Risk prediction model construction for post-icu syndrome and personalized ai intervention method and system

Abstract

This invention discloses a risk prediction model construction and personalized AI intervention method and system for post-ICU syndrome. The method includes: collecting multimodal patient data; fusing text and speech features to generate a comprehensive cognitive feature score, then incorporating clinical assessment scores and physiological features using a topology-preserving dimensionality reduction algorithm to integrate clinically operable parameters, generating a patient state feature representation; constructing an actionable decision tree model based on this, using an improved information gain criterion that integrates operable parameters for node splitting, and employing an economically cost-sensitive pruning criterion to optimize the estimated intervention cost; deploying a three-level trigger analysis mechanism to achieve hierarchical decision-making from rapid screening to deep personalized path optimization; executing tiered interventions based on the analysis results and dynamically adjusting parameters, while incrementally updating the model using feedback data. This invention achieves closed-loop generation and dynamic optimization of personalized intervention plans, from accurate risk prediction to clinically executable and cost-optimized solutions.

Description

Technical Field

[0001] This invention relates to the intersection of medical and health information technology and artificial intelligence, specifically to a risk prediction model construction and personalized AI intervention method and system for post-ICU syndrome. Background Technology

[0002] Post-ICU syndrome (PICS) is a long-term series of physical, cognitive, and psychological dysfunctions that critically ill patients experience after being transferred out of the ICU, severely impacting their quality of life and resulting in a significant social and medical burden. Accurately predicting PICS risk and implementing early, personalized interventions are crucial for improving prognosis, but current clinical practice faces multiple challenges.

[0003] In risk prediction, traditional methods rely heavily on limited, structured physiological indicators and rating scales, making it difficult to comprehensively capture the complex information reflecting cognitive and psychological disorders. While machine learning models have been applied in medical risk prediction, mainstream models (such as logistic regression, support vector machines, or black-box deep networks) typically stop at outputting risk probabilities, failing to establish a direct and interpretable link with specific, clinically actionable interventions. Clinicians struggle to develop clear and quantifiable intervention plans based on the abstract outputs of these models.

[0004] During the intervention phase, existing guidelines are mostly general recommendations, lacking the ability to precisely and dynamically adjust based on individual multi-dimensional risk characteristics. Furthermore, building high-precision predictive models requires large-scale, multi-center data, but the privacy sensitivity of medical data and the problem of data silos between institutions severely hinder collaborative model training and knowledge sharing. In addition, the economic cost considerations of intervention measures are often overlooked in existing automated decision-making systems, failing to meet the pursuit of health economics benefits in modern healthcare management.

[0005] Therefore, there is an urgent need for an intelligent system that can integrate multimodal data, generate interpretable and clinically executable personalized intervention plans, protect privacy and co-evolve, and balance intervention effectiveness and cost-effectiveness. Summary of the Invention

[0006] To address the deficiencies of the prior art as described in the background section, this invention aims to construct an intelligent diagnosis and treatment system for post-ICU syndrome that can integrate multimodal data, generate clinically feasible and cost-effective personalized intervention plans, and possess privacy-preserving collaborative learning, intelligent hierarchical response, and dynamic closed-loop optimization capabilities.

[0007] In a first aspect, embodiments of this application provide a method for constructing a risk prediction model and providing personalized AI intervention for post-ICU syndrome, the method comprising:

[0008] S1. Collect structured clinical data, unstructured text data and temporal speech data of patients, and extract features respectively to obtain physiological feature vectors, text semantic feature vectors and acoustic feature vectors;

[0009] S2. The text semantic feature vector and acoustic feature vector are fused to generate a comprehensive cognitive feature score; the score is then fused with the standardized clinical assessment score and physiological feature vector, and clinical interventionability parameters are incorporated through a topology-preserving dimensionality reduction algorithm to generate a patient state feature representation containing intervention guidance information.

[0010] S3. Based on the patient state feature representation, construct an actionable decision tree model, wherein: when splitting nodes, an improved information gain criterion that incorporates clinical interventionability parameters is adopted, which simultaneously considers the purity improvement of feature splitting and the feasibility score of the clinical intervention corresponding to the feature; when pruning the model, an economic cost-sensitive pruning criterion is adopted, which is optimized by minimizing the estimated total intervention cost function;

[0011] S4. During the model deployment phase, a three-level trigger analysis is performed: First, a first-level rapid risk screening is conducted based on key physiological indicators and simplified cognitive scores; when the screening exceeds the threshold, a second-level standard analysis is triggered, and the complete actionable decision tree model is run; when the preset deep analysis trigger conditions are met, a third-level deep analysis is initiated. The deep analysis trigger conditions include the occurrence of specific clinical events in the patient, poor response to standard intervention, or excessive uncertainty in model prediction, and personalized path optimization based on reinforcement learning is performed.

[0012] S5. Based on the results of the multi-level analysis, implement tiered interventions, collect feedback data through real-time monitoring to dynamically adjust intervention parameters, and use follow-up data streams to incrementally update the actionable decision tree model.

[0013] Secondly, embodiments of this application provide a risk prediction model construction and personalized AI intervention system for post-ICU syndrome, applied to the risk prediction model construction and personalized AI intervention method for post-ICU syndrome as described in the first aspect, the system comprising:

[0014] The multimodal data acquisition and feature extraction module is used to acquire patients' structured clinical data, unstructured text data, and temporal speech data, and to extract features from them to obtain physiological feature vectors, text semantic feature vectors, and acoustic feature vectors.

