Anesthesia recovery state determination method and device based on time-aware quantization large language model

Abstract

This invention relates to a method for determining anesthesia recovery status based on a time-aware quantitative large language model, comprising: collecting and preprocessing non-EEG physiological signals of the patient during the perioperative period; constructing temporal context information characterizing the dynamic changes in the anesthesia recovery process; obtaining and training a time-aware quantitative large language model; inputting a multi-dimensional input sequence into the trained time-aware quantitative large language model, and outputting the patient's current anesthesia recovery status determination result and the remaining awakening time prediction result. This invention effectively reduces clinical monitoring costs and improves the system's scalability; it significantly reduces the model parameter storage scale and computational overhead, enabling the model to be efficiently deployed on computationally limited clinical monitoring equipment or edge computing terminals, improving the system's real-time performance and application flexibility; and it achieves joint prediction of recovery status and remaining awakening time, further improving the accuracy and stability of anesthesia recovery monitoring.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to a method and device for determining the anesthesia recovery status based on a time-aware quantitative large language model. Background Technology

[0002] The recovery period from general anesthesia is one of the highest-risk and most critical stages in perioperative management. Real-time, accurate, and continuous monitoring of the patient's anesthetic recovery status is of significant clinical importance. Current clinical practice typically uses bispectral index monitoring based on electroencephalogram (EEG) signal analysis to assess anesthetic depth. While widely used, this method relies on specialized electrodes and complex signal acquisition equipment, resulting in high costs and susceptibility to interference from electromyography (EMG), electrosurgical noise, and electrode contact quality. This makes it difficult to widely apply in pre-hospital emergency care, patient transport, primary healthcare institutions, and resource-constrained environments.

[0003] With the development of wearable monitoring devices, researchers have gradually tried to use non-EEG physiological signals such as electrocardiogram (ECG) signals and photoplethysmography (PPG) signals to indirectly reflect the activity state of the autonomic nervous system, and to establish anesthesia status determination models with the help of machine learning or deep learning methods. Although such methods have reduced monitoring costs and improved the convenience of data acquisition to some extent, most existing technologies still rely on manually designed statistical features or short-term window analysis strategies, which are difficult to effectively capture the complex long-term temporal dependencies and dynamic evolution patterns in physiological signals. At the same time, because the anesthesia recovery process itself has significant time non-stationarity, traditional discriminant models usually simplify it to a static classification problem, which cannot accurately describe the continuous dynamic characteristics of physiological characteristics changing over time during the process from deep anesthesia to awakening.

[0004] In recent years, large language models have shown excellent performance in sequence modeling and context understanding, providing a new technical path for high-level semantic modeling of complex physiological signals. However, these models generally have problems such as large parameter scale, high computational resource requirements and large inference latency, making it difficult to deploy them directly in real-time medical monitoring systems or edge devices. In particular, their application is significantly limited in perioperative monitoring scenarios where timeliness and stability requirements are extremely high. Summary of the Invention

[0005] To address the problems of expensive monitoring equipment, high model computational complexity, and difficulty in depicting the dynamic process of resuscitation in existing technologies, the primary objective of this invention is to provide a time-aware quantitative large language model-based method for determining the state of anesthesia resuscitation. This method effectively reduces the cost of clinical monitoring and improves the scalability of the system, significantly reduces the storage scale of model parameters and computational overhead while ensuring the expressive power of the model, and improves the accuracy and stability of anesthesia resuscitation monitoring.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for determining the anesthesia recovery status based on a time-aware quantitative large language model, the method comprising the following sequential steps:

[0007] (1) Collect non-EEG physiological signals of patients during the perioperative period and preprocess them to obtain multi-channel continuous physiological time series data;

[0008] (2) Perform time window segmentation on multi-channel continuous physiological time series data to obtain multidimensional input sequences arranged in time order;

[0009] (3) Based on the multidimensional input sequence, construct temporal context information representing the dynamic changes in the anesthesia resuscitation process, and convert the temporal context information into a temporal embedding vector;

[0010] (4) Perform low-bit quantization on the pre-trained large language model to obtain a quantized large language model; introduce a time-aware gating mechanism module into the quantized large language model to obtain a time-aware quantized large language model; dynamically adjust the low-rank adapter weights of the time-aware quantized large language model through the time embedding vector; the dynamic adjustment includes scaling and feature selection control of the low-rank adapter weights.

