Method and system for constructing prediction model for postoperative cognitive decline of old patient

Abstract

The invention discloses a prediction model construction method and system for postoperative cognitive decline of an old patient. The method comprises the following steps: generating a preoperative feature matrix; evaluating a POCD preliminary risk probability by using a static risk prediction model; a variational auto-encoder is introduced to carry out modeling on the preoperative feature matrix, and personalized preoperative cognitive state distribution of a patient is generated by learning implicit distribution of the preoperative feature matrix, so that initial construction of the digital twin is completed; dynamically collecting cognitive scale scores, physiological indexes, voice features and behavior activity data at different postoperative time points, and generating a time sequence feature matrix; inputting the preoperative feature matrix and the time sequence feature matrix into a dynamic time sequence model, and updating the POCD risk probability and cognitive state distribution of the patient; and simulating a personalized cognitive recovery trajectory, predicting the time for the patient to recover to the baseline state, and outputting an optimal intervention strategy to reduce the POCD risk. According to the method, parameters can be finely adjusted in real time according to newly-added postoperative data, and a rapidly-updated prediction result is provided, so that the problem that the prediction result of an existing model is difficult to update in real time is solved.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a prediction model construction method and system for postoperative cognitive impairment in elderly patients. Background Art

[0002] With the rapid development of artificial intelligence technology, digital twin technology has gradually been introduced into the medical industry, providing innovative solutions for personalized diagnosis and treatment and dynamic health status prediction. In the study of postoperative cognitive impairment (POCD) in elderly patients, existing methods usually rely on risk assessment models based on preoperative static characteristics, such as traditional machine learning algorithms based on cognitive scale scores (such as MoCA, MMSE), inflammatory factors, brain imaging features, etc. These models can provide preliminary risk predictions in the preoperative stage by integrating preoperative physiological and cognitive characteristics, providing a reference for the optimization of surgical plans. However, they have difficulty capturing the dynamic changes in the cognitive state of patients after surgery, ignoring the nonlinear evolution of time series features and the complex dependencies of multimodal data, thus limiting the comprehensive evaluation of the postoperative recovery process.

[0003] In recent years, with the maturity of time series modeling technology, it has shown good performance in dealing with dynamic changes in cognitive scores and trends in physiological indicators. These technologies are usually limited to the modeling of single modal features and fail to effectively combine multimodal data, resulting in the lack of multi-dimensional information expression in the model when describing the postoperative recovery process. In addition, these models usually only focus on the predictive ability of time series, while ignoring the interpretability and operability of the prediction results in clinical applications, especially in formulating personalized intervention paths for patients. For example, most current risk prediction models can only provide static risk scores, cannot generate the recovery trajectory of patients' future cognitive status, and cannot simulate the specific effects of different intervention strategies on the recovery process.

[0004] More importantly, existing models often show lag and non-adaptability when processing new postoperative data, making it difficult to update prediction results in real time to reflect the latest status of the patient's postoperative dynamic changes. This lag makes it impossible for the model to provide timely and effective reference for early postoperative interventions, which in turn affects the patient's long-term recovery. Therefore, how to build a new POCD prediction model to accurately predict POCD risk and generate personalized recovery trajectories has become an important technical challenge that needs to be solved in the current field. Summary of the invention

[0005] In order to solve the problem that the existing model exhibits lag and non-adaptivity when processing new postoperative data, and is difficult to update the prediction results in real time, the present invention provides a method and system for constructing a prediction model for postoperative cognitive decline in elderly patients.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] In one aspect, the present invention discloses a method for constructing a prediction model for postoperative cognitive decline in elderly patients, comprising the following steps:

[0008] Collect preoperative cognitive scale scores, physiological indicators, imaging features, and multimodal data of speech features to generate the preoperative feature matrix X Pre ;

[0009] Using static risk prediction model to estimate the initial risk probability of POCD Introduce the variational autoencoder to the preoperative feature matrix X Pre Modeling is performed by learning the preoperative feature matrix X Pre The implicit distribution of the patient's preoperative cognitive state is generated to complete the initial construction of the digital twin.

[0010] Dynamically collect cognitive scale scores, physiological indicators, voice characteristics, and behavioral activity data at different time points after surgery to generate a time series feature matrix X Dynamic ;

[0011] The preoperative feature matrix X Pre and the time series feature matrix X Dynamic The dynamic time series model is input to update the patient’s POCD risk probability and cognitive status distribution;

[0012] Simulate personalized cognitive recovery trajectories, predict the time it takes for patients to return to baseline, and output optimal intervention strategies to reduce the risk of POCD.

[0013] Further: the preoperative cognitive scale score, physiological index, imaging features and voice feature multimodal data are collected to generate a preoperative feature matrix X Pre include:

[0014] The cognitive scale scores include the total score and sub-scores of the Montreal Cognitive Assessment Scale and the Mini-Mental State Examination Scale;

[0015] The physiological indicators include inflammatory factors, oxidative stress markers (MDA), and the age and gender of the patient;

[0016] The imaging features include hippocampal volume, gray matter volume, and white matter lesion volume extracted from brain MRI;

[0017] The speech features include speech rate, language pause time, and vocabulary diversity indicators extracted from speech recordings of cognitive tests.

[0018] Further: The use of a static risk prediction model to assess the initial risk probability of POCD Introduce the variational autoencoder to the preoperative feature matrix X Pre Modeling is performed by learning the preoperative feature matrix X Pre The implicit distribution of the patient's preoperative cognitive state distribution is generated to complete the initial construction of the digital twin, including:

[0019] Using static risk prediction model to estimate the initial risk probability of POCD

[0020] The variational autoencoder includes an encoder, a latent variable distribution, and a decoder, where the encoder transforms the preoperative feature matrix X Pre Mapped to the mean μ and variance σ of the latent variable 2 ;

[0021] Use the reparameterization technique to sample the latent variable z;

[0022] The decoder is reconstructed by the hidden variable z The latent variable z is stored in the digital twin as a personalized representation of the patient’s preoperative cognitive state:

[0023] Store POCD preliminary risk probability and hidden variable z as the initial state of the digital twin.