[0015] The clinically interventionable feature fusion module is used to fuse the text semantic feature vector and acoustic feature vector to generate a comprehensive cognitive feature score; this score is then fused with standardized clinical assessment scores and physiological feature vectors, and clinically interventionable parameters are incorporated through a topology-preserving dimensionality reduction algorithm to generate a patient state feature representation containing intervention guidance information;

[0016] An actionable decision tree construction and optimization module is used to construct an actionable decision tree model based on the patient state feature representation, wherein: when splitting nodes, an improved information gain criterion that incorporates clinical interventionability parameters is adopted, which simultaneously considers the purity improvement of feature splitting and the feasibility score of the clinical intervention corresponding to the feature; when pruning the model, an economic cost-sensitive pruning criterion is adopted, which is optimized by minimizing the estimated total intervention cost function;

[0017] The multi-level intelligent triggering and dynamic analysis module is used to perform three-level triggering analysis during the model deployment phase: First, a first-level rapid risk screening is performed based on key physiological indicators and simplified cognitive scores; when the screening exceeds the threshold, a second-level standard analysis is triggered, and the complete actionable decision tree model is run; when the preset deep analysis triggering conditions are met, a third-level deep analysis is initiated. The deep analysis triggering conditions include the occurrence of specific clinical events in patients, poor response to standard interventions, or excessive uncertainty in model predictions, and personalized path optimization based on reinforcement learning is performed.

[0018] The personalized intervention execution and closed-loop optimization module is used to execute tiered interventions based on the multi-level analysis results, dynamically adjust intervention parameters by collecting feedback data through real-time monitoring, and incrementally update the actionable decision tree model using follow-up data streams.

[0019] Thirdly, embodiments of this application provide an electronic device, including:

[0020] processor;

[0021] Memory used to store processor-executable instructions;

[0022] The processor is configured to implement the risk prediction model construction and personalized AI intervention method for post-ICU syndrome as described in the first aspect when executing the instructions.

[0023] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the risk prediction model construction and personalized AI intervention method for post-ICU syndrome as described in the first aspect.

[0024] Compared with existing technologies, this invention has the following significant advantages: By constructing an actionable decision tree model, risk prediction is directly linked to specific, cost-quantifiable intervention packages, making the AI ​​output highly clinically operable and interpretable, truly achieving a closed loop in assisting clinical decision-making. Introducing economic cost criteria and Pareto front analysis into key stages such as model pruning ensures that the final model is not only accurate in prediction but also pursues optimal cost-effectiveness in the generated intervention plans, aligning with the concepts of modern precision medicine and value-based healthcare. Through a three-level trigger analysis mechanism, the system can intelligently differentiate case complexity and rationally allocate resources. For complex and difficult cases, deep analysis based on reinforcement learning and Monte Carlo tree search is initiated to generate multi-path comparison reports, supporting high-level clinical decision-making. By integrating federated learning and differential privacy technologies, while strictly protecting patient data privacy, data barriers between institutions are broken down, enabling the model to utilize a wider range of real-world data for collaborative training and continuous optimization, improving the model's generalization ability and timeliness. By utilizing real-time monitoring feedback and long-term follow-up data, the system can dynamically adjust intervention parameters and incrementally update the core prediction model, enabling the entire system to self-optimize and adapt to changes in clinical practice, thus overcoming the limitations of static models. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the process for constructing a risk prediction model and a personalized AI intervention method for post-ICU syndrome, as provided in an embodiment of this application.

[0026] Figure 2 The system architecture diagram for the risk prediction model of post-ICU syndrome and the personalized AI intervention system provided in this application.

[0027] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0029] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0030] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] Example 1

[0032] Figure 1 This is a schematic diagram illustrating a method for constructing a risk prediction model and providing personalized AI intervention for post-ICU syndrome, as provided in one embodiment of this application. Figure 1 As shown, a risk prediction model for post-ICU syndrome and a personalized AI intervention method are constructed, including:

[0033] S1. Collect structured clinical data, unstructured text data, and temporal speech data from patients, and extract features from each to obtain physiological feature vectors, textual semantic feature vectors, and acoustic feature vectors. Comprehensively collect three types of data from patients: structured (e.g., vital signs), unstructured (e.g., medical records), and temporal (e.g., speech), and transform them into standardized feature vectors (physiological, textual semantic, and acoustic) that can be processed by machines. This forms the data foundation for the system to perceive and understand the patient's condition.

[0034] Specifically, in this embodiment, the feature extraction of multimodal data in S1 includes:

[0035] The structured clinical data undergoes preprocessing, including long-tail truncating of abnormal physiological indicators based on clinical reversibility, filling missing values ​​using multiple imputation or the K-nearest neighbor algorithm, and Z-score standardization of continuous variables to obtain the physiological feature vector. Taking 72 hours of continuous heart rate monitoring data as an example, extreme high values ​​(e.g., instantaneous >220 beats / min) that are not physiologically consistent due to brief equipment interference are truncated and deleted according to clinical guidelines as "irreversible abnormal noise" rather than "reversible disease fluctuations." Subsequently, for blood pressure values ​​at individual time points with missing records, the K-nearest neighbor (K=5) algorithm is used to intelligently fill in the missing values ​​based on the blood pressure patterns of adjacent time points and other similar patients. Finally, for continuous variables such as age, mean arterial pressure, and blood oxygen saturation of all patients, their mean and standard deviation are calculated across the entire training set, and Z-score standardization is performed to eliminate the influence of dimensions, ultimately outputting the standardized physiological feature vector.

