A cognitive intervention system and method for children and adolescents with ath-associated sdb

Abstract

The application discloses a cognitive intervention system for children and adolescents with SDB accompanied by ATH. The cognitive intervention system comprises a data acquisition module for collecting physiological data of patients at night and cognitive behavior data of patients during the day; a cognitive evaluation module for evaluating the cognition of patients based on the collected cognitive behavior data during the day; a risk prediction module for predicting the cognition of patients through a pre-trained cognitive risk prediction model according to the cognitive evaluation results, the physiological data of patients at night and clinical static data; and a personalized intervention module for outputting the optimal candidate intervention strategy through forward-looking simulation and multi-target evaluation on a digital mirror based on the cognitive prediction results of patients, the cognitive evaluation results and the multi-modal data of patients. Thus, the quantitative causal bridge between the pathological physiological state at night and the cognitive behavior function during the day is established, so that accurate cognitive evaluation, scientific prognosis and personalized intervention can be realized.

Description

Technical Field

[0001] This invention relates to a cognitive intervention system for children and adolescents with acute thrombosis (ATH) and comorbid SDB, as well as corresponding cognitive intervention methods, belonging to the field of cognitive training technology. Background Technology

[0002] Adenoid and / or tonsillar hypertrophy (ATH) is a common condition in children, accounting for 42% to 70% of pediatric visits to ENT / Head and Neck Surgery clinics. Upper airway obstruction caused by ATH can lead to sleep-disordered breathing (SDB). The vast majority of pediatric SDB cases are caused by ATH, ranging from mild primary snoring (PS) to severe obstructive sleep apnea (OSA).

[0003] SDB (Sleep Deprivation Byproducts) can cause cognitive impairment and emotional problems in children, primarily including impaired intelligence, attention / executive function, memory, language, and visuospatial discrimination, as well as anxiety and depressive symptoms. At the behavioral level, ATH (Awake Thrombocytosis) often leads to mouth breathing in children, causing language disorders and intermittent hypoxemia, which in turn impairs cognitive function. At the hormonal level, on the one hand, ATH-induced SDB can trigger sleep deprivation, leading to damage to the hypothalamus-pituitary-thyroid axis, thereby lowering thyroid hormone levels. Thyroid dysfunction affects hippocampal synaptic function, impairing children's learning and memory. On the other hand, children with ATH have elevated levels of inflammatory factors (e.g., pro-inflammatory factors), and similar chronic inflammatory diseases can also lead to cognitive decline. Studies have found significant differences in cognitive function between children with SDB and healthy controls. SDB is associated with impaired executive function as early as preschool age, with the greatest impact on its inhibitory dimension. During school age, children with habitual snoring experience declines in general intelligence, visual attention, visuospatial and directional perception, and the degree of neuropsychological deficits in children is also correlated with the severity of OSA (Obstructive Sleep Apnea). A recent systematic review and meta-analysis that incorporated 66 previous related studies found that children with both PS and OSA had significant impairments in all cognitive domains.

[0004] Adenoidectomy / tonsillectomy (AT) is currently the first-line treatment for acute myocardial infarction (ATH) complicated by spontaneous adenomatous dysplasia (SDB) in children. However, previous studies do not appear to support the claim that conventional treatment can improve cognitive function in these children. In cognitive assessments of children with obstructive sleep apnea (OSA) 7 months after AT, AT did not significantly improve attention and executive function compared to the observational waiting group. Although subsequent studies have found significant improvements in cognitive function in OSA children 7 months after AT compared to the observational waiting group, these improvements were limited to nonverbal reasoning, fine motor skills, and selective attention, and the effect size was small (Cohen's d = 0.20–0.24), suggesting that the effect of AT on improving cognitive function in OSA children may be very limited. Recently, a study published in the top medical journal Pediatrics found that one year after AT, the Brief Intellectual Ability (BIA) scores (including comprehension-knowledge, fluid reasoning, and processing speed) in OSA children were not significantly different from those who did not undergo AT, indicating that AT does not improve cognitive function in children with SDB. A study by Biggs et al., examining the long-term cognitive impact (3 years) of SDB symptom relief in preschool children, extended this conclusion to other treatment states and methods besides surgery. That is, regardless of whether treatment is received, the type of treatment used (AT, nasal steroids), or whether SDB is relieved, the impact on cognitive deficits is minimal. It is noteworthy that AT has been the most common surgical procedure performed on children under general anesthesia for nearly a century, but the anesthesia itself can affect cognitive function. Children who underwent general anesthesia before age 4 scored lower on the Wechsler Intelligence Scale for Children (WISC) between ages 5 and 18 compared to children who did not undergo general anesthesia.

[0005] In conclusion, children with acute myocardial infarction (ATH) and secondary cognitive impairment (SDB) have cognitive impairment, but acupuncture and anesthetic (AT) is unlikely to help them improve these deficits, and the anesthesia during surgery can also affect cognitive function. Cognitive impairment negatively impacts the physical and mental health, academic performance, and daily life of children and adolescents, placing a heavy burden on families and society. Furthermore, cognitive impairment may prevent children and adolescents with ATH and SDB from facing subsequent systemic treatment in a good physical and mental state, and may also weaken the effectiveness of treatment.

[0006] Therefore, there is an urgent need to build and validate an intervention system that covers multiple cognitive domains and is designed for children and adolescents with ATH and SDB, including assessment, prediction and intervention functions. Summary of the Invention

[0007] The primary technical problem to be solved by this invention is to provide a cognitive intervention system for children and adolescents with ATH and SDB.

[0008] Another technical problem to be solved by the present invention is to provide a cognitive intervention method for children and adolescents with ATH and SDB.

[0009] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: According to a first aspect of the present invention, a cognitive intervention system for children and adolescents with acute thromboangiitis obliterans (ATH) and comorbid dysplasia of the kidney (SDB) is provided, comprising: The data acquisition module is used to collect patients' nighttime physiological data and daytime cognitive behavioral data around the clock, extract nighttime features from the nighttime physiological data, extract behavioral features from the daytime cognitive behavioral data, and store them in the feature database along with the corresponding metadata. The cognitive assessment module is connected to the data acquisition module. Based on the behavioral characteristics corresponding to the daytime cognitive behavior data collected synchronously when the patient performs a preset gamified cognitive assessment task, the module performs multimodal fusion analysis through the Transformer model to generate a multimodal cognitive feature vector representing the patient's cognitive function status. The risk prediction module is connected to the data acquisition module to receive the patient's nighttime physiological data and daytime cognitive behavior data, and to input the patient's clinical static data into a pre-trained cognitive risk prediction model to output the patient's current multidimensional cognitive risk index and the predicted range of the cognitive development trajectory under different intervention conditions. The personalized intervention module is connected to the risk prediction module, cognitive assessment module and data acquisition module to construct and continuously update a dynamic cognitive digital mirror for the patient based on the patient’s current multidimensional cognitive risk index and cognitive development trajectory prediction range. Specifically, multiple candidate intervention strategies are generated by performing prospective simulations and multi-objective assessments on the digital mirror; and the optimal candidate intervention strategy is dynamically selected and executed from the multiple candidate intervention strategies to provide personalized cognitive training intervention for the patient.

[0010] Preferably, the data acquisition module includes: The nighttime data acquisition unit includes at least a polysomnography monitor and a flexible head-mounted EEG cap for acquiring the patient's nighttime physiological data; wherein the nighttime physiological data includes at least polysomnography data and nighttime EEG signals; The daytime data acquisition unit includes at least a touch-screen tablet, a front-facing camera, a built-in microphone, a flexible head-mounted EEG cap, and a wearable wristband. The touch-screen tablet is used to acquire and present gamified cognitive assessments and cognitive training tasks, and to record the patient's behavioral response data. The front-facing camera, located on the touch-screen tablet, is used to capture the patient's facial expressions in real time. The built-in microphone is located within the touch-screen tablet to collect the patient's speech data. The flexible head-mounted EEG cap is used to collect the patient's daytime EEG signals. The wearable wristband is worn on the patient's wrist to collect skin conductivity signals. The behavioral response data, facial expressions, speech data, daytime EEG signals, and skin conductivity signals collectively constitute the patient's daytime cognitive-behavioral data. The feature extraction engine is connected to the nighttime data acquisition unit and the daytime data acquisition unit to extract nighttime features from the nighttime physiological data and behavioral features from the daytime cognitive behavior data, and stores them in the feature database along with the corresponding metadata.

[0011] Preferably, the cognitive assessment module includes: The gamified cognitive assessment task library has a variety of pre-set gamified cognitive assessment tasks covering core cognitive domains, which are used to push cognitive assessment tasks to the patients; wherein, different types of gamified assessment tasks correspond to different cognitive domains. The difficulty dynamic adjustment unit has a built-in DeepFM-based algorithm to dynamically adjust the difficulty parameters of the next test based on the patient's real-time accuracy and reaction time, so that the cognitive assessment is always in the patient's zone of proximal development and can detect the "ceiling" and "breakdown point" of their cognitive ability. The data synchronization acquisition unit is communicatively connected to the data acquisition module to perform millisecond-level timestamp synchronization when the patient performs a preset gamified cognitive assessment task, thereby capturing the behavioral characteristics corresponding to the patient's daytime cognitive behavior data. A cognitive computing unit is connected to the data synchronization acquisition unit to receive the patient's behavioral characteristics and perform multimodal fusion analysis through the Transformer model to generate a multimodal cognitive feature vector representing the patient's cognitive function state. The real-time feedback unit is connected to the cognitive computing unit and has a preset positive encouragement logic to provide different forms of positive feedback to the patient based on different behavioral response data, and output a visual assessment report based on the patient's multimodal cognitive feature vector.