[0011] (5) Input multi-channel continuous physiological time series data into the time-aware quantized large language model for training, and optimize the parameters of the time-aware quantized large language model by combining loss functions to obtain the trained time-aware quantized large language model.

[0012] (6) Input the multidimensional input sequence into the trained time-aware quantization large language model and output the current anesthesia recovery status of the patient and the prediction of the remaining awakening time.

[0013] In step (1), the non-EEG physiological signals include ECG signals, photoplethysmography (PPG) signals, and respiratory signals; the preprocessing includes filtering and noise reduction, normalization, and time synchronization alignment.

[0014] In step (3), the time context information includes the relative anesthesia end time, time location code, and stage status identifier; the time embedding vector is generated through a nonlinear mapping network.

[0015] In step (4), the low-bit quantization processing refers to determining the quantization parameters by minimizing the quantization error and maximizing the activation information entropy; the time-aware gating mechanism module includes a time embedding mapping unit, a gating coefficient generation unit, and an adapter weight adjustment unit;

[0016] The time embedding mapping unit is used to embed the time vector. Input to a nonlinear mapping network to generate temporal feature representations :

[0017] ;

[0018] in, and These represent the weight parameters and bias parameters of the nonlinear mapping network, respectively. Represents a non-linear activation function;

[0019] The gating coefficient generation unit represents based on time characteristics. Generation time gating coefficient :

[0020] ;

[0021] in, and These represent the weight parameters and bias parameters of the gated network, respectively. This represents the Sigmoid activation function;

[0022] The adapter weight adjustment unit adjusts according to the time gating coefficient. The weights of the low-rank adapter are dynamically adjusted to obtain the adjusted low-rank adapter weight matrix. :

[0023]

[0024] in, This represents the low-rank adapter weight matrix. This represents element-wise multiplication;

[0025] The time-aware gating mechanism module is set in the adapter layer of the quantized large language model and is used to dynamically adjust the weights of the low-rank adapter. The quantized large language model includes an input embedding layer, a multi-layer Transformer encoding layer and an output prediction layer. Each Transformer encoding layer includes a self-attention sub-layer and a feedforward neural network sub-layer. Low-rank adapters are introduced in both the self-attention sub-layer and the feedforward neural network sub-layer to achieve efficient parameter adjustment.

[0026] Step (5) specifically refers to: inputting the multidimensional input sequence into the input embedding layer of the time-aware quantized large language model for feature mapping to obtain the input feature representation. ;

[0027] Represent the input features Feature extraction is performed by sequentially inputting multiple Transformer encoding layers. Each Transformer encoding layer includes a self-attention sub-layer and a feedforward neural network sub-layer. The self-attention sub-layer is calculated as follows:

[0028] ;

[0029] in, , , These represent the query vector, key vector, and value vector, respectively. Indicates the dimension of a vector; This represents the computational process of the self-attention mechanism, used to calculate based on the query vector. With key vector The correlation between the values ​​of the vectors A weighted summation is performed to obtain a feature representation that includes global context information; ;

[0030] In the low-rank adapter of the Transformer encoding layer, a time-aware gating mechanism module is used to generate gating coefficients based on the time embedding vector and dynamically adjust the weight matrix of the low-rank adapter to enhance the ability of the time-aware quantized large language model to express the time features of different anesthesia recovery stages.

[0031] The combined loss function includes a resuscitation status classification loss function and a resuscitation time prediction error loss function. The resuscitation status classification loss function measures the difference between the model-predicted current anesthesia resuscitation status and the true label. This resuscitation status classification loss function is optimized using a cross-entropy loss function for the classification task. Let the probability distribution of the resuscitation status predicted by the model be:

[0032] ;

[0033] in, Indicates the number of recovery status categories; This indicates that the model predicts the sample belongs to the first... The probability of a class; This indicates that the model predicts the sample belongs to the first... The probability of a class;

[0034] Authentic Labels Represented as:

[0035]

[0036] in: When the sample belongs to the first When class, Otherwise ; Indicates the tag category index; Indicates the first The class's actual label indicates the variable;

[0037] Recovery status classification loss function for:

[0038] ;

[0039] The recovery time prediction error loss function Represented as:

[0040] ;

[0041] In the formula, For the first The actual recovery time for each sample; The recovery time predicted by the model; Sample size; recovery time prediction error loss function The mean squared error loss function is used.

