Clinical operation anesthesia administration decision-making system based on artificial intelligence

Abstract

The invention discloses a clinical operation anesthesia administration decision-making system based on artificial intelligence, which not only comprehensively considers the basic information of a patient, but also considers the real-time change condition of vital signs at the same time, and combines the past medical history of the patient and used vasoactive drugs to determine the clinical operation anesthesia administration decision-making system. The real-time doses of the propofol and the remifentanil are given, so that data reference is provided for anesthetists; therefore, an anesthetist can judge whether to accept the drug dosage decision according to the drug dosage decided by the system and the vital sign fluctuation condition made by the system according to the existing vital signs, the medication condition and the characteristics of a patient.

Description

Technical Field

[0001] The present invention relates to the field of clinical surgical auxiliary tools, and particularly to an artificial intelligence-based clinical surgical anesthesia drug administration decision-making system. Background Art

[0002] Anesthesia is a reversible functional inhibition of the central nervous system and / or peripheral nervous system produced by drugs or other methods. The main feature of this inhibition is the loss of sensation, especially pain, for surgical treatment without pain, while ensuring the life safety of patients during the perioperative period. This process must be completed by anesthesiologists. The anesthesia effect mainly depends on the work experience of anesthesiologists, and anesthesia work is directly related to the life safety of patients, making anesthesiologists in a highly tense state for a long time.

[0003] In order to relieve the work pressure of anesthesiologists, existing automatic decision-making systems for anesthetic drugs include:

[0004] ① The closed-loop system based on the pharmacokinetics / pharmacodynamics (PK / PD) model automatically calculates the infusion rate of propofol according to the patient's weight, age, and physiological status.

[0005] ② The deep RL anesthesia system: uses reinforcement learning algorithms to control the infusion of propofol.

[0006] ③ The McSleepy system: uses BIS as a feedback signal to automatically adjust the infusion of propofol and remifentanil.

[0007] The above existing research and development results either only consider the control of the injection rate of a single drug (propofol); or only use a single vital sign (such as BIS) as the decision basis for drug dosage use; and do not make personalized drug decisions based on the patient's past medical history, resulting in poor clinical effects. Summary of the Invention

[0008] Aiming at the above deficiencies in the prior art, the artificial intelligence-based clinical surgical anesthesia drug administration decision-making system provided by the present invention solves the problem of poor clinical effects of existing automatic decision-making systems for anesthetic drugs.

[0009] In order to achieve the above invention purpose, the technical solution adopted by the present invention is:

[0010] Provide an artificial intelligence-based clinical surgical anesthesia drug administration decision-making system, which includes:

[0011] A basic data acquisition module for collecting the patient's basic information, surgical ASA grade, vital sign change data, drug usage dose data, and past medical history; wherein the drug usage dose data includes anesthetic induction period drug usage dose data, anesthetic pump output drug data, and vasoactive drug usage dose data;

[0012] A data preprocessing module, which is used to complete and encode the data obtained by the basic data acquisition module to obtain encoded data corresponding to different types of data;

[0013] A data mapping module, which is used to splice the encoded data corresponding to the patient's basic information, the encoded data corresponding to the surgical ASA grade, the encoded data corresponding to the vital sign change data, the encoded data corresponding to the past medical history, the encoded data corresponding to the dosage data of vasoactive drugs, and the encoded data corresponding to the dosage data of drugs during the anesthesia induction period, and map the spliced encoded data to a high-dimensional hidden space through an encoder to obtain a state vector;

[0014] A data reprocessing module, which is used to map the encoded data corresponding to the drug data output by the anesthesia pump to the same dimension as the state vector obtained by the data encoding module and splice them to obtain a prediction vector;

[0015] A data prediction module, which is used to use the prediction vector as the input of the MLP model, and predict the patient's vital signs in the next time period and the drug data that the anesthesia pump needs to output in the next time period through the MLP model;

[0016] A data display module, which is used to display the patient's vital signs predicted by the MLP model and the drug data that the anesthesia pump needs to output in the next time period.

