computer program or instructions being executed by the processor to implement the steps of the aforementioned propofol dosing prediction method based on a neural network.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The propofol dosing prediction model of this invention improves the prediction accuracy of propofol anesthetic dosage by introducing multi-task learning and attention mechanisms, and constructing a loss function to improve the branch interaction constraint mechanism. It enables personalized dynamic dosing recommendations based on individual patient differences, ensuring the safety and comfort of patients undergoing painless gastroscopy and colonoscopy. Compared with traditional regression models, it can capture nonlinear relationships and significantly improve prediction accuracy.

[0018] 2. The propofol dosing prediction model of the present invention also introduces the SHAP algorithm to quantify the importance of input variables, generate feature contribution value ranking and visualization results for each patient's dosing prediction results, and clearly indicate the positive or negative impact of the patient's basic information on the prediction results. It can not only provide numerical predictions, but also provide clinically interpretable evidence, which facilitates anesthesiologists to make clinical judgments and manual corrections based on the interpretation results of the model of the present invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0020] Figure 1 This is a network structure diagram of the drug administration prediction model of the present invention; Figure 2 The characteristic importance of the output induced dose of the drug delivery prediction model of the present invention; Figure 3 The training / validation loss and validation R of the output induced dose of the drug delivery prediction model of this invention. 2 curve; Figure 4 The importance of the characteristic of the dosing prediction model of the present invention in terms of maintenance dose output per hour per kilogram of body weight; Figure 5 The training / validation loss and validation R of the dosing prediction model of this invention output maintenance dose per hour per kilogram of body weight. 2 curve; Figure 6 The frequency of propofol administration under traditional methods; Figure 7 The dosing frequency of propofol is the dosing prediction model of this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. Example

[0022] Reference Figures 1-7 This invention provides a method for constructing a propofol dosing prediction model based on neural networks, comprising the following steps: The collected patient basic information and propofol medication data were used as input variables, and the input variables were preprocessed. A neural network model with an interactive constraint mechanism and a weighted multi-task loss function was constructed to linearly fuse the induction dose and predicted dose of propofol administration. The preprocessed input variables were then input into the neural network model for training. The trained neural network model was interpreted using the SHAP interpretability analysis module, and then the neural network model was tested using preprocessed input variables to obtain the propofol dosing prediction model.

[0023] Specifically, the patient baseline information and propofol dosage data used in this embodiment are based on a study conducted at the Endoscopic Day Surgery Center of Peking University Third Hospital from July to September 2023. This study complied with the Declaration of Helsinki and was approved by the Ethics Committee of Peking University Third Hospital (Approval No.: M2024956). Furthermore, this prospective study has been registered with the Chinese Clinical Trial Registry (Registration No.: ChiCTR2400093328).

[0024] This study recruited 162 patients scheduled for gastrointestinal examinations. Adults aged 18 to 65 years with an American Society of Anesthesiologists (ASA) classification of I or II were enrolled. ASA I represents good health except for surgery; ASA II represents mild systemic disease with good tolerability. Patients classified as ASA III and IV were excluded, as they had severe systemic disease and significant functional impairment or ongoing life-threatening conditions, respectively. These patients underwent diagnostic esophago-gastroduodenoscopy and colonoscopy under general anesthesia (without intubation). Informed consent was obtained from all patients prior to enrollment, and each participant signed an informed consent form, receiving a full explanation of the study's purpose and nature.

[0025] Exclusion criteria are as follows: a history of allergic reactions to propofol or opioids, a history of opioid or propofol abuse, having received general anesthesia within the previous 30 days, requiring therapeutic procedures during gastrointestinal examinations, anesthesia lasting more than 60 minutes, incomplete colonoscopy due to inadequate bowel preparation, and unwillingness to participate in the study.

[0026] Prior to the gastroscopy and colonoscopy, a preoperative interview was conducted with the patient to collect basic information, including demographic information such as age, sex, height, and weight. The patient's history of chronic cardiovascular and cerebrovascular diseases, including hypertension, diabetes, cerebrovascular disease, hyperlipidemia, and coronary heart disease, was also inquired about. In addition, the interview also collected historical records of other factors that may affect the propofol dosage, including depression, anxiety, insomnia, recent smoking, and alcohol abuse. The gastroscopy and colonoscopy were performed consecutively, from 8:00 a.m. to 12:00 p.m. and from 1:00 p.m. to 5:00 p.m., with a one-hour lunch break from 12:00 p.m. to 1:00 p.m.