[0012] Preferably, the risk prediction module includes: The sequence construction unit is used to construct a structured input sequence for cognitive risk prediction based on the patient's nighttime characteristics and the daytime multimodal cognitive feature vector; wherein, the structured input sequence includes: a nighttime physiological spatiotemporal map sequence constructed based on the nighttime physiological data, and a daytime behavior-physiological event sequence constructed based on the multimodal cognitive feature vector and daytime behavior data; The cognitive risk prediction unit has a built-in spatiotemporal graph attention memory network (ST-GAMNet) model, which is used to receive the nighttime physiological spatiotemporal graph sequence, the daytime behavior-physiological event sequence and the patient's clinical static data, and output the patient's current multidimensional cognitive risk index and the predicted range of future cognitive development trajectory. The prediction output unit is connected to the cognitive risk prediction unit to output visualized prediction results and explanations.

[0013] Preferably, the cognitive risk prediction model is ST-GAMNet, which includes: An input layer is used to receive the structured input sequence and the patient's static clinical data; The parallel coding layer is pre-set with multiple types of encoders to encode the structured input sequence and the clinical static data in parallel, thereby outputting nighttime pathological feature vectors, daytime cognitive function feature vectors and clinical static feature vectors respectively. A cross-modal fusion layer, connected to the parallel coding layer, is used to receive the nighttime pathological feature vector, the daytime cognitive function feature vector, and the clinical static feature vector, and to perform cross-modal alignment and deep fusion to output the final comprehensive state vector containing historical state and prior knowledge. The multi-task prediction layer, connected to the cross-modal fusion layer, is used to receive the comprehensive state vector and output the patient's current multidimensional cognitive risk index and classify the risk level through the Softmax activation function in conjunction with the fully connected layer; it is also used to receive the comprehensive state vector and output the prediction range of the future cognitive development trajectory under different intervention conditions through the decoder of the conditional variational autoencoder.

[0014] Preferably, the plurality of encoders includes at least: A fully connected network is used to receive the static clinical data and output a static clinical feature vector. A spatiotemporal convolutional network is used to receive the nighttime physiological spatiotemporal map sequence and perform spatial and temporal convolution operations to output the nighttime pathological feature vector. A multi-head self-attention Transformer encoder is used to receive the daytime behavior-physiological event sequence to output the daytime cognitive function feature vector.

[0015] Preferably, the prediction output unit outputs at least: Personalized prediction reports are used to display the multidimensional cognitive risk index, risk level type, and predicted cognitive development trajectory under different intervention conditions. Interpretable attribution analysis is used to demonstrate nocturnal pathological features and / or daytime cognitive function features that contribute significantly to the high-risk index. Clinical decision support recommendations are provided to demonstrate quantitative suggestions on interventions and durations based on the personalized prediction reports and interpretable attribution analysis, as well as the predicted probability of achieving the target therapeutic effect using the quantitative suggestions.

[0016] Preferably, the personalized intervention module includes: The gamified cognitive training task library has a variety of pre-set cognitive training tasks covering core cognitive domains, which are used to push gamified cognitive training tasks to the patient; wherein, different types of gamified cognitive training tasks correspond to different cognitive domains. The adaptive difficulty adjustment unit is used to dynamically adjust the task difficulty during training by using a preset contextual multi-armed slot machine algorithm based on the patient's real-time physiological and behavioral data. The real-time feedback and incentive unit is used to analyze the patient's physiological and behavioral characteristics in real time, provide positive feedback to the patient, and control the task parameters of the gamified cognitive training task. The dynamic cognitive digital mirror unit is used to pre-construct a continuously evolving dynamic cognitive digital mirror based on the patient's historical data and by using a recurrent neural network combined with neural frequent differential equations, in order to simulate possible responses to different intervention stimuli.

[0017] Preferably, the multi-objective evaluation includes at least two of the following: Short-term gains are calculated based on the improvement in cognitive state at recent simulated time points; Mid-term stability is calculated based on the volatility of cognitive states at mid-term simulated time points; Long-term convergence is calculated based on the closeness between the simulated endpoint cognitive state and the expected optimal trajectory; The efficiency index is calculated based on the ratio of the total improvement in cognitive state during the simulation period to the total training time. Fatigue risk is calculated based on the predicted peak cognitive fatigue during the simulation period.

[0018] Preferably, the fast non-dominated sorting genetic algorithm II is used to optimize the indicators of the multi-objective evaluation in order to identify the Pareto optimal solution from multiple candidate intervention strategies.

[0019] According to a second aspect of the present invention, a cognitive intervention method utilizing the above-described cognitive intervention system is provided, comprising the following steps: Acquire patients' nighttime physiological data and daytime cognitive behavioral data; Cognitive function assessment is performed based on the daytime cognitive behavior data, and the cognitive assessment results are output. Based on the nighttime physiological data, the daytime cognitive behavior data, and the cognitive assessment results, the patient's cognitive risk and development trajectory are predicted. Based on the patient's cognitive risk and development trajectory, a personalized cognitive training program is generated and optimized to provide personalized cognitive intervention for the patient.

[0020] Compared with the prior art, the present invention has the following technical effects: (1) Holographic data profile: For the first time, nighttime breathing, blood oxygen, electroencephalogram, skin conductance level, sleep, body movement and daytime behavior, eye movement, speech, micro-expression, and skin conductance level were correlated across time and space to construct an unprecedented "holographic data profile" of cognitive impairment in children with ATH accompanied by SDB.

[0021] (2) Ecological and non-intrusive: Wearable and non-contact technologies are used to the maximum extent to complete data collection in the natural home environment, avoiding the "white coat effect" and the tension of laboratory sleep, making the data more authentic and reliable.

[0022] (3) Deep correlation with pathophysiological mechanisms: Directly collect biomarkers (e.g., θ / β ratio, pupil changes) that are closely related to the core pathology of the disease (e.g., nocturnal hypoxia, sleep fragmentation) and the core cognitive mechanisms (e.g., attention control, cognitive effort) to make the data analysis more accurate.

[0023] (4) Laying the foundation for advanced AI models: Rich and multimodal time series data provide a perfect data foundation for training high-performance deep learning prediction models, making accurate and individualized cognitive risk prediction possible.

[0024] (5) It comprehensively monitors the entire cognitive process of "how to think, how to strive, and how to get frustrated," and establishes a causal bridge from neuroelectrophysiology to cognitive behavior through millisecond-level synchronization and fusion analysis, giving the assessment results unprecedented interpretive depth. In addition, the adaptive and supportive assessment environment itself is a lightweight cognitive training, providing a perfect foundation and adaptive preparation for subsequent intervention modules.

[0025] (6) The ST-GAM Net model, through spatiotemporal graph network and cross-modal aligned attention mechanism, is no longer satisfied with discovering statistical associations between features, but actively explores their inherent physiological causal paths, making the prediction more biologically reasonable.

[0026] (7) Dynamic and conditional trajectory prediction: For the first time in the field of pediatric ATH with SDB, conditional prediction in the form of "if...then..." is realized, allowing doctors to simulate the long-term cognitive benefits that different treatment options may bring, and realize true personalized medicine.

[0027] Furthermore, the introduction of memory enhancement networks and continuous learning imbues the model with a "medical history concept," enabling it to understand the dynamic evolution of a child's cognitive state rather than making isolated snapshot-like judgments. A powerful interpretability engine transforms the black-box model into a "smart advisor" for doctors, making its decision-making process transparent and traceable, greatly enhancing the willingness and safety of clinical adoption.

[0028] (8) The leap from static prescription to dynamic programming: the intervention scheme is no longer fixed, but a dynamic programming process that is continuously optimized by the fast non-dominated sorting genetic algorithm II and the digital mirror engine.

[0029] (9) Prospective simulation intervention (digital mirror): Allows the system to "trial and error" in virtual space, selects the absolutely optimal rehabilitation path for children in the real world, maximizes efficiency and avoids training risks. In addition, it uses neurophysiological signals to achieve unprecedentedly accurate feedback, upgrading from a closed loop of "intervention behavior" itself to a closed loop of "neurophysiological effects of intervention".

[0030] (10) It achieves true “individualization” rather than “personalization”: the system learns not a general pattern of a group of patients, but the unique cognitive learning characteristics and dynamics of each child, thus achieving true “training according to the brain”. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the structure of a cognitive intervention system for children and adolescents with ATH and SDB, provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the data acquisition module in the first embodiment of the present invention; Figure 3 This is a schematic diagram of the cognitive assessment module in the first embodiment of the present invention; Figure 4 This is a schematic diagram of the risk prediction module in the first embodiment of the present invention; Figure 5 This is a schematic diagram of the personalized intervention module in the first embodiment of the present invention; Figure 6 A flowchart illustrating a cognitive intervention method for children and adolescents with ATH and SDB as provided in the second embodiment of the present invention; Figure 7This is a schematic diagram of the structure of an electronic device provided in the third embodiment of the present invention. Detailed Implementation

[0032] The technical content of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0033] This invention aims to construct a data-driven, closed-loop system for the diagnosis and treatment of cognitive function in children and adolescents with adenoid and tonsillar hypertrophy accompanied by sleep-disordered breathing (ATH with SDB). Unlike existing cognitive training software, this system deeply integrates the pathophysiological mechanisms of the disease, principles of cognitive neuroscience, and advanced artificial intelligence technology to achieve proactive management throughout the entire process, from objective assessment to mechanism prediction and then to prospective intervention.

[0034] In various embodiments of the present invention, the core content includes the following parts: First, cross-temporal and spatial multimodal data association and fusion. By establishing a quantitative causal bridge between "nighttime pathophysiological state" and "daytime cognitive behavioral function," the system can understand the biological mechanisms behind cognitive impairment (e.g., attributing daytime inattention to EEG disturbances during specific sleep periods at night), rather than merely describing the phenomenon.