[0017] The beneficial effects of the present invention are as follows: This system not only comprehensively considers the patient's basic information (such as age, gender, height, weight, BMI value), but also takes into account the real-time changes in vital signs, and combines the patient's past medical history and the vasoactive drugs used to give the real-time dosages of two drugs, propofol and remifentanil, providing data reference for anesthesiologists, enabling anesthesiologists to make a judgment on whether to accept the drug dosage decision based on the drug dosage determined by this system and the vital sign fluctuations made by this system according to the existing vital signs, medication conditions and the patient's own characteristics. Brief Description of the Drawings

[0018] Figure 1 is the structural block diagram of this system;

[0019] Figure 2 is the schematic diagram of the clinical effect of this system. Detailed Embodiment

[0020] The specific embodiments of the present invention will be described below to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0021] As Figure 1 shown, the clinical surgical anesthesia drug administration decision-making system based on artificial intelligence includes:

[0022] A basic data acquisition module for collecting the patient's basic information, surgical ASA grade, vital sign change data, drug usage dose data, and past medical history; among which the drug usage dose data includes drug usage dose data during the anesthesia induction period, drug output data of the anesthesia pump, and vasoactive drug usage dose data;

[0023] A data preprocessing module for complementing and encoding the data obtained by the basic data acquisition module to obtain encoded data corresponding to different types of data;

[0024] A data mapping module for splicing the encoded data corresponding to the patient's basic information, the encoded data corresponding to the surgical ASA grade, the encoded data corresponding to the vital sign change data, the encoded data corresponding to the past medical history, the encoded data corresponding to the vasoactive drug usage dose data, and the encoded data corresponding to the drug usage dose data during the anesthesia induction period, and mapping the spliced encoded data to a high-dimensional hidden space through an encoder to obtain a state vector;

[0025] A data reprocessing module for mapping the encoded data corresponding to the drug output data of the anesthesia pump to the same dimension as the state vector obtained by the data encoding module and splicing them to obtain a prediction vector;

[0026] A data prediction module for taking the prediction vector as the input of the MLP model and predicting the patient's vital signs in the next time period and the drug data that the anesthesia pump needs to output in the next time period through the MLP model;

[0027] A data display module for displaying the patient's vital signs in the next time period predicted by the MLP model and the drug data that the anesthesia pump needs to output in the next time period.

[0028] The patient's basic information includes the current time, age, height, weight, gender, and BMI value; the vital signs include heart rate, electroencephalogram BIS, systolic blood pressure, diastolic blood pressure, and mean arterial pressure; the drugs during the anesthesia induction period include propofol and sufentanil; the vasoactive drugs include methoxamine, ephedrine, and norepinephrine; the past medical history includes heart disease, kidney disease, hypertension, and anemia; the drugs output by the anesthesia pump include propofol and remifentanil.

[0029] In the basic data acquisition module, record the corresponding data in minutes and integrate it into a csv file to form a file in the specified format; among them, the vital sign change data and drug dosage data are collected for at least 10 minutes; the vital sign change data is rounded to the nearest integer.

[0030] When recording the corresponding data, for the vital sign change data and vasoactive drug dosage data, perform the following operations:

[0031] Fill in the missing values according to the types recorded in the metadata file, and generate feature data and mask data; among them, the mask data is used to mark the missing feature data;

[0032] Perform one-hot encoding on the generated feature data, that is, change the feature values in the feature data into discrete binary vectors according to the types in the metadata;

[0033] Concatenate the one-hot encoded feature data and the mask data to complete the corresponding data recording;

[0034] When recording the corresponding data, for the drug dosage data, perform the following operations:

[0035] Extract the action features of propofol and sufentanil during anesthesia induction and the action features of propofol and remifentanil output by the anesthesia pump respectively, and fill in the missing values with the mean filling scheme to complete the corresponding data recording;

[0036] When recording the corresponding data, for the past medical history, perform the following operations:

[0037] If suffering from the corresponding disease, record the corresponding data value as 1, otherwise record the corresponding data value as 0 to complete the corresponding data recording.