[0024] Further: the preoperative feature matrix X Pre and the time series feature matrix X Dynamic Input the dynamic time series model to update the patient's POCD risk probability and cognitive status distribution including:

[0025] Use positional encoding to add temporal information to input features Among them, X t is the feature vector at the tth time point, PE(t) is the position information at the tth time point; the input data containing time information Input into the self-attention mechanism of the Transformer model to obtain H t , where H t It is the output representation of the Transformer model at each time point;

[0026] Use the fully connected layer to transform the H output of Transformer t Mapped to POCD risk probability P POCD (t), and obtain the risk probability P that changes with time POCD (t);

[0027] Use a recurrent neural network to update the patient's cognitive state distribution z(t).

[0028] Further: The simulated personalized cognitive recovery trajectory predicts the time for the patient to return to the baseline state, and outputs the best intervention strategy to reduce the risk of POCD, including:

[0029] Use a recurrent neural network to update the patient's cognitive state distribution z(t) to represent the patient's personalized cognitive state, which serves as the basis for generating a personalized recovery trajectory;

[0030] Generate cognitive recovery trajectories at future time points through generative AI models Simulate the patient's recovery process;

[0031] Based on the generated cognitive recovery trajectory Predict the time it takes for patients to return to their preoperative baseline status;

[0032] Define states t is the patient's cognitive state at time t, action a t For intervention measures, reward r t For comprehensive evaluation of recovery progress and risk reduction:

[0033]

[0034] Among them, ω1 and ω2 are weight coefficients, balancing recovery progress and risk reduction, C(t) represents the cognitive score at time point t, Preoperative baseline cognitive score, P POCD (t) is the POCD risk probability at time t;

[0035] Use DNQ to train the agent model and learn different intervention strategies t Impact on patient recovery:

[0036]

[0037] Q(s t , a t ) is in state s t Take action a t The expected return, α is the learning rate, and γ is the discount factor;

[0038] Through reinforcement learning optimization, the optimal intervention strategy a* is selected to minimize POCD risk and maximize cognitive recovery:

[0039]

[0040] On the other hand, the present invention also discloses a prediction model construction system for postoperative cognitive decline in elderly patients, comprising:

[0041] Preoperative data collection and multimodal feature extraction module: collect preoperative cognitive scale scores, physiological indicators, imaging features, and multimodal data of speech features to generate the preoperative feature matrix X Pre ;

[0042] Preoperative preliminary digital twin building block: using static risk prediction model to assess the preliminary risk probability of POCD Introduce the variational autoencoder to the preoperative feature matrix X Pre Modeling is performed by learning the preoperative feature matrix X Pre The implicit distribution of the patient's preoperative cognitive state is generated to complete the initial construction of the digital twin.

[0043] Postoperative multi-time point data collection and feature generation module: Dynamically collect cognitive scale scores, physiological indicators, voice features and behavioral activity data at different time points after surgery to generate a time series feature matrix X Dynamic ;

[0044] Postoperative dynamic twin update and adaptive learning module: the preoperative feature matrix X Pre and the time series feature matrix X Dynamic The dynamic time series model is input to update the patient’s POCD risk probability and cognitive status distribution;

[0045] Generate personalized recovery trajectory and optimize path module: simulate personalized cognitive recovery trajectory, predict the time it takes for patients to return to baseline state, and output the best intervention strategy to reduce the risk of POCD.

[0046] Compared with the prior art, the present invention has the following technical advances:

[0047] The solution of the present invention integrates preoperative and postoperative multimodal data, including cognitive scale scores, physiological indicators, imaging features, voice features, and behavioral activity data. The comprehensive integration of these data significantly improves the model's ability to describe the patient's preoperative baseline state and dynamic changes in postoperative recovery. The Transformer model is used to dynamically model postoperative time series data. By capturing the short-term dependencies and long-term trends of time series features, the patient's POCD risk probability and cognitive state distribution are updated in real time, solving the problem that traditional static risk prediction models cannot dynamically capture postoperative changes in patients. At the same time, the introduction of incremental learning mechanisms ensures that the model can fine-tune parameters in real time according to newly added postoperative data, providing rapidly updated prediction results, thereby overcoming the problem that existing models are difficult to update prediction results in real time, greatly improving clinical applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0049] In the attached picture:

[0050] Figure 1 It is a flow chart of the present invention.

[0051] Figure 2 It is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0052] The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0053] Example 1

[0054] like Figure 1 As shown, the present invention discloses a method for constructing a prediction model for postoperative cognitive decline in elderly patients, comprising:

[0055] Step 1: Collect preoperative cognitive scale scores, physiological indicators, imaging features, and multimodal data of speech features to generate the preoperative feature matrix X Pre ;

[0056] Step 2: Use a static risk prediction model to estimate the initial risk probability of POCD Introduce the variational autoencoder to the preoperative feature matrix X Pre Modeling is performed by learning the preoperative feature matrix X Pre The implicit distribution of the patient's preoperative cognitive state is generated to complete the initial construction of the digital twin.

[0057] Step 3: Dynamically collect cognitive scale scores, physiological indicators, voice features, and behavioral activity data at different time points after surgery to generate a time series feature matrix X Dynamic ;

[0058] Step 4: Convert the preoperative feature matrix X Pre and the time series feature matrix X Dynamic The dynamic time series model is input to update the patient’s POCD risk probability and cognitive status distribution;

[0059] Step 5: Simulate a personalized cognitive recovery trajectory, predict the time it takes for the patient to return to baseline, and output the best intervention strategy to reduce the risk of POCD.