[0035] Second, mechanism-oriented specialized prediction models and simulations. Utilizing the "Spatiotemporal Graph Attention Memory Network (ST-GAM Net)" specifically designed for ATH accompanied by SDB, a leap from correlation analysis to mechanism inference and prognostic prediction is achieved, thereby upgrading AI from a "black box" prediction tool to an interpretable "decision support system" that provides doctors with quantitative treatment suggestions.

[0036] Third, proactive and dynamic intervention based on individual digital mirrors. By constructing a continuously evolving "dynamic cognitive digital mirror" for each child, the intervention plan can be "simulated first, then implemented." This upgrades cognitive rehabilitation from "static prescription" and "trial and error adjustment" to "dynamic planning" and "proactive optimization," maximizing intervention efficiency and safety, and achieving true "brain-specific training."

[0037] The following details how the cognitive intervention system provided in this embodiment of the invention achieves full-process management of "assessment-prediction-screening-intervention".

[0038] First Embodiment like Figure 1As shown, the first embodiment of this invention provides a cognitive intervention system for children and adolescents with acute myocardial infarction (ATH) and secondary cognitive deficit hyperplasia (SDB), comprising at least a data acquisition module 1, a cognitive assessment module 2, a risk prediction module 3, and a personalized intervention module 4. The data acquisition module 1 collects patients' nighttime physiological data and daytime cognitive behavioral data around the clock, providing a data foundation for subsequent steps. The cognitive assessment module 2 performs cognitive assessments on patients based on the collected daytime cognitive behavioral data. The risk prediction module 3, based on the patients' cognitive assessment results, combined with their nighttime physiological data and clinical static data, uses a pre-trained cognitive risk prediction model to predict patients' cognition, thus providing direction for personalized intervention. The personalized intervention module 4, based on the patients' cognitive prediction results, cognitive assessment results, and the patients' multimodal data, performs prospective simulations and multi-objective assessments on a constructed digital mirror to output the optimal candidate intervention strategy for personalized cognitive training intervention.

[0039] The functions and working relationships of each module are explained in detail below.

[0040] I. Data Acquisition Module Data acquisition module 1 is the cornerstone of the entire system, such as... Figure 2 As shown, the data acquisition module 1 comprises three parts: a nighttime data acquisition unit 11, a daytime data acquisition unit 12, and a feature extraction engine 13. Its core innovation lies in realizing all-weather, multimodal, and non-invasive synchronous monitoring and feature extraction of the "nighttime pathophysiological characteristics" and "daytime cognitive and behavioral performance" of children and adolescents with ATH and SDB, providing an unparalleled data foundation for subsequent accurate prediction and intervention.

[0041] (1) Nighttime data acquisition unit The nighttime data acquisition unit 11 is designed to capture key physiological signals directly related to SDB and cognitive impairment without interference in a natural home sleep environment. Specifically, it includes: a. Portable Polysomnography (PSG) Monitor: Integrates medical-grade biosensors and is widely used in homes, sleep centers, and medical institutions. The device features 9 leads monitoring parameters: nasal airflow, chest breathing, oxygen saturation (SpO2), pulse, pulse waveform, body position, body movement, chest / abdominal movement, and heart rate. Among these: Heart rate variability (HRV), particularly the ratio of nighttime low-frequency power to high-frequency power (LF / HF), was collected and calculated as an indicator of autonomic dysfunction and sleep structure fragmentation.

[0042] By collecting and calculating the chest and abdominal breathing effort, respiratory flow waveform, and pulse signal, and analyzing the amplitude and periodicity of the respiratory waveform, apnea (stopped airflow) and hypoventilation (significantly reduced airflow) events can be automatically identified, and the apnea-hypopnea index (AHI) can be calculated.

[0043] Continuous monitoring of SpO2 allows for accurate calculation of the oxygen reduction index (ODI).

[0044] The collected and calculated body position and body movement data are used for preliminary sleep stage assessment and sleep posture identification (supplementary sleep is more likely to cause airway obstruction). b. Flexible Headband EEG Cap: Utilizing flexible dry electrode technology, this cap is specially designed for children, offering comfort and being glue-free. Its primary function is to collect EEG signals throughout the night and accurately identify non-rapid eye movement (NREM) sleep. During this stage, the system focuses on calculating the power ratio (θ / β ratio) of theta waves (4-7Hz) to beta waves (15-30Hz) in the prefrontal cortex. Elevated θ / β ratios during NREM have been shown to be highly correlated with daytime inattention and impaired executive function, serving as a "neural bridge" connecting nighttime sleep quality and daytime cognitive function.

[0045] (2) Daytime data acquisition unit The daytime data acquisition unit 12 simultaneously captures the behavioral and physiological manifestations of the child's higher-order cognitive functions during systematic cognitive assessments or daily human-computer interactions. Specifically, this includes: a. High-precision touch tablet: As the main carrier for presenting gamified cognitive assessment and training tasks, it accurately records behavioral data such as reaction time, accuracy, and touch trajectory.

[0046] b. Front-facing high-definition camera: The tablet's front-facing camera is used to record the child's facial expressions during tasks. Through computer vision models, facial movement units are analyzed to identify micro-expression changes caused by cognitive load, such as frowning and drooping corners of the mouth, serving as objective quantitative indicators of task engagement and emotional frustration. The front-facing camera, combined with a near-infrared light source, enables eye-tracking without external devices. The system extracts indicators such as fixation point, saccade speed, and pupil diameter changes. For example, in attention tasks, sustained pupil dilation is an important indicator of cognitive effort, while inattention is accompanied by rapid fluctuations in pupil diameter.

[0047] c. Built-in microphone: Captures speech responses in tasks requiring verbal feedback (such as digit span forward and backward tests). Speech analysis algorithms extract acoustic features such as pitch fluctuations, speech rate, and pause frequency. Children with ATH often exhibit mouth breathing, which may lead to abnormal oral muscle function, thus affecting the fluency and clarity of their speech. These features can reflect their cognitive-motor integration abilities.

[0048] d. Flexible headband EEG cap: Utilizing flexible dry electrode technology, this cap is specially designed for children, offering comfort and being glue-free. Its primary function is to collect EEG signals during cognitive assessments. During this phase, the system will focus on acquiring task-related event-related potentials (ERPs) and power variations across different frequency bands.

[0049] e. Wearable wristband: Used to record changes in children's skin conductance and electrical activity during cognitive assessments. Skin conductance is an auxiliary indicator of emotional arousal and cognitive engagement.

[0050] (3) Feature extraction engine The feature extraction engine 13 is essentially the "brain" of the data acquisition module 1, responsible for transforming raw sensor data into feature indicators with clinical and cognitive scientific significance. Specifically, it includes: a. Automated Data Processing System: This system establishes an automated data processing pipeline on cloud or edge computing devices. Nighttime data is automatically analyzed after synchronization in the morning, calculating key nighttime features such as AHI, ODI, average SpO2, minimum SpO2, NREM θ / β ratio, average HRV, and LF / HF. Daytime data is processed in real time after each assessment, generating behavioral features such as comprehensive reaction time, eye-tracking attention concentration index, facial expression fluctuation index, and speech fluency score. For example, the "eye-tracking attention concentration index" (the percentage of time spent fixating on a target area) is extracted from eye-tracking data, the "prefrontal cortex θ-band event-related synchronization" intensity is extracted from EEG signals, and "reaction time volatility" is calculated from behavioral data.

[0051] b. Feature Database: All extracted features will be standardized and stored in a structured feature database along with their corresponding metadata (such as timestamps, patient IDs, demographic information, clinical information, etc.) for subsequent modules to access.

[0052] II. Cognitive Assessment Module Cognitive Assessment Module 2 serves as the "cognitive function detector" of the entire system. Its core function is to transcend the subjectivity and fragmentation of traditional scales. Through controlled gamified tasks, it simultaneously stimulates, captures, and integrates multi-dimensional signals from behavior, neurophysiology, and computer vision, thereby creating a high-resolution profile of an individual's cognitive function. In its design, this module transforms the assessment process from an "exam" into an "immersive gamified interactive experience," allowing children to undergo precise, objective, and quantitative evaluations of their core cognitive domains (perception, attention, memory, executive function, etc.) without any preconceived notions. Figure 3 As shown, this module integrates four units: task generation, data acquisition, data calculation, and result feedback. Details are as follows: (1) Adaptive task recommendation unit This unit forms the foundation of cognitive assessment and comprises two parts: a gamified cognitive assessment task library and a dynamic difficulty adjustment unit. The gamified cognitive assessment task library contains a variety of pre-set gamified cognitive assessment tasks covering core cognitive domains, which are then pushed to patients. The dynamic difficulty adjustment unit incorporates the DeepFM algorithm to dynamically adjust the difficulty parameters of the next test based on the patient's real-time accuracy and reaction time, ensuring that the cognitive assessment remains within the patient's zone of proximal development. This allows for dynamic task adjustment based on the child's real-time performance.

[0053] a. Gamified Cognitive Assessment Task Library: Construct a game task library covering core cognitive domains, for example: "Interstellar Diplomacy" (Sustained Attention and Inhibition Control): Children and teenagers are required to tap on hostile planets on the screen according to rules, but must resist tapping on suddenly appearing friendly planets. A vibrant starry sky background enhances immersion.

[0054] "Memory Maze" (Working Memory and Spatial Perception): Children and adolescents need to remember the path a virtual pet takes in a maze and then recreate that path later. The complexity of the maze changes dynamically.

[0055] "Weather Playground" (Cognitive Flexibility): Children need to switch between different classification rules (classify by shape / classify by color) according to changes in weather conditions (e.g., sunny / rainy).

[0056] b. Dynamic Adjustment of Task Difficulty Based on DeepFM: This unit incorporates a DeepFM-based algorithm. This algorithm dynamically adjusts the difficulty parameters of the next trial (e.g., stimulus presentation time, number of distractions, rule complexity, etc.) based on the child's real-time accuracy and reaction time. This not only ensures the assessment stays within the child's "zone of proximal development," maintaining their focus and motivation, but also accurately detects the "peaks" and "troughs" of their cognitive abilities.