[0038] In the specific implementation process, the surgical records of all patients are organized according to the unique ID of the patient, and there is one csv file for each patient's each operation. This file contains records in minutes. Each record has information with the features of [time, heart rate, systolic blood pressure, diastolic blood pressure, electroencephalogram signal, age, weight, height, gender, BMI value, mean arterial pressure, hypertension disease marker, heart disease marker, kidney disease marker, anemia marker, surgical ASA grade, ephedrine dose, methoxamine dose, norepinephrine dose, propofol dose for single administration during induction, sufentanil dose for single administration during induction].

[0039] The MLP model predicts the vital signs of the patient in the next time period as electroencephalogram BIS and mean arterial pressure.

[0040] In the data mapping module and the data reprocessing module, the mapping operations are both carried out by the cascaded encoder of the Transformer model.

[0041] In the data prediction module, the MLP model includes 4 multi-layer perceptrons, which are respectively used to predict the electroencephalogram BIS of the patient in the next time period, predict the mean arterial pressure of the patient in the next time period, decide the propofol dose that the anesthesia pump needs to output in the next time period, and decide the remifentanil dose that the anesthesia pump needs to output in the next time period.

[0042] When deciding the propofol dose that the anesthesia pump needs to output in the next time period and the remifentanil dose that the anesthesia pump needs to output in the next time period, the corresponding multi-layer perceptron will output the selection probability of discrete points with a unit of 1 in the range of [0, 100]. Each discrete point represents a dose. The finally output drug dose is to perform an argmax maximization on this probability in the range of [0, 100], that is, take the value with the maximum probability, and this value represents the action that will bring the maximum reward.

[0043] During the training process of this system, the MLP model also includes a multi-layer perceptron for predicting the value corresponding to the current state, and the system is trained based on the loss values of 5 multi-layer perceptrons; among them:

[0044] The expression of the loss function of the multi-layer perceptron for predicting the drug data that the anesthesia pump needs to output in the next time period is:

[0045]

[0046] Where is the loss value of the multi-layer perceptron for predicting the drug data that the anesthesia pump needs to output in the next time period; N is the total number of samples used in a single training; a is a reference parameter, bbf represents propofol, and rftn represents remifentanil; is the prediction result of the multi-layer perceptron for the i-th sample; is the true drug administration dose corresponding to the i-th sample; CE(.) represents the cross-entropy loss function;

[0047] The expression of the loss function of the multi-layer perceptron for predicting the electroencephalogram BIS of the patient in the next time period is:

[0048]

[0049] Where is the loss value of the multi-layer perceptron for predicting the electroencephalogram BIS of the patient in the next time period; is the prediction result of the multi-layer perceptron for predicting the electroencephalogram BIS of the patient in the next time period for the i-th sample; is the true electroencephalogram BIS data corresponding to the i-th sample;

[0050] The loss function expression of the multi - layer perceptron for predicting the mean arterial pressure of a patient in the next time period is:

[0051]

[0052] where is the loss value of the multi - layer perceptron for predicting the mean arterial pressure of a patient in the next time period; is the prediction result of the multi - layer perceptron for predicting the mean arterial pressure of a patient in the next time period for the \(i\) - th sample; is the true mean arterial pressure data corresponding to the \(i\) - th sample;

[0053] The loss function expression of the multi - layer perceptron for predicting the value corresponding to the current state is:

[0054]

[0055] where is the loss value of the multi - layer perceptron for predicting the value corresponding to the current state; is the prediction result of the multi - layer perceptron for predicting the value corresponding to the current state for the value corresponding to the \(i\) - th sample; is the true value corresponding to the \(i\) - th sample; where the prediction result of the value corresponding to the \(i\) - th sample is the sum of the reward values corresponding to the 1st sample to the \(i\) - th sample;

[0056] In this embodiment, and are both in the form of probability distributions, The expression of

[0057]

[0058] where \(P\) is the prediction object; \(dim\) is the parameter of the softmax function; \(\log\) is the logarithm to the base 10;

[0059] The expression of the total loss value is:

[0060]

[0061] where is the total loss value of the system during training.