[0060] Specifically, in step 1, the purpose is to generate a preoperative feature matrix X by collecting multimodal preoperative data Pre, providing rich and structured input for subsequent model building.

[0061] Step 1.1: Extraction of cognitive scale scores

[0062] 1. Collect data:

[0063] The preoperative cognitive scale scores were collected, including the total score and sub-score of MoCA and MMSE, respectively. MoCA and C MMSE Indicates: C MoCA =[M Attention , M Memory , M Executive , M Lantguage ]、C MMSE =[S Orientation , S Memory , S Lantguage , S Calculation ], where each sub-item represents the score of a certain functional area in the scale.

[0064] 2. Structured Representation:

[0065] The sub-item scores of MoCA and MMSE were combined to form a cognitive feature matrix: C = [C MoCA ,C MMSE ].

[0066] Step 1.2: Extraction of physiological indicators

[0067] 1. Collect data:

[0068] The patient's preoperative physiological indicators were recorded, including inflammatory factors (IL-6, CRP), oxidative stress markers (MDA), and the patient's age, gender and other characteristics, which were represented by vector R: R = [R IL6 , R CRP , R MAD , R Age , R Cender ]

[0069] 2. Data Standardization:

[0070] Normalize the numerical features to eliminate dimensionality differences: Among them, μ R and σ R are the mean and variance of the physiological indicators, respectively.

[0071] Step 1.3: Extraction of imaging features

[0072] 1. Collect data:

[0073] Extract key brain features from patients’ preoperative MRI or CT images, including hippocampal volume (VHippocampus ), gray matter volume (V GrayMatter ) and white matter lesion volume (V WhiteMatter ).

[0074] 2. Feature structuring:

[0075] The above features are directly expressed as image feature vectors, which are convenient for merging with other modality data. Hippocampus , V GrayMatter , V WhiteMatter ].

[0076] Step 1.4: Extraction of speech features

[0077] The speech features, including speech rate (S speechRate ), pause time (S PauseTime ), language fluency (S Fluency ) and lexical diversity (S LexicalDiversity ).

[0078] S=[S speechRate , S PauseTime ,S Fluency ,S LexicalDiversity ]

[0079] The formula for calculating speech speed is:

[0080] The formula for calculating the pause time ratio is:

[0081] Step 1.5: Integration of feature matrix

[0082] 1. Multimodal data merging:

[0083] Combine cognitive features C, physiological indicators R, imaging features F, and speech features S to form the preoperative feature matrix:

[0084] X Pre =[C, R, F, S]

[0085] 2. Output preoperative feature matrix:

[0086] The final preoperative feature matrix is ​​structurally represented as:

[0087]

[0088] 1. Each modality data (cognitive score, physiological index, imaging features, speech features) is collected and processed to generate an independent feature vector (such as C, R, F, S).

[0089] 2. After the data of each modality is standardized, they are merged and interactively generated to form a unified preoperative feature matrix X Pre .

[0090] 3. The structuring of the feature matrix provides comprehensive and high-quality input for subsequent risk prediction and twin modeling.

[0091] In this embodiment, MoCA (Montreal Cognitive Assessment) and MMSE (Mini-Mental State Examination) are two commonly used cognitive function assessment tools, which assess the overall cognitive status of patients by scoring multiple cognitive domains of patients. Collecting the total score and sub-score of MoCA and MMSE means:

[0092] Total score: Provides a quantitative indicator of overall cognitive function by adding up the scores of each sub-item.

[0093] Item scoring: records the performance of specific dimensions of cognitive function (such as attention, memory, language ability, etc.), and deeply analyzes the patient's specific problems from multiple cognitive areas. MoCA is more sensitive to the detection of mild cognitive impairment (MCI), especially in the early stages, when MoCA is more likely to identify subtle cognitive dysfunction.

[0094] MMSE is mainly used to screen for moderate to severe cognitive impairment (such as late-stage Alzheimer's disease) and has a strong ability to assess the overall cognitive state. MoCA includes assessments of more advanced cognitive areas, such as executive function and abstract thinking ability, and is suitable for detecting cognitive deficits in complex tasks. MMSE focuses more on the assessment of attention, memory, and language abilities, and the questions are relatively simple, making it suitable for rapid screening in a shorter period of time.

[0095] MoCA is often used to detect mild cognitive impairment (MCI), which is one of the early manifestations of POCD. Therefore, it is very valuable for the early prediction of postoperative cognitive decline. MMSE provides a more stable comprehensive cognitive status assessment, which can make up for the shortcomings of MoCA in screening deep cognitive deficits. The combination of MoCA and MMSE can more comprehensively capture the overall state and detailed characteristics of patients' preoperative cognitive function and improve the accuracy of POCD risk prediction. When generating the preoperative feature matrix, the sub-scores of MoCA and MMSE can be used as different feature inputs to provide multi-dimensional information and support personalized twin modeling and the construction of dynamic prediction models. The combination of sensitivity to mild cognitive impairment (MoCA) and comprehensive coverage of moderate and severe disorders (MMSE) can adapt to patient groups with different cognitive states.

[0096] Specifically, in step 2, the purpose of step 2 is to use a static risk prediction model to predict the initial risk probability of POCD. Evaluation is performed and a variational autoencoder (VAE) is introduced to the preoperative feature matrix X Pre Modeling is performed to learn the implicit distribution of patients’ preoperative characteristics, thereby generating personalized cognitive state distribution and completing the initial construction of the digital twin.