[0042] Constructing the combined loss function :

[0043] ;

[0044] in: The weighting coefficients for the recovery status classification task are used to control the recovery status classification loss function. The degree of contribution to the overall optimization goal; The weighting coefficients for the recovery time prediction task are used to control the recovery time prediction error loss function. The degree of contribution to the overall optimization goal;

[0045] The feature representation after multi-layer Transformer encoding is input into the output prediction layer. The classification prediction head outputs the patient's current anesthesia recovery status, and the regression prediction head outputs the remaining time prediction result for the patient to reach a state of consciousness.

[0046] Another object of the present invention is to provide an electronic device comprising:

[0047] Processor; and

[0048] The memory stores computer program instructions that, when executed by the processor, cause the processor to perform the anesthesia resuscitation status determination method based on a time-aware quantitative large language model as described above.

[0049] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the anesthesia resuscitation status determination method based on a time-aware quantized large language model as described above.

[0050] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, by introducing a time-aware quantized large language model, the present invention models and analyzes the non-EEG physiological signals of perioperative patients, enabling continuous dynamic determination of the anesthesia recovery process without relying on EEG monitoring equipment. This reduces the dependence of traditional anesthesia depth monitoring systems on dedicated hardware equipment, thereby effectively reducing clinical monitoring costs and improving the system's scalability. Second, by performing low-bit quantization processing on the pre-trained large language model, the present invention significantly reduces the model parameter storage scale and computational overhead while ensuring the model's expressive power, enabling the model to operate on computationally limited clinical monitoring equipment or edge computing environments. The invention enables efficient deployment on computing terminals, improving the system's real-time performance and application flexibility. Third, by constructing temporal context information and introducing a time-aware gating mechanism, the model can dynamically adjust its weights based on time changes during the patient's anesthesia recovery process. This enhances the model's ability to model time-dependent relationships during the recovery phase and improves the accuracy of identifying changes in the recovery status. Compared to traditional methods based on single features or static models, this invention can more accurately depict the dynamic changes in physiological signals during the patient's transition from deep anesthesia to wakefulness, enabling joint prediction of the recovery status and remaining awakening time, further improving the accuracy and stability of anesthesia recovery monitoring. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0052] like Figure 1 As shown, a method for determining the anesthesia recovery status based on a time-aware quantitative large language model is presented. This method includes the following sequential steps:

[0053] (1) Collect non-EEG physiological signals of patients during the perioperative period and preprocess them to obtain multi-channel continuous physiological time series data;

[0054] (2) The continuous physiological time series data of multiple channels is divided into time windows to obtain a multidimensional input sequence arranged in time order; the continuous signal is divided into data segments of fixed length by using a sliding window method, and the data segments are arranged in time order to form a multidimensional input sequence to characterize the dynamic changes in the patient's anesthesia recovery process.

[0055] (3) Based on the multidimensional input sequence, construct temporal context information representing the dynamic changes in the anesthesia resuscitation process, and convert the temporal context information into a temporal embedding vector;

[0056] (4) Perform low-bit quantization on the pre-trained large language model to obtain a quantized large language model; introduce a time-aware gating mechanism module into the quantized large language model to obtain a time-aware quantized large language model; dynamically adjust the low-rank adapter weights of the time-aware quantized large language model through the time embedding vector; the dynamic adjustment includes scaling and feature selection control of the low-rank adapter weights.

[0057] (5) Input multi-channel continuous physiological time series data into the time-aware quantized large language model for training, and optimize the parameters of the time-aware quantized large language model by combining loss functions to obtain the trained time-aware quantized large language model.

[0058] (6) Input the multidimensional input sequence into the trained time-aware quantization large language model and output the current anesthesia recovery status of the patient and the prediction of the remaining awakening time.

[0059] In step (1), the non-EEG physiological signals include ECG signals, PEP pulse wave signals, and respiratory signals; the preprocessing includes filtering and noise reduction, normalization, and time synchronization alignment. Specifically, the patient's ECG signals, PEP pulse wave signals, and respiratory signals are simultaneously acquired using a multi-parameter monitoring device, and the signals are bandpass filtered to remove power frequency interference and high-frequency noise. Simultaneously, amplitude normalization and time alignment are performed to obtain multi-channel continuous physiological time-series data.

[0060] In step (3), the time context information includes the relative anesthesia end time, time location encoding, and stage status identifier; the time variables are converted into high-dimensional time embedding vectors through a nonlinear mapping network and injected into the model input feature space to enhance the model's ability to perceive dynamic time information. The time embedding vectors are generated through a nonlinear mapping network.