[0057] (2) Multimodal data synchronous acquisition unit This unit seamlessly integrates with the aforementioned data acquisition module 1, performing millisecond-level timestamp synchronization during task execution and capturing the following data streams: a. Behavioral data stream: reaction time, accuracy, and touch trajectory slope (reflecting decision uncertainty) from high-precision touch tablets.

[0058] b. Eye movement data stream: fixation heatmap, saccade speed, pupil diameter change (as the gold standard for cognitive load), etc.

[0059] c. EEG signal data stream: Using a lightweight head-mounted device, focusing on the prefrontal and parietal lobes, task-related ERPs are extracted in real time, such as the amplitude and latency of P300 (related to attention and decision-making), and the power changes of the theta band (related to cognitive control).

[0060] d. Skin conductance response data stream: Micro-changes in skin conductance levels are measured using a wearable wristband as an auxiliary indicator of emotional arousal and cognitive engagement.

[0061] e. Facial data stream: Real-time analysis of facial movement units, such as frowning (cognitive effort), downturned corners of the mouth (frustration), and wandering gaze (distraction).

[0062] (3) Cognitive computing and fusion unit This unit is the "brain" of the multimodal cognitive assessment module, responsible for transforming raw data into meaningful cognitive indicators.

[0063] a. Feature extraction and standardization: Automatically summarize the standardized feature data of each data stream from the feature database of the aforementioned data acquisition module.

[0064] b. Cognitive Feature Integration Engine (Primary Fusion): Perform time-aligned multimodal data correlation analysis. For example, when an erroneous response occurs, we not only see the behavioral "error," but also simultaneously observe: the state of the pupils just before the decision (e.g., abnormal dilation, i.e., cognitive overload out of control), the state of the theta band power in the prefrontal cortex (e.g., failure to effectively increase, i.e., cognitive control failure), and the subtle facial expressions corresponding to the timestamp (e.g., frowning, i.e., experiencing frustration). This correlation can deepen a simple erroneous response into an interpretation of the failure mechanism of the cognitive process.

[0065] c. Cross-modal feature vectorization (advanced fusion and cognitive metric generation): A multimodal Transformer model is employed, treating features from different modalities as a unified network for learning. This model outputs a multimodal cognitive vector, which is a comprehensive quantification of an individual's cognitive state during the task. For example, the final "working memory" score is no longer simply the number of items recalled, but a composite score integrating behavioral accuracy, pupillary change patterns during memory retrieval, relevant neural oscillation patterns in EEG, changes in skin conductance, and facial change patterns.

[0066] (4) Real-time feedback unit To ensure the children's immersion and the effectiveness of the assessment, some calculations need to be completed in real time on the terminal.

[0067] a. Edge computing nodes: Deploy lightweight AI models on tablets to process eye movement and facial expression data in real time and dynamically adjust game difficulty.

[0068] b. Immersive Feedback Mechanism: Abandoning traditional "right / wrong" visual cues. When the system detects frustration in a child through real-time data (e.g., elevated skin conductance, frequent frowning), it triggers an in-game "encouragement mechanism," where the game character provides a prompt: "This is a bit difficult, shall we try again?" When high concentration is detected (e.g., stable pupil signals, ERP waveform), richer positive feedback is provided, such as starting to play cheerful music.

[0069] c. Visualized results report: Multimodal cognitive profile: Outputs a visual report that not only shows the standardized scores of each cognitive dimension, but also uses multimodal data (such as ERP waveform and time-frequency graph, eye-tracking thermogram, facial expression change curve, and skin conductance level change curve) as evidence to intuitively reveal the physiological and behavioral mechanisms behind cognitive strengths and weaknesses.

[0070] Interpretable diagnostic clues: The output report will explain abnormal cognitive phenomena based on cognitive assessment scores and multimodal processing mechanisms during the assessment process. For example: "The child's inattention is mainly manifested as weak inhibitory control function, the neurophysiological basis of which is a decrease in the amplitude of the prefrontal P300 wave during inhibitory tasks, accompanied by abnormal pupillary dilation patterns." III. Risk Prediction Module Risk prediction module 3 serves as the "decision brain" of the entire system, constructing a digital mirror capable of understanding the complex causal relationships of "nighttime pathology - daytime function - current state - developmental trajectory and long-term prognosis," generating a unique "digital mirror of cognitive prognosis" for each child. Its core function is to bridge the gap between "data description" and "future prediction." Utilizing massive amounts of multimodal data collected and fused from preceding modules, it employs advanced deep learning models to quantitatively assess the severity of current cognitive deficits in children with ATH and SDB, accurately predicting their future cognitive development trajectory and recovery potential after intervention, providing data-driven recommendations for clinical decision-making. Figure 4 As shown, this module integrates three parts: construction of multimodal spatiotemporal high-dimensional features, training and optimization of cognitive deficit risk prediction models, and visualization and interpretation of prediction results.

[0071] (1) Construction of multimodal spatiotemporal high-dimensional features The multimodal spatiotemporal high-dimensional feature construction module automatically generates nighttime and daytime high-order features from the multimodal data stream after feature extraction and standardization of the "cognitive computing and fusion unit" in the aforementioned module, based on timestamps. Specifically, the nighttime high-order feature is a nighttime physiological spatiotemporal map sequence constructed based on nighttime physiological data, and the daytime high-order feature is a daytime behavior-physiological event sequence constructed based on multimodal cognitive feature vectors and daytime behavioral data.

[0072] For example, combining nocturnal blood oxygen saturation with sleep stages, a "NREM hypoxic load" is constructed; combining reaction time with pupil diameter changes during the "Weather Paradise" task, a "Weather Paradise-Cognitive Flexibility-Cognitive Effort Efficiency Index 1" is constructed; combining skin conductance levels with prefrontal theta band power changes during the "Weather Paradise" task, a "Weather Paradise-Cognitive Flexibility-Cognitive Effort Efficiency Index 2" is constructed; combining this with all cognitive flexibility and cognitive effort efficiency indices of the "Weather Paradise" task, a "Weather Paradise-Cognitive Flexibility-Cognitive Effort Efficiency Overall Index" is constructed; and combining the "Cognitive Effort Efficiency Overall Index" of all cognitive flexibility tasks, an "Individual Cognitive Flexibility-Cognitive Effort Efficiency Overall Index" is constructed.

[0073] This process continues until a "panoramic feature vector of an individual" covering 150 dimensions is formed.

[0074] (2) Cognitive Risk Prediction Unit The core of this unit is a multimodal, multi-task deep learning architecture, which is preferably a "Spatiotemporal Graph Attention Memory Network" (ST-GAM Net) in this embodiment.

[0075] a. Model architecture: Encoder end: ① Fully connected network: Used to process static information. The child's demographic and clinical information will be converted into a static feature vector through a fully connected layer.

[0076] ② Spatiotemporal Graph Convolutional Network (ST-GCN): Used to process multiple physiological signals at night. It treats the human body as a graph structure (nodes: heart, brain, skin, blood, respiration, body movement, sleep, sleeping posture; edges: physiological relationships between nodes), simultaneously capturing the temporal dynamics within signals and the spatial relationships between signals. For example, it analyzes the spatiotemporal transmission path of "how a sleep apnea event leads to a decrease in blood oxygenation, which in turn causes an increase in the theta / β ratio."

[0077] ③ Multi-head self-attention Transformer encoder: Used to process sequential data from daytime assessments (e.g., eye movement trajectories, continuous responses). It can capture distant dependencies in assessment tasks and accurately locate key moments in cognitive behavioral states such as attentional distraction or memory retrieval failure.

[0078] Fusion and Memory End: ① Cross-modal aligned attention mechanism: This mechanism allows the feature vectors produced by the night encoder and the day encoder to "communicate". For example, the model will automatically learn the association weights between two cross-temporal and cross-modal features: "abnormal θ / β ratio during the nighttime NREM period" and "reduced P300 amplitude in the prefrontal cortex during the daytime task".

[0079] ② Gated Recurrent Unit and Memory Enhancement Network: An external memory matrix is ​​introduced to simulate human long-term memory. This network not only analyzes single test data, but also processes the data from each follow-up visit of the child as a continuous time series, thereby learning the individual pattern of cognitive function evolution over time and distinguishing between temporary fluctuations and trends of increase / decrease.

[0080] Decoder end (multi-task output): ① Function 1: Current Cognitive Deficit Risk Score (Classification and Regression): Outputs a multidimensional cognitive risk index ranging from 0 to 100, corresponding to core domains such as perception, attention, memory, and executive function. It also categorizes these as "low risk," "medium risk," or "high risk."

[0081] ② Function 2: Prediction of future cognitive trajectory (sequence generation): Under the premise of giving different treatment opinions (e.g., observation only, AT surgery only, AT surgery combined with n months of cognitive training, etc.), the conditional variational autoencoder is used to generate the possible development range of each cognitive dimension of the child's ability in the next 3, 6 and 12 months, and present it in the form of probability range (e.g., after AT surgery combined with 5 months of cognitive training, there is a 95% probability that the attention index will improve to the range of [66, 76] 6 months later).

[0082] b. Data integration and processing flow of the model Data reception and structuring: The multimodal spatiotemporal high-dimensional feature building unit outputs a structured "individual panoramic feature vector" containing 150 dimensions. This vector is not a simple planar list, but a multi-level, structured data object with spatiotemporal semantics, which is precisely input into different parts of the ST-GAM Net model.

[0083] The “Individual Panoramic Feature Vector” (V) is structurally defined as three main components: V = {V static V nocturnal V diurnal}

[0084] ① V static(Static Clinical Feature Vector): Includes 25 dimensions: age, gender, education level, BMI, adenoid hypertrophy grade (I-IV), tonsil hypertrophy grade (I-IV), OSA-18 score (5 dimensions + total score), PSQ-SRBD score (4 dimensions + total score), disease duration, whether the baby was born at full term, birth weight, birth length, feeding method after birth, past medical history, allergy history, and family history. This data will be directly input into the fully connected layer of the model for initial encoding, and then fused with high-level features extracted from dynamic data.