[0097] Step 2.1: Static risk prediction model to assess the initial risk probability of POCD

[0098] 1. Input preoperative feature matrix:

[0099] Obtain the preoperative feature matrix X from step 1 Pre , where each row represents the preoperative multimodal features of a patient:

[0100] X Pre =[C, R, F, S]

[0101] 2. Build a static risk prediction model:

[0102] Use XGBoost model to train X Pre After training, the model outputs the patient's preliminary risk probability of POCD before surgery:

[0103]

[0104] Among them, f static :Static risk prediction model, is the probability of the patient developing POCD before surgery.

[0105] Step 2.2: Introduce variational autoencoder (VAE) to learn preoperative feature distribution

[0106] 1. Model input:

[0107] Using the preoperative feature matrix X Pre As input to the VAE, the latent distribution of patients’ preoperative characteristics is learned.

[0108] 2. Coding stage:

[0109] X Pre Mapped to latent variable space, generating potential distribution parameters (mean μ and variance σ 2 ):

[0110] μ,logσ 2 =f encoder (X Pre )

[0111] f encoder represents the encoder neural network of VAE; μ is the mean of the latent variable distribution, which represents the central position of the patient's preoperative cognitive state; σ 2 The variance of the latent variable distribution represents the uncertainty of the patient's status.

[0112] 3. Hidden variable sampling:

[0113] Use the reparameterization trick to sample from a latent variable distribution:

[0114] z=μ+σ⊙∈,∈~N(O,I)

[0115] z represents the latent variable representing the patient's personalized preoperative cognitive state, and ⊙ represents the multiplication of elements one by one.

[0116] 4. Decoding stage:

[0117] The latent variable z is restored to the original feature space through the decoder to generate the reconstructed preoperative feature matrix

[0118]

[0119] Among them, f decoder is the decoder neural network of VAE.

[0120] Step 2.3: Optimize the VAE model to generate personalized cognitive state distribution

[0121] 1. Reconstruction loss:

[0122] Measure the original preoperative characteristics X Pre With reconstruction features The differences:

[0123]

[0124] 2. KL divergence loss:

[0125] Ensure that the latent variable distribution N(μ, σ 2 ), close to the standard normal distribution N(0,I):

[0126]

[0127] Where N(0, I) is a multivariate normal distribution with a mean vector of 0, and the mean of all variables is 0. I is the covariance matrix, the variance of each variable is 1 (the value on the main diagonal), and the variables are independent of each other (the off-diagonal elements are 0). is the variance of the encoding distribution, μ i is the mean.

[0128] 3. Total loss function:

[0129] By jointly optimizing the reconstruction loss and the KL divergence loss, the implicit distribution of the patient’s preoperative features is learned:

[0130] L VAE =L reconstruction+β·L KL

[0131] β adjusts the weight of the two parts of the loss.

[0132] Step 2.4: Build a preliminary digital twin

[0133] 1. Personalized cognitive state distribution:

[0134] The patient's personalized cognitive state distribution is represented by the latent variable z:

[0135] z~N(μ,σ 2 )

[0136] 2. Preliminary Digital Twin:

[0137] The patient's initial risk of POCD before surgery The latent variable z is stored in the digital twin to form the patient’s personalized preoperative state.

[0138] Specifically, in step 3, the purpose of step 3 is to generate a time series feature matrix X by dynamically collecting cognitive scale scores, physiological indicators, voice features, and behavioral activity data at different time points after surgery. Dynamic , providing input for subsequent dynamic model updates and POCD risk prediction.

[0139] Step 3.1: Dynamic collection of postoperative multimodal data

[0140] 1. Time point definition:

[0141] The time point set for postoperative data collection is set to T = {t1, t2, ..., t T}, for example, the 1st day, 7th day, and 30th day after surgery.

[0142] 2. Collect cognitive scale scores:

[0143] At each time point t, the total scores and sub-item scores of MoCA and MMSE were recorded and expressed as:

[0144] C(t)=[C MoCA (t), C MMSE (t)]

[0145] 3. Collect physiological indicators:

[0146] At each time point t, inflammatory factors (such as IL-6 and CRP) and other key physiological indicators (such as blood oxygen saturation SpO2) are recorded to form a physiological indicator vector:

[0147]

[0148] 4. Collect voice features:

[0149] Extract speech features (such as speech rate S) from speech recordings of postoperative cognitive tests SpeechRate , pause time S PauseTime , Fluency S Fluency ), expressed as:

[0150] S(t)=[S SpeechRate (t), S PauseTime (t), S Fluency (t)]

[0151] 5. Collecting behavioral activity data:

[0152] Recording the average number of steps per day using wearable devices Steps and sleep duration B Sleep , which constitutes the characteristics of behavioral activities:

[0153] B(t)=[B Steps (t), B Sleep (t)]

[0154] Step 3.2: Feature integration

[0155] 1. Single time point feature representation:

[0156] The characteristics of each time point t are integrated by cognitive scores, physiological indicators, speech characteristics and behavioral activity characteristics:

[0157] X t =[C(t), R(t), S(t), B(t)]

[0158] Step 3.3: Generate time series feature matrix

[0159] 1. Time series construction:

[0160] According to the time point set T, each time point X t The features of are stacked to form a time series feature matrix:

[0161]

[0162] Specifically, in step 4, the purpose of step 4 is to transform the preoperative feature matrix X Pre and the postoperative time series feature matrix X Dynamic Input a dynamic time series model (such as Transformer or LSTM) and update the patient’s POCD risk probability P through the model POCD (t) and cognitive state distribution z(t), and adaptively adjust the model parameters through an incremental learning mechanism.