[0036] The unstructured text data is processed, including word segmentation using medical-domain word segmentation tools and semantic encoding using a clinically pre-trained language model enhanced with a medical knowledge graph, to obtain the text semantic feature vector. For doctor-recorded descriptions of patient journeys, such as "The patient is depressed today, refuses rehabilitation training, and complains of frequent nightmares," word segmentation and entity recognition are first performed using a medical-specific word segmentation tool. Then, the segmented sequence is input into a clinically pre-trained language model (e.g., a variant of Clinical-BERT) trained on large-scale medical literature and electronic medical records and incorporating a disease-symptom relationship knowledge graph. The model extracts sentence-level contextual semantic representations, encoding the text into a high-dimensional, dense vector containing the semantics of "depressive mood" and "post-traumatic stress," i.e., the text semantic feature vector.

[0037] The temporal speech data is processed, including extracting Mel-frequency cepstral coefficients (MFCCs) and speech rate variation coefficients (PRCs) as basic features, and extracting temporal patterns using a one-dimensional convolutional neural network to obtain the acoustic feature vector. Three minutes of speech was recorded during a standardized psychological assessment interview with the patient. First, the Mel-frequency cepstral coefficients (MFCCs, 13-dimensional) of each frame of the original audio were extracted as basic features characterizing the spectral properties of the sound. Simultaneously, the speech rate (syllables / second) and its PRC over time were calculated as auxiliary features reflecting fluency and emotional state. Next, this MFCC temporal sequence was input into a one-dimensional convolutional neural network (1D-CNN). This network automatically learns and extracts deeper temporal patterns related to emotion and cognitive load (such as subtle changes in intonation and pause patterns) through multiple convolutional and pooling operations. The output of the final layer of the network serves as the high-level acoustic feature vector.

[0038] S2. The text semantic feature vector and acoustic feature vector are fused to generate a comprehensive cognitive feature score. This score is then fused with standardized clinical assessment scores and physiological feature vectors, and clinical interventionability parameters are incorporated through a topology-preserving dimensionality reduction algorithm to generate a patient state feature representation containing intervention guidance information. The text and speech features are fused to generate a "comprehensive cognitive feature score" reflecting cognitive psychological state. This score is then deeply fused with physiological features and clinical assessment scores, with clinical interventionability parameters pre-introduced in the process, ultimately generating a "patient state feature representation" that not only characterizes the patient's overall state but also naturally indicates which aspects can be intervened and how. This is the core information preprocessing for achieving "actionable" decision-making.

[0039] Specifically, in this embodiment, generating a comprehensive cognitive feature score in S2 includes:

[0040] The acoustic feature vector is processed by a first classification network to obtain the probability of depressive tendency. Specifically, the acoustic feature vector extracted by S1, which represents the high-level temporal pattern of the patient's speech, is input into a pre-trained depressive tendency classification network (e.g., a deep neural network containing fully connected layers and softmax output). This network is trained on a large amount of clinical speech data labeled with depression assessment scores. The network analyzes the input acoustic features and outputs a scalar value between 0 and 1, for example, 0.72, as the quantified probability that the patient's speech contains a depressive tendency.

[0041] The text semantic feature vector is processed by a second classification network to obtain the anxiety tendency probability. Specifically, the text semantic feature vector containing patient semantic information extracted in S1 is input into a pre-trained anxiety tendency classification network (e.g., a similar deep neural network). This network is trained on a large amount of clinical text data labeled with anxiety assessment scores. The network analyzes the input text semantic features and outputs a scalar value between 0 and 1, for example, 0.58, as the quantified probability that the patient's text contains an anxiety tendency.

[0042] A fusion weight is dynamically generated based on a gated attention mechanism to weight and fuse the probabilities of depressive tendency and anxiety tendency to generate the comprehensive cognitive feature score. Specifically, a gated attention network module is designed that simultaneously receives the original acoustic feature vector and the text semantic feature vector as input. This module calculates the interaction and importance between these two sets of features through a small neural network and dynamically generates a fusion weight value (e.g., weight = 0.6). This weight indicates the relative contribution of depressive tendency (from speech) and anxiety tendency (from text) to the comprehensive cognitive state score in the current assessment of a specific patient. Finally, based on this dynamic weight, the probabilities of depressive tendency (0.72) and anxiety tendency (0.58) are weighted and summed to calculate the final comprehensive cognitive feature score (e.g., 0.72*0.6 + 0.58*0.4 = 0.664). This score fuses bimodal information, and the weights are adaptively determined by data.

[0043] Furthermore, the clinically modifiable parameters incorporated into S2 through a topology-preserving dimensionality reduction algorithm include:

[0044] The physiological feature vectors, comprehensive cognitive feature scores, and clinical assessment scores are used as node features, and a patient similarity graph is constructed based on feature similarity. Taking 100 post-ICU patients as an example, each patient is considered as a node. The initial feature vector of each node is composed of the physiological feature vector extracted by S1 (such as heart rate variability, inflammatory markers), the comprehensive cognitive feature score generated by S2 (such as 0.664), and the standardized clinical assessment score (such as physical function score, cognitive screening score). The cosine similarity of feature vectors between all patient node pairs is calculated, and the edges connecting each node to its K most similar nodes (such as K=10) are retained. The weight of the edge is the similarity value, thereby constructing a patient similarity graph that reflects the multidimensional state similarity relationship between patients.