[0085] ② V nocturnal (Nighttime Physiological Spatiotemporal Graph Sequence): This is a time-series graph structure (high-dimensional spatiotemporal tensor). The graph node features at each time point (e.g., every 100 milliseconds) include: key nighttime features of each node (heart, brain, skin, blood, respiration, body movement, sleep, sleep posture), and high-dimensional spatiotemporal features of nighttime multimodalities. The edges of the graph represent potential connections between nodes (e.g., heart-brain coupling). This data is a dedicated input for "ST-GCN".

[0086] ③ V diurnal (Daytime Behavioral-Physiological Event Sequence): This is a sequence of events occurring during a cognitive assessment task. The length of the sequence can vary (depending on the task content and the child's performance). Each event (e.g., receiving a stimulus, performing a key press response) is associated with a rich feature data package, including: reaction time, correctness, mean pupil diameter during stimulus presentation, eye movement saccade speed, and P300 amplitude, theta band power, and other daytime multimodal spatiotemporal high-dimensional features extracted from the EEG. This data serves as the dedicated input for the "Multi-Head Self-Attention Transformer Encoder."

[0087] Dedicated processing and feature abstraction within the model: The encoder of the model starts working in parallel, processing the data types that each is good at.

[0088] ① Fully connected network processing V static A simple fully connected layer transforms static features into a more expressive static feature vector (E). static ), output the encoded E static .

[0089] ② ST-GCN processing V nocturnalST-GCN slides across the graph structure, simultaneously learning temporal patterns (e.g., a typical apnea lasts 20 seconds) and spatial patterns (e.g., when apnea occurs, the blood oxygen node signal begins to decline with a 5-second delay, followed by a burst of micro-awakening theta waves in the EEG node). ST-GCN ultimately generates a highly abstract, fixed-dimensional nocturnal pathological feature vector (E) for the entire nighttime data. nocturnal This vector encapsulates the severity, pattern, and impact on physiological systems of nocturnal pathology.

[0090] ③ Multi-head self-attention Transformer encoder processes V diurnal The Transformer utilizes its self-attention mechanism to analyze the entire task sequence. It can detect long-range dependency patterns such as "after three consecutive correct responses, the reaction time to the fourth stimulus is significantly prolonged, accompanied by pupillary dilation saturation, indicating cognitive fatigue." The Transformer generates a highly abstract, fixed-dimensional feature vector (E) of daytime cognitive function for the entire daytime assessment sequence. diurnal This vector encapsulates the overall performance, stability, and neural efficiency during the cognitive assessment process.

[0091] Cross-modal fusion and deep memory: combining three high-level feature vectors: E static E nocturnal and E diurnal Deep integration will be carried out.

[0092] ① Clinical context injection and cross-modal alignment attention mechanism: In this step, E static It plays the role of "clinical prior knowledge." It is simultaneously provided to those processing E. nocturnal and E diurnal A cross-modal alignment attention mechanism.

[0093] Cross-modal aligned attention mechanism allows E nocturnal and E diurnal Mutual queries. Specifically, "daytime query" refers to E diurnal Will to E nocturnal The question is: "My poor attention control is primarily caused by which nighttime pathological phenomenon?" At this point, the model will consider the child's demographic and clinical background when making its judgment. For a child with grade IV adenoid hypertrophy, the attention weight might clearly point to "severe apnea and hypoxia"; while for a younger child, the attention weight might clearly point to "fragmented sleep structure." "Nighttime query" refers to E nocturnal Also to E diurnalQuestion: "What are the most obvious manifestations of severe hypoxia I've detected during the day?" Attention weights might point to "slow pupillary response during the encoding phase of a working memory task." Through this cross-conversation constrained by clinical context, a novel, deeply fused cross-modal feature vector (E) is generated. cross This directly establishes a quantitative bridge between nocturnal pathology and daytime functional impairment.

[0094] ② Gated recurrent units and memory enhancement networks: If the current visit is the nth follow-up of the child, then the E values ​​from the previous (n-1) visits... cross The vector is sequentially fed into a gated recurrent unit (GRU). The GRU acts like a memory stream, capturing the evolution of cognitive states over time. Simultaneously, a memory-enhancing network, as an external knowledge base, stores typical cognitive evolution patterns learned from large-scale data (e.g., "fast recovery," "slow recovery"). The GRU outputs a vector that considers the temporal context of the individual's historical states. The memory network provides a priori reference pattern. The combination of these two ultimately generates a comprehensive state vector (E) that includes historical states and prior knowledge. final ).

[0095] Multi-task prediction output: E with complete information final The decoder is fed into the multi-task output.

[0096] ① Function 1: E final The system uses a fully connected layer, Softmax (for classification), and linear activation (for regression) to directly output the current multidimensional cognitive risk index and classify the risk level.

[0097] ② Function 2: E final The vector is concatenated with an encoding representing the intervention condition (e.g., [0,1,0,0] represents AT surgery only, [0,1,1,5] represents AT surgery combined with 5 months of cognitive training). This concatenated vector is then fed into the decoder of a Conditional Variational Autoencoder (CVAE). The CVAE generates not a fixed value, but a probability distribution through random sampling, thus outputting a predicted range for future cognitive scores. This visually demonstrates that "after a certain intervention, 3 / 6 / 12 months, a certain cognitive index has a 95% probability of improving / decreasing to a certain range."

[0098] c. Model training and optimization Training data: The system is pre-trained using a large database containing multimodal data from thousands of children with ATH and SDB throughout their diagnosis, treatment and long-term follow-up, and then adapted to the specific data format of this system through transfer learning.

[0099] Privacy protection technology: Differential privacy technology is used in model training to add noise to the data, ensuring that no sensitive personal information can be inferred from the model parameters.

[0100] Continuous learning mechanism: After system deployment, as new data is continuously input, the model will be dynamically fine-tuned and optimized through Bayesian Online Learning algorithms to continuously improve its predictive performance in the real world.

[0101] (3) Prediction result visualization and interpretation unit a. Personalized Prediction Report: Automatically generates richly illustrated clinical reports that intuitively display multidimensional cognitive risk indices, risk levels and types, cognitive development prediction trajectories (the most likely developmental path of the child's future cognitive function and the corresponding confidence interval under different intervention strategies), and key evidence.

[0102] b. Interpretable Attribution Analysis: Integrating interpretable technologies such as SHAP, it accurately identifies which nighttime or daytime features contribute most to the current high-risk score, making the prediction results verifiable and helping doctors understand the model's decision-making logic.

[0103] c. Clinical decision support recommendations: Based on the prediction results, provide quantitative recommendations, such as: "AT surgery combined with 6 months of cognitive training is strongly recommended, which is expected to increase the probability of executive function returning to normal level by 66%."

[0104] IV. Personalized Intervention Module Personalized Intervention Module 4 acts as the "digital therapist" of the entire system. Its core function is to transform the results and suggestions from the aforementioned cognitive deficit risk prediction module into precise, dynamic, and concrete intervention actions. Based on the child's own multidimensional cognitive risk index, risk level type, cognitive development prediction trajectory, and interpretability attribution analysis, it pre-generates a unique digital intervention plan. During implementation, it uses real-time multimodal data for closed-loop feedback and dynamic optimization, achieving proactive reshaping of cognitive function tailored to each individual. Figure 5 As shown, the personalized intervention module 4 specifically includes: (1) Gamified cognitive training task library This unit constructs a game task library covering core cognitive domains, the content of which is similar to the "gamified cognitive assessment task library" in the adaptive task recommendation unit of the aforementioned multimodal cognitive assessment module.

[0105] (2) Personalized prescription generation unit This unit is the starting point for intervention, responsible for "translating" the prediction results of the aforementioned cognitive deficit risk prediction module into an executable training plan.

[0106] a. Input: Receive the core output from the aforementioned cognitive deficit risk prediction module, including: multidimensional cognitive risk index, risk level type, cognitive development prediction trajectory, and interpretability attribution analysis (e.g., "working memory deficits are mainly caused by low encoding efficiency").

[0107] b. Core Algorithm - Constrained Optimization Model: Decision variables include training task type, initial difficulty, weekly training frequency, number of training tasks per session, and the ratio of training intensity (number of tasks and training duration) for each cognitive domain.

[0108] Objective function: Maximize expected cognitive gains (derived from the cognitive development prediction trajectory) while minimizing the cognitive load and frustration experienced by the child.

[0109] Constraints: Based on interpretable attribution analysis, priority should be given to training the most critical cognitive domains with the most deficits; reasonable upper limits for training difficulty and intensity should be set according to the child's age, gender, and baseline level.

[0110] c. Output: A personalized cognitive intervention prescription that clearly specifies the initial training content, intensity, and frequency.

[0111] (3) Adaptive difficulty adjustment unit This unit is the "autopilot system" of intervention, ensuring that the difficulty of intervention training is always at the "balance point between skills and challenges".

[0112] a. Input: The high-precision touch tablet and built-in microphone in the aforementioned data acquisition module - daytime data acquisition unit are used to collect the child's behavioral data (accuracy, reaction time, touch trajectory and voice response, etc.) in real time during the training task. The embedded eye-tracking device and front-facing high-definition camera are used to collect the child's physiological data (fixation point, saccade speed, pupil diameter, micro-expression changes, etc.) in real time during the training task.

