Multi-source data fusion ai health status comprehensive evaluation method

Abstract

The embodiment of the application relates to a kind of multi-source data fusion AI health state comprehensive evaluation method, comprising: collecting and AD disease condition assessment and management related multidimensional data;Multi-dimensional data are aligned in time and space, cleaning and standardization processing, and from it extract and AD disease evolution related time sequence characteristics, behavior pattern characteristics and clinical text features, and form multi-source fusion feature vector by fusion;Multi-source fusion feature vector is input into the preset deep learning evaluation model, and the quantitative evaluation result of current patient is obtained;Quantitative evaluation result at least includes the quantitative output of patient overall cognitive function level, daily life activity ability and neuropsychiatric symptom severity;According to quantitative evaluation result, match reasoning is carried out in combination with preset AD diagnosis and treatment specification and evidence-based medicine knowledge base, and personalized health management decision report is generated;Health management decision report includes at least one of cognitive intervention suggestion, non-drug therapy scheme, care guide points and disease risk early warning.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology and artificial intelligence, and in particular to a comprehensive assessment method for health status based on multi-source data fusion and artificial intelligence (AI). Background Technology

[0002] Alzheimer's disease (AD) is a progressive neurodegenerative disease, and its management requires long-term, comprehensive, and dynamic monitoring and assessment. Currently, the health management of AD patients mainly faces the following challenges: Assessment is delayed and incomplete: Clinical diagnosis and condition assessment rely heavily on periodic outpatient neuropsychological scale tests and imaging examinations. These assessments are spaced far apart and fail to reflect the continuous and dynamic changes in the patient's cognitive, behavioral, and functional status in daily life, especially with a significant lag in capturing behavioral and psychological symptoms of dementia (BPSD).

[0003] In recent years, although some studies have attempted to use wearable devices or single data sources for AD-assisted assessment, they generally suffer from problems such as limited data dimensions, one-sided model evaluation indicators, and disconnect from clinical practice. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a comprehensive AI health status assessment method based on multi-source data fusion, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, this invention provides a comprehensive AI-based health status assessment method based on multi-source data fusion, the method comprising: Collect multi-dimensional data related to AD disease assessment and management; the multi-dimensional data includes: physiological and behavioral time-series data collected by sensing devices and smart wearable devices deployed in the patient's living environment; electronic medical record data obtained from the medical institution's information system, the electronic medical record data including at least neuropsychological scale scores and observational records of daily cognitive performance, self-care ability and mental and behavioral symptoms entered by the patient or caregiver through a terminal; The multi-dimensional data is spatiotemporally aligned, cleaned, and standardized, and temporal features, behavioral pattern features, and clinical text features related to the evolution of AD are extracted and fused to form a multi-source fusion feature vector. The multi-source fusion feature vector is input into a preset deep learning evaluation model to obtain the quantitative evaluation result of the current patient; the quantitative evaluation result includes at least the quantitative output of the patient's overall cognitive function level, daily living activities ability and the severity of neuropsychiatric symptoms; Based on the quantitative assessment results, a personalized health management decision report is generated by matching and reasoning with the preset AD diagnosis and treatment guidelines and evidence-based medicine knowledge base. The health management decision report includes at least one of the following: cognitive intervention recommendations, non-drug therapy plans, care guidance points, and disease risk warnings.

[0006] In one possible implementation, the physiological and behavioral time-series data includes at least heart rate variability data and sleep structure stage data to reflect cognitive load and diurnal rhythm, as well as gait parameters and indoor activity trajectory data to reflect motor function and behavioral abnormalities. The electronic medical record data also includes imaging report texts and medication history records; The observational records include the frequency and severity of at least one of the following symptoms: apathy, anxiety, agitation, hallucinations, or repetitive behaviors.

[0007] In one possible implementation, the extraction of temporal features, behavioral pattern features, and clinical text features related to the progression of Alzheimer's disease specifically includes: From the aforementioned physiological and behavioral time-series data, long-term time-series statistical features representing the diurnal fluctuation pattern of cognitive function and the decline trend of physical activity are extracted; From the text data of the electronic medical records, key diagnostic entities and historical trend features of neuropsychological scale scores are extracted using natural language processing technology.

[0008] In one possible implementation, the evaluation model is a neural network based on a multi-task learning architecture, comprising: A shared feature encoder is used to perform high-level abstraction and cross-modal correlation modeling on the input multi-source fusion feature vector to generate a shared feature representation; At least three parallel task-specific output layers, each receiving the shared feature representation and executing it separately: The first output layer performs regression or classification tasks and outputs assessment results used to quantitatively evaluate the patient's overall cognitive function level and the degree of impairment in at least one sub-cognitive domain. The second output layer performs a classification task and outputs assessment results for evaluating the patient's basic activities of daily living and / or instrumental activities of daily living abilities. The third output layer performs classification or regression tasks and outputs assessment results used to evaluate the current severity or future risk level of at least one neuropsychiatric symptom in the patient.