[0045] The similarity graph is dimensionality reduced using a graph Laplacian feature mapping algorithm. This algorithm minimizes a loss function that includes a graph Laplacian regularization term, ensuring that nodes similar (connected by edges) in the high-dimensional space remain close in the reduced-dimensional space, thus preserving the original data manifold topology. During dimensionality reduction, an "interventionibility score" assessed by clinical experts is further incorporated as a constraint. For example, the interventionibility score for the "serum potassium level" feature (easily adjustable with medication) is 0.9, while the score for the "age" feature (immutable) is 0.1. By modifying the objective function, a constraint term based on a clinical interventionibility parameter is added while minimizing the graph Laplacian regularization term (preserving the topology). Specifically, each edge in the graph is assigned a weight adjustment factor related to the similarity of the nodes at both ends on the interventionibility feature. This ensures that when optimizing the low-dimensional embedding, the algorithm not only considers the original feature similarity but also strengthens the proximity of nodes with highly interventionibility features in the low-dimensional space. Ultimately, a low-dimensional patient state feature representation is generated for each patient that simultaneously encodes their original state information and clinical intervention potential.

[0046] S3. Based on the patient state feature representation, construct an actionable decision tree model. The core of this step is to address the problems pointed out in the background art: abstract model output, lack of correlation with specific interventions, and neglect of the economic costs of interventions. Specifically: when splitting nodes, an improved information gain criterion incorporating clinical interventionability parameters is adopted. This criterion simultaneously considers the purity improvement of feature splitting and the feasibility score of the corresponding clinical intervention; when pruning the model, an economically cost-sensitive pruning criterion is adopted, optimized by minimizing the estimated total intervention cost function. Improved information gain is employed. The formula can be:

[0047] ,

[0048] in, , , Represents the dataset of the current node. For the sample size, Indicates candidate splitting features, Its set of possible values; The feature value is represented as A subset of; This represents the entropy function, which measures uncertainty. A comprehensive score representing the clinical interveneability of the characteristics; The sample belongs to the category The probability (risk level); The interventionibility weight coefficient (0≤ ≤1) For feature values The composite interventionability score; For the first Clinical expert weights (normalized); For experts For feature values Intervention feasibility score (0-10); The number of experts participating in the evaluation; Total number of categories. Select to make The most significant feature is used for splitting.

[0049] Economically cost-sensitive pruning criteria include leaf node cost models, for leaf nodes cost :

[0050] ,

[0051] in, , , leaf node The number of associated interventions; Indication of measures Basic unit cost For falling into leaf nodes The sample average risk score, The overall average risk score. measures The intensity adjustment coefficient, measures Minimum / maximum intervention period, leaf node The probability of readmission. To periodically adjust the sensitivity parameters, This represents the medical insurance reimbursement ratio (0-1).

[0052] Total cost function The pruning optimization formula is as follows:

[0053] ,

[0054] A pruning evaluation function is used in the pruning evaluation. Optimization is performed. This function combines model accuracy, total predicted intervention cost, and model complexity into a single objective through tradeoff coefficients. By adjusting these coefficients, the preference for cost control and model simplicity can be managed, thereby achieving a balance between predictive performance and economic benefits. It is expressed as:

[0055] ,

[0056] in, Describe the set of leaf nodes of tree T. Indicates falling into a leaf node The number of samples, The regularization coefficient is . The cost variance between leaf nodes. For the tree after pruning, For model accuracy, As a weighting factor, This is the complexity function.

[0057] Based on the aforementioned feature representations, a specialized decision tree model is constructed. Through node splitting, an improved criterion is adopted, prioritizing splits based on features that can both differentiate risks and facilitate clinical intervention, ensuring the tree's growth aligns with clinical action logic. Through model pruning, an economic criterion is applied to remove branches that would lead to excessively high intervention costs while maintaining predictive accuracy, ensuring the final model-recommended interventions are cost-effective. The output of this step is an actionable model that maps patient features to a pre-defined, cost-optimized intervention package.

[0058] Specifically, in this embodiment, the construction of the actionable decision tree model in step S3 further includes:

[0059] Each leaf node of the decision tree is associated with a pre-defined clinical intervention package, which includes one or more specific clinical interventions and their estimated economic costs. After the actionable decision tree is trained, the system automatically generates an initial intervention package suggestion based on the common characteristics of all training samples falling into each leaf node (such as the average value of each characteristic and the main risk category). This suggestion is then reviewed, revised, and finally confirmed by a team of clinical experts, forming a standardized clinical intervention package fixedly associated with that leaf node. For example, for a specific leaf node, its associated package might include three core measures: ① Computerized cognitive training three times a week for 30 minutes each time (estimated cost: X yuan / session); ② Psychological counseling once a week (estimated cost: Y yuan / session); ③ Family health education manual and regular telephone follow-ups (estimated cost: Z yuan / month). Any patient assigned to that leaf node by the model will automatically receive this complete, specific, and cost-quantified recommended plan.

[0060] The economically cost-sensitive pruning criterion employs Pareto front analysis for multi-objective optimization during pruning evaluation, aiming to minimize the model's total estimated intervention cost while controlling the decline in model predictive performance within a preset threshold. Specifically, the system first constructs a Pareto front (i.e., a set of non-dominated solutions) with "prediction error" and "total estimated cost" as coordinate axes by evaluating different pruning schemes. Then, starting from a preset performance degradation tolerance threshold (e.g., a decrease in prediction accuracy not exceeding 2%), the system filters out all pruning schemes that satisfy this performance constraint on the Pareto front. Finally, from these feasible schemes, the one with the lowest total estimated intervention cost is selected as the final pruning strategy. This ensures that the final model, while maintaining sufficient predictive accuracy, achieves optimal economic benefits at the population level for its recommended intervention path.