[0113] b. Algorithm Core - Contextual Multi-Arm Slot Machine: The algorithm treats the different difficulty levels of each training task as "slot machine arms." Based on the current "context" (i.e., the child's real-time physiological data such as fatigue level and historical performance), the algorithm dynamically selects the optimal difficulty for the next trial. Its goal is to balance "skill" and "challenge": most of the time (randomly within 80%–90%), it pushes the difficulty level (skill) that is currently estimated to bring the greatest cognitive benefit, but occasionally (randomly within 10%–20%), it pushes slightly more difficult or extremely difficult levels (challenge) to more accurately calibrate the boundaries of the child's cognitive abilities.

[0114] c. Output: Real-time, automated adjustment of parameters such as the difficulty of the training task, the speed of stimulus presentation, and the number of distractors.

[0115] (4) Multimodal real-time feedback and excitation This unit is essential for maintaining the child's dependence on interventions, ensuring their participation and motivation.

[0116] a. Physiological state perception: By analyzing eye movements (e.g., gaze dispersion) and micro-expressions (e.g., drooping upper eyelids) in real time, the system can predict the inattention of distraction tens of milliseconds to several seconds in advance.

[0117] b. Dynamic Narrative and Reward Mechanism: Cognitive training tasks are embedded in different gamified narratives (e.g., "Spring Outing"). When the system detects that the child is about to become distracted, it triggers a plot event (e.g., "Your friends are having a kite-flying competition; you need to concentrate to help Soyorin's kite fly the highest!") to re-attract attention. Furthermore, when the system detects high cognitive effort through physiological signals, it awards "effort points" and provides encouraging animations even if the task is failed.

[0118] c. Physiological state control mechanism: Transforming the child's real-time physiological state data into direct control parameters within training tasks and scenario events. For example, in the aforementioned "flying a kite" scenario event, the child is required to maintain focus (e.g., focusing the gaze point, frowning, etc.) to keep the kite rising, transforming abstract cognitive activities into intuitive visual feedback.

[0119] (5) Dynamic cognitive digital mirror unit This unit will create an ever-evolving digital mirror image for each patient.

[0120] a. Digital Mirror Construction: Based on all the child's historical data (from assessments to each training session), a dynamic model of their "cognitive system" is constructed using a recurrent neural network (RNN) combined with neural ODEs. This digital mirror can simulate their possible responses to different intervention stimuli.

[0121] Strategy generator: Based on the current cognitive state, historical assessment results, historical intervention effects and long-term prognostic goals, it automatically generates a diverse and feasible set of candidate training strategies {U_i}.

[0122] ① Parameterized encoding: Each policy U_i is a structured vector that encodes the core parameters of the next intervention. Wherein, U_i = [task type sequence, difficulty gradient curve, training duration, rest interval between tasks, cognitive domain intensity ratio].

[0123] ② Generation method: i. Goal-oriented mutation: Using the current optimal policy as a baseline, a new policy with asymptotic optimization is generated through Gaussian perturbation.

[0124] ii. Diversity injection: Latin hypercube sampling (LHS) is used to generate a batch of uniformly distributed random policies in the policy space to ensure the breadth of exploration.

[0125] iii. Constraint Filter: All generated strategies must pass a feasibility check, such as ensuring that the total cognitive load does not exceed the individual child's tolerance threshold.

[0126] Multi-timescale parallel simulation engine: For each candidate policy U_i, simulates the evolution of cognitive state at multiple key time points in the future (e.g., 3 days, 1 week, 1 month, 3 months) after the policy is executed in parallel on a digital mirror.

[0127] ① Initialization: Set the current state [h(t0), z(t0)] of the digital mirror to the initial state of the simulation.

[0128] h(t0) is a multidimensional cognitive state vector, representing the observable and quantifiable multidimensional cognitive function state of the child at time t0. Examples of typical dimensions constituting h(t0) are as follows: i. Attention Stability Index h1(t0): Derived from tasks involving sustained attention in the aforementioned multimodal cognitive assessment module. It integrates behavioral data (e.g., reaction time volatility and omission errors in sustained performance tasks) and physiological data (e.g., pupil diameter stability during the task and EEG θ / β ratio) to quantify the child's ability to maintain attention.

[0129] ii. Working Memory Capacity Index h2(t0): Derived from tasks related to working memory ability in the aforementioned multimodal cognitive assessment module. Based on task accuracy, memory span, and combined with ERP (e.g., P300) amplitude during the memory encoding and retrieval stages, it quantifies the child's ability to temporarily store and process information.

[0130] iii. Executive Function Index h3(t0): Derived from tasks involving executive function in the aforementioned multimodal cognitive assessment module. It integrates the accuracy loss during task switching, reaction time for inhibitory control, and EEG features related to the prefrontal cortex to quantify the child's abilities in planning, problem-solving, and cognitive flexibility.

[0131] iv. Information Processing Speed ​​Index h4 (t0): Derived from the simple reaction time tasks involved in the aforementioned multimodal cognitive assessment module. Based on the average reaction time of the task, and adjusted for age and task difficulty, it quantifies the speed at which the brain performs basic cognitive operations.

[0132] v. Verbal Learning and Memory Index h5(t0): Derived from the aforementioned multimodal cognitive assessment module involving auditory or visual vocabulary learning tasks. Based on scores of immediate recall, delayed recall, and recognition, it quantifies language memory ability closely related to academic performance.

[0133] z(t0) is a multidimensional hidden dynamic state vector, representing the potential, intrinsic neurophysiological processes driving changes in the cognitive state h(t0) at time t0. Examples of typical dimensions constituting z(t0) include: i. Neuroplasticity potential z1 (t0): Calculated from the child's historical data (e.g., learning curves, response speed to previous training). It characterizes the intrinsic potential of the brain's neural networks to change due to experience and training. A higher value indicates a better expected intervention outcome.

[0134] ii. Cognitive Fatigue Cumulative Value z2(t0): This value integrates fatigue indicators from nighttime sleep quality, daytime reaction time drift, and heart rate variability (HRV). It characterizes the current level of fatigue in the nervous system. A higher value indicates a need to reduce training intensity or increase rest.

[0135] iii. Neural metabolic efficiency z3(t0): Inferred from effort-related physiological signals in cognitive tasks, such as the "cognitive effort efficacy index" (pupil dilation or EEG load changes associated with a unit improvement in accuracy). It characterizes the brain's energy utilization efficiency when processing information. A higher value indicates less cognitive resources are required to process the same task, and higher efficiency.

[0136] iv. Sleep repair efficacy z4 (t0): derived from nighttime data, such as the proportion of slow-wave sleep (deep sleep), the characteristics of sleep spindles, and the degree of improvement in cognitive baseline upon waking compared to before bedtime. It characterizes the effectiveness of sleep in restoring cognitive function. A higher value indicates a better restorative effect of sleep.

[0137] v. Stress / Arousal Level z5(t0): This integrates skin conductance response level, heart rate variability (e.g., LF / HF ratio), and EEG power at specific frequencies. It characterizes the arousal level of the autonomic and central nervous systems. Both excessively high and low levels can impair cognitive performance.

[0138] ② Dynamic solution: For each policy-time pair (U_i, t_target), call the ODE solver (e.g., Dopri5) to numerically integrate the system of the divine ordinary differential equations.

[0139] ③ Uncertainty modeling: A random noise term is introduced on the right side of the ODE to extend the model into a stochastic differential equation. The expected value and variance of the predicted state are calculated through Monte Carlo simulation (run 100 times) to quantify the uncertainty of the prediction.

[0140] ④ Anti-interference test: Inject simulated real-world interferences (e.g., poor sleep quality on a certain day - abnormal LF / HF, abnormal sleep stage, excessive body movement, etc.) into the simulation to evaluate the robustness of the candidate strategies.

[0141] Multi-objective policy evaluator: Performs multi-dimensional and quantitative evaluation of the simulation results of each candidate policy U_i, and identifies the optimal trade-off solution under multiple (potentially conflicting) objectives.

[0142] ① Calculation of evaluation indicators: i. Short-term gains = w1·Δh (3 days) + w2·Δh (1 week), used to assess the short-term improvement. Where: Δh (3 days): The change in the cognitive state vector h from the current time point t0 to the next 3 days; Δh(3 days) = h(3 days) - h(t0); w1, w2: weighting coefficients, satisfying w1 + w2 = 1; usually w2 > w1, because the change in 1 week is more statistically significant than the change in 3 days.

[0143] ii. Medium-term stability = -Var(h(2 weeks), h(1 month)), used to assess state volatility; a negative value indicates greater stability. h(2 weeks): Predicted cognitive state vector after 2 weeks of policy execution; h(1 month): Predicted cognitive state vector after 1 month of policy execution; Var(): Variance calculation function, specifically a measure of the volatility of cognitive state vectors at two points in time; Var(h(2 weeks), h(1 month)) = ||h(1 month) - h(2 weeks)||2 (Euclidean norm); The Var() function takes a negative value because smaller variance is better; taking a negative value makes the value larger, which indicates better stability.

[0144] iii. Long-term convergence = -Distance(h(3 months), T_optimal), used to assess the closeness to the expected optimal trajectory. Wherein, T_optimal: The target state vector of the expected optimal cognitive development trajectory at 3 months; Distance(): Distance metric function; Distance(h(3 months), T_optimal) = || h(3 months) - T_optimal ||2; The Distance() function takes a negative value because smaller distances are better; taking a negative value results in a larger value, indicating better convergence.

[0145] iv. Efficiency index = Σ(Δh) / Σ(training duration), used to evaluate cognitive gains per unit time. Wherein, Σ(Δh): The cumulative sum of the total improvement in cognitive state during the simulated time period. Σ(Δh) [(t + Δt) - h(t)] Σ (Training Duration): The total cumulative training time within the same simulation period. v. Fatigue risk = max(z2(t)), used to assess the predicted hidden peak fatigue level. z2(t): The dimension in the hidden dynamic state vector z(t) specifically represents "cognitive fatigue." It is a scalar time series, and the larger the value, the higher the degree of fatigue. t: A time variable that iterates through the entire simulation time period [t0, t_end]. max(): A function that retrieves the maximum value in a time series, identifying the peak value. ② Multi-objective optimization method: i. The strategy selection problem is formalized as a multi-objective optimization problem, with the goal of maximizing [short-term returns, medium-term stability, long-term convergence, and efficiency index], while constraining [fatigue risk < threshold].