[0061] In step (4), the low-bit quantization process refers to determining the quantization parameters by minimizing the quantization error and maximizing the activation information entropy, so as to retain the feature expression ability of the original model as much as possible while reducing the model storage and computational complexity; the time-aware gating mechanism module includes a time embedding mapping unit, a gating coefficient generation unit, and an adapter weight adjustment unit;

[0062] The time embedding mapping unit is used to embed the time vector. Input to a nonlinear mapping network to generate temporal feature representations :

[0063] ;

[0064] in, and These represent the weight parameters and bias parameters of the nonlinear mapping network, respectively. Represents a non-linear activation function;

[0065] The gating coefficient generation unit represents based on time characteristics. Generation time gating coefficient :

[0066] ;

[0067] in, and These represent the weight parameters and bias parameters of the gated network, respectively. This represents the Sigmoid activation function. The Sigmoid function enables continuous and smooth gating control, making the adjustment of model parameters by time information differentiable, thus facilitating model training using gradient descent.

[0068] The adapter weight adjustment unit adjusts according to the time gating coefficient. The weights of the low-rank adapter are dynamically adjusted to obtain the adjusted low-rank adapter weight matrix. :

[0069]

[0070] in, This represents the low-rank adapter weight matrix. This represents element-wise multiplication;

[0071] The time-aware gating mechanism module is set in the adapter layer of the quantized large language model and is used to dynamically adjust the weights of the low-rank adapter. The quantized large language model includes an input embedding layer, a multi-layer Transformer encoding layer and an output prediction layer. Each Transformer encoding layer includes a self-attention sub-layer and a feedforward neural network sub-layer. Low-rank adapters are introduced in both the self-attention sub-layer and the feedforward neural network sub-layer to achieve efficient parameter adjustment.

[0072] Step (5) specifically refers to: inputting the multidimensional input sequence into the input embedding layer of the time-aware quantized large language model for feature mapping to obtain the input feature representation. ;

[0073] Represent the input features Feature extraction is performed by sequentially inputting multiple Transformer encoding layers. Each Transformer encoding layer includes a self-attention sub-layer and a feedforward neural network sub-layer. The self-attention sub-layer is calculated as follows:

[0074] ;

[0075] in, , , These represent the query vector, key vector, and value vector, respectively. Indicates the dimension of a vector; This represents the computational process of the self-attention mechanism, used to calculate based on the query vector. With key vector The correlation between the values ​​of the vectors Weighted summation is performed to obtain a feature representation that includes global context information; softmax represents the softmax normalization function, which maps any real number to a value between 0 and 1, and ensures that the sum of all outputs is 1, thus representing the importance weight of features at each location, used to adjust the value vector. Perform weighted calculations.

[0076] In the low-rank adapter of the Transformer encoding layer, a time-aware gating mechanism module is used to generate gating coefficients based on the time embedding vector and dynamically adjust the weight matrix of the low-rank adapter to enhance the ability of the time-aware quantized large language model to express the time features of different anesthesia recovery stages.

[0077] The combined loss function includes a resuscitation status classification loss function and a resuscitation time prediction error loss function. The resuscitation status classification loss function measures the difference between the model-predicted current anesthesia resuscitation status and the true label. In this invention, the anesthesia resuscitation status is typically divided into several discrete categories, such as deep anesthesia, transitional resuscitation, and conscious state. The resuscitation status classification loss function uses a cross-entropy loss function to optimize the classification task. Let the probability distribution of the resuscitation status predicted by the model be:

[0078] ;

[0079] in, Indicates the number of recovery status categories; This indicates that the model predicts the sample belongs to the first... The probability of a class; This indicates that the model predicts the sample belongs to the first... The probability of a class;

[0080] Authentic Labels Represented as:

[0081]

[0082] in: When the sample belongs to the first When class, Otherwise ; Indicates the tag category index; Indicates the first The class's actual label indicates the variable;

[0083] Recovery status classification loss function for:

[0084] ;

[0085] Recovery status classification loss function By minimizing the difference between the true class and the predicted probability, the model is able to more accurately identify the patient's current stage of anesthesia recovery.

[0086] The resuscitation time prediction error loss function measures the deviation between the model's predicted remaining time for the patient to reach consciousness and the actual time. In this invention, this task is a continuous numerical prediction problem, therefore, the mean squared error loss function is used for optimization. The resuscitation time prediction error loss function... Represented as:

[0087] ;

[0088] In the formula, For the first The actual recovery time for each sample; The recovery time predicted by the model; Sample size; recovery time prediction error loss function The mean squared error loss function is used.