[0009] In one possible implementation, the cleaning and standardization of multi-dimensional data specifically includes: For the aforementioned physiological and behavioral time-series data, an outlier detection method based on interquartile range is used to identify and remove noisy data points; The scores of the neuropsychological scales were processed using a standardized method based on norms of the patient's age and education level. One-hot encoding is performed on clinical features extracted from the text, and min-max normalization is performed on numerical clinical features.

[0010] In one possible implementation, generating personalized health management decision reports specifically includes: The quantitative evaluation results are matched with preset threshold conditions in the knowledge base; If the cognitive function assessment result is below the first threshold and the impairment in a specific cognitive domain is prominent, then a cognitive intervention suggestion is generated, including digital training tasks for that specific cognitive domain, recommendation frequency and duration parameters. If the assessment result of the daily living activities ability is lower than the second threshold, care guidance points are generated, including a list of specific environmental safety hazards, corresponding modification measures, and step-by-step daily care process guidelines. If the risk level of neuropsychiatric symptoms exceeds the third threshold, a warning message is generated, including a description of the non-drug intervention scenario, specific communication and behavioral guidance steps, and a reminder to review the medication plan.

[0011] In one possible implementation, the method further includes: Collect feedback data on the effects of the decision report after it is implemented in real care scenarios; The effect feedback data is correlated with the original quantitative evaluation results and knowledge base rules that triggered the decision to form an optimization sample; The deep learning evaluation model is incrementally learned using the optimized samples to update the model parameters; and / or the trigger threshold or confidence level of the corresponding rule in the knowledge base is adaptively adjusted based on the statistical results of the optimized samples.

[0012] In one possible implementation, the method further includes: The quantitative assessment results and the personalized health management decision report are structured and packaged according to the standard format of clinical document architecture; Through the authorized interface, the encapsulated data packet is pushed to the designated attending physician's electronic medical record system or regional medical information platform; Receive and integrate online feedback from clinicians, and confirm, revise, or add annotations to the decision report.

[0013] In one possible implementation, the training process of the evaluation model includes: A training set was obtained consisting of the multi-source fusion feature vectors of multiple AD patients and their corresponding clinical gold standard labels, wherein the gold standard labels were determined by clinical experts based on neuropsychological testing and clinical assessment. With the goal of minimizing the joint loss function of multiple task output layers of the model, the parameters of the shared feature encoder and each task output layer are trained using an end-to-end supervised learning approach.

[0014] In one possible implementation, the evaluation model is trained using a transfer learning and personalized fine-tuning strategy: The model was pre-trained using a large-scale multi-source fusion feature vector dataset across patient populations to obtain a general benchmark evaluation model; For specific target patients, the general benchmark assessment model is fine-tuned using their personal historical accumulated multi-source fusion feature vector data to obtain a personalized assessment model that is more suitable for the individual's condition.

[0015] By applying the AI-based comprehensive health status assessment method based on multi-source data fusion provided in this invention, a continuous digital profile of the patient is constructed by integrating data from multiple sources, including family environment, medical institutions, and caregiver observations, overcoming the limitations of traditional intermittent assessments. Utilizing a deep learning model based on multi-task learning, the method can simultaneously and accurately quantify and assess multiple core dimensions of the patient, such as overall cognitive function, impairment in specific cognitive domains, activities of daily living, and the severity / risk of neuropsychiatric symptoms. The assessment results are more comprehensive and closer to the real, complex condition. Through spatiotemporal alignment, feature engineering, and multi-source fusion techniques, heterogeneous and multimodal data are transformed into a unified multi-source fusion feature vector, allowing data from different sources to complement and corroborate each other. The deep learning model can uncover complex patterns and correlations from these fused features that cannot be revealed by single-source data, significantly improving the accuracy and robustness of the assessment. Furthermore, based on the accurate quantitative assessment results and combined with a structured AD diagnosis and treatment knowledge base, a customized health management decision report can be automatically generated. This report not only includes assessment conclusions but also provides specific and actionable recommendations, making care based on evidence and greatly improving management efficiency and scientific rigor. Furthermore, by incorporating feedback data on the effectiveness of decision reports in practical applications back into the system, the evaluation model can be incrementally learned, and the knowledge base rules can be adaptively adjusted, enabling the entire system to have continuous learning and optimization capabilities, thereby better aligning with the individual patient's disease progression and actual care needs. Attached Figure Description

[0016] Figure 1 Flowchart of the AI-based comprehensive health status assessment method based on multi-source data fusion provided by this invention; Figure 2 This is a structural diagram of a deep learning evaluation model provided in an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0019] Figure 1 Flowchart of AI-based comprehensive health status assessment method based on multi-source data fusion. The main body executing this method is a system with processing and computing capabilities, such as... Figure 1 As shown, the method includes the following steps: Step 110: Collect multi-dimensional data related to AD disease assessment and management.