[0146] ii. Use the Non-dominated Sorting Genetic Algorithm II to find the Pareto optimal front – the set of strategies that cannot be improved on any one objective without harming the others.

[0147] Model Predictive Controller: Based on the multi-objective evaluation results, it executes the final strategy selection and manages the "rolling optimization" process.

[0148] ① Value Function: Define a comprehensive value function V(U) i This summarizes multi-objective assessments into a single score, making it easier to rank them.

[0149] V(U) i = α·Short-term returns + β·Medium-term stability + γ·Long-term convergence + η·Efficiency index - λ·Fatigue risk.

[0150] ② Dynamic weight adjustment: The weights [α, β, γ, η, λ] are not fixed and will be dynamically adjusted according to the current stage of the child.

[0151] i. Early recovery phase: Increase α (short-term returns) and η (efficiency) to quickly build confidence.

[0152] ii. Plateau period: Increase γ (long-term convergence) to encourage breakthroughs.

[0153] iii. Fatigue signs period: Increase λ (fatigue penalty) and focus on recovery.

[0154] ③ Rolling optimization: Select V(U) from the Pareto front. i The highest-level strategy U* is used only as the execution plan for the next intervention. After the next intervention, the state is updated with new real data, and the entire optimization loop is rerun to achieve closed-loop feedback and dynamic adjustment.

[0155] Online calibrator: Continuously narrows the gap between digital mirror predictions and real-world observations, ensuring the fidelity of the simulation.

[0156] ① Gap Monitoring: Weekly comparison with predicted status h predicted（t1） Compared with the actual state h obtained from the assessment real（t1） .

[0157] ② Triggered learning: When the moving average of the prediction error exceeds a predetermined threshold, the online learning process is triggered.

[0158] ③ Flexible fine-tuning: Using the flexible weight consolidation technique, while calibrating the model parameters (θ, φ) with new data, it penalizes the forgetting of learned old knowledge, thereby improving the accuracy of personalized prediction for specific children without destroying the generalization ability of the original model.

[0159] b. Forward-looking simulation and scheme selection: ① Before each adjustment to the intervention plan, the system will perform at least 10,000 simulations on a dynamic digital mirror.

[0160] ② Dynamic digital mirroring tests various potential combinations of training strategies (e.g., "Should we increase working memory load or consolidate attention first in the next training session?").

[0161] ③ Through simulation, predict the impact of each strategy on the child's real cognitive state (its digital mirror image) at several key time points in the future (e.g., 3 days, 1 week, 1 month, 3 months, etc.).

[0162] ④ Finally, the system will select a strategy that can effectively guide the child's cognitive state toward the expected optimal trajectory in the simulation, and use it as the formal plan for the next training session.

[0163] c. Output: A dynamically updated, simulation-validated optimal intervention strategy for the next iteration.

[0164] (6) Results Visualization and Interpretation Unit The core function of this unit is to build trust between humans and machines and to provide clinical action guidelines. Through multi-level, multi-perspective visualization and natural language explanations, it ensures that doctors and parents can understand "why this plan was chosen" and clearly see "the effects and progress after the plan was implemented".

[0165] a. Cognitive Rehabilitation Outcome Interface: A homepage overview for all patients, providing readily accessible core information.

[0166] Key metrics: ① Current Total Cognitive Score: A single score that integrates multiple cognitive dimensions, which intuitively displays the overall level of cognitive health, much like a "credit score".

[0167] ② Training Completion Status: The progress bar shows the comparison between the planned training volume and the actual training volume completed.

[0168] ③ Fatigue Index Indicator: A traffic light-style icon (green / yellow / red) that displays the child's current cognitive fatigue state as estimated by the model in real time.

[0169] ④ Compliance score: a compliance score based on the physiological state during training and the quality of completion.

[0170] Cognitive Dimension Radar Chart: A dynamic radar chart that clearly shows the child's current level in core dimensions such as attention, memory, executive function, processing speed, and language ability (compared to age-appropriate norms and their own baseline).

[0171] Core Trend Line: A concise time trend line that shows the changes in "Total Cognitive Integral" over the past (which can vary on time scales of days, weeks, months, and years), making progress readily apparent.

[0172] b. Intervention Plan Explanation Interface: When the system recommends a plan for the next training session, this component is responsible for explaining "why this plan is chosen".

[0173] Decision Influence Factors Plot: A horizontal bar chart showing which factors (e.g., "slow improvement in recent working memory", "improvement in nighttime blood oxygen levels", "fatigue levels within a safe range") contribute the most to the decision on the next course of action.

[0174] Option comparison card: Present 2-3 alternative options in card format, clearly indicating the advantages (e.g., "rapid short-term improvement") and risks (e.g., "higher risk of fatigue accumulation") of each option, and highlighting why the next option is the optimal trade-off.

[0175] Natural Language Summary: Automatically generates a concise text description, such as: "The next program focuses on improving your child's attentional stability in distracting environments. We've increased the weight of 'selective attention' tasks but appropriately reduced the frequency of task switching to strike a balance between improving effectiveness and controlling fatigue." c. Multi-timescale prediction trajectory interface: Displays digital mirror simulations and predictions of the future.

[0176] Cognitive Development Trajectory Chart: A set of line graphs, with the X-axis representing time (from now to the next 3 months) and the Y-axis representing cognitive scores and scores for each key cognitive dimension.

[0177] ① Curve: Shows the predicted cognitive development path under the recommended scheme.

[0178] ② Confidence interval: The shaded area represents the range of uncertainty in the prediction (95% confidence interval).

[0179] ③ Reference line: Shows the expected optimal trajectory and the natural development trajectory (assuming no intervention) for comparison.

[0180] ④ Interactive function: Allows patients to click on any time point within the next 3 months to view the predicted status details for that point.

[0181] The "What-If" scenario simulator is an advanced interactive feature. Doctors or parents can manually adjust training parameters (e.g., adding one training session per week), and the system will draw new simulation trajectories on the graph in real time, visually demonstrating the possible consequences of different choices.

[0182] d. Attribution Analysis and Progress Tracing Interface: Used for in-depth analysis of training effectiveness, answering "What made the difference?".

[0183] Intervention-Effect Correlation Diagram: A Sankey diagram connecting different training tasks (source) with improvements in different cognitive abilities (endpoint), with the thickness of the lines indicating the strength of the correlation. It clearly shows which specific cognitive abilities each training task primarily improved.

[0184] Progress Attribution Timeline: A chart combining a timeline and markers of significant events. It marks when the training program was adjusted, what significant events the child experienced (e.g., colds, insomnia), and aligns these events with the curve of changes in cognitive scores, helping to identify key drivers of progress or external causes of setbacks.

[0185] Second Embodiment like Figure 6 As shown, based on the first embodiment described above, the second embodiment of the present invention provides a cognitive intervention method for children and adolescents with ATH and SDB, specifically including the following steps: S1: Obtain the patient's nighttime physiological data and daytime cognitive behavioral data.

[0186] Specifically, the data acquisition module 1 collects patients' nighttime physiological data and daytime cognitive behavior data around the clock, thus providing a data foundation for subsequent steps.

[0187] The specific data acquisition process is described in the first embodiment above regarding data acquisition module 1, and will not be repeated here.

[0188] S2: Perform cognitive function assessment based on daytime cognitive behavior data and output the cognitive assessment results.

[0189] Specifically, after data collection is completed through step S1, the cognitive assessment module 2 conducts a cognitive assessment on the patient based on the collected daytime cognitive behavior data, thereby outputting the cognitive assessment results for the patient.

[0190] The specific cognitive assessment process is described in the first embodiment above regarding cognitive assessment module 2, and will not be repeated here.

[0191] S3: Predict the patient's cognitive risk and development trajectory.

[0192] Specifically, after data collection is completed based on step S1 and cognitive assessment of the patient is completed based on step S2, the cognitive prediction module 3 is used to predict the patient's cognitive risk and development trajectory based on nighttime physiological data, daytime cognitive behavior data and cognitive assessment results, thereby providing direction and guidance for subsequent personalized intervention.

[0193] The specific cognitive prediction process is described in the first embodiment above regarding cognitive prediction module 3, and will not be repeated here.

[0194] S4: Provide personalized cognitive interventions for patients.

[0195] Specifically, based on the predicted cognitive risk and development trajectory of the patient in step S3, prospective simulations and multi-objective assessments are conducted on a digital mirror to output multiple candidate intervention strategies. On this basis, the optimal candidate intervention strategy is selected to provide personalized cognitive training intervention for the patient, thereby achieving the best intervention effect.

[0196] The specific cognitive intervention process is described in the first embodiment above regarding the personalized intervention module 4, and will not be repeated here.

[0197] Third Embodiment like Figure 7 As shown in the second embodiment, the third embodiment of the present invention provides an electronic device. This electronic device includes one or more processors 100 and a memory 200. The memory 200 is coupled to the processor 100 and is used to store one or more programs. When the programs are executed by the processor 100, the processor 100 implements the cognitive intervention method for children and adolescents with ATH and SDB as described in the above embodiments.

[0198] The processor 100 controls the overall operation of the electronic device to complete all or part of the steps of the cognitive intervention method for children and adolescents with ATH and SDB. The processor 100 can be a central processing unit (CPU), graphics processing unit (GPU), field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), digital signal processing (DSP), etc. The memory 200 stores various types of data to support the operation of the electronic device. This data may include, for example, instructions for any application or method operating on the electronic device, and application-related data. The memory 200 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, etc.