[0062] The calculation expression of the reward value corresponding to any sample is:

[0063]

[0064] risk final = α1 * risk bis + α2 * riskmap

[0065] where risk bis is the reward value corresponding to the electroencephalogram BIS; clip(.) represents a truncation function, which truncates objects less than 0 to 0, truncates objects greater than 100 to 100, and does not process objects within the range of 0 to 100; c0, c1, and c2 are all constants; ln(.) is the natural logarithm; bis is the electroencephalogram BIS value; risk map is the reward value corresponding to the mean arterial pressure; map is the mean arterial pressure value; height 20 is 1.2 times the standard mean arterial pressure value; low 20 is 0.8 times the standard mean arterial pressure value; height 10 is 1.1 times the standard mean arterial pressure value; low 10 is 0.9 times the standard mean arterial pressure value; map = 0 indicates that there is no mean arterial pressure value, corresponding to the feature data marked by the masked data; α1 and α2 are weight parameters; risk final is the reward value at the corresponding moment.

[0066] When inputting data into this system for training, each csv file will be traversed and checked first to ensure that its length meets at least start_len (the minimum input in the model inference stage, set to 10 records in this embodiment) + pre_len (the result inferred by the model based on the input 10 records, set to 5 results in this embodiment). There must be at least 15 records in a csv file to be used as training data.

[0067] Secondly, we will determine a termination position end_index for each csv training data (and the length of this termination position is less than the length of the csv record to ensure no out-of-bounds), and end_index is generated by the random method of numpy from the interval (start_len, len_df - pre_len) (where len_df represents the length of a single csv file).

[0068] After that, enter the judgment logic. If the length of the dataframe of [0:end_index] is greater than max_len = 64, it will be directly truncated to ensure its length is 64. If the length is insufficient, the padding scheme will be used to fill it with all 0s, and a bool matrix named selected_padding will be generated synchronously, where the data filled with 0s manually by us is marked as True, and the unfilled part is marked as false.

[0069] Finally, directly concatenate the information of [end_index:end_index+pre_len] in the csv record behind this dataframe, which can be understood as the real record (labels_data) that our inference needs to refer to.

[0070] When applying this system, directly use this system after training, which only contains the first four multi-layer perceptrons, as Figure 2 shown, this system will automatically predict and display the patient's vital signs in the next time period and the drug data that the anesthesia pump needs to output in the next time period during clinical use.

[0071] In an embodiment of the present invention, during the process of training this system, the following processing is performed:

[0072] (1) Optimization in training:

[0073] ① Dual-stream Transformer architecture: Adopt the Encoder-Decoder structure to separate state representation learning and dynamic modeling. And perform multi-scale feature fusion, and integrate physiological state, drug administration actions and clinical decisions through feature concatenation operations.

[0074] ② Hybrid supervision:

[0075] Supervise the dose decisions of propofol and remifentanil through the physician's drug administration records, perform immediate supervision, and achieve the cloning of doctors' medication behaviors.

[0076] Use the cumulative reward data to constrain the value function prediction, perform delayed supervision, and achieve the policy gradient optimization based on the cumulative reward.

[0077] (2) Training loop:

[0078] Forward propagation: Input the historical observation sequence, and output the action policy, value estimation, reward prediction and physiological state prediction.

[0079] Loss calculation stage: Calculate the losses of each component through the loss functions listed in this embodiment.

[0080] Backward propagation: Calculate the parameter gradients.

[0081] Parameter update: Apply AdamW update after performing gradient clipping.

[0082] (3) Evaluation metrics:

[0083] Action decision: Mean absolute error (MAE) of dose selection.

[0084] Vital sign prediction: MAE between the prediction and the real vital signs.