[0163] Step 4.1: Input data preparation

[0164] 1. Input feature matrix:

[0165] The preoperative characteristic moment X Pre and the postoperative time series feature matrix X Dynamic Combine or separate processing:

[0166] X input =[X Pre ,X Dynamic ]

[0167] 2. Time series length and dimension:

[0168] Assuming there are T time points and the feature dimension of each time point is D, the shape of the input feature matrix is ​​[T, D], where T is the number of time points and D is the feature dimension of each time point.

[0169] Step 4.2: Dynamic time series model selection and input processing

[0170] 1. Select the model:

[0171] Use Transformer to process the data through the following steps.

[0172] 2. Positional encoding:

[0173] Since the Transformer model does not have built-in time information, it is necessary to use positional encoding to add time information to the input features: X t : The feature vector at the tth time point; PE(t): The position information at the tth time point.

[0174] 3. Input to Transformer:

[0175] Input data containing time information Input into Transformer's self-attention mechanism: H t It is the output representation of Transformer at each time point.

[0176] Step 4.3: Update POCD risk probability

[0177] 1. POCD risk probability mapping:

[0178] Use the fully connected layer to transform the H output of Transformer t Mapped to POCD risk probability P POCD (t):

[0179] P POCD (t) = σ(W POCD ·Ht +b POCD )

[0180] σ is the Sigmoid function, which compresses the output value to the interval [0, 1], indicating the POCD risk probability, W POcD and b POCD are the weights and biases of the fully connected layers.

[0181] 2. Dynamic risk updates:

[0182] The POCD risk probability P at each time point POCD (t) will change over time. For example, the risk probability at day t will be updated based on the preoperative characteristics and data from all previous time points.

[0183] Step 4.4: Update the cognitive state distribution

[0184] 1. Hidden variable update (GRU or LSTM):

[0185] To capture the implicit state in the time series, we use a recurrent neural network (such as GRU or LSTM) to update the patient’s cognitive state distribution z(t) = GRU(z(t-1), H t ), z(t) represents the distribution of cognitive status at the tth time point, represents the patient's cognitive function status, and z(t-1) represents the distribution of cognitive status at the previous time point.

[0186] 2. The implicit state of hidden variables:

[0187] z(t) represents the patient’s cognitive state on day t, and after being updated by GRU, it will serve as the updated state of the digital twin.

[0188] Step 4.5: Incremental learning and adaptive adjustment

[0189] 1. Incremental Learning:

[0190] In the incremental learning process, whenever new data (such as the features at time point t+1) arrives, the parameters of the model are fine-tuned according to the new data, and the parameters θ of the model are optimized by gradient descent:

[0191]

[0192] η is the learning rate, is the gradient of the loss function, and the loss function L comprehensively considers the errors of POCD risk prediction and latent variable update.

[0193] 2. Joint Optimization:

[0194] The loss function L can consist of two parts:

[0195] POCD risk prediction loss: Optimize risk prediction using cross entropy loss:

[0196]

[0197] Hidden variable update loss: Use KL divergence or other methods to optimize the cognitive state distribution:

[0198]

[0199] 3. Overall loss:

[0200] Jointly optimize two losses: L total =L POCD +λL KL , λ is the weight coefficient, which is used to balance the optimization of POCD risk prediction and cognitive status distribution.

[0201] Step 4.6: Output the results

[0202] 1. Dynamic POCD risk probability:

[0203] After dynamic updating, the POCD risk probability P at each time point is output POCD (t).

[0204] 2. Distribution of cognitive status:

[0205] The cognitive state distribution z(t) at each time point represents the patient's personalized cognitive state, which changes over time.

[0206] 3. Model parameter update:

[0207] The model parameters are continuously optimized through incremental learning, so that the model can adapt to changes in postoperative data.

[0208] Specifically, in step 5, the purpose of step 5 is to use generative AI and reinforcement learning technology to simulate the patient's personalized cognitive recovery trajectory, predict the time it takes for the patient to recover to the baseline state (such as preoperative cognitive function level), and output the best intervention strategy (such as cognitive training, anti-inflammatory treatment, etc.) to reduce the risk of POCD and accelerate the patient's cognitive recovery.

[0209] Step 5.1: Generate a personalized cognitive recovery trajectory

[0210] 1. Hidden variable update:

[0211] The patient latent variable z(t) obtained in step 4 is used to represent the patient's personalized cognitive state as the basis for generating a personalized recovery trajectory.

[0212] Generate cognitive states at future time points through generative AI models such as variational autoencoders (VAE) or generative adversarial networks (GANs) to simulate the patient’s recovery process: represents the predicted cognitive score (such as MoCA or MMSE total score), which is a generative model used to generate the predicted cognitive score based on the latent variable z(t).

[0213] 2. Recovery trajectory generation:

[0214] Simulate the process of patients recovering from surgery to preoperative baseline status based on the formula:

[0215]

[0216] Preoperative baseline cognitive score (such as preoperative MoCA total score);

[0217] Initial postoperative cognitive score (such as MoCA total score on day 1).

[0218] λ: Recovery rate parameter, controls the speed at which the patient recovers.

[0219] The formula is based on an exponential decay model, which indicates that patients gradually move toward baseline levels during recovery.

[0220] 3. Recovery time prediction:

[0221] Based on the generated cognitive recovery trajectory, predict the time t for the patient to return to the preoperative baseline state recovery , that is, when hour:

[0222]

[0223] t current is the recovery progress at the current time point, is the current cognitive score.

[0224] Step 5.2: Output the best intervention strategy

[0225] 1. Reinforcement learning to optimize intervention strategies:

[0226] Reinforcement learning (RL) methods were used to simulate the effects of different intervention strategies on patients' cognitive recovery and optimize the strategies to accelerate recovery and reduce the risk of POCD.