[0089] Recovery time prediction error loss function By minimizing the squared error between the predicted time and the actual time, the model can more accurately predict the remaining time for the patient to reach a state of consciousness.

[0090] To simultaneously optimize the tasks of determining recovery status and predicting recovery time, this invention constructs a combined loss function. :

[0091] ;

[0092] in: The weighting coefficients for the recovery status classification task are used to control the recovery status classification loss function. The degree of contribution to the overall optimization goal; The weighting coefficients for the recovery time prediction task are used to control the recovery time prediction error loss function. The degree of contribution to the overall optimization goal;

[0093] when When the value is large, the model pays more attention to the accuracy of anesthesia recovery status classification; when When the value is large, the model focuses more on the accuracy of predicting the remaining time for the patient to reach consciousness. This is achieved by setting weighting coefficients. and It can balance the recovery status classification task and the recovery time prediction task, thereby achieving multi-task joint optimization and improving the overall prediction performance of the model.

[0094] By minimizing the above combined loss function The parameters of the time-aware quantification large language model are optimized to improve the accuracy of patient resuscitation time prediction while ensuring the accuracy of resuscitation status identification, thereby enhancing the overall effect of anesthesia resuscitation status assessment.

[0095] The feature representation after multi-layer Transformer encoding is input into the output prediction layer. The classification prediction head outputs the patient's current anesthesia recovery status, and the regression prediction head outputs the remaining time for the patient to reach a state of consciousness, thereby achieving continuous dynamic monitoring of the anesthesia recovery process.

[0096] In summary, this invention constructs a time-aware quantized large language model to model and analyze non-EEG physiological signals of perioperative patients. It enables continuous dynamic assessment of the anesthesia recovery process without relying on EEG monitoring equipment, reducing the dependence of traditional anesthesia depth monitoring systems on dedicated hardware, thereby effectively reducing clinical monitoring costs and improving system scalability. Furthermore, by performing low-bit quantization on the pre-trained large language model, this invention significantly reduces the model parameter storage size and computational overhead while maintaining model expressive power. This allows for efficient deployment on computationally limited clinical monitoring equipment or edge computing terminals, improving system real-time performance and application flexibility. By constructing temporal context information and introducing a time-aware gating mechanism module, this invention enables the model to dynamically adjust model weights based on time changes during the patient's anesthesia recovery process, thereby enhancing the model's ability to model time dependencies in the recovery phase and improving the accuracy of identifying changes in recovery status. Compared to traditional methods based on single features or static models, this invention can more accurately depict the dynamic changes in physiological signals during the transition from deep anesthesia to wakefulness, achieving joint prediction of recovery status and remaining awakening time, further improving the accuracy and stability of anesthesia recovery monitoring.

[0097] 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 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 claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for determining the anesthesia recovery status based on a time-aware quantitative large language model, characterized in that: The method includes the following steps in sequence: (1) Collect non-EEG physiological signals of patients during the perioperative period and preprocess them to obtain multi-channel continuous physiological time series data; (2) Perform time window segmentation on multi-channel continuous physiological time series data to obtain multidimensional input sequences arranged in time order; (3) Based on the multidimensional input sequence, construct temporal context information representing the dynamic changes in the anesthesia resuscitation process, and convert the temporal context information into a temporal embedding vector; (4) Perform low-bit quantization on the pre-trained large language model to obtain a quantized large language model; introduce a time-aware gating mechanism module into the quantized large language model to obtain a time-aware quantized large language model; dynamically adjust the low-rank adapter weights of the time-aware quantized large language model through the time embedding vector; the dynamic adjustment includes scaling and feature selection control of the low-rank adapter weights. (5) Input multi-channel continuous physiological time series data into the time-aware quantized large language model for training, and optimize the parameters of the time-aware quantized large language model by combining loss functions to obtain the trained time-aware quantized large language model. (6) Input the multidimensional input sequence into the trained time-aware quantization large language model and output the current anesthesia recovery status of the patient and the prediction of the remaining awakening time.

2. The method for determining anesthesia recovery status based on a time-aware quantitative large language model according to claim 1, characterized in that: In step (1), the non-EEG physiological signals include ECG signals, photoplethysmography (PPG) signals, and respiratory signals; the preprocessing includes filtering and noise reduction, normalization, and time synchronization alignment.

3. The method for determining anesthesia recovery status based on a time-aware quantitative large language model according to claim 1, characterized in that: In step (3), the time context information includes the relative anesthesia end time, time location code, and stage status identifier; the time embedding vector is generated through a nonlinear mapping network.