[0085] (4) Implementation details:

[0086] The PyTorch Lightning framework is adopted to organize the training process. The training device is a 128-core CPU and a 4090 graphics card server. A total of 1000 epochs are trained, which takes about 3 hours.

[0087] (5) optim configuration:

[0088] Optimizer: AdamW algorithm, with an initial learning rate of 0.0001.

[0089] Batch_Size: 128.

[0090] Regularization: Gradient clipping (threshold = 1.0) is used to prevent gradient explosion.

[0091] In summary, the present invention trains the MLP model by selecting specific data, performing specific data processing, choosing specific loss functions and training methods, so that the trained system can fully express the states of patients of different ages, genders, suffering from different diseases, and using different types / doses of drugs, and can accurately evaluate the impact of drug doses on the vital signs of different patients to the greatest extent, providing data support for anesthesiologists.

Claims

1. A clinical surgical anesthesia medication decision-making system based on artificial intelligence, characterized in that: include: Basic data acquisition module, used to collect basic patient information, surgical ASA grade, vital signs change data, drug dosage data and past medical history; The drug dosage data include the drug dosage data during anesthesia induction, the drug output data of anesthesia pump and the vasoactive drug dosage data; A data preprocessing module is used to complete and encode the data acquired by the basic data acquisition module to obtain encoded data corresponding to different types of data; A data mapping module is used to concatenate the coded data corresponding to the patient's basic information, the coded data corresponding to the surgical ASA grade, the coded data corresponding to the vital signs change data, the coded data corresponding to the medical history, the coded data corresponding to the vasoactive drug dosage data, and the coded data corresponding to the anesthesia induction period drug dosage data, and map the concatenated coded data to a high-dimensional latent space through an encoder to obtain a state vector; A data reprocessing module is used to map the coded data corresponding to the drug data output by the anesthesia pump to the same dimension as the state vector obtained by the data encoding module and concatenate them to obtain a prediction vector; The data prediction module is used to use the prediction vector as the input of the MLP model, and predict the patient's vital signs in the next period and the drug data that the anesthesia pump needs to output in the next period through the MLP model; The data display module is used to display the patient's vital signs in the next period predicted by the MLP model and the drug data that the anesthesia pump needs to output in the next period.

2. The artificial intelligence-based clinical surgical anesthesia medication decision-making system according to claim 1 is characterized in that: The patient's basic information includes the current time, age, height, weight, gender and BMI value; vital signs include heart rate, EEG BIS, systolic blood pressure, diastolic blood pressure and mean arterial pressure; anesthesia induction drugs include propofol and sufentanil; vasoactive drugs include methoxamine, ephedrine and norepinephrine; past medical history includes heart disease, kidney disease, hypertension and anemia; anesthesia pump output drugs include propofol and remifentanil.

3. The artificial intelligence-based clinical surgical anesthesia medication decision-making system according to claim 2 is characterized in that: In the basic data acquisition module, the corresponding data are recorded in units of minutes and integrated into a csv file to form a file in a specified format; the vital signs change data and drug dosage data are collected for at least 10 minutes; the vital signs change data are rounded off.

4. The artificial intelligence-based clinical surgical anesthesia medication decision-making system according to claim 3 is characterized in that: When recording the corresponding data, for the data on changes in vital signs and the dosage of vasoactive drugs, perform the following operations: Fill missing values ​​according to the type of metadata file record, and generate feature data and mask data; the mask data is used to mark missing feature data; Perform one-hot encoding on the generated feature data, that is, change the feature values ​​in the feature data into discrete binary vectors according to the type in the metadata; The feature data processed by the one-hot code is concatenated with the mask data to complete the corresponding data recording; When recording the corresponding data, for the drug dosage data, perform the following operations: The motion characteristics of propofol and sufentanil during anesthesia induction, and the motion characteristics of propofol and remifentanil output by the anesthesia pump were extracted respectively, and the missing values ​​were filled with the mean filling scheme to complete the corresponding data records; When recording the corresponding data, perform the following operations for the past medical history: If the patient suffers from the corresponding disease, the corresponding data value is recorded as 1, otherwise the corresponding data value is recorded as 0, and the corresponding data recording is completed.