[0020] The multidimensional data includes physiological and behavioral time-series data collected by sensing devices and smart wearable devices deployed in the patient's living environment; electronic medical record data obtained from the medical institution's information system, which includes at least neuropsychological scale scores; and observational records of daily cognitive performance, self-care ability, and mental and behavioral symptoms entered by the patient or caregiver through a terminal.

[0021] The physiological and behavioral time-series data include at least heart rate variability data and sleep structure stage data to reflect cognitive load and diurnal rhythm, as well as gait parameters and indoor activity trajectory data to reflect motor function and behavioral abnormalities. The electronic medical record data also includes imaging report texts and medication history records; The observational records include the frequency and severity of at least one of the following symptoms: apathy, anxiety, agitation, hallucinations, or repetitive behaviors.

[0022] Specifically, in one example, home smart sensing data, such as the smart bracelet worn by Mr. Li, continuously collects his heart rate variability (HRV) and nighttime sleep structure stages (light sleep, deep sleep, and REM sleep). Millimeter-wave radar sensors installed in the home record his daytime indoor activity trajectory, location movement speed, and mark abnormally long periods of stillness or aimless wandering events.

[0023] After authorization, the system retrieved Li's historical data from the hospital's electronic medical record system, including: Montreal Cognitive Assessment (MoCA) scores from one year ago and six months ago, which were 20 / 30 and 17 / 30 respectively; the description text of the Magnetic Resonance Imaging (MRI) report, which mentioned "mild hippocampal atrophy"; and the long-term medication record "donepezil tablets, 10mg, once daily".

[0024] Li's daughter uses a mobile app to record her daily instrumental activities of daily living (IADL) progress, such as: "I failed to use the microwave to heat food independently today." She also records her BPSD log when she encounters specific behaviors, such as: "I felt irritable in the evening and asked the same question twice."

[0025] The IADL (Independent Activities of Daily Living) can be assessed using the Lawton IADL scale, which is divided into 8 levels, from Level 1 (completely independent) to Level 8 (completely dependent). Basic Activities of Daily Living (BADL) is assessed using the Barthel Index, which ranges from 0 to 100 points, corresponding to different levels of functional dependence.

[0026] Step 120: Perform spatiotemporal alignment, cleaning and standardization on the multi-dimensional data, and extract temporal features, behavioral pattern features and clinical text features related to the evolution of AD disease from it, and fuse them to form a multi-source fusion feature vector; Specifically, the extraction of temporal features, behavioral pattern features, and clinical text features related to the progression of Alzheimer's disease includes: From the aforementioned physiological and behavioral time-series data, long-term time-series statistical features representing the diurnal fluctuation pattern of cognitive function and the decline trend of physical activity are extracted; From the text data of the electronic medical records, key diagnostic entities and historical trend features of neuropsychological scale scores are extracted using natural language processing technology.

[0027] The cleaning and standardization of multi-dimensional data specifically includes: For the aforementioned physiological and behavioral time-series data, an outlier detection method based on interquartile range is used to identify and remove noisy data points; The scores of the neuropsychological scales were processed using a standardized method based on norms of the patient's age and education level. One-hot encoding is performed on clinical features extracted from the text, and min-max normalization is performed on numerical clinical features.

[0028] The statistical window for long-term time-series statistical features is, for example, the past 30 days, which can be dynamically adjusted to 14 or 90 days based on the accumulated patient data. Specific features include: the linear regression slope of the proportion of nocturnal REM sleep over 30 days, the standard deviation of the ratio of low-frequency to high-frequency HRV during the day, and the rate of change of the average daily activity radius. The norms can adopt the "Montreal Cognitive Assessment Chinese Norms (Peking Union Medical College Hospital, 2019)," which provides the mean and standard deviation of MoCA scores stratified by age (e.g., 50-59, 60-69, 70-79, ≥80 years) and education level (e.g., illiterate, primary school, junior high school, senior high school and above). The standardization formula is: Standardized score = (Raw score - Norm mean) / Norm standard deviation. For electronic medical record text, clinical features such as key diagnostic entities and historical trends are extracted using natural language processing technology. For numerical features, such as historical scale scores, min-max normalization is performed; for categorical features, such as imaging findings, one-hot encoding or label encoding is performed.