[0027] General anesthesia was administered by two anesthesiologists according to institutional procedures and current guidelines. Intravenous puncture was performed in the preparation room prior to the gastroscopy and colonoscopy, followed by infusion of lactated Ringer's solution. Dyclonine hydrochloride (10 ml: 0.1 g) was administered orally before the gastroscopy for mucosal lubrication and anesthesia. From the time the patient entered the operating room until the end of the procedure, heart rate (HR), heart rhythm, oxygen saturation (SpO2), noninvasive blood pressure (NIBP), and bispectral index (BIS) were continuously monitored. These parameters were recorded at five time points: before induction (T0), after the endoscope passed through the esophageal orifice (T1), after the endoscope passed through the cardia (T2), at the start of the colonoscopy (T3), and before leaving the operating room (T4). The patient received oxygen via nasal cannula at a flow rate of 5 L / min while in the left lateral decubitus position.

[0028] Propofol administration during gastrointestinal endoscopy is divided into two phases: an initial bolus dose during induction (i.e., the induction dose), followed by continuous infusion via a continuous infusion pump (i.e., the maintenance dose). In clinical practice, to determine the accurate dosage, propofol administration follows a pattern of starting with a relatively low dose within the guidelines and then gradually increasing to meet clinical needs. Specifically, during induction, propofol was titrated at an infusion rate of 800 mL / h until a hypnotic effect was achieved, while fentanyl was administered at 0.001 mg / kg. For maintenance, propofol was continuously infused at a rate of 4 to 5 mg / kg to maintain the bispectral index (BIS) between 40 and 65. Each time the patient exhibited a BIS ≥ 65 and / or signs such as a gag reflex, cough, or body movement, an additional dose of propofol was administered at a rate of 0.5 to 1 mg / kg. If two consecutive additional doses of propofol were required, the maintenance infusion rate was increased by 0.5 mg / kg / h. If the reflex could not be suppressed after three doses of propofol, an additional dose of fentanyl was administered, and the patient was excluded from the data analysis to eliminate the interference of fentanyl dose variations on propofol administration. If the BIS was < 40, the maintenance infusion rate of propofol was reduced by 0.5–1 mg / kg / h.

[0029] Of the 162 patients initially recruited, 146 were ultimately eligible for data analysis. The data were divided into four groups based on the time of anesthesia onset: Group 1 (8:00 - 10:00), Group 2 (10:00 - 12:00), Group 3 (13:00 - 15:00), and Group 4 (15:00 - 17:00).

[0030] All endoscopic procedures were performed by skilled endoscopists with at least five years of colonoscopy experience. Propofol infusion was stopped immediately when the colonoscope reached the ileum and was ready to be withdrawn. After the endoscope was removed from the anus, the patient's shoulder was gently tapped and verbally alerted to awakening. The awakening time was defined as the time interval between the cessation of propofol infusion and the patient's modified Aldrete score being >8. Once the modified Aldrete score reached 8, the patient was transferred to the post-anesthesia care unit for further recovery.

[0031] Adverse reactions are defined as oxygen saturation (SpO2) below 90%, heart rate (HR) below 50 bpm, systolic blood pressure (SBP) below 90 mmHg, or bispectral index (BIS) below 40. When SpO2 drops below 90%, interventions such as mandibular maneuver, oropharyngeal or nasopharyngeal airway placement, or mask ventilation are implemented according to operational requirements and patient condition. If heart rate drops below 50 bpm, atropine (0.5 mg) is administered. When systolic blood pressure drops below 90 mmHg, ephedrine (5-10 mg) is administered, and 200 ml of lactated Ringer's solution is supplemented. If an adverse reaction occurs again after the first intervention, the propofol maintenance infusion rate is reduced by 0.5-1 mg / kg / hour.

[0032] After organizing the collected patient basic information and propofol medication data, the following characteristics were identified: age, gender, weight, height, anesthesia start time, ASA classification, chronic diseases, other factors, duration, historical induction dose per unit body weight, historical maintenance dose per unit body weight per unit time, and total dosage. These characteristics were used as input variables, and the input variables were standardized or normalized.