[0227] Define states t is the patient's cognitive state at time t, action a t For intervention measures (such as cognitive training, anti-inflammatory treatment, etc.), reward t For comprehensive evaluation of recovery progress and risk reduction:

[0228]

[0229] Among them, ω1 and ω2 are weight coefficients, balancing recovery progress and risk reduction, C(t) represents the cognitive score at time point t, Preoperative baseline cognitive score, P POCD (t) is the POCD risk probability at time point t.

[0230] 2. Reinforcement Learning Model:

[0231] Use DNQ to train the agent model and learn different intervention strategies t Impact on patient recovery:

[0232]

[0233] Q(s t , a t ) is in state s t Take action a t is the expected return, α is the learning rate, and γ is the discount factor.

[0234] 3. Intervention strategy output:

[0235] Through reinforcement learning optimization, the optimal intervention a* is selected to minimize POCD risk and maximize cognitive recovery:

[0236]

[0237] Intervention strategiesa* may include increasing the frequency of cognitive training, adjusting medication doses, optimizing pain management, etc.

[0238] Step 5.3: Optimize intervention pathways and effect evaluation

[0239] 1. Intervention path simulation:

[0240] Through the optimized intervention strategy * , simulate the patient's recovery path at different time points and predict the patient's cognitive recovery progress under optimal intervention.

[0241] Based on the optimal intervention path output by the model, the intervention effect is evaluated. The effect evaluation may include:

[0242] The risk of POCD was reduced on postoperative day 30;

[0243] The time to return to preoperative cognitive status was shortened.

[0244] Step 5.4: Output the final result

[0245] 1. Personalized recovery trajectory:

[0246] Output simulated personalized recovery trajectory And predict the time t to return to the baseline state recovery .

[0247] 2. Best intervention strategy:

[0248] Output the best intervention strategy optimized by reinforcement learning, including cognitive training, medication adjustment, behavioral activity management, etc.

[0249] Through these steps, it is possible to provide patients with personalized recovery plans and predict the optimal recovery time and optimized intervention strategies, thereby effectively reducing the risk of POCD and accelerating postoperative cognitive recovery.

[0250] An example is:

[0251] Patient Information:

[0252] Gender: Male;

[0253] Age: 75 years old

[0254] Type of surgery: Hip replacement;

[0255] Preoperative status:

[0256] MoCA total score: 25 (attention 5 / 6, memory 4 / 5);

[0257] MMSE total score: 28 (orientation 10 / 10, memory 5 / 5);

[0258] IL-6: 6pg / mL, CRP: 8mg / L;

[0259] Initial status after surgery:

[0260] Day 1 MoCA total score: 20, MMSE total score: 25;

[0261] IL-6: 12 pg / mL, speech rate: 70 words / min;

[0262] 1. Preoperative feature matrix:

[0263] X Pre =[25,28,6,8,3100]

[0264] Including preoperative multimodal features such as MoCA, MMSE, IL-6, CRP, and hippocampal volume.

[0265] 2. Time series feature matrix:

[0266] X Dynamic : Dynamic feature matrix of postoperative time points (taking the 1st, 7th and 30th days after surgery as examples):

[0267]

[0268] The rows represent the time points (day 1, day 7, day 30), and the columns represent multimodal features such as MoCA, MMSE, IL-6, speech rate, and number of steps.

[0269] 3. Hidden variable update:

[0270] Update latent variables through a dynamic time series model:

[0271] The latent variable on the 30th day is z(30)=[0.4,-0.2,0.6].

[0272] 4. Generate a personalized cognitive recovery trajectory

[0273] 4.1. Model generates cognitive scores:

[0274] Use the patient latent variable z(t) and the generative model to predict future cognitive scores:

[0275]

[0276] 4.2 Recovery trajectory formula:

[0277] Generate the patient's MoCA prediction trajectory using the exponential decay model:

[0278]

[0279] Preoperative baseline Initial value after surgery The recovery rate parameter λ=0.05.

[0280] Simulate the recovery trajectory from day 1 to day 90:

[0281]

[0282] 4.3 Recovery trajectory calculation results:

[0283]

[0284] 4.4 Predicting the time to return to the baseline:

[0285] when hour:

[0286]

[0287] Assume the current rating is calculate:

[0288]

[0289] Recovery to baseline is expected by day 74.

[0290] 5. Output the best intervention strategy

[0291] 5.1 Reinforcement Learning Model:

[0292] Defining states Define action a t For different intervention strategies:

[0293] a1: Increase cognitive training to 3 times per week;

[0294] a2: increase the dose of anti-inflammatory drugs by 10%;

[0295] a3: Walking goal increased to 6,000 steps per day.

[0296] 5.2 Reward Function:

[0297] Combined with improved cognitive scores and reduced POCD risk:

[0298]

[0299] 5.3. Reinforcement learning strategy optimization:

[0300] Using DNQ to optimize strategy:

[0301]

[0302] Optimal action a*Choose a3: Increase walking goal to 6,000 steps per day.

[0303] 6. Optimized recovery path

[0304] 6.1 Recovery time after optimization:

[0305] Through intervention measures (increasing walking goals), the recovery rate λ is increased to 0.06, and the recovery time is recalculated:

[0306]

[0307] 6.2 Risk of POCD after optimization:

[0308] The risk of POCD was reduced to P POCD (30) = 35% (45% before optimization).

[0309] Final result:

[0310] Time to return to baseline: Through the optimal intervention strategy, the time it took for patients to return to baseline was shortened from 74 days to 62 days;

[0311] Best intervention strategy: Increase walking goal to 6,000 steps per day while maintaining current dose of anti-inflammatory therapy;

[0312] Optimized POCD risk: The POCD risk at postoperative day 30 was reduced from 45% to 35%.