[0029] Specifically, continuing the previous example, for sensor data, anomaly detection based on the interquartile range (IQR) is used to remove noise points such as abnormal instantaneous heart rate peaks. For the MoCA score, it is standardized according to Mr. Li's age (e.g., 72 years) and years of education (e.g., 6 years), referencing normative data to obtain a relative cognitive score. From the HRV and sleep data of the past 30 days, statistical features such as the decreasing trend of the weekly average nighttime REM sleep ratio and the increased diurnal variability of the daytime HRV low-frequency to high-frequency ratio (LF / HF) are calculated. From the activity trajectory, features such as the daily effective activity area radius and the frequency of abnormal wandering events are extracted. Using a natural language processing model, the key entity hippocampus and the attribute of mild atrophy are extracted from the MRI report; the trend feature of a 3-point decrease in the MoCA score over six months is extracted from historical records. All the extracted numerical and vectorized features are concatenated and min-max normalized to form a fixed-dimensional multi-source fusion feature vector, representing Mr. Li's current and recent comprehensive state.

[0030] In one example, for instance, regarding Mr. Li, based on his HRV and sleep data from the past 30 days, the following were calculated: a decreasing trend in the weekly average nighttime REM sleep ratio and an increasing diurnal variability in the daytime HRV low-frequency to high-frequency ratio (LF / HF). From his activity trajectory, the daily effective activity area radius and the frequency of abnormal wandering events were extracted. Using a natural language processing model, the key entity "hippocampus" and the attribute "mild atrophy" were extracted from the MRI report; and a 3-point decrease in his MoCA score over six months was extracted from historical records. All features were concatenated and min-max normalized to obtain Mr. Li's multi-source fusion feature vector.

[0031] This natural language processing model can extract key diagnostic entities and historical trends from electronic medical record text, and uses a BERT-base-Chinese pre-trained language model for named entity recognition and relation extraction. For example, the model, fine-tuned on the CCKS-2019 electronic medical record dataset, achieved an F1 score of 0.92 for recognizing anatomical entities such as the hippocampus, frontal lobe, and ventricles.

[0032] Step 130: Input the multi-source fusion feature vector into a preset deep learning evaluation model to obtain the quantitative evaluation result of the current patient; the quantitative evaluation result includes at least the quantitative output of the patient's overall cognitive function level, daily living activities ability and the severity of neuropsychiatric symptoms; Here, examples are provided rather than limitations. The overall cognitive function level is output as a continuous score from 0 to 100, with higher scores indicating better cognitive function. The sub-cognitive domains include five sub-items: memory, executive function, visuospatial function, language, and attention, with each sub-item outputting a score from 0 to 20.

[0033] In one example, the evaluation model is a neural network based on a multi-task learning architecture, such as... Figure 2 As shown, the model includes: A shared feature encoder is used to perform high-level abstraction and cross-modal correlation modeling on the input multi-source fusion feature vector to generate a shared feature representation; at least three parallel task-specific output layers, each receiving the shared feature representation and performing the following: a first output layer performs a regression or classification task and outputs an assessment result for quantitatively evaluating the patient's overall cognitive function level and the degree of impairment in at least one sub-cognitive domain; a second output layer performs a classification task and outputs an assessment result for evaluating the patient's basic activities of daily living and / or instrumental activities of daily living abilities; a third output layer performs a classification or regression task and outputs an assessment result for evaluating the current severity or future risk level of at least one neuropsychiatric symptom in the patient.

[0034] In one example, the input layer can receive a multi-source fusion feature vector with a fixed dimension, such as 512.

[0035] The shared feature encoder consists of three stacked fully connected layers with 512, 256, and 128 neurons per layer, respectively. Each layer is followed by a batch normalization layer and a ReLU activation function, and Dropout is introduced with a dropout rate of 0.3 to prevent overfitting. The encoder outputs a 128-dimensional shared feature representation.

[0036] The three task branches include: Branch 1 is the cognitive branch, which includes two fully connected layers (128→64→1), outputting the overall cognitive score (regression) and sub-scores (5 output nodes, using sigmoid activation to output scores of 0-20).

[0037] Branch 2 is the Activities of Daily Living (ADL) branch, which includes two fully connected layers (128→64→4) and outputs the ADL level (4 categories, softmax activation).

[0038] Branch 3 is the NPS branch, which includes two fully connected layers (128→64→3) and outputs the risk level of each symptom (3-class classification, softmax activation).