[0061] S4. In the model deployment phase, a three-level trigger analysis is performed: First, a rapid risk screening is conducted based on key physiological indicators and a quickly calculated simplified cognitive score (e.g., the probability of depressive tendency based on short speech fragments or an emotion score based on short text, or a simplified calculation version of the comprehensive cognitive feature score in S2). When the screening exceeds a threshold, a second-level standard analysis is triggered, running the complete actionable decision tree model. When the preset deep analysis trigger conditions are met, a third-level deep analysis is initiated. These deep analysis trigger conditions include the patient experiencing a specific clinical event, poor response to standard intervention, or excessive uncertainty in model prediction, leading to personalized path optimization based on reinforcement learning. A flexible analysis strategy is designed in the application phase. The first-level screening is used to quickly filter low-risk patients, saving resources. The second-level analysis is used for medium- and high-risk patients, enabling the complete actionable decision tree model and providing standard intervention recommendations. The third-level deep analysis is used for complex cases with special conditions (e.g., sudden changes in condition, ineffective intervention), initiating reinforcement learning for deep path exploration and generating personalized optimization solutions. This step achieves adaptive intelligent scheduling of analysis resources.

[0062] Specifically, in this embodiment, the three-level depth analysis in S4 includes:

[0063] The system continuously monitors at least one of the following deep analysis trigger conditions: Clinical event trigger condition: The patient experiences a predefined specific adverse clinical event; Intervention response trigger condition: Based on real-time feedback data analysis of the standard intervention protocol, it is determined that the patient's physiological, cognitive, or psychological indicators have not reached a preset expected response threshold; Model uncertainty trigger condition: The confidence level of the actionable decision tree model's risk prediction for the current patient state is lower than a preset confidence threshold. Specifically, the system continuously monitors the patient's state and initiates deep analysis when at least one preset condition is triggered. For example: ① Clinical event trigger: The patient is recorded to have a "new-onset lung infection"; ② Intervention response trigger: After two weeks of implementing the standard rehabilitation protocol, the improvement rate of the patient's "6-minute walking distance" has not reached the expected 50%; ③ Model uncertainty trigger: The actionable decision tree model's classification of the patient's current state has a sample assignment probability (confidence level) of only 0.55 for its leaf nodes, which is lower than the preset threshold of 0.7.

[0064] When any triggering condition is met, the current patient's state feature representation is used as the initial state, and the actionable decision tree model is deconstructed into a state-action space. After deep analysis is initiated, the system sets the current patient's "patient state feature representation" as the initial state of the reinforcement learning agent. Simultaneously, the trained actionable decision tree model is deconstructed: each decision node and its splitting rule define the dimension and partition of the state space, while the intervention package associated with each leaf node is defined as the action the agent can choose in the current state. Thus, a structured state-action exploration space corresponding to clinical decision-making logic is constructed.

[0065] A deep reinforcement learning algorithm is employed to explore multiple alternative personalized intervention paths in the state-action space, and the long-term efficacy and cost-effectiveness of each path are extrapolated using Monte Carlo tree search. In this space, a deep reinforcement learning agent (e.g., based on DQN or A3C algorithms) begins its exploration. Instead of strictly adhering to the single path of the original decision tree, it attempts to select different actions (i.e., attempts to associate intervention packages with other leaf nodes), simulating post-intervention state transitions. To evaluate the long-term value of each simulated path, the system uses Monte Carlo tree search for extrapolation: starting from each simulated decision point, multiple random extrapolations are performed until the end of the simulated follow-up period. The simulated efficacy gains (e.g., the gain in quality-of-life-adjusted years, QALY) and total intervention costs of each path are cumulatively calculated to assess its long-term benefits.

[0066] The output includes a deep decision-making report with multi-path comparative analysis. After exploration and simulation, the system generates a deep decision-making report. The report not only provides the baseline path recommended by the original standard decision tree but also lists 2-3 optimized alternative paths explored by deep reinforcement learning, demonstrating higher simulated long-term benefits. Each path clearly displays its core intervention combination, simulated long-term efficacy and cost curves, and provides a multi-dimensional quantitative comparison (such as efficacy, cost, and feasibility). This report provides clinical teams with a data-driven, in-depth reference for high-level decision-making and protocol adjustments for complex cases.

[0067] S5. Execute a tiered intervention based on the multi-level analysis results. Collect feedback data through real-time monitoring to dynamically adjust intervention parameters, and incrementally update the actionable decision tree model using follow-up data streams. Execute the intervention plan generated in S4 and collect feedback to form a closed loop. Adjust intervention parameters based on the patient's real-time response to the intervention (e.g., changes in physiological indicators). Utilize long-term follow-up data to incrementally learn the actionable decision tree model in S3, enabling the system to continuously evolve and adapt to new knowledge and case patterns.