[0033] Next, construct a neural network model, such as Figure 1 As shown, the neural network model includes an input layer, a shared feature extraction layer, a branching layer, and an output layer, specifically: The input layer is used to input preprocessed input variables; The input variables are further fed into the shared feature extraction layer to extract input features and calculate the importance weights of each variable. The shared feature extraction layer includes a fully connected layer, a batch normalization layer, an activation function, and an attention mechanism. The network structure is Dense(64) → BatchNorm → ReLU → Dense(32) → Dropout(0.2) → Attention. After the shared feature extraction layer, the branch layer includes an induced dose branch layer and a maintenance dose branch layer. Both the induced dose branch layer and the maintenance dose branch layer contain multi-layer fully connected networks. The weighted features are input to the induced dose branch layer and the maintenance dose branch layer, respectively. The induced dose branch layer performs induced dose prediction, and the maintenance dose branch layer performs maintenance dose prediction. The network structure of the induced dose branch is: Dense(32) → Dense(16) → output induced dose. The network structure of the maintenance dose branch is: Dense(32) → Dense(16) → output maintenance dose. To reflect the synergistic and restrictive relationship between the induction dose and the maintenance dose, an interactive constraint mechanism is introduced between the induction dose branch layer and the maintenance dose branch layer. This interactive constraint mechanism adopts a Cross-Stitch structure and incorporates a weighted multi-task loss function to linearly fuse the characteristics of the induction dose and maintenance dose. Cross-Stitch provides a flexible and learnable mechanism to model the intrinsic relationship between the two tasks of induction and maintenance doses, replacing fixed rules that rely solely on manual presets. Cross-Stitch can be understood as installing a "smart regulating valve" within the dosing prediction model; this "smart regulating valve" can automatically learn and coordinate the induction and maintenance doses.

[0034] The introduced weighted multi-task loss function consists of the following parts: ① mean squared error loss for induced dose prediction; ② mean squared error loss for maintenance dose prediction; ③ a regularization constraint term based on medical prior knowledge, used to maintain the balance between induced and maintenance doses. The expression for the weighted multi-task loss function is as follows:

[0035] In the formula: and These represent the actual and predicted values ​​of the induced amount per unit body weight, respectively. and These represent the actual and predicted values ​​of the maintenance duration per unit body weight per unit time, respectively. This represents the mean squared error, used to measure the regression accuracy for each task; The loss weight coefficients for different subtasks control the contribution of different subtasks and regularization terms to the total loss; This is a cross-task regularization term designed to constrain the dynamic balance between the inducing and maintaining variables.

[0036] Considering the "inverse relationship" adjustment logic between induction and maintenance doses in clinical practice, where a higher induction dose may lead to a lower maintenance dose, this regularization term can promote coordinated adjustment of induction and maintenance doses during model learning, preventing clinically unreasonable dose combinations and thus improving the medical usability and clinical interpretability of the model output.

[0037] The Cross-Stitch structure used in the interaction constraint mechanism is represented as follows: , ,

[0038] Where: parameter matrix Given a 2×2 trainable weight matrix, f induce and f maintain f′ represents the features of the induced and maintenance branches before the interaction, respectively. induce and f′ maintain These represent the features after being processed by the interaction constraint mechanism; α represents the proportion of the induced quantity branch features that retain their own features during interaction updates, 1-α represents the proportion obtained from the maintenance quantity branch features, β represents the proportion of the maintenance quantity branch features that retain their own features during interaction updates, and 1-β represents the proportion obtained from the induced quantity branch features.

[0039] The essence of the Cross-Stitch structure is linear combination, allowing the features of the induced and maintenance branches to influence each other, thus modeling the cooperative or restrictive relationship between the induced and maintenance quantities; parameter matrix The parameters α and β in A are used to control the degree of information sharing between two tasks. They are learnable and are continuously adjusted during training to optimize information flow. Introducing structural constraints can achieve dynamic coupling of information between the induction task and the maintenance task, reflecting the "ebb and flow" mechanism in clinical practice.

[0040] Finally, the predicted values ​​of induced weight per unit body weight and maintenance weight per unit time are output through the output layer.