[0039] Specifically, and not limited to, the time window for predicting the future risk level can be the next 7 days. Risk levels are divided into three levels: low, medium, and high, corresponding to probabilities of occurrence of <30%, 30%-70%, and >70%, respectively. The training process of the evaluation model includes: obtaining a training set consisting of the multi-source fusion feature vectors of multiple AD patients and their corresponding clinical gold standard labels, wherein the gold standard labels are determined by clinical experts based on neuropsychological testing and clinical assessment; and training the shared feature encoder and the parameters of each task output layer using an end-to-end supervised learning method, with the goal of minimizing the joint loss function of the model's multiple task output layers.

[0040] Continuing with the previous example, the model processes and outputs quantitative results in three dimensions in parallel. The first output layer is a cognitive function assessment, with an overall cognitive score of 35 / 100, indicating moderate impairment. The model further shows that the scores for the "delayed recall" and "visuospatial / executive function" sub-items are significantly lower.

[0041] The second output layer is the assessment of daily living skills, which outputs "Instrumental Activities of Daily Living (IADL) Level: 3 (Requires a lot of help)".

[0042] The third output layer is for neuropsychiatric symptom assessment, outputting "Repetitive behavior risk level: high (probability 80%)" and "Dusk agitation risk level: medium (probability 60%)".

[0043] The specific training process for evaluating the model is as follows: Training set construction involved obtaining a training set consisting of the multi-source fusion feature vectors of multiple AD patients and their corresponding clinical gold standard labels. The gold standard labels were determined by clinical experts based on standardized neuropsychological testing and clinical assessment results, and included at least: quantitative scores of overall cognitive function level and the degree of impairment in specific cognitive domains; basic / instrumental activities of daily living ability levels; and the current severity or future risk level of various neuropsychiatric symptoms.

[0044] To define the model training objective, this application aims to minimize the joint loss function of the output layers for multiple tasks. It employs an end-to-end supervised learning approach, simultaneously training the shared feature encoder and the parameters of each task's output layer. An example of the joint loss function is shown below: L = λ1*L1 + λ2*L2 + λ3*L3 Where L is the joint loss function, L1 is the loss for the cognitive assessment task (mean squared error is commonly used for regression tasks, and cross-entropy is commonly used for classification tasks), L2 is the loss for the daily living skills assessment task, and L3 is the loss for the neuropsychiatric symptom assessment task; λ1, λ2, and λ3 are the weight coefficients for each task, which can be adjusted according to the performance on the validation set. Specifically, if the output layer performs a regression task, the mean squared error (MSE) loss function is used; if it performs a classification task, the cross-entropy loss function is used.

[0045] To improve the model's generalization ability and adaptability across different patients, the following two-stage training strategy was adopted: General pre-training utilizes a large-scale, multi-source fused feature vector dataset across patient populations to pre-train the model, resulting in a general benchmark evaluation model. This stage can employ unsupervised or self-supervised pre-training tasks, such as masked feature reconstruction and contrastive learning, to fully explore shared disease progression patterns across patients.

[0046] Personalized fine-tuning involves using multi-source fusion feature vector data accumulated from the individual's personal history for specific target patients. Typically, at least two weeks of daily monitoring data are required to fine-tune all or part of the parameters of the general benchmark assessment model to obtain a personalized assessment model that is more suitable for the individual's condition.

[0047] In practical applications, the system collects feedback data on the effects of the decision report after its execution, and associates this data with the original quantitative evaluation results and knowledge base rules that triggered the decision, forming optimization samples. These optimization samples are then used to incrementally learn the current model, for example, by employing experience replay or elastic weight consolidation methods to update model parameters, enabling the model to adapt to the dynamic evolution of the patient's condition. Simultaneously, based on the statistical results of the optimization samples, the trigger thresholds or confidence levels of the corresponding rules in the knowledge base are adaptively adjusted.

[0048] Step 140: Based on the quantitative assessment results, a personalized health management decision report is generated by matching and reasoning with the preset AD diagnosis and treatment guidelines and evidence-based medicine knowledge base; the health management decision report includes at least one of the following: cognitive intervention recommendations, non-drug therapy plans, care guidance points, and disease risk warnings.

[0049] The pre-defined AD diagnosis and treatment guidelines and evidence-based medicine knowledge base are constructed based on the following authoritative guidelines: the "2018 Chinese Guidelines for the Diagnosis and Treatment of Dementia and Cognitive Impairment," the "Guidelines for the Diagnosis and Treatment of Alzheimer's Disease (2020 Edition)" (National Health Commission), and the "Expert Consensus on Non-Drug Intervention for Dementia Mental and Behavioral Symptoms (2022)." The rules in the knowledge base adopt the production rule (IF-THEN) format and contain a total of 128 rules, covering four major categories: cognitive intervention (32 rules), daily life care (45 rules), neuropsychiatric symptom management (38 rules), and medication reminders (13 rules). Each rule includes prerequisites, namely a comparison of quantitative assessment results with thresholds, and conclusions, namely specific recommendations.