[0068] Specifically, in this embodiment, the dynamic adjustment of intervention parameters in S5 includes:

[0069] Based on the intervention results, the system matches the clinical intervention knowledge base to generate an electronic intervention prescription that includes training content, intensity, frequency, and duration. Specifically, the system automatically matches and calls upon the clinical intervention knowledge base based on the final plan derived from the three-level analysis (e.g., a recommendation from the second-level analysis or a path selected from the third-level in-depth analysis). For example, for the "severe cognitive impairment with moderate risk of depression" plan, the system generates an executable electronic intervention prescription that clearly specifies: ① Training content: The "Nonverbal Memory and Executive Functions" computer training module is used; ② Intensity: Each training session includes tasks at 5 difficulty levels; ③ Frequency: Once daily, 5 days a week; ④ Duration: 8 weeks. This prescription is directly pushed to the patient and rehabilitation therapist via a mobile application.

[0070] The intervention prescription is optimized using a reinforcement learning framework. The system uses patient task completion data, physiological indicator changes, and emotional feedback data collected through real-time monitoring as state inputs. Intervention parameter adjustments serve as action outputs, and the degree of improvement in risk indicators and cost control are used as reward signals. During the execution of the prescription, an embedded online reinforcement learning framework is activated. Its state inputs are updated daily, including: the patient's completion rate of the training task (e.g., 85%), the change in daily resting heart rate variability from baseline (e.g., an increase of 5%), and the average daily mood score (e.g., 2 / 5 point) obtained from a simplified mood scale. Based on this state, the reinforcement learning policy network outputs adjustment actions daily, such as "reducing the difficulty of tomorrow's task by one level" or "adding a reminder for breathing relaxation training." Reward signals are calculated weekly, comprehensively considering the degree of improvement in core risk indicators (e.g., depression score) and the total cost of the intervention program during the week, thereby driving continuous optimization of the policy network with the goal of maximizing efficacy while minimizing costs.

[0071] The system collects real-time voice and text feedback data through a multimodal monitoring agent, calculates the probability of psychological risk, and automatically adjusts the intensity of psychological intervention when the probability exceeds a threshold. Deployed on smart terminals, the system continuously collects voice tone, speech rate, and short written log text data from patients as they perform intervention tasks or engage in daily interactions. This data is transmitted in real-time to a backend analysis module, which uses a pre-trained acoustic-semantic bimodal model to quickly calculate the immediate probability of psychological risk (e.g., a real-time probability of depression of 0.75). Once this probability exceeds a preset warning threshold (e.g., 0.7), the system automatically executes preset rules without manual intervention: ① immediately sends a warning notification to the attending physician; ② automatically adjusts the intensity of psychological intervention, for example, temporarily increasing the scheduled psychological counseling session for the next day from 20 minutes to 30 minutes and pushing a customized psychological reassurance message.

[0072] The method also includes model co-training and privacy protection steps:

[0073] A federated learning framework is constructed, comprising a central server and multiple local hospital nodes. To improve the model's versatility, a federated learning framework is built. This framework includes a central coordination server (deployed in the cloud or by a trusted third party) and multiple local hospital nodes connected to the framework (e.g., servers of Hospital A, Hospital B, and Hospital C, respectively). Each local node has a complete local training environment and communicates securely with the central server via an encrypted channel.

[0074] Each local node independently executes steps S1 through S3 to train its local model using local data. At the start of each federated training cycle, each local node uses only anonymized historical patient data from its own hospital to independently complete the entire training process, from feature extraction (S1) and feature fusion (S2) to building an actionable decision tree model (S3), obtaining a local model based on its own hospital data. During this process, any original patient data remains within the hospital and is not sent out.

[0075] Before uploading model parameter updates to the central server, each local node adds noise perturbation that meets privacy budget requirements. After completing local training, each hospital node does not upload the complete model or data; instead, it calculates the parameter differences between the local model and the previous global model (updating the gradient). Before uploading this update information, the node runs a differential privacy protection algorithm, adding precisely calibrated random Laplacian noise to the gradient according to a preset privacy budget (ε - the ε value in differential privacy, such as ε=2.0). The perturbed gradient information can no longer be used to deduce the original data of any specific patient.

[0076] The central server aggregates the perturbed information from each node to update and distribute the global model parameters, enabling cross-institutional collaborative training without sharing the original data. The central server collects noisy, perturbed model update gradients from all participating hospital nodes. The server uses a federated averaging algorithm to securely aggregate these gradients and calculate the updated global model parameters. Subsequently, the central server distributes the updated global model parameters to all participating hospital nodes. Each hospital node uses these parameters to update its local model, thus enabling the collaborative training of a more powerful global model using data from multiple hospitals without sharing any original patient data.

[0077] The model parameter update information includes parameter update gradients or model output information after knowledge distillation. The federated learning framework of this invention supports two secure information upload modes: (1) Parameter update gradient mode. This is the default mode. After each hospital node completes training locally, it calculates the gradient information (i.e., direction and magnitude of change) of the parameters (such as node splitting threshold and weight) of its local actionable decision tree model compared to the previous round of global model. For example, hospital node A calculates the gradient vector of its model on the splitting feature of "heart rate variability". After adding differential privacy noise, it only uploads the gradient vector to the central server for the average update of the global model. (2) Knowledge distillation output information mode. This mode is enabled when it is necessary to perform heterogeneous processing of the model structure or further compress information. The central server first distributes a set of representative public calibration datasets (without privacy information) to each node. Each node uses its local model to infer the dataset and generates a series of soft labels (i.e., the probability distribution output of the model for each sample belonging to different risk categories, for example: [low risk: 0.2, medium risk: 0.7, high risk: 0.1]). Each node uploads these highly condensed model outputs (rather than the model parameters themselves) after adding noise. The central server aggregates these soft tags to guide the training of a lightweight "student" global model, achieving knowledge fusion in the same way, but with less transmitted information and further reducing the risk of privacy exposure.