5. The artificial intelligence-based clinical surgical anesthesia drug administration decision system according to claim 4 is characterized in that: The MLP model predicts the patient's vital signs in the next period as EEG BIS and mean arterial pressure.

6. The artificial intelligence-based clinical surgical anesthesia medication decision-making system according to claim 5, characterized in that: In both the data mapping module and the data reprocessing module, the mapping operations are performed by the cascade encoders of the Transformer model.

7. The artificial intelligence-based clinical surgical anesthesia drug administration decision system according to claim 6, characterized in that: In the data prediction module, the MLP model includes 4 multi-layer perceptrons, which are used to predict the patient's EEG BIS in the next period, predict the patient's mean arterial pressure in the next period, decide the propofol dose that the anesthesia pump needs to output in the next period, and decide the remifentanil dose that the anesthesia pump needs to output in the next period.

8. The artificial intelligence-based clinical surgical anesthesia drug administration decision system according to claim 7, characterized in that: During the training process of this system, the MLP model also includes a multi-layer perceptron for predicting the value corresponding to the current state. The system is trained based on the loss values ​​of 5 multi-layer perceptrons; among them: The loss function expression of the multilayer perceptron used to predict the drug data that the anesthesia pump needs to output in the next period is: in is the loss value of the multilayer perceptron used to predict the drug data that the anesthesia pump needs to output in the next period; N is the total number of samples used in a single training; a is a reference parameter, bbf represents propofol, and rftn represents remifentanil; is the prediction result of the multilayer perceptron for the i-th sample; is the actual dosage corresponding to the i-th sample; CE(.) represents the cross entropy loss function; The loss function expression of the multilayer perceptron used to predict the patient's EEG BIS in the next period is: in is the loss value of the multilayer perceptron used to predict the patient's EEG BIS in the next period; is the prediction result of the multilayer perceptron for the i-th sample used to predict the patient's EEG BIS in the next period; is the real EEG BIS data corresponding to the i-th sample; The loss function expression of the multilayer perceptron used to predict the patient's mean arterial pressure in the next period is: in is the loss value of the multilayer perceptron used to predict the patient's mean arterial pressure in the next period; is the prediction result of the multilayer perceptron for the i-th sample used to predict the patient's mean arterial pressure in the next period; is the true mean arterial pressure data corresponding to the i-th sample; The loss function expression of the multilayer perceptron used to predict the value corresponding to the current state is: in is the loss value of the multilayer perceptron used to predict the value corresponding to the current state; is the prediction result of the multilayer perceptron used to predict the value corresponding to the current state for the value corresponding to the i-th sample; is the true value corresponding to the i-th sample; the predicted value corresponding to the i-th sample is the sum of the reward values ​​corresponding to the 1st sample to the i-th sample; The expression of the total loss value is: in is the total loss value of the system during the training process.

9. The artificial intelligence-based clinical surgical anesthesia drug administration decision system according to claim 8, characterized in that: The calculation expression of the reward value corresponding to any sample is: risk final =α1*risk bis +α2*risk map The risk bis is the reward value corresponding to the EEG BIS; clip(.) represents the clipping function, which clips objects less than 0 to 0, clips objects greater than 100 to 100, and does not process objects between 0 and 100; c0, c1, and c2 are all constants; ln(.) is the natural logarithm; bis is the EEG BIS value; risk map is the reward value corresponding to the mean arterial pressure; map is the mean arterial pressure value; height 20 1.2 times the standard mean arterial pressure; low 20 0.8 times the standard mean arterial pressure value; height 10 1.1 times the standard mean arterial pressure value; low 10 is 0.9 times the standard mean arterial pressure value; map = 0 means there is no mean arterial pressure value, corresponding to the feature data marked by the mask data; α1 and α2 are weight parameters; risk final is the reward value at the corresponding moment.

Images