[0050] The non-pharmacological treatment plan is generated based on the specific type of neuropsychiatric symptoms, for example: For agitated / aggressive behavior, "personalized music therapy" is recommended, which involves selecting soothing music familiar to the patient, 15 minutes each time, twice a day; at the same time, it is also recommended to adjust the environment, such as reducing noise and lowering the intensity of lighting.

[0051] To address apathy, we recommend “structured daily activities,” such as simple crafts, pet interaction, and organizing personal belongings, with 20 minutes for each activity and 3 activities per day.

[0052] For sleep disorders, "light therapy" is recommended, which involves receiving 30 minutes of 10,000 lux white light every morning; at the same time, a nighttime sleep hygiene plan should be developed, such as a fixed bedtime and avoiding caffeine.

[0053] Specifically, generating personalized health management decision reports includes: matching the quantitative assessment results with preset threshold conditions in the knowledge base; if the cognitive function assessment result is below the first threshold and the impairment in a specific cognitive domain is prominent, then generating cognitive intervention suggestions including digital training tasks for that specific cognitive domain, recommendation frequency and duration parameters; if the daily living activities assessment result is below the second threshold, then generating care guidance points including a list of specific environmental safety hazards, corresponding modification measures, and step-by-step daily care process guidelines; if the neuropsychiatric symptom risk level exceeds the third threshold, then generating warning information including a description of non-pharmacological intervention scenarios, specific communication and behavioral guidance steps, and medication regimen review reminders.

[0054] Continuing with the previous example, the matching rules could be: "IF Instrumental Activities of Daily Living (ADL) level ≥ 3 THEN trigger 'Home Safety and Assisted Living' intervention package"; "IF Specific cognitive sub-item (delayed recall) score low AND overall cognitive score < 40 THEN trigger 'Memory Enhancement Training' recommendation".

[0055] The system automatically generates "Li's Health Management Recommendations for This Week," which includes: Cognitive Intervention Recommendations: "For weak delayed recall ability, we recommend using the 'Picture Association Memory' training module, 5 times a week, 15 minutes each time." Care Guidance Points: "Given the decline in IADL ability, we recommend: 1) Posting simple illustrated meal reheating process cards in the kitchen; 2) Considering introducing a smart pillbox with reminder function." Risk Warning and Intervention: "Warning: High repetitive behaviors and risk of twilight agitation. Recommendations: 1) Establish a regular daytime activity schedule and reduce idle time; 2) Increase soft lighting and play soothing music before dusk; 3) Ask family members to record the specific triggering situations for the doctor to review the medication plan at the next follow-up visit." The report is then pushed to caregivers, such as the daughter, via the app.

[0056] If a specification is required, the first threshold, second threshold, and third threshold involved in this application can be determined according to any of the following methods: Based on clinical guidelines and referring to the thresholds recommended in the "Guidelines for the Diagnosis and Treatment of Alzheimer's Disease" or relevant evidence-based medicine, for example, the first threshold is set at 20 points on the Montreal Cognitive Assessment (MoCA) (out of 30), the second threshold is set at level 3 in instrumental activities of daily living (IADL) (out of 8, level 3 indicates a need for a lot of help), and the third threshold is set at 50% of the probability of the occurrence of neuropsychiatric symptoms; Based on training data statistics, the distribution of assessment results of the patient population in the training set was used, and the 25th percentile or 75th percentile was used as the threshold. Based on configurable parameters, the threshold is preset as an adjustable parameter in the system, which can be adjusted by clinicians according to the individual patient's condition.

[0057] Furthermore, the method also includes: collecting effect feedback data after the decision report is executed in a real care scenario; associating the effect feedback data with the original quantitative evaluation results and knowledge base rules that triggered the decision to form an optimized sample; using the optimized sample to incrementally learn the deep learning evaluation model to update the model parameters; and / or adaptively adjusting the trigger threshold or confidence level of the corresponding rule in the knowledge base based on the statistical results of the optimized sample.

[0058] Specifically, incremental learning can employ the Elastic Weight Consolidation (EWC) algorithm, which imposes constraints on important parameters of the old task when fine-tuning new samples to prevent catastrophic forgetting. For example, during each incremental learning iteration, 20% of the historical optimized samples can be randomly sampled and mixed with the current samples for training.

[0059] Specifically, for example, two weeks later, caregivers reported that "the success rate of reheating meals improved after using the process card," and new sensor data showed that their activity patterns during the evening hours became more regular. This feedback data was correlated with the original evaluation results and triggering rules to form optimization samples, used to fine-tune the model and improve the confidence of the rules.