[0199] In one exemplary embodiment, the electronic device may be implemented by a computer chip or physical entity, or by a product with certain functions, to perform the cognitive intervention method described above for children and adolescents with ATH accompanied by SDB, and achieve the same technical effect as the method described above. A typical embodiment is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, an in-vehicle human-machine interaction device, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0200] In another exemplary embodiment, the present invention also provides a computer-readable storage medium including program instructions that, when executed by a processor, implement the steps of the cognitive intervention method for children and adolescents with ATH and SDB as described in any of the above embodiments. For example, the computer-readable storage medium may be the memory including the program instructions described above, which may be executed by a processor of an electronic device to complete the cognitive intervention method for children and adolescents with ATH and SDB described above, and achieve the same technical effects as the method described above.

[0201] It should be noted that the above embodiments are merely illustrative examples. The technical solutions of each embodiment can be combined, and all are within the protection scope of this invention.

[0202] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0203] The cognitive intervention system and method for children and adolescents with ATH and SDB provided by this invention have been described in detail above. Any obvious modifications made to this invention by those skilled in the art without departing from its essential content will constitute an infringement of the patent rights of this invention and will incur corresponding legal liability.

Claims

1. A cognitive intervention system for children and adolescents with acute myocardial infarction (ATH) and comorbid dysplasia of the kidney (SDB), characterized in that... include: The data acquisition module is used to collect patients' nighttime physiological data and daytime cognitive behavioral data around the clock, extract nighttime features from the nighttime physiological data, extract behavioral features from the daytime cognitive behavioral data, and store them in the feature database along with the corresponding metadata. The cognitive assessment module is connected to the data acquisition module. Based on the behavioral characteristics corresponding to the daytime cognitive behavior data collected synchronously when the patient performs a preset gamified cognitive assessment task, the module performs multimodal fusion analysis through the Transformer model to generate a multimodal cognitive feature vector representing the patient's cognitive function status. The risk prediction module is connected to the data acquisition module to receive the patient's nighttime physiological data and daytime cognitive behavior data, and to input the patient's clinical static data into a pre-trained cognitive risk prediction model to output the patient's current multidimensional cognitive risk index and the predicted range of the cognitive development trajectory under different intervention conditions. The personalized intervention module is connected to the risk prediction module, cognitive assessment module and data acquisition module to construct and continuously update a dynamic cognitive digital mirror for the patient based on the patient’s current multidimensional cognitive risk index and cognitive development trajectory prediction range. Specifically, multiple candidate intervention strategies are generated by performing prospective simulations and multi-objective assessments on the digital mirror; and the optimal candidate intervention strategy is dynamically selected and executed from the multiple candidate intervention strategies to provide personalized cognitive training intervention for the patient.

2. The cognitive intervention system as described in claim 1, characterized in that... The data acquisition module includes: The nighttime data acquisition unit includes at least a polysomnography monitor and a flexible head-mounted EEG cap for acquiring the patient's nighttime physiological data; wherein the nighttime physiological data includes at least polysomnography data and nighttime EEG signals; The daytime data acquisition unit includes at least a touch-screen tablet, a front-facing camera, a built-in microphone, a flexible head-mounted EEG cap, and a wearable wristband. The touch-screen tablet is used to acquire and present gamified cognitive assessments and cognitive training tasks, and to record the patient's behavioral response data. The front-facing camera, located on the touch-screen tablet, is used to capture the patient's facial expressions in real time. The built-in microphone is located within the touch-screen tablet to collect the patient's speech data. The flexible head-mounted EEG cap is used to collect the patient's daytime EEG signals. The wearable wristband is worn on the patient's wrist to collect skin conductivity signals. The behavioral response data, facial expressions, speech data, daytime EEG signals, and skin conductivity signals collectively constitute the patient's daytime cognitive-behavioral data. The feature extraction engine is connected to the nighttime data acquisition unit and the daytime data acquisition unit to extract nighttime features from the nighttime physiological data and behavioral features from the daytime cognitive behavior data, and stores them in the feature database along with the corresponding metadata.

3. The cognitive intervention system as described in claim 2, characterized in that... The cognitive assessment module includes: The gamified cognitive assessment task library has a variety of pre-set gamified cognitive assessment tasks covering core cognitive domains, which are used to push cognitive assessment tasks to the patients; wherein, different types of gamified assessment tasks correspond to different cognitive domains. The difficulty dynamic adjustment unit has a built-in DeepFM algorithm to dynamically adjust the difficulty parameters of the next test based on the patient's real-time accuracy and reaction time, so that the cognitive assessment is always in the patient's zone of proximal development. The data synchronization acquisition unit is communicatively connected to the data acquisition module to perform millisecond-level timestamp synchronization when the patient performs a preset gamified cognitive assessment task, thereby capturing the behavioral characteristics corresponding to the patient's daytime cognitive behavior data. A cognitive computing unit is connected to the data synchronization acquisition unit to receive the patient's behavioral characteristics and perform multimodal fusion analysis through the Transformer model to generate a multimodal cognitive feature vector representing the patient's cognitive function state. The real-time feedback unit is connected to the cognitive computing unit and has a preset positive encouragement logic to provide different forms of positive feedback to the patient based on different behavioral response data, and output a visual assessment report based on the patient's multimodal cognitive feature vector.

4. The cognitive intervention system as described in claim 3, characterized in that... The risk prediction module includes: A sequence construction unit is used to construct a structured input sequence for cognitive risk prediction based on the patient's nighttime characteristics and the multimodal cognitive feature vector; wherein, the structured input sequence includes: a nighttime physiological spatiotemporal map sequence constructed based on the nighttime physiological data, and a daytime behavior-physiological event sequence constructed based on the multimodal cognitive feature vector and daytime behavior data; The cognitive risk prediction unit has a built-in spatiotemporal graph attention memory network model, which is used to receive the nighttime physiological spatiotemporal graph sequence, the daytime behavior-physiological event sequence and the patient's clinical static data, and output the patient's current multidimensional cognitive risk index and the predicted range of future cognitive development trajectory. The prediction output unit is connected to the cognitive risk prediction unit to output visualized prediction results and explanations.

5. The cognitive intervention system as described in claim 4, characterized in that... The cognitive risk prediction model is a spatiotemporal graph attention memory network, which includes: An input layer is used to receive the structured input sequence and the patient's static clinical data; The parallel coding layer is pre-set with multiple types of encoders to encode the structured input sequence and the clinical static data in parallel, thereby outputting nighttime pathological feature vectors, daytime cognitive function feature vectors and clinical static feature vectors respectively. A cross-modal fusion layer, connected to the parallel coding layer, is used to receive the nighttime pathological feature vector, the daytime cognitive function feature vector, and the clinical static feature vector, and to perform cross-modal alignment and fusion to output the final comprehensive state vector. The multi-task prediction layer, connected to the cross-modal fusion layer, is used to receive the comprehensive state vector and output the patient's current multidimensional cognitive risk index through a Softmax activation function in conjunction with a fully connected layer; it is also used to receive the comprehensive state vector and output the predicted range of future cognitive development trajectory under different intervention conditions through a decoder of a conditional variational autoencoder.

6. The cognitive intervention system as described in claim 5, characterized in that... The various types of encoders include at least: A fully connected network is used to receive the static clinical data and output a static clinical feature vector. A spatiotemporal convolutional network is used to receive the nighttime physiological spatiotemporal map sequence and perform spatial and temporal convolution operations to output the nighttime pathological feature vector. A multi-head self-attention Transformer encoder is used to receive the daytime behavior-physiological event sequence to output the daytime cognitive function feature vector.

7. The cognitive intervention system as described in claim 1, characterized in that... The personalized intervention module includes: The gamified cognitive training task library has a variety of pre-set cognitive training tasks covering core cognitive domains, which are used to push gamified cognitive training tasks to the patient; wherein, different types of gamified cognitive training tasks correspond to different cognitive domains. An adaptive difficulty adjustment unit is used to dynamically adjust the task difficulty during training by using a preset contextual multi-armed slot machine algorithm based on the patient's real-time physiological and performance data. The real-time feedback and incentive unit is used to analyze the patient's behavioral characteristics in real time, provide positive feedback to the patient, and control the task parameters of the gamified cognitive training task. The dynamic cognitive digital mirror unit is used to pre-construct a continuously evolving dynamic cognitive digital mirror based on the patient's historical data and by using a recurrent neural network combined with neural frequent differential equations, in order to simulate possible responses to different intervention stimuli.

8. The cognitive intervention system as described in claim 1, characterized in that... The multi-objective assessment includes at least two of the following: Short-term gains are calculated based on the improvement in cognitive state at recent simulated time points; Mid-term stability is calculated based on the volatility of cognitive states at mid-term simulated time points; Long-term convergence is calculated based on the closeness between the simulated endpoint cognitive state and the expected optimal trajectory; The efficiency index is calculated based on the ratio of the total improvement in cognitive state during the simulation period to the total training time. Fatigue risk is calculated based on the predicted peak cognitive fatigue during the simulation period.

9. The cognitive intervention system as described in claim 8, characterized in that: The fast nondominated sorting genetic algorithm II is used to optimize the indicators of the multi-objective evaluation in order to identify Pareto optimal solutions from multiple candidate intervention strategies.

10. A cognitive intervention method for children and adolescents with acute thromboangiitis obliterans (ATH) and comorbid dysplasia of the common genital dysplasia (SDB), implemented using the cognitive intervention system described in any one of claims 1 to 9, characterized in that... include: Acquire patients' nighttime physiological data and daytime cognitive behavioral data; Cognitive function assessment is performed based on the daytime cognitive behavior data, and the cognitive assessment results are output. Based on the nighttime physiological data, the daytime cognitive behavior data, and the cognitive assessment results, the patient's cognitive risk and development trajectory are predicted. Based on the patient's cognitive risk and development trajectory, a personalized cognitive training program is generated and optimized to provide personalized cognitive intervention for the patient.

Images