[0078] Example 2

[0079] like Figure 2 As shown, this application provides a system architecture diagram for risk prediction model construction and personalized AI intervention for post-ICU syndrome, which is applied to the risk prediction model construction and personalized AI intervention system for post-ICU syndrome as described in Embodiment 1. It includes: a multimodal data acquisition and feature extraction module 210, a clinically interventionable feature fusion module 220, an actionable decision tree construction and optimization module 230, a multi-level intelligent triggering and dynamic analysis module 240, and a personalized intervention execution and closed-loop optimization module 250.

[0080] The multimodal data acquisition and feature extraction module 210 is used to acquire structured clinical data, unstructured text data and temporal speech data of patients, and to extract features from them to obtain physiological feature vectors, text semantic feature vectors and acoustic feature vectors.

[0081] The clinically interventionable feature fusion module 220 is used to fuse the text semantic feature vector and the acoustic feature vector to generate a comprehensive cognitive feature score; the score is then fused with the standardized clinical assessment score and the physiological feature vector, and the clinically interventionable parameters are incorporated through a topology-preserving dimensionality reduction algorithm to generate a patient state feature representation containing intervention guidance information.

[0082] The actionable decision tree construction and optimization module 230 is used to construct an actionable decision tree model based on the patient state feature representation, wherein: when splitting nodes, an improved information gain criterion that incorporates clinical interventionability parameters is adopted, which simultaneously considers the purity improvement of feature splitting and the feasibility score of the clinical intervention corresponding to the feature; when pruning the model, an economic cost-sensitive pruning criterion is adopted, which is optimized by minimizing the estimated total intervention cost function.

[0083] The multi-level intelligent triggering and dynamic analysis module 240 is used to perform three-level triggering analysis during the model deployment phase: firstly, a first-level rapid risk screening is performed based on key physiological indicators and simple cognitive scores; when the screening exceeds the threshold, a second-level standard analysis is triggered, and the complete actionable decision tree model is run; when the preset deep analysis triggering conditions are met, a third-level deep analysis is started, including when the patient experiences a specific clinical event, responds poorly to standard intervention, or the model prediction uncertainty is too high, and personalized path optimization based on reinforcement learning is performed.

[0084] The personalized intervention execution and closed-loop optimization module 250 is used to execute tiered interventions based on the multi-level analysis results, dynamically adjust intervention parameters by collecting feedback data through real-time monitoring, and incrementally update the actionable decision tree model using follow-up data streams.

[0085] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0086] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0087] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0088] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0089] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0090] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0091] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0092] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0093] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A risk prediction model construction and personalized AI intervention method for post-ICU syndrome, characterized in that, Includes the following steps: S1. Collect structured clinical data, unstructured text data and temporal speech data of patients, and extract features respectively to obtain physiological feature vectors, text semantic feature vectors and acoustic feature vectors; S2. The text semantic feature vector and acoustic feature vector are fused to generate a comprehensive cognitive feature score; the score is then fused with the standardized clinical assessment score and physiological feature vector, and clinical interventionability parameters are incorporated through a topology-preserving dimensionality reduction algorithm to generate a patient state feature representation containing intervention guidance information. S3. Based on the patient state feature representation, construct an actionable decision tree model, wherein: when splitting nodes, an improved information gain criterion that incorporates clinical interventionability parameters is adopted, which simultaneously considers the purity improvement of feature splitting and the feasibility score of the clinical intervention corresponding to the feature; when pruning the model, an economic cost-sensitive pruning criterion is adopted, which is optimized by minimizing the estimated total intervention cost function; S4. During the model deployment phase, a three-level trigger analysis is performed: First, a first-level rapid risk screening is conducted based on key physiological indicators and simplified cognitive scores; when the screening exceeds the threshold, a second-level standard analysis is triggered, and the complete actionable decision tree model is run; when the preset deep analysis trigger conditions are met, a third-level deep analysis is initiated. The deep analysis trigger conditions include the occurrence of specific clinical events in the patient, poor response to standard intervention, or excessive uncertainty in model prediction, and personalized path optimization based on reinforcement learning is performed. S5. Based on the results of the multi-level analysis, implement tiered interventions, collect feedback data through real-time monitoring to dynamically adjust intervention parameters, and use follow-up data streams to incrementally update the actionable decision tree model.

2. The method according to claim 1, characterized in that, The feature extraction of multimodal data in S1 specifically includes: The structured clinical data is preprocessed, including performing long-tail truncation on abnormal physiological indicators based on clinical reversibility judgment, filling missing values ​​using multiple imputation or K nearest neighbor algorithm, and standardizing continuous variables using Z-score to obtain the physiological feature vector. The unstructured text data is processed, including segmenting words using a medical domain word segmentation tool and semantically encoding the text using a clinical pre-trained language model based on medical knowledge graph enhancement, to obtain the text semantic feature vector. The temporal speech data is processed by extracting Mel frequency cepstral coefficients and speech rate variation coefficients as basic features, and extracting temporal patterns through a one-dimensional convolutional neural network to obtain the acoustic feature vector.