[0041] Reference Figures 2-5 After the neural network model is built, the preprocessed input variables are randomly divided into training and test sets in a ratio of 8:2. The neural network model is trained using the training set. After training, the neural network model is further interpreted using the SHAP interpretability analysis module. The neural network model is then tested using the test set to obtain the final propofol dosing prediction model.

[0042] like Figure 3 , Figure 5As shown, in addition to displaying the loss function curve for the training set, the loss curve for the validation set and the performance metric (R² value as the accuracy metric) for the validation set are also plotted simultaneously. This not only provides a more intuitive reflection of the convergence of the drug delivery prediction model during training but also effectively demonstrates the generalization performance of the drug delivery prediction model on the validation set. Figure 3 , Figure 5 The horizontal axis represents the number of training epochs, and the vertical axis uses a dual-axis approach, representing the loss value (MSE) and the validation set R². The same graph clearly shows the changing trends of training loss, validation loss, and validation R², thus better illustrating the convergence and generalization ability of the drug delivery prediction model.

[0043] In this embodiment, the specific process of interpreting the neural network model using the SHAP interpretability analysis module is as follows: For each predicted value of induced weight per unit body weight and predicted value of maintenance weight per unit body weight per unit time, the SHAP value is used to calculate the marginal contribution ranking and visualization results of each input variable. The marginal contribution order of factors such as weight, age, and circadian rhythm in this prediction is shown through SHAP visualization (such as force graphs and bar charts); If the marginal contribution of certain features significantly conflicts with prior medical knowledge, it may indicate that the neural network model needs further optimization or human intervention. Example

[0044] This invention also provides a propofol dosing prediction system based on neural networks. The prediction system includes a data collection and processing module, a data input module, a calculation module, an output module, and a SHAP interpretability analysis module. The data collection and processing module collects basic patient information and propofol medication data, and uses the collected information and data as input variables. The input variables are preprocessed. The data input module inputs the preprocessed input variables into the propofol dosing prediction model. The calculation module uses the propofol dosing prediction model to calculate and linearly fuse the input variables, performing induction dose prediction and maintenance dose prediction respectively. The output module outputs the predicted values ​​of the induction dose per unit body weight and the maintenance dose per unit body weight per unit time. The SHAP interpretability analysis module calculates the marginal contribution of the input variables in each prediction result. Example

[0045] This invention also provides a propofol dosing prediction method based on a neural network. This prediction method is implemented based on the aforementioned propofol dosing prediction system based on a neural network, and includes the following steps: Basic information of individual patients in clinical practice is obtained as input variables. In this embodiment, the basic information includes age, gender, weight, height, anesthesia start time, ASA classification, chronic diseases, other factors, and duration. The input variables are fed into the propofol dosing prediction system. After input variable standardization or normalization preprocessing, feature extraction, branch prediction and interaction constraint calculation, the marginal contribution of the input variables to the prediction results is calculated by combining the SHAP algorithm. The system clearly indicates the positive or negative impact of a certain input variable in the patient's basic information (such as age, gender, anesthesia start time, etc.) on the prediction results. Finally, the system outputs the predicted value of the induction amount per unit weight and the predicted value of the maintenance amount per unit weight per unit time for individual patients.

[0046] Table 1. Dosing frequencies of propofol administered via conventional methods in four groups.

[0047] Table 2. Dosing frequencies of propofol administered using the dosing prediction model of this invention in four groups.

[0048] In Tables 1 and 2 above, the time period for Group 1 is 8:00-10:00; the time period for Group 2 is 10:00-12:00; the time period for Group 3 is 13:00-15:00; and the time period for Group 4 is 15:00-17:00.

[0049] Table 3 Evaluation of the output values ​​of the traditional model and the model of this invention

[0050] In Table 3, the evaluation indicators used are root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 As shown in Table 3, compared with the traditional linear regression model, the dosing prediction model of this invention outperforms the traditional linear regression model in terms of RMSE, MAE, and R2, demonstrating superior prediction accuracy. Figure 7 As shown, the frequency of patient movement and the number of times medication was administered were significantly reduced during gastrointestinal endoscopy.

[0051] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a propofol dosing prediction model based on neural networks, characterized in that, Includes the following steps: The collected patient basic information and propofol medication data were used as input variables, and the input variables were preprocessed. A neural network model with an interactive constraint mechanism and a weighted multi-task loss function was constructed to linearly fuse the induction dose and predicted dose of propofol administration. The preprocessed input variables were then input into the neural network model for training. The trained neural network model was interpreted using the SHAP interpretability analysis module, and then the neural network model was tested using preprocessed input variables to obtain the propofol dosing prediction model.