[0060] The feedback data includes at least one of the following three categories: Caregiver subjective ratings obtain caregivers' assessment of the ease of implementation of each suggestion in the decision report via the terminal, such as a score of 1-5, and their satisfaction with the outcome, such as a score of 1-5.

[0061] Changes in objective physiological and behavioral indicators, after the implementation of the decision report, the sensors detected a decrease in the frequency of target symptoms, such as a decrease in wandering events from 5 times a day to 2 times a day, or an improvement in sleep quality, such as an increase in the proportion of REM sleep of more than 10%.

[0062] Clinicians confirm that the attending physician has categorized the decision report in the electronic medical record system as reasonable or unreasonable, and may add revision suggestions.

[0063] Furthermore, the method also includes: The quantitative assessment results and the personalized health management decision report are structured and packaged according to the standard format of clinical document architecture; the packaged data package is pushed to the designated attending physician's electronic medical record system or regional medical information platform through an authorized interface; online feedback from clinicians is received and integrated to confirm, revise or supplement the decision report.

[0064] Specifically, the system packaged Li's assessment report using the HL7 FHIR standard and pushed it to the management platform of his community-contracted doctor. After reviewing it online, the doctor confirmed the non-pharmacological intervention recommendations and added a remote medical order to "review liver and kidney function at the next outpatient visit." This order was simultaneously updated in the patient's decision report.

[0065] Furthermore, the evaluation model is trained using transfer learning and personalized fine-tuning strategies: The model was pre-trained using a large-scale multi-source fusion feature vector dataset across patient populations to obtain a general benchmark evaluation model; For specific target patients, the general benchmark assessment model is fine-tuned using their personal historical accumulated multi-source fusion feature vector data to obtain a personalized assessment model that is more suitable for the individual's condition.

[0066] By applying the AI-based comprehensive health status assessment method based on multi-source data fusion provided in this invention, a continuous digital profile of the patient is constructed by integrating data from multiple sources, including family environment, medical institutions, and caregiver observations, overcoming the limitations of traditional intermittent assessments. Utilizing a deep learning model based on multi-task learning, the method can simultaneously and accurately quantify and assess multiple core dimensions of the patient, such as overall cognitive function, impairment in specific cognitive domains, activities of daily living, and the severity / risk of neuropsychiatric symptoms. The assessment results are more comprehensive and closer to the real, complex condition. Through spatiotemporal alignment, feature engineering, and multi-source fusion techniques, heterogeneous and multimodal data are transformed into a unified multi-source fusion feature vector, allowing data from different sources to complement and corroborate each other. The deep learning model can uncover complex patterns and correlations from these fused features that cannot be revealed by single-source data, significantly improving the accuracy and robustness of the assessment. Furthermore, based on the accurate quantitative assessment results and combined with a structured AD diagnosis and treatment knowledge base, a customized health management decision report can be automatically generated. This report not only includes assessment conclusions but also provides specific and actionable recommendations, making care based on evidence and greatly improving management efficiency and scientific rigor. Furthermore, by incorporating feedback data on the effectiveness of decision reports in practical applications back into the system, the evaluation model can be incrementally learned, and the knowledge base rules can be adaptively adjusted, enabling the entire system to have continuous learning and optimization capabilities, thereby better aligning with the individual patient's disease progression and actual care needs.

[0067] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0068] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A comprehensive AI-based health status assessment method based on multi-source data fusion, characterized in that, The method includes: Collect multi-dimensional data related to AD disease assessment and management; the multi-dimensional data includes: physiological and behavioral time-series data collected by sensing devices and smart wearable devices deployed in the patient's living environment; electronic medical record data obtained from the medical institution's information system, the electronic medical record data including at least neuropsychological scale scores and observational records of daily cognitive performance, self-care ability and mental and behavioral symptoms entered by the patient or caregiver through a terminal; The multi-dimensional data is spatiotemporally aligned, cleaned, and standardized, and temporal features, behavioral pattern features, and clinical text features related to the evolution of AD are extracted and fused to form a multi-source fusion feature vector. The multi-source fusion feature vector is input into a preset deep learning evaluation model to obtain the quantitative evaluation result of the current patient; the quantitative evaluation result includes at least the quantitative output of the patient's overall cognitive function level, daily living activities ability and the severity of neuropsychiatric symptoms; Based on the quantitative assessment results, a personalized health management decision report is generated by matching and reasoning with the preset AD diagnosis and treatment guidelines and evidence-based medicine knowledge base. The health management decision report includes at least one of the following: cognitive intervention recommendations, non-drug therapy plans, care guidance points, and disease risk warnings.