4. The method for determining anesthesia recovery status based on a time-aware quantitative large language model according to claim 1, characterized in that: In step (4), the low-bit quantization processing refers to determining the quantization parameters by minimizing the quantization error and maximizing the activation information entropy; the time-aware gating mechanism module includes a time embedding mapping unit, a gating coefficient generation unit, and an adapter weight adjustment unit; The time embedding mapping unit is used to embed the time vector. Input to a nonlinear mapping network to generate temporal feature representations : ； in, and These represent the weight parameters and bias parameters of the nonlinear mapping network, respectively. Represents a nonlinear activation function; The gating coefficient generation unit represents based on time characteristics. Generation time gating coefficient : ； in, and These represent the weight parameters and bias parameters of the gated network, respectively. This represents the Sigmoid activation function; The adapter weight adjustment unit adjusts according to the time gating coefficient. The weights of the low-rank adapter are dynamically adjusted to obtain the adjusted low-rank adapter weight matrix. : ； in, This represents the low-rank adapter weight matrix. This represents element-wise multiplication; The time-aware gating mechanism module is set in the adapter layer of the quantized large language model and is used to dynamically adjust the weights of the low-rank adapter. The quantized large language model includes an input embedding layer, a multi-layer Transformer encoding layer and an output prediction layer. Each Transformer encoding layer includes a self-attention sub-layer and a feedforward neural network sub-layer. Low-rank adapters are introduced in both the self-attention sub-layer and the feedforward neural network sub-layer to achieve efficient parameter adjustment.

5. The method for determining anesthesia recovery status based on a time-aware quantitative large language model according to claim 1, characterized in that: Step (5) specifically refers to: inputting the multidimensional input sequence into the input embedding layer of the time-aware quantized large language model for feature mapping to obtain the input feature representation. ; Represent the input features Feature extraction is performed by sequentially inputting multiple Transformer encoding layers. Each Transformer encoding layer includes a self-attention sub-layer and a feedforward neural network sub-layer. The self-attention sub-layer is calculated as follows: ； in, , , These represent the query vector, key vector, and value vector, respectively. Indicates the vector dimension; This represents the computational process of the self-attention mechanism, used to calculate based on the query vector. With key vector The correlation between the value vectors A weighted summation is performed to obtain a feature representation that includes global context information; ; In the low-rank adapter of the Transformer encoding layer, a time-aware gating mechanism module is used to generate gating coefficients based on the time embedding vector and dynamically adjust the weight matrix of the low-rank adapter to enhance the ability of the time-aware quantized large language model to express the time features of different anesthesia recovery stages. The combined loss function includes a resuscitation status classification loss function and a resuscitation time prediction error loss function. The resuscitation status classification loss function measures the difference between the model-predicted current anesthesia resuscitation status and the true label. This resuscitation status classification loss function is optimized using a cross-entropy loss function for the classification task. Let the probability distribution of the resuscitation status predicted by the model be: ； in, Indicates the number of recovery status categories; This indicates that the model predicts the sample belongs to the first... The probability of a class; This indicates that the model predicts the sample belongs to the first... The probability of a class; Authentic Labels Represented as: ； in: When the sample belongs to the first When class, Otherwise ; Indicates the tag category index; Indicates the first The class's actual label indicates the variable; Recovery state classification loss function for: ； The recovery time prediction error loss function Represented as: ； In the formula, For the first The actual recovery time for each sample; The recovery time predicted by the model; Sample size; recovery time prediction error loss function The mean squared error loss function is used. Constructing the combined loss function : ； in: The weighting coefficients for the recovery status classification task are used to control the recovery status classification loss function. The degree of contribution to the overall optimization goal; These are the weighting coefficients for the recovery time prediction task, used to control the recovery time prediction error loss function. The degree of contribution to the overall optimization goal; The feature representation after multi-layer Transformer encoding is input into the output prediction layer. The classification prediction head outputs the patient's current anesthesia recovery status, and the regression prediction head outputs the remaining time prediction result for the patient to reach a state of consciousness.

6. An electronic device, comprising: processor; as well as A memory storing computer program instructions, which, when executed by the processor, cause the processor to perform the anesthesia resuscitation status determination method based on a time-aware quantitative large language model as described in any one of claims 1-5.

7. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the anesthesia resuscitation status determination method based on a time-aware quantitative large language model as described in any one of claims 1-5.

Images