2. The method for constructing a propofol dosing prediction model based on a neural network according to claim 1, characterized in that, A neural network model consists of an input layer, a shared feature extraction layer, a branching layer, and an output layer, wherein: The input layer is used to input preprocessed input variables; The shared feature extraction layer is used to extract weighted features from the preprocessed input variables; The branch layer includes an induced dose branch layer and a maintenance dose branch layer. Both the induced dose branch layer and the maintenance dose branch layer contain multi-layer fully connected networks. The weighted features are input into the induced dose branch layer and the maintenance dose branch layer, respectively, for induced dose prediction and maintenance dose prediction. The output layer is used to output the predicted value of the induced amount per unit body weight and the predicted value of the maintenance amount per unit body weight per unit time, respectively.

3. The method for constructing a propofol dosing prediction model based on a neural network according to claim 2, characterized in that, An interaction constraint mechanism is introduced between the induced quantity branch layer and the maintenance quantity branch layer. This mechanism employs a Cross-Stitch structure to linearly fuse the weighted feature tasks input to the induced quantity and maintenance quantity branch layers, determining the synergistic and constraining relationships between the induced and maintenance quantities. The Cross-Stitch structure is expressed as follows: , , Where: parameter matrix Given a 2×2 trainable weight matrix, f induce and f maintain f′ represents the features of the induced and maintenance branches before the interaction, respectively. induce and f′ maintain These represent the features after being processed by the interaction constraint mechanism; α represents the proportion of the induced quantity branch features that retain their own features during interaction updates, 1-α represents the proportion obtained from the maintenance quantity branch features, β represents the proportion of the maintenance quantity branch features that retain their own features during interaction updates, and 1-β represents the proportion obtained from the induced quantity branch features.

4. The method for constructing a propofol dosing prediction model based on a neural network according to claim 3, characterized in that, A weighted multi-task loss function is introduced between the induced dose branch layer and the maintenance dose branch layer to linearly fuse the induced dose and maintenance dose characteristics, further determining the synergistic and restrictive relationships between the induced and maintenance doses. The expression of the weighted multi-task loss function is as follows: In the formula: and These represent the actual and predicted values ​​of the induced amount per unit body weight, respectively. and These represent the actual and predicted values ​​of the maintenance duration per unit body weight per unit time, respectively. Indicates mean square error; , , These are the loss weight coefficients for different subtasks; This is a cross-task regular expression.

5. A propofol dosing prediction system based on neural networks, characterized in that, The prediction system includes: The data collection and processing module collects patients' basic information and propofol medication data as input variables, and preprocesses the input variables. The data input module is used to input the preprocessed input variables into the propofol dosing prediction model established by the construction method of any one of claims 1-4; The solution module uses the propofol dosing prediction model to solve and linearly fuse the input variables, and performs induction dose prediction and maintenance dose prediction respectively. The output module outputs the predicted values ​​of induced weight per unit body weight and maintenance weight per unit time.

6. The propofol dosing prediction system based on a neural network according to claim 5, characterized in that, The prediction system also includes a SHAP interpretability analysis module, which is used to calculate the marginal contribution of the input variables in each prediction result.

7. A propofol dosing prediction method based on neural networks, characterized in that, This prediction method is based on a propofol dosing prediction system based on a neural network as described in any one of claims 5-6, and the prediction method includes the following steps: Basic information and propofol usage data of individual patients were obtained as input variables; Input variables are fed into the propofol dosing prediction system, and standardized or normalized preprocessing, feature extraction, branch prediction and interaction constraint calculation are performed. Finally, the predicted values ​​of induced dose per unit weight and maintenance dose per unit weight per unit time for individual patients are output. The marginal contribution ranking and visualization results of the input variables to the predicted values ​​of induced and sustained weight per unit of body weight per unit time were calculated using the SHAP algorithm.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program or instructions that cause a processor to execute the steps of the propofol dosing prediction method based on a neural network as described in claim 7.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores computer programs or instructions, which, when executed by the processor, implement the steps of the propofol dosing prediction method based on a neural network as described in claim 7.

Images