[0313] Another example is:

[0314] Patient Information:

[0315] Gender: Female;

[0316] Age: 72

[0317] Type of surgery: Knee replacement;

[0318] Preoperative status:

[0319] MoCA total score: 23 (attention 4 / 6, memory 3 / 5);

[0320] MMSE total score: 27 (orientation 9 / 10, memory 5 / 5);

[0321] IL-6: 7pg / mL, CRP: 10mg / L;

[0322] MRI shows: hippocampal volume 3050mm 3 ;

[0323] Initial status after surgery:

[0324] Day 1 MoCA total score: 19, MMSE total score: 24;

[0325] IL-6: 13 pg / mL, speech rate: 65 words / min;

[0326] 1. Preoperative feature matrix:

[0327] X Pre =[23,27,7,10,3050]

[0328] Including preoperative multimodal features such as MoCA, MMSE, IL-6, CRP, and hippocampal volume.

[0329] 2. Time series feature matrix:

[0330] Dynamic feature matrix on the 1st, 7th and 30th day after surgery:

[0331]

[0332] 3. Hidden variable update:

[0333] Update the patient's hidden variables through GRU:

[0334] On the 30th day, the latent variable z(30) = [0.3, -0.1, 0.5].

[0335] 4. Generate a personalized cognitive recovery trajectory

[0336] 4.1 Generative model predicts future cognitive scores:

[0337] Use latent variables z(t) to predict the patient's future MoCA score trajectory:

[0338]

[0339] 4.2 Recovery trajectory formula:

[0340] Generate predicted trajectories using an exponential decay model:

[0341]

[0342] Preoperative baseline Initial value after surgery The recovery rate parameter λ = 0.04. 4.3 Calculate the future trajectory:

[0343]

[0344] 4.4 Predict the time to return to baseline:

[0345] Dangdang hour::

[0346]

[0347] Current rating is calculate:

[0348]

[0349] 5 Output the best intervention strategy

[0350] 5.1 Reinforcement Learning Model:

[0351] Status: The status of the patient at time point t:

[0352]

[0353] Define action a t For different intervention strategies:

[0354] Action: Intervention strategy collection:

[0355] a1: Increase cognitive training to 3 times per week;

[0356] a2: Adjust the dose of anti-inflammatory drugs (reduce IL-6 levels);

[0357] a3: Behavioral management: Increase daily walking steps by 500 steps.

[0358] 5.2 Reward Function:

[0359] Reinforcement learning rewards combined with cognitive score improvement and POCD risk reduction:

[0360]

[0361] 5.3 Optimizing intervention strategies:

[0362] Update the policy using DQN:

[0363]

[0364] 4. Choose the best strategy:

[0365] The optimal action a* is obtained through the reinforcement learning model:

[0366] a*=a2: Adjust the dose of anti-inflammatory drugs to control the IL-6 level to below 6pg / mL.

[0367] 6. Optimized recovery path

[0368] 6.1 Recovery time after optimization:

[0369] After adjusting the dosage of anti-inflammatory drugs, the recovery rate λ is increased to 0.05, and the recovery time is recalculated:

[0370]

[0371] 6.2 Risk of POCD after optimization:

[0372] The risk of POCD at day 30 increased from P POCD (30) = 40% down to P POCD (30) = 30%.

[0373] Final Result

[0374] Time to return to baseline: Through optimization of anti-inflammatory treatment, the time to return to baseline was shortened from 82 days to 66 days;

[0375] The best intervention strategy: adjust the dose of anti-inflammatory drugs (control IL-6 to below 6pg / mL) and maintain the current intensity of cognitive training;

[0376] Optimized POCD risk: The POCD risk at postoperative day 30 was reduced from 40% to 30%.

[0377] Example 2

[0378] like Figure 2As shown, the present invention discloses a prediction model construction system for postoperative cognitive decline in elderly patients, comprising:

[0379] Preoperative data collection and multimodal feature extraction module: collect preoperative cognitive scale scores, physiological indicators, imaging features, and multimodal data of speech features to generate the preoperative feature matrix X Pre ;

[0380] Preoperative preliminary digital twin building block: using static risk prediction model to assess the preliminary risk probability of POCD Introduce the variational autoencoder to the preoperative feature matrix X Pre Modeling is performed by learning the preoperative feature matrix X Pre The implicit distribution of the patient's preoperative cognitive state is generated to complete the initial construction of the digital twin.

[0381] Postoperative multi-time point data collection and feature generation module: Dynamically collect cognitive scale scores, physiological indicators, voice features and behavioral activity data at different time points after surgery to generate a time series feature matrix X Dynamic ;

[0382] Postoperative dynamic twin update and adaptive learning module: the preoperative feature matrix X Pre and the time series feature matrix X Dynamic The dynamic time series model is input to update the patient’s POCD risk probability and cognitive status distribution;

[0383] Generate personalized recovery trajectory and optimize path module: simulate personalized cognitive recovery trajectory, predict the time it takes for patients to return to baseline state, and output the best intervention strategy to reduce the risk of POCD.

[0384] The module in the embodiment 2 is used to implement the function in the embodiment 1. This embodiment can be implemented by a system, the system includes a processor and a memory, the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the prediction model construction method for postoperative cognitive decline in elderly patients according to the embodiment 1 of the present application is implemented. The system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface, and their settings and functions are known in the art, so they will not be repeated here.

[0385] In the present application, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. For example, a computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by an application, a module or both. Any such computer storage medium may be part of a device or accessible or connectable to a device. Any application or module described in this application may be implemented using computer-readable / executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

[0386] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of protection of the claims of the present invention.

Claims

1. A method for constructing a prediction model for postoperative cognitive decline in elderly patients, characterized in that: include: Collect preoperative cognitive scale scores, physiological indicators, imaging features, and multimodal data of speech features to generate the preoperative feature matrix X Pre ; Using static risk prediction model to estimate the initial risk probability of POCD Introduce the variational autoencoder to the preoperative feature matrix X Pre Modeling is performed by learning the preoperative feature matrix X Pre The implicit distribution of the patient's preoperative cognitive state is generated to complete the initial construction of the digital twin. Dynamically collect cognitive scale scores, physiological indicators, voice characteristics, and behavioral activity data at different time points after surgery to generate a time series feature matrix X Dynamic ; The preoperative feature matrix X Pre and the time series feature matrix X Dynamic The dynamic time series model is input to update the patient’s POCD risk probability and cognitive status distribution; Simulate personalized cognitive recovery trajectories, predict the time it takes for patients to return to baseline, and output optimal intervention strategies to reduce the risk of POCD.

2. The method for constructing a prediction model for postoperative cognitive decline in elderly patients according to claim 1, characterized in that: The preoperative cognitive scale score, physiological index, imaging features and voice feature multimodal data are collected to generate a preoperative feature matrix X Pre include: The cognitive scale scores include the total score and sub-scores of the Montreal Cognitive Assessment and the Mini-Mental State Examination; The physiological indicators include inflammatory factors, oxidative stress markers (MDA), and the age and gender of the patient; The imaging features include hippocampal volume, gray matter volume, and white matter lesion volume extracted from brain MRI; The speech features include speech rate, language pause time, and vocabulary diversity indicators extracted from speech recordings of recognition tests.

3. The method for constructing a prediction model for postoperative cognitive decline in elderly patients according to claim 2, characterized in that: The use of a static risk prediction model to assess the initial risk probability of POCD Introduce the variational autoencoder to the preoperative feature matrix X Pre Modeling is performed by learning the preoperative feature matrix X Pre The implicit distribution of the patient's preoperative cognitive state distribution is generated to complete the initial construction of the digital twin, including: Using static risk prediction model to estimate the initial risk probability of POCD The variational autoencoder includes an encoder, a latent variable distribution, and a decoder, where the encoder transforms the preoperative feature matrix X Pre Mapped to the mean μ and variance σ of the latent variable 2 ; Use the reparameterization technique to sample the latent variable z; The decoder is reconstructed by the hidden variable z The latent variable z is stored in the digital twin as a personalized representation of the patient’s preoperative cognitive state: Store POCD preliminary risk probability and hidden variable z as the initial state of the digital twin.

4. The method for constructing a prediction model for postoperative cognitive decline in elderly patients according to claim 3, characterized in that: The preoperative feature matrix X Pre and the time series feature matrix X Dynamic Input the dynamic time series model to update the patient's POCD risk probability and cognitive status distribution including: Use positional encoding to add temporal information to input features Among them, X t is the feature vector at the tth time point, PE(t) is the position information at the tth time point; the input data containing time information Input into the self-attention mechanism of the Transformer model to obtain H t , where H t It is the output representation of the Transformer model at each time point; Use the fully connected layer to transform the H output of Transformer t Mapped to POCD risk probability P POCD (t), obtain the risk probability P that changes at any time POCD (t); Use a recurrent neural network to update the patient's cognitive state distribution z(t).

5. The method for constructing a prediction model for postoperative cognitive decline in elderly patients according to claim 4, characterized in that: The simulation of personalized cognitive recovery trajectory, predicting the time for patients to return to baseline status, and outputting the best intervention strategy to reduce the risk of POCD include: Use a recurrent neural network to update the patient's cognitive state distribution z(t) to represent the patient's personalized cognitive state, which serves as the basis for generating a personalized recovery trajectory; Generate cognitive recovery trajectories at future time points through generative AI models Simulate the patient's recovery process; Based on the generated cognitive recovery trajectory Predict the time it takes for a patient to return to their preoperative baseline status; Define states t is the patient's cognitive state at time t, action a t For intervention measures, reward r t For comprehensive evaluation of recovery progress and risk reduction: Among them, ω1 and ω2 are weight coefficients, balancing recovery progress and risk reduction, C(t) represents the cognitive score at time point t, Preoperative baseline cognitive score, p POCD (t) is the POCD risk probability at time t; Use DNQ to train the agent model and learn different intervention strategies t Impact on patient recovery: Q(s t , a t ) is in state s t Take action a t The expected return, α is the learning rate, and γ is the discount factor; Through reinforcement learning optimization, the optimal intervention strategy a is selected * To minimize POCD risk and maximize cognitive recovery:

6. A prediction model construction system for postoperative cognitive decline in elderly patients, characterized in that: include: Preoperative data collection and multimodal feature extraction module: collect preoperative cognitive scale scores, physiological indicators, imaging features, and multimodal data of speech features to generate the preoperative feature matrix X Pre ; Preoperative preliminary digital twin building block: using static risk prediction model to assess the preliminary risk probability of POCD Introduce the variational autoencoder to the preoperative feature matrix X Pre Modeling is performed by learning the preoperative feature matrix X Pre The implicit distribution of the patient's preoperative cognitive state is generated to complete the initial construction of the digital twin. Postoperative multi-time point data collection and feature generation module: Dynamically collect cognitive scale scores, physiological indicators, voice features and behavioral activity data at different time points after surgery to generate a time series feature matrix X Dynamic ; Postoperative dynamic twin update and adaptive learning module: the preoperative feature matrix X Pre and the time series feature matrix X Dynamic The dynamic time series model is input to update the patient’s POCD risk probability and cognitive status distribution; Generate personalized recovery trajectory and optimize path module: simulate personalized cognitive recovery trajectory, predict the time it takes for patients to return to baseline state, and output the best intervention strategy to reduce the risk of POCD.

Images