2. The method according to claim 1, characterized in that, The physiological and behavioral time-series data include at least heart rate variability data and sleep structure stage data to reflect cognitive load and diurnal rhythm, as well as gait parameters and indoor activity trajectory data to reflect motor function and behavioral abnormalities. The electronic medical record data also includes imaging report texts and medication history records; The observational records include the frequency and severity of at least one of the following symptoms: apathy, anxiety, agitation, hallucinations, or repetitive behaviors.

3. The method according to claim 1, characterized in that, The extraction of temporal features, behavioral pattern features, and clinical text features related to the progression of Alzheimer's disease specifically includes: From the aforementioned physiological and behavioral time-series data, long-term time-series statistical features representing the diurnal fluctuation pattern of cognitive function and the decline trend of physical activity are extracted; From the text data of the electronic medical records, key diagnostic entities and historical trend features of neuropsychological scale scores are extracted using natural language processing technology.

4. The method according to claim 1, characterized in that, The evaluation model is a neural network based on a multi-task learning architecture, including: A shared feature encoder is used to perform high-level abstraction and cross-modal correlation modeling on the input multi-source fusion feature vector to generate a shared feature representation; At least three parallel task-specific output layers, each receiving the shared feature representation and executing it separately: The first output layer performs regression or classification tasks and outputs assessment results used to quantitatively evaluate the patient's overall cognitive function level and the degree of impairment in at least one sub-cognitive domain. The second output layer performs a classification task and outputs assessment results for evaluating the patient's basic activities of daily living and / or instrumental activities of daily living abilities. The third output layer performs classification or regression tasks and outputs assessment results used to evaluate the current severity or future risk level of at least one neuropsychiatric symptom in the patient.

5. The method according to claim 1, characterized in that, The cleaning and standardization of multi-dimensional data specifically includes: For the aforementioned physiological and behavioral time-series data, an outlier detection method based on interquartile range is used to identify and remove noisy data points; The scores of the neuropsychological scales were processed using a standardized method based on norms of the patient's age and education level. One-hot encoding is performed on clinical features extracted from the text, and min-max normalization is performed on numerical clinical features.

6. The method according to claim 1, characterized in that, The generation of personalized health management decision reports specifically includes: The quantitative evaluation results are matched with preset threshold conditions in the knowledge base; If the cognitive function assessment result is below the first threshold and the impairment in a specific cognitive domain is prominent, then a cognitive intervention suggestion is generated, including digital training tasks for that specific cognitive domain, recommendation frequency and duration parameters. If the assessment result of the daily living activities ability is lower than the second threshold, care guidance points are generated, including a list of specific environmental safety hazards, corresponding modification measures, and step-by-step daily care process guidelines. If the risk level of neuropsychiatric symptoms exceeds the third threshold, a warning message is generated, including a description of the non-drug intervention scenario, specific communication and behavioral guidance steps, and a reminder to review the medication plan.

7. The method according to claim 1, characterized in that, The method further includes: Obtain feedback data on the effects of the decision report after it is implemented in a real care scenario; The effect feedback data is correlated with the original quantitative evaluation results and knowledge base rules that triggered the decision to form an optimization sample; The deep learning evaluation model is incrementally learned using the optimized samples to update the model parameters; and / or the trigger threshold or confidence level of the corresponding rule in the knowledge base is adaptively adjusted based on the statistical results of the optimized samples.

8. The method according to claim 1, characterized in that, The method further includes: The quantitative assessment results and the personalized health management decision report are structured and packaged according to the standard format of clinical document architecture; Through the authorized interface, the encapsulated data packet is pushed to the designated attending physician's electronic medical record system or regional medical information platform; Receive and integrate online feedback from clinicians, and confirm, revise, or add annotations to the decision report.

9. The method according to claim 1 or 4, characterized in that, The training process of the evaluation model includes: A training set was obtained consisting of the multi-source fusion feature vectors of multiple AD patients and their corresponding clinical gold standard labels, wherein the gold standard labels were determined by clinical experts based on neuropsychological testing and clinical assessment. With the goal of minimizing the joint loss function of the output layers of multiple tasks in the model, the parameters of the shared feature encoder and each task output layer are trained using an end-to-end supervised learning approach.

10. The method according to claim 9, characterized in that, The evaluation model is trained using a transfer learning and personalized fine-tuning strategy. The model was pre-trained using a large-scale multi-source fusion feature vector dataset across patient populations to obtain a general benchmark evaluation model; For specific target patients, the general benchmark assessment model is fine-tuned using their personal historical accumulated multi-source fusion feature vector data to obtain a personalized assessment model that is more suitable for the individual's condition.

Images