3. The method according to claim 1, characterized in that, The comprehensive cognitive feature score generated in S2 includes: The acoustic feature vector is processed by a first classification network to obtain the probability of depressive tendency; The text semantic feature vector is processed by a second classification network to obtain the anxiety tendency probability; Based on the gating attention mechanism, the fusion weights are dynamically generated, and the probabilities of depressive tendency and anxiety tendency are weighted and fused to generate the comprehensive cognitive feature score.

4. The method according to claim 1, characterized in that, The clinically modifiable parameters incorporated into S2 through the topology-preserving dimensionality reduction algorithm include: The physiological feature vector, comprehensive cognitive feature score, and clinical assessment score are used as node features, and a patient similarity map is constructed based on feature similarity. The similarity graph is reduced in dimensionality using the graph Laplacian feature mapping algorithm, and clinically modifiable parameters assessed by clinical experts are incorporated into the features or feature combinations during the dimensionality reduction process to generate the patient state feature representation.

5. The method according to claim 1, characterized in that, The construction of the actionable decision tree model in S3 also includes: Each leaf node of the decision tree is associated with a pre-defined clinical intervention package, which includes one or more specific clinical intervention measures and their estimated economic costs. The economically cost-sensitive pruning criterion employs Pareto front analysis for multi-objective optimization during pruning evaluation, aiming to minimize the model's total estimated intervention cost and control the decline in model predictive performance within a preset threshold.

6. The method according to claim 1, characterized in that, The three-level depth analysis in S4 specifically includes: Real-time monitoring of at least one of the following deep analysis trigger conditions: Clinical event triggering conditions: A patient experiences a predefined specific adverse clinical event; Intervention response triggering conditions: Based on real-time feedback data analysis of the standard intervention protocol, it is determined that the patient's physiological, cognitive, or psychological indicators have not reached the preset response expectation threshold. Model uncertainty triggering condition: The confidence level of the actionable decision tree model in predicting the risk of the current patient status is lower than the preset confidence threshold; When any triggering condition is met, the patient state characteristics of the current patient are used as the initial state, and the actionable decision tree model is deconstructed into a state-action space. A deep reinforcement learning algorithm is used to explore multiple alternative personalized intervention paths in the state-action space, and the long-term efficacy and cost-effectiveness of each path are deduced through Monte Carlo tree search. The output includes an in-depth decision report with multi-path comparison analysis.

7. The method according to claim 1, characterized in that, The dynamic adjustment of intervention parameters in S5 includes: Based on the intervention results, a clinical intervention knowledge base is matched to generate an electronic intervention prescription that includes training content, intensity, frequency, and cycle. The intervention prescription is optimized using a reinforcement learning framework, wherein: the patient's intervention task completion data, physiological indicator change data, and emotional feedback data collected through the real-time monitoring are used as state inputs, the intervention parameter adjustment is used as action outputs, and the degree of improvement of risk indicators and cost control are used as reward signals. The multimodal monitoring agent collects voice and text feedback data in real time, calculates the probability of psychological risk, and automatically adjusts the intensity of psychological intervention when the threshold is exceeded.

8. The method according to claim 1, characterized in that, The method also includes model co-training and privacy protection steps: Construct a federated learning framework that includes a central server and multiple local nodes in hospitals; Each local node independently executes steps S1 to S3 to train its local model using local data. Before uploading the model parameter update information to the central server, each local node adds noise perturbation that meets privacy budget requirements; The central server aggregates the perturbation information from each node to update and distribute the global model parameters, enabling cross-institutional collaborative training without sharing the original data.

9. The method according to claim 8, characterized in that, The model parameter update information includes parameter update gradients or model output information after knowledge distillation.

10. A risk prediction model construction and personalized AI intervention system for post-ICU syndrome, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The multimodal data acquisition and feature extraction module is used to acquire patients' structured clinical data, unstructured text data, and temporal speech data, and to extract features from them to obtain physiological feature vectors, text semantic feature vectors, and acoustic feature vectors. The clinically interventionable feature fusion module is used to fuse the text semantic feature vector and acoustic feature vector to generate a comprehensive cognitive feature score; this score is then fused with standardized clinical assessment scores and physiological feature vectors, and clinically interventionable parameters are incorporated through a topology-preserving dimensionality reduction algorithm to generate a patient state feature representation containing intervention guidance information; An actionable decision tree construction and optimization module is used to construct an actionable decision tree model based on the patient state feature representation, wherein: when splitting nodes, an improved information gain criterion that incorporates clinical interventionability parameters is adopted, which simultaneously considers the purity improvement of feature splitting and the feasibility score of the clinical intervention corresponding to the feature; when pruning the model, an economic cost-sensitive pruning criterion is adopted, which is optimized by minimizing the estimated total intervention cost function; The multi-level intelligent triggering and dynamic analysis module is used to perform three-level triggering analysis during the model deployment phase: First, a first-level rapid risk screening is performed based on key physiological indicators and simplified cognitive scores; when the screening exceeds the threshold, a second-level standard analysis is triggered, and the complete actionable decision tree model is run; when the preset deep analysis triggering conditions are met, a third-level deep analysis is initiated. The deep analysis triggering conditions include the occurrence of specific clinical events in patients, poor response to standard interventions, or excessive uncertainty in model predictions, and personalized path optimization based on reinforcement learning is performed. The personalized intervention execution and closed-loop optimization module is used to execute tiered interventions based on the multi-level analysis results, dynamically adjust intervention parameters by collecting feedback data through real-time monitoring, and incrementally update the actionable decision tree model using follow-up data